BOOK - Advances in Land Remote Sensing PDF

Shunlin Liang
Editor
Advances in
Land Remote
Sensing
System, Modeling, Inversion
and Application
Advances in Land Remote Sensing
Advances in Land Remote

Sensing
System, Modeling, Inversion and Application
Shunlin Liang
Editor
Department of Geography, University of Maryland,
College Park, MD, USA
123
Shunlin Liang
University of Maryland
College Park, MD
USA
ISBN 978-1-4020-6449-4
e-ISBN 978-1-4020-6450-0
Library of Congress Control Number: 2007940919

c 2008 Springer Science+Business Media B.V.
No part of this work may be reproduced, stored in a retrieval system, or transmitted in any form or by
any means, electronic, mechanical, photocopying, microfilming, recording or otherwise, without written
permission from the Publisher, with the exception of any material supplied specifically for the purpose of
being entered and executed on a computer system, for exclusive use by the purchaser of the work.
Cover illustration: From sensors and platforms, to information extraction and applications (compilation
by S. Liang)
Printed on acid-free paper
9 8 7 6 5 4 3 2 1
springer.com
Preface
This book is primarily based on presentations in the three reviewing panels of the
9th International Symposium on Physical Measurements and Signatures in Remote
Sensing held at the Institute for Geographical Sciences and Natural Resource Research, Chinese Academy of Sciences, China, in October 2005. It presents a collection of review papers on remote sensing sensor systems, radiation modeling and
inversion of land surface variables, and remote sensing applications. Each chapter
summarizes the progress in the past few years and also identifies the research issues
for the near future.
Acknowledgements
This symposium series is affiliated with the International Society for Photogrammetry and Remote Sensing (ISPRS) Commission VII/I Working Group on Fundamental Physics and Modeling led by Professor Michael Schaepman (Chair, Wageningen
University, the Netherlands), Professor Shunlin Liang (co-Chair, University of
Maryland, USA), and Dr. Mathias Kneubuehler (Secretary, University of Zurich,
CH, Switzerland) (20042012). It was sponsored and/or financially supported by
the Institute of Geographical Sciences and Natural Resources Research (IGSNRR)
of Chinese Academy of Sciences (CAS), Institute of Remote Sensing Applications
(IRSA) of CAS, Chinese 973 Project Quantitative Remote Sensing of Major Factors for Spatio-temporal Heterogeneity on the Land Surface undertaken by Beijing
Normal University, US National Aeronautics and Space Administration (NASA),
International Society for Photogrammetry and Remote Sensing, IEEE Geoscience
and Remote Sensing Society, and Scientific Data Center for Resources and Environment, CAS.
This symposium would not be successful without scientific leadership by the International Scientific Committee and effective organization by the local Organizing
Committee.
vii
International Scientific Committee
Professor Guanhua Xu, Minister of the Chinese Ministry of Science and Technology, China (honorary Chair)
Professor Shunlin Liang, University of Maryland, USA (Chair)
Dr. Frederic Baret, INRA, Avignon, France
Professor Mike Barnsley, University of Wales Swansea, UK
Professor Marvin Bauer, University of Minnesota, USA
Professor Jon Benediktsson, University of Iceland, Iceland
Professor Jing Chen, University of Toronto, Canada
Professor Peng Gong, University of California at Berkeley, USA
Dr. David Goodenough, Pacific Forestry Centre, Natural Resources Canada
Dr. Tom Jackson, USDA /ARS at Beltsville, Maryland, USA
Dr. David Jupp, CSIRO Earth Observation Centre, Australia
Dr. Yann Kerr, CNES/CESBIO, France
Dr. Marc Leroy, MEDIAS, France
Dr. Philip Lewis, University College London, UK
Professor Deren Li, Wuhan University, China
Professor Xiaowen Li, Beijing Normal University, China
Professor Jiyuan Liu, IGSNRR, CAS, China
Dr. John V. Martonchik, Jet Propulsion Laboratory, USA
Professor Ranga Myneni, Boston University, USA
Professor Ziyuan Ouyang, Institute of Geochemistry, CAS, China
Dr. Jeff Privette, NASA /GSFC, USA
Dr. Jon Ranson, NASA /GSFC, USA
Professor Michael Schaepman, Wageningen University, The Netherlands
Professor Jose Sobrino, University of Valencia, Spain
Dr. Karl Staenz, Canadian Centre for Remote Sensing, Canada
Professor Alan Strahler, Boston University, USA
Professor Qingxi Tong, Institute of Remote Sensing Applications, CAS, China
Dr. Michel Verstraete, JRC, Ispra, Italy
Dr. Charlie Walthall, USDA /ARS at Beltsville, Maryland, USA
Dr. Zhengming Wan, University of California at Santa Barbara, USA
ix
Organizing Committee
Professor Jiyuan Liu, Director of IGSNRR, CAS (co-Chair)
Professor Xiaowen Li, Director of Center for Remote Sensing and GIS, Beijing
Normal University and Director of IRSA, CAS (co-Chair)
Professor Dafang Zhuang, Execute Vice-director of Scientific Data Center for
Resources and Environment, CAS (Vice Chair)
Professor Mingkui Cao, IGSNRR, CAS
Professor Changqing Song, Chinese National Foundation of Sciences
Professor Renhua, Zhang, IGSNRR, CAS
Professor Lixin Zhang, Beijing Normal University
Dr. Keping Du, Beijing Normal University
Professor Mengxue Li, National Remote Sensing Center of China
Professor Boqin Zhu, Institute of Remote Sensing Applications, CAS
Dr. Ronggao Liu, IGSNRR, CAS (General Secretary)
Shunlin Liang
University of Maryland, College Park, MD, USA
Reviewers
Each chapter is anonymously reviewed by at least one reviewer. Their valuable comments and suggestions have greatly helped to improve the quality of the volume.
Jing M. Chen
University of Toronto, Canada
Jan Clevers
Wageningen University, The Netherlands
Ruth DeFries
University of Maryland, USA
Alan R. Gillespie
University of Washington, USA
Hongliang Fang
Xiuping Jia
The University of New South Wales, Australia
David L.B. Jupp
CSIRO Marine and Atmospheric Research, Australia
Yann H. Kerr
CNES/CESBIO, France
Yuri Knyazikhin
Boston University, USA
Randy Koster
NASA, USA
Eric F. Lambin
University of Louvain, Belgium
Tiit Nilson
Tartu Observatory, Estonia
xi
xii
Paolo Pampaloni
IFAC-CNR, Italy
Bernard Pinty
EC Joint Research Centre, Italy
Jeff Privette
NOAA, USA
Reviewers
Contents
Recent Advances in Land Remote Sensing: An Overview . . . . . . . . . .

Shunlin Liang
Part I Remote Sensing Systems

2
Passive Microwave Remote Sensing for Land Applications . . . . . . . . .

Thomas J. Jackson
Active Microwave Remote Sensing Systems and Applications

to Snow Monitoring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
Jiancheng Shi
Multi-angular Thermal Infrared Observations

of Terrestrial Vegetation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
Massimo Menenti, Li Jia, and Zhao-Liang Li
Terrestrial Applications of Multiangle Remote Sensing . . . . . . . . . . . . 95

Mark J. Chopping
Part II Physical Modeling and Inversion Algorithms

6
Modeling the Spectral Signature of Forests: Application of Remote

Sensing Models to Coniferous Canopies . . . . . . . . . . . . . . . . . . . . . . . . . 147
Pauline Stenberg, Matti Mottus, and Miina Rautiainen
Estimating Canopy Characteristics from Remote Sensing

Observations: Review of Methods and Associated Problems . . . . . . . . 173
Frederic Baret and Samuel Buis
Knowledge Database and Inversion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203

Jindi Wang and Xiaowen Li
xiii
xiv
Contents
Retrieval of Surface Albedo from Satellite Sensors . . . . . . . . . . . . . . . . 219

Crystal Schaaf, John Martonchik, Bernard Pinty, Yves Govaerts,
Feng Gao, Alessio Lattanzio, Jicheng Liu, Alan Strahler, and Malcolm
Taberner
10
Modeling and Inversion in Thermal Infrared Remote Sensing

over Vegetated Land Surfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245
Frederic Jacob, Thomas Schmugge, Albert Olioso, Andrew French,
Dominique Courault, Kenta Ogawa, Francois Petitcolin,
Ghani Chehbouni, Ana Pinheiro, and Jeffrey Privette
11
Spectrally Consistent Pansharpening . . . . . . . . . . . . . . . . . . . . . . . . . . . 293

Ari Vesteinsson, Henrik Aanaes, Johannes R. Sveinsson,
and Jon Atli Benediktsson
12
Data Assimilation Methods for Land Surface Variable Estimation . . 313

Shunlin Liang and Jun Qin
13
Methodologies for Mapping Land Cover/Land Use and its Change . . 341
Nina Siu-Ngan Lam
14
Methodologies for Mapping Plant Functional Types . . . . . . . . . . . . . . 369

Wanxiao Sun and Shunlin Liang
Part III Remote Sensing Applications

15
Monitoring and Management of Agriculture with Remote Sensing . . 397

Zhongxin Chen, Sen Li, Jianqiang Ren, Pan Gong, Mingwei Zhang,
Limin Wang, Shenliang Xiao, and Daohui Jiang
16
Remote Sensing of Terrestrial Primary Production

and Carbon Cycle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 423
Maosheng Zhao and Steven W. Running
17
Applications of Terrestrial Remote Sensing to Climate Modeling . . . . 445

Robert E. Dickinson
18
Improving the Utilization of Remotely Sensed Data . . . . . . . . . . . . . . . 465

John R. Townshend, Stephen Briggs, Roy Gibson, Michael Hales,
Paul Menzel, Brent Smith, Yukio Haruyama, Chu Ishida, John Latham,
Jeff Tschirley, Deren Li, Mengxue Li, Liangming Liu,
and Gilles Sommeria
19
Emerging Issues in Land Remote Sensing . . . . . . . . . . . . . . . . . . . . . . . 485

Shunlin Liang, Michael Schaepman, Thomas J. Jackson, David Jupp,
Xiaowen Li, Jiyuan Liu, Ronggao Liu, Alan Strahler,
John R. Townshend, and Diane Wickland
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 495
CD-ROM included
Contributors
Henrik Aanaes
Informatics and Mathematical Modelling, Technical University of Denmark,
Denmark
haa@imm.dtu.dk
Frederic Baret
UMR1114, INRA-CSE, 84914 Avignon, France
baret@avignon.inra.fr
Jon Atli Benediktsson
Department of Electrical and Computer Engineering, University of Iceland,
Hjardarhaga 2-6, 107 Reykjavik, Iceland
benedikt@hi.is
Stephen Briggs
European Space Agency, Via Galileo Galilei, 00644 Frascati, Rome, Italy
stephen.briggs@esa.int
Sanuel Buis
UMR1114, INRA-CSE, 84914 Avignon, France
Ghani Chehbouni
Institute of Research for the Development, Center for Spatial Studies of the
Biosphere, UMR CESBio, Toulouse, France
ghani.chehbouni@cesbio.cnes.fr
Zhongxin Chen
Key Laboratory of Resource Remote Sensing & Digital Agriculture, Ministry
of Agriculture, Beijing 100081, China
chzhx@yahoo.com
Mark J. Chopping
Department of Earth and Environmental Studies, Montclair State University,
1 Normal Ave, Montclair, NJ 07043, USA
chopping@pegasus.montclair.edu
xv
xvi
Contributors
Dominique Courault
National Institute for Agronomical Research, Climate Soil Environment Unit,
UMR CSE INRA / UAPV, 84914 Avignon, France
courault@avignon.inra.fr
Robert E. Dickinson
School of Earth and Atmospheric Sciences, Georgia Institute of Technology,
311 Ferst Drive, Atlanta, GA 30332-0340, USA
robted@eas.gatech.edu
Andrew French
United States Department of Agriculture/Agricultural Research Service, US Arid
Land Agricultural Research Center, 21881 North Cardon Lane, Maricopa,
AZ 85238, USA
andrew.french@ars.usda.gov
Feng Gao
Earth Resources Technology, Inc., 8106 Stayton Dr., Jessup, MD 20794, USA
Roy Gibson
EUMETSAT
roy.gibson@wanadoo.fr
Pan Gong
Institute of Agricultural Resources & Regional Planning, Chinese Academy
of Agricultural Sciences, Beijing 100081, China
Yves Govaerts
EUMETSAT, Am Kavalleriesand 31, D-64295 Darmstadt, Germany
yves.govaerts@eumetsat.int
Michael Hales
NOAA, Silver Spring, MD 20910, USA
michael.hales@noaa.gov
Yukio Haruyama
JAXA, 1-8-10 Harumi, Chuo-ku, Tokyo 104-6023, Japan
haruyama.yukio@jaxa.jp
Chu Ishida
JAXA, 1-8-10 Harumi, Chuo-ku, Tokyo 104-6023, Japan
ishida.chu@jaxa.jp
Thomas J. Jackson
USDA ARS Hydrology and Remote Sensing Lab, 104 Bldg. 007 BARC-West,
Beltsville, MD 20705,
tjackson@hydrolab.arsusda.gov
Frederic Jacob
Institute of Research for the Development, Laboratory for studies on Interactions
between Soils Agrosystems Hydrosystems, UMR LISAH SupAgro/INRA/IRD,
Montpellier, France
frederic.jacob@supagro.inra.fr
Contributors
xvii
Li Jia
Alterra Green World Research, Wageningen University and Research Centre,
The Netherlands
li.jia@wur.nl
Daohui Jiang
David Jupp
CSIRO Marine and Atmospheric Research, Canberra ACT 2601 Australia
David.Jupp@csiro.au
Nina Siu-Ngan Lam
Department of Environmental Studies, Louisiana State University, Baton Rouge,
LA 70808, USA
nlam@lsu.edu
John Latham
Food and Agriculture Organization, Rome, Italy
john.latham@fao.org
Alessio Lattanzio
Makalumedia gmbh, Robert-Bosch Strasse 7, 64296 Darmstadt, Germany
Deren Li
Wuhan University, 39 Loyu Road, Wuhan, 430070, China
dli@wtusm.edu.cn
Mengxue Li
NRSCC, 15B, Fuxing Road, Beijing, 100862, China
mengxueli@hotmail.com
Sen Li
Xiaowen Li
Research Center for Remote Sensing and GIS, Beijing Normal University, No.19
XieJieKouWaiDaJie Street, Beijing 100875, China
lix@bnu.edu.cn
Zhao-Liang Li
Institute of Geographic Sciences and Natural Resources Research, Beijing, China
Shunlin Liang
Department of Geography, University of Maryland, College Park, USA
sliang@umd.edu
Jicheng Liu
Department of Geography and Environment, Boston University,
675 Commonwealth Ave., Boston, MA 02215, USA
xviii
Contributors
Jiyuan Liu
Institute for Geographical Sciences and Natural Resource Research, Chinese
Academy of Sciences, Beijing, China
liujy@igsnrr.ac.cn
Liangming Liu
National Remote Sensing Center of China
Ronggao Liu
Institute for Geographical Sciences and Natural Resource Research, Chinese
Academy of Sciences, Beijing, China
liurg@lreis.ac.cn
John Martonchik
Jet Propulsion Laboratory, Mail Stop 169-237, 4800 Oak Grove Dr., Pasadena,
CA 91109, USA
john.v.martonchik@jpl.nasa.gov
Massimo Menenti
TRIO/LSIIT, University Louis Pasteur (ULP), Strasbourg, France and Istituto per i
Sistemi Agricoli e Forestali del Mediterraneo (ISAFOM), Naples, Italy
m.menenti@isafom.cnr.it
Paul Menzel
University of Wisconsin, Space Science and Engineering Center, Madison,
WI 53706, USA
paulm@ssec.wisc.edu
Matti Mottus
Department of Forest Ecology, FI-00014 University of Helsinki, Finland; Tartu
Observatory, 61602 Toravere, Tartumaa, Estonia
mottus@ut.ee
Kenta Ogawa
Department of Geo-system Engineering, University of Tokyo and Hitachi Ltd,
Tokyo, Japan
Albert Olioso
National Institute for Agronomical Research, Climate Soil Environment Unit,
UMR CSE INRA/UAPV, Avignon, France
olioso@avignon.inra.fr
Francois Petitcolin
ACRI-ST, Sophia Antipolis, France
ptc@acri-st.fr
Ana Pinheiro
Biospheric Sciences Branch, NASAs GSFC, Greenbelt, MD 20771, USA
ana.pinheiro@gsfc.nasa.gov
Contributors
xix
Bernard Pinty
Global Environment Monitoring Unit, IES, EC Joint Research Centre, TP 440,
via E. Fermi, I-21020 Ispra (VA), Italy
bernard.pinty@jrc.it
Jeffrey Privette
NOAAs National Climatic Data Center, Asheville, NC 28801-5001, USA
jeffrey.privette@noaa.gov
Jun Qin
Institute for Geographical Science and Natural Resource Research, Beijing, China
shuairenqin@gmail.com
Miina Rautiainen
Department of Forest Resource Management, FI-00014 University of Helsinki,
Finland
miina.rautiainen@helsinki.fi
Jianqiang Ren
Key Laboratory of Resource Remote Sensing & Digital Agriculture, Ministry of
Agriculture, Beijing 100081, China and Institute of Agricultural Resources &
Regional Planning, Chinese Academy of Agricultural Sciences, Beijing 100081,
China
Steven W. Running
Numerical Terradynamic Simulation Group, Department of Ecosystem and
Conservation Science, University of Montana, Missoula, MT 59812, USA
swr@ntsg.umt.edu
Crystal Schaaf
schaaf@bu.edu
Michael Schaepman
Centre for Geo-Information, Wageningen University, Wageningen, The Netherlands
Michael.Schaepman@wur.nl
Thomas Schmugge
Gerald Thomas Professor of Water Resources, College of Agriculture New Mexico
State University, Las Cruces, NM, USA
schmugge@nmsu.edu
Jiancheng Shi
Institute for Computational Earth System Science, University of California,
Santa Barbara, CA 93106-3060, USA
shi@icess.ucsb.edu
Brent Smith
NOAA, Silver Spring, MD 20910, USA
brent.smith@noaa.gov
xx
Contributors
Gilles Sommeria
World Climate Research Programme, WMO, CH-1211 Geneva, Switzerland
gsommeria@wmo.int
Pauline Stenberg
Department of Forest Resource Management, FI-00014 University of Helsinki,
Finland
Pauline.Stenberg@helsinki.fi
Alan Strahler
alan@bu.edu
Wanxiao Sun
Department of Geography and Planning, Grand Valley State University, USA
sunwa@gvsu.edu
Johannes R. Sveinsson
sveinsso@hi.is
Malcolm Taberner
Global Environment Monitoring Unit, IES, EC Joint Research Centre, TP 440,
via E. Fermi, I-21020 Ispra (VA), Italy
malcolm.taberner@jrc.it
John R. Townshend
Department of Geography, University of Maryland, College Park, MD 20742, USA
jtownshe@umd.edu
Jeff Tschirley
Food and Agriculture Organization, 00153 Rome, Italy
jeff.tschirley@fao.org
Ari Vesteinsson
Jindi Wang
Research Center for Remote Sensing and GIS, Beijing Normal University, No.19
XieJieKouWaiDaJie Street, Beijing 100875, China
wangjd@bnu.edu.cn
Limin Wang
Key Laboratory of Resource Remote Sensing & Digital Agriculture, Ministry
of Agriculture, Beijing 100081, China
Diane Wickland
NASA Headquarters, Washington, DC, USA
diane.e.wickland@nasa.gov
Contributors
Shenliang Xiao
Mingwei Zhang
Maosheng Zhao
Numerical Terradynamic Simulation Group, Department of Ecosystem
and Conservation Science, University of Montana, Missoula, MT 59812, USA
zhao@ntsg.umt.edu
xxi
Chapter 1
Recent Advances in Land Remote Sensing:

An Overview
Shunlin Liang
Earths surface is undergoing rapid changes due to urbanization, industrialization

and globalization. Environmental problems such as water shortages, desertification,
soil depletion, greenhouse gas emissions warming the atmosphere, deforestation,
elevated coastal waterway sediment and nutrient fluxes, among other environmental
problems, are increasingly common and troubling consequences of human activities. Policy decisions about the environment rely on accurate and reliable information, especially data and understanding leading to better predictions of natural
hazards, epidemics, impacts of energy choices, and climate variations. Comprehensive, systematic Earth observations are key to forecasting Earth system dynamics.
Predicting future scenarios of our planets habitability requires analysis of what has
transpired in the past along with observations of present conditions and processes.
Timely, quality long-term global data acquired through remote sensing is essential
for the ongoing viability and enhancement of human society on Earth.
The field of remote sensing (Earth observation) has developed rapidly. Many
publications have documented its progress (e.g., the special issue of Remote Sensing Reviews with a set of papers reviewing the modeling and inversion of surface
bidirectional reflectance distribution function (BRDF) (Liang and Strahler, 2000),
and the edited or authored books on similar subjects (Liang, 2004; Myneni and Ross,
1991). To systematically summarize the achievements of terrestrial remote sensing
in recent years and to set the research agenda for the near future, the 9th International Symposium on Physical Measurements and Signatures in Remote Sensing
(ISPMSRS), held in October 2005 in Beijing, organized three review panels. The
papers compiled in this book are largely from these panels and are organized into
three parts, respectively.
Part I of the book (corresponding to the first panel, chaired by Drs. David Jupp
and Tom Jackson, with Drs. Ralph Dubayah, Michael Schaepman, Jianchen Shi,
Stephen Ungar, and David Le Vine) focuses on remote sensing systems and sensors. As there are many different remote sensing systems (the result of various
Shunlin Liang
S. Liang (ed.), Advances in Land Remote Sensing, 16.
c Springer Science + Business Media B.V., 2008
S. Liang
permutations of passive vs. active, specific domain of the electromagnetic spectrum sampled, and spectral channel bandwidth) currently in operation or planned to
be operational in the future, this panel evaluated the capabilities of these systems
for estimating key land surface variables and how they can best be improved and
combined effectively. The panel presentations covered microwave, thermal-infrared
(IR), hyperspectral optical, and Lidar remote sensing. Based on these discussions,
four chapters are compiled on this subject, including passive microwave, active
microwave, multiangle thermal IR, and multiangle optical remote sensing.
In Chapter 2 on passive microwave remote sensing, Jackson identifies six factors in passive microwave sensor design that affect the retrieval of land surface
properties: frequency, polarization, view geometry, spatial resolution, temporal coverage, and signal-to-noise ratio. While summarizing the features of three current
relevant satellite instruments and two other satellite sensors, he points out that the
low frequency observations (<6 GHz) and coarse spatial resolution are obstaclesto-land applications. He further discusses three approaches, currently under investigation, for solving these problems. The first is synthetic aperture radiometry, such
as the planned Soil Moisture Ocean Salinity mission. The second is use of large,
lightweight antennas and the Hydros satellite mission (recently cancelled) which
could have demonstrated this approach. The third is disaggregation by integrating
with higher resolution visible to thermal remote sensing data through data fusion.
The next decade will see several exploratory missions using new technologies,
and innovative approaches to integrating passive microwave with active microwave
measurements.
Shi summarizes, in Chapter 3 on active microwave remote sensing, the significant
developments in mapping snow, inferring snow wetness and snow water equivalence
using active microwave sensors (mainly synthetic aperture radar SAR). Following
a brief introduction to different SAR systems, including current and future systems,
he discusses three types of SAR measurements: backscattering measurements at a
given frequency and polarization, polarization properties obtained by a fully polarimetric instrument, and interferometry from repeat-pass sampling. Shi devotes more
space to reviewing various techniques that use these three measurements to map
snow properties, followed by a discussion of the need for future SAR systems to
map snow properties.
Multiangle thermal-IR remote sensing is covered in Chapter 4. Menenti et al.
first present experimental evidence of angular signatures of thermal-IR observations, then demonstrate that two soil and foliage temperatures normally can account
for most thermal IR observations at the top of the canopy. (further division into
sunlit and shadowed components are needed only under extreme conditions). They
review various geometric-optics and radiative transfer modeling techniques, and
finally describe inversion techniques to estimate component temperatures from multiangle thermal-IR observations. The related subject on estimating land surface skin
temperature is given in Chapter 10.
In Chapter 5 on multiangle optical remote sensing, Chopping first outlines different sensor systems for acquiring multiangle data, and then reviews a variety of
land applications using multiangle observations, such as mapping canopy openness,
1 Recent Advances in Land Remote Sensing: An Overview
clumping index, structural scattering index, land cover and community type, snow
and ice, and dust emissions. He also outlines different modeling techniques for
angular signatures in the optical domain and discusses other factors (e.g., nearsimultaneous and accumulated angular sampling, scale) related to multiangle data
acquisition.
Although these four chapters cover a variety of sensors and applications, two
other remote sensing techniques deserve mention. Lidar has a unique ability to measure vertical and spatial heterogeneity across multiple scales. Many studies have
demonstrated that lidar can be used for estimating forest structure variables such as
canopy height and profile, canopy cover, growth/successional state, foliar profile,
stem density, basal area, biomass, and canopy base height. Except for IceSat, currently there are no spaceborne lidar remote sensing systems available. Hyperspectral
remote sensing is another approach that has great potential.
A more challenging issue faced by the remote sensing community is how to
integrate data from different sensor systems in such a way that derivation of land
surface variables is maximized.
Part II of the book presents surface radiation modeling and inversion methods
corresponding to the second panel chaired by Drs. Alan Strahler and Xiaowen Li,
and contributed by Drs. Frederic Baret, Frederic Jacob, Philip Lewis, Massimo
Menenti, Pauline Stenberg, and Jindi Wang. There have been considerable investments on developing physical models to understand the surface radiation regimes.
Some of these models have been incorporated into very useful algorithms for
estimating land surface variables from satellite observations. This panel assessed
the progress of surface radiation modeling and satellite inversion algorithm development in many different aspects. The papers in this part are composed of panel
presentations and two additional contributions. Note that Chapters 4 and 5 also contain modeling and inversion mainly for multiangle observations.
Stenberg et al. in Chapter 6 describe different geometric-optical models and radiative transfer models for simulating the reflectances of heteorogenerous canopies,
particularly forests. In radiative transfer modeling, they review mainly the threedimensional models. They also discuss the structural and spectral characteristics of
coniferous forests, the clumping effects and reflectance scaling using the p-theory,
and the inputs for simulating the forest reflectance.
Baret and Buis in Chapter 7 review various methods for estimating canopy properties from remotely sensed data. They first describe the empirical regression-type
algorithms, using either experimental data or radiative transfer simulated data, and
different model inversion algorithms, including optimization algorithms, look-up
table algorithms, and two stochastic inversion algorithms. They then discuss the
nature of ill-posed problem in the inversion of canopy properties, and finally summarize two methods to address the ill-posed issue by using the a priori information
or posing additional spatial, temporal and model coupling constraints.
Wand and Li in Chapter 8 present a knowledge database and discuss its applications to improving the inversion of land surface information from remotely sensed
data. They first discuss the need for a prior knowledge due to the ill-posed problem
in remote sensing inversion that is also discussed by Baret and Buis in Chapter 7,
S. Liang
then describe the database they have developed and finally demonstrate how it has
been used in different inversion studies. They also address the parameter scaling and
validation issues.
Schaaf et al. in Chapter 9 summarize the current algorithms for estimating
land surface broadband albedo and anisotropy from satellite observations. After
providing some general background, they present the specific algorithms for
three representative sensors: MODIS (multispectral), MISR (multiangle) and Meteosat (geostationary) and also discuss how to integrate historic and current albedo
products.
In Chapter 10, Jacob et al. comprehensively review various modeling methods
of thermal signatures and inversion techniques for estimating land surface skin temperature from thermal-IR remotely sensed data. They start with the definitions of
temperatures and emissivities, and analysis of spatial, spectral, angular and temporal
signatures in thermal-IR remote sensing. Various modeling techniques and inversion
methods are then evaluated. They further examine current research issues, including
land surface brightness temperature from atmospheric correction or model simulation, ensemble emissivity and radiometric temperature of mixed surface types from
different observation capabilities, aerodynamic temperature, and directional effects
of temperature and emissivity. The potential applications of thermal-IR remote sensing are also discussed.
Vesteinsson et al. in Chapter 11 present a new method to integrate remote sensing
data of different spatial resolutions. After reviewing various image fusion methods
in four major types (Frequency, color transformation, statistical, and hierarchical),
they present a new data fusion framework. They model the fusion issue as a minimization problem where the objective function is the energy of a Gibbs distribution
consisting of three energy terms: spectral consistency constraint, smoothness constraint, and imaging physics constraint. Minimization of the objective function is
sought using stochastic optimization. This framework is finally implemented using
two IKONOS satellite images by integrating the panchromatic band with multispectral bands.
Data assimilation methods for estimating land surface variables are presented
by Liang and Qin in Chapter 12. After introducing the basic principles of data
assimilation, they identify a series of critical issues, such as data and products to
be assimilated, parameters to be estimated, assimilation algorithms to be used, error
matrices to be determined, and imperfect numerical models. The latest applications
of data assimilation methods to various fields (e.g., hydrology, carbon cycle, agricultural productivity) are also reviewed.
Lam in Chapter 13 introduces the use of textural/spatial measures to automated
land cover classification and change detection. She first identifies the major criteria for evaluating textural measures and then shows different examples in using
them to improve classification accuracy. She also summarizes the existing methods into a framework, and then argues that the textural approach has potential for
rapid change detection. She finally demonstrates the use of textural measures solely
(not including spectral information) for change detection in New Orleans before and
after Hurricane Katrina.
1 Recent Advances in Land Remote Sensing: An Overview
In Chapter 14, Sun and Liang review the existing methods for mapping plant
functional types (PFTs) at regional to global scale from remote sensing data. Traditionally, land surface models represent vegetation as discrete biomes that are not
natural vegetation units but are products of classification. Since most land models
are expanding beyond their traditional biogeophysical roots to include biogeochemistry, especially photosynthesis and the carbon cycle, the land modeling community
has started using PFTs to represent land surface. There is an urgent need for mapping PFT using remote sensing. The authors finally present a multisource evidential
reasoning data fusion framework using high-level land products for improved mapping of PFTs from satellite observations.
These chapters in Part II of the book cover a variety of topics related to land
surface radiation modeling and inversion. I must emphasize the importance of validation here. The new modeling techniques, new inversion methods, and newly generated products require thorough validation. Though a specific chapter on validation
is not presented here, a recently published special issue on this topic (Morisette
et al., 2006) provides many relevant details.
Part III of the book discusses remote sensing applications corresponding to the
last panel chaired by Drs. John Townshend and Jiyuan Liu, and with contributions
by Drs. Zhongxin Chen, David Goodenough, and Diane Wickland. Remote sensing
science driven by applications will have more widespread use and benefit larger user
communities. Significant disconnections between remote sensing development and
applications persist. Some products developed by remote sensing scientists have
not been widely utilized. Many variables required by surface process models and
decision support systems have not been generated. Product accuracy and application
requirements may not always be consistent. Even the same variables may be defined
differently in remote sensing algorithms and application models. This panel provided a useful dialogue between remote sensing scientists and application experts.
This part of the book includes some of the discussions and also presents two more
chapters on this topic.
Chen et al. in Chapter 15 review the remote sensing applications pertinent to agricultural monitoring and management. They discuss progress in four subjects: crop
identification and mapping, crop yield estimation and prediction, crop phenology
monitoring, and soil moisture estimation from optical to microwave remote sensing.
They demonstrate the need for multisensor data fusion, use of multiple signatures,
and integration of data and numerical models.
Zhao and Running in Chapter 16 review the historical development and the recent
advances in the application of satellite remote sensing data for estimating terrestrial
productivity and monitoring carbon cycle-related ecosystem dynamics and changes.
They first outline the development of using vegetation index for estimating land surface biophysical variables, then describe the MODIS GPP and NPP products and the
findings from the vegetation-related products. They further discuss the application
of long-term satellite data to the study of terrestrial ecosystems, including phenology
monitoring, changes in regional carbon storage resulting from land use change, carbon flux changes induced by climate change, disturbance detection, and validation
of ecosystem models. They also propose a scheme for an integrated study of carbon
dynamics.
S. Liang
In Chapter 17, Dickinson discusses applications of terrestrial remote sensing to

climate modeling. Beginning with an introduction to the formulation of the climate
models, he then describes land surface models, their components, and energy and
water balance requirements. He reviews the contributions from many land remote
sensing products (e.g., LAI, albedo and temperature) in terms of roles of solar
radiation in climate model terrestrial system, and summarizes various assessments
of the climate models when using these products. He further considers how terrestrial remote sensing can better support climate models and eventually be a component of climate prediction through data assimilation.
Townshend et al. in Chapter 18 identify and analyze factors either facilitating or
hampering the increased use of satellite data, products and services by compiling
25 case studies. These factors include technical, education and capacity building,
financial and policy. They discover that successful adoption of remote sensing products always involves balanced cooperation between space agencies and users. They
derive a set of principles that will lead to enhanced usage of space-based data and
products and summarize the recommendations to the space agencies.
Liang et al. in Chapter 19 summarize the key questions and issues discussed
by three review panels and identify some emerging issues in land remote sensing,
including sensor networks, modeling complex landscapes, machine learning techniques for inversion, and spatial scaling.
Finally, note that two special issues in Photogrammetric Engineering and Remote
Sensing (October 2007), and Journal of Remote Sensing (September 2007) from the
9th ISPMSRS add to and complement the review papers in this book. Readers are
referred to these special issues for details.
References
Liang S (2004) Quantitative Remote Sensing of Land Surfaces. Wiley, New York, 534pp
Liang S, Strahler A (guest eds) (2000) Land surface bi-directional reflectance distribution function
(BRDF): recent advances and future prospects. Special issue of Remote Sens. Rev. 18:83511
Morisette J, Baret F, Liang S (guest eds) (July 2006) Global land product validation. Special issue
of IEEE Trans. Geosci. Remote Sens. 44(7):16951937
Myneni RB, Ross J (eds) (1991) Photon-vegetation Interactions: Applications in Optical Remote
Sensing and Plant Physiology. Springer, New York
Part I
Remote Sensing Systems
Chapter 2
Passive Microwave Remote Sensing for Land

Applications
Thomas J. Jackson
Abstract Land applications, in particular soil moisture retrieval, have been hampered by the lack of low frequency passive microwave observations and the coarse
spatial resolution of existing sensors. The next decade could see several improved
operational and exploratory missions using new technologies as well as innovative
disaggregation and data fusion approaches that could lead the way to an order of
magnitude improvement in spatial resolution.
Keywords: Passive microwave soil moisture land surface
2.1 Introduction
Passive microwave remote sensing has made major contributions in atmospheric and
oceanic sciences. These applications have exploited higher frequencies and used low
frequencies to establish background conditions. Land applications have been hampered by the availability of low frequency observations (<6 GHz) and coarse spatial
resolution. Conventional technologies and approaches to retrievals have limited spatial resolution to the 50 + km range, which has in turn limited the potential usage
to only very large-scale studies. Three factors will affect land applications of passive microwave remote sensing over the next decade: the operational low frequency
instruments available, exploratory missions using new technologies, and innovative
approaches to disaggregating the coarser resolution passive microwave observations
with active microwave measurements. The current satellite sensors, criteria for mission design, and future missions are presented in this chapter. Because soil moisture
is one of the most important and challenging problems the discussion will focus on
this topic.
Thomas J. Jackson
USDA ARS Hydrology and Remote Sensing Lab, Beltsville, MD, USA
tjackson@hydrolab.arsusda.gov
10
T.J. Jackson
2.2 Satellite Passive Microwave Sensor Systems

Low frequency passive microwave remote sensing provides information on the
dielectric and temperature properties of the Earths atmosphere and surface as well
as some characterization of its geometric features. The dielectric properties are
dependent upon the water content of the target. As a result, passive microwave remote sensing has been useful in studies of the atmosphere, oceans, snow, ice, and
land. The breadth of parameters and variables that can be retrieved is illustrated in
Table 2.1 for the Advanced Microwave Scanning Radiometer (AMSR-E) instrument
on the National Space and Aeronautics Administration (NASA) Aqua satellite.
Land applications of passive microwave remote sensing have included classification, temperature, vegetation characteristics, and soil moisture. Current operational
sensors such as the Special Sensor Microwave/Imager (SSM/I) have been used in
land cover type classification (Neale et al., 1990), land surface temperatures estimation and in deriving a general surface wetness parameter (Basist et al., 1998).
System design for land variable retrieval includes at least six considerations:
Frequency
Polarization
Viewing geometry
Spatial resolution
Temporal coverage
Signal to Noise
2.2.1 Frequency Selection

Lower frequencies provide greater sensitivity to changes in soil moisture through
a greater range of vegetation cover conditions. For soil moisture retrieval these are
very important considerations (Jackson and Schmugge, 1991). Figure 2.1 illustrates
this point using a sensitivity parameter (degrees K/% soil moisture) and identifies
several key satellite instruments. For bare soil there is little change in sensitivity as a
Table 2.1 Earth variables derived from AMSR-E

Category
Variable
Atmosphere
Total integrated water vapor over the ocean

Rainfall over ocean
Sea surface temperature
Ocean surface wind speed
Sea ice concentration
Snow depth over sea ice
Sea ice temperature
Snow-cover water equivalent over land
Surface soil wetness
Oceans
Sea ice
Snow
Land
2 Passive Microwave Remote Sensing for Land Applications

High
SMOS
Low
AMSR-E Meteorological
Satellites
Bare
Sensitivity
Fig. 2.1 Sensitivity of

brightness temperature to
soil moisture as a function of
microwave frequency
11
Vegetated
3
5
10
20 30
Frequency (GHz)
50
function of frequency (although) there is a decrease in sensing depth with increasing

frequency. In the presence of vegetation the impact of frequency is much more significant. Surface roughness can also decrease sensitivity. Another aspect of Fig. 2.1
to note is that research and applications were limited to the lowest frequencies of the
meteorological satellites for many years. When AMSR-E and other recent satellite
instruments were launched a significant improvement in soil moisture retrieval was
anticipated.
Another key issue in frequency selection is radio frequency interference. This
has become a significant problem with AMSR-E, primarily at 6.9 GHz in the USA.
Contamination of the signal has rendered the data unusable and as a result applications have had to adapt to using a less desirable higher frequency (Li et al., 2004;
Njoku et al., 2005). As shown in Fig. 2.1, the shift in frequency from 6.9 GHz to
10.7 GHz for soil moisture retrieval results in a reduction in sensitivity and loss of
information. It should be noted that the 6.9 GHz band is not a protected radio frequency, however, 10.7 GHz is protected and still is contaminated in some regions of
the world.
2.2.2 Polarization
It is well known that horizontal polarization is much more sensitive to changes in
the soil dielectric constant than vertical polarization measurements. Simple computations using radiative transfer equations can readily verify this fact (Ulaby et al.,
1982). However, vertical polarization can be very useful in normalizing the horizontal measurements. Information on the polarization difference has been used to characterize the vegetation in several studies and algorithms (Becker and Choudhury,
1988; Paloscia et al., 2001; Owe et al., 2001; Njoku et al., 2003). With the recent
launch of the WindSat instrument (Gaiser et al., 2004) with a fully polarimetric microwave radiometer, investigators are beginning to examine what new information
about the land may be contained in the additional channels the sensor provides
(Narvekar et al., 2007).
12
T.J. Jackson
2.2.3 Viewing Geometry

Nearly all satellite microwave radiometers used for land studies have employed
conical scanning. Conical scanning provides constant incidence angle observations, which simplifies the retrieval process. Higher incidence angles result in wider
swaths and increased temporal coverage. So, there are good reasons for choosing
this design.
However, the highest sensitivity to soil moisture will be at low incidence angles.
Radiative transfer relationships include the angle in computation. In addition, the
path length through the vegetation canopy will increase with angle. At low incidence
angles the path attenuation will be smaller. Some algorithm designs actually exploit
these changes with incidence angle in their retrievals (Wigneron et al., 2000). In
this approach, a correction for vegetation is made in performing the soil moisture
retrieval by using several incidence angle observations.
2.2.4 Spatial Resolution

Spatial resolution has been the challenge for passive microwave remote sensing
of land. Unlike atmospheres and oceans, the land exhibits heterogeneity at several
scales that may be significant. With low frequencies it is extremely difficult to design an antenna system necessary for a high resolution observing system. This topic
will be covered in depth in a later section.
2.2.5 Temporal Coverage

With conical scanning low-resolution sensors, the temporal coverage has not been a
major issue for applications. Wide swaths of current and future sensors result in 13
day global coverage.
2.2.6 Signal to Noise

To date, using conventional antenna technology, signal to noise has not been an issue
for land applications (soil moisture) because the range of response and sensitivity to
the geophysical parameters are quite large.
13
2.3 Selected Conical Scanning Instruments

Three current satellite instruments are of interest in passive microwave sensing of
land; SSM/I, AMSR-E, and WindSat. Features of each mission are summarized in
Table 2.2. There have been several other important satellites for land remote sensing
that include the Scanning Multifrequency Microwave Radiometer (SMMR) (Owe
et al., 1992) and the Tropical Rainfall Measurement Mission Microwave Imager
(TMI) (Bindlish et al., 2003).
Worth noting in Table 2.2 is the fact that the SSM/I instrument now has an over
20-year period of record. Interpreting data from the SSM/I to extract surface information requires accounting for atmospheric effects on the measurement. When one
considers the atmospheric correction, the significance of vegetation attenuation, and
the shallow contributing depth of soil for these high frequencies, it becomes apparent that the data are of limited value for estimating soil water As a result, SSM/I
data has been applied to only a limited set of land applications including a soil wetness product (Basist et al., 1998). Some successful attempts have been made at soil
moisture retrieval under selected conditions (Jackson, 1997).
Aqua was launched in May 2002 and ADEOS-II was also launched in the same
year. Each included an AMSR instrument. ADEOS-II was lost after a few months
of operation but Aqua is still providing data. As shown in Table 2.1 AMSR-E was to
provide significantly lower frequency measurements than the SSM/I with the same
level of spatial resolution. AMSR-E holds great promise for estimating soil water
content in sparsely vegetated regions and is the best possibility in the near term for
mapping soil water.
As opposed to previous passive microwave satellite missions, both NASA (Njoku
et al., 2003) and the Japanese Aerospace Exploration Agency (JAXA) (Njoku et al.,
2000) have included soil moisture as a mission product from AMSR-E. As noted
previously, it is not expected that AMSR-E can provide a globally reliable soil moisture product. However, it will work in some regions and the efforts at operational
implementation and validation of products are providing valuable lessons for future
missions.
Table 2.2 Selected satellite conical scanning microwave radiometers
Frequency (GHz)
85/91
37
22/23
19
10.8
6.9/6.8
1.4
37 IFOV (km)
Observing time
Period of record
Satellite Instrument
SSM/I
AMSR-E
WindSat
V, H
V, H
V
V, H
27 38
Various
1985
V, H
V, H
V, H
V, H
V, H
V, H
8 14
1:30 p.m.
2002
V, H, U, F
V, H
V, H, U, F
V, H, U, F
V, H
8 13
6:00 a.m.
2003
14
T.J. Jackson
The third satellite instrument is WindSat that launched in 2003 and includes a
multifrequency passive microwave radiometer system with a C band channel. This
is a risk reduction mission for one component of the next generation of operational
polar orbiting satellites that the U.S. will be implementing.
Regarding operational sensors, the Advanced Microwave Scanning Radiometer
(AMSR-E) on Aqua will transition to a component of the long term Japanese Global
Climate Observing Mission (GCOM) and WindSat may evolve to the Conical Microwave Imaging System (CMIS) on National Polar Orbiting Earth Satellite System
(NPOESS). The lower frequencies and commitments to both data and algorithms
should result in better products for land studies despite the rather coarse spatial resolution.
Japan provided the AMSR-E instrument to the Aqua platform and has plans to
follow this up with an improved version on the Global Climate Observing MissionsWater (GCOM-W) satellites possibly in 2010. It is uncertain what the USA will
be doing as a follow on to WindSat and SSM/I. Until recently plans called for a
Conical Microwave Imaging System (CMIS) as part of the National Polar Orbiting Earth Satellite System (NPOESS). CMIS would be similar to WindSat, offering
multifrequency (including C and X band) as well as fully polarimetric observations.
However, issues with instruments and costs are expected to impact the final design
and delay the current launch date. Another source of passive microwave measurements will be satellites involved in the Global Precipitation Mission that follows the
successful TMI. This mission is scheduled for 2012.
2.4 New Directions and Missions

As noted previously, achieving a spatial resolution that can satisfy land application
requirements (or even demonstrate an approach that could open up future options)
has been the challenge to passive microwave remote sensing. The problem has been
achieving high spatial resolution at lower frequencies requires a large antenna. Getting large and heavy antennas into space is obviously difficult and costly. There are
at least three approaches to solving this problem that are currently under investigation:
Synthetic aperture radiometry
Large lightweight antennas
Data fusion-disaggregation using higher resolution remote sensing techniques
Synthetic aperture radiometry attempts to solve the spatial resolution dilemma by
replacing the traditional large filled array with a sparse array. It can achieve what
a very large antenna would, but with a much smaller mass and size. Earlier studies
established one dimensional synthetic aperture methods (Le Vine et al., 1994) and
demonstrated that this approach could be used for soil moisture retrieval (Jackson
et al., 1995). One-dimensional methods use a real aperture along track and synthetic
aperture methods across track, essentially a set of long sticks.
15
The European Space Agency (ESA) will take synthetic aperture radiometry a
step further in the Soil Moisture Ocean Salinity (SMOS) mission that uses a two dimensional design (Kerr et al., 2001). SMOS includes an interferometric radiometer
operating at L band (1.4 GHz) and uses a technique based on the cross-correlation
of observations from all possible combinations of receiver pairs. SMOS is currently
scheduled for launch in late 2008. At an altitude of 763 km, the antenna will view
a swath of almost 3,000 km providing a 40 km global soil moisture product every
23 days.
Another key element of SMOS that is tied to its design is the soil moisture retrieval algorithm. Each SMOS 40 km footprint is the average of 1.2 s. As the satellite
moves along its orbital path each pixel is observed at many different viewing angles,
which permits the observation of brightness temperature as a function of viewing angle. These multiple angles are used with radiative transfer equations to retrieve soil
moisture as well as vegetation information (Wigneron et al., 2000).
An alternative solution to the size-mass issues of real aperture antennas is the use
of lightweight mesh. This approach is compatible with designs that can be packaged
into smaller volumes for launch and also have much lower mass. A large mesh
antenna was one of the key features of the Hydros satellite mission that was studied
by NASA under its Earth System Science Pathfinder program (Entekhabi et al.,
2004) but was recently cancelled by NASA. Using a 6 m mesh antenna operating
at 1.4 GHz, Hydros would be able to provide a 40 km brightness temperature/soil
moisture product. It would employ conical scanning over a wide swath that results
in 23 day global coverage. Conical scanning facilitates some aspects of retrieval.
However, the spinning of the large antenna would require careful engineering. This
concept is currently being re-considered for priority implantation.
The passive microwave instruments of SMOS and Hydros would have demonstrated two potential paths to achieving higher spatial resolution using only passive
sensors. The 40 km products would satisfy the demands of climatology studies.
There are many additional applications that could utilize soil moisture if it was available at higher spatial resolution. If successful, each of these technologies could lead
to future missions with resolutions of better than 10 km.
Another approach to better spatial resolution is disaggregation. The concept of
disaggregation to achieve a higher resolution product is based on the assumption that
the passive microwave instrument provides a reliable soil moisture product and that
alternative remote sensing measurements while at higher spatial resolutions have
lower accuracies. This has been explored in the past using visible-near infrared remote sensing. In regions with uniform soil properties and small amounts of vegetation this might work. However at these wavelengths the high-resolution data only
represents an extremely thin surface soil layer that easily disconnects from the lower
profile. If there is vegetation this approach will not work unless the vegetation exhibits a response related to the soil moisture state. There is some justification for
this; however, it is also compounded by lags in time between changes in soil moisture and changes in canopy visible-near infrared characteristics.
Using thermal infrared for disaggregation has also been proposed. It too has limitations. However, it is expected that the lag time between soil moisture changes and
16
T.J. Jackson
temperature changes will be shorter than with physiographic features observed with
visible-near infrared.
Disaggregation was explored by Chauhan et al. (2003) as a solution for NPOESS.
It was proposed that a combination of microwave, visible-near infrared, and thermal infrared could be used. It was demonstrated using 25 km SSM/I products and
1 km Advanced Very High Resolution Radiometer (AVHRR) data. This might work
sometimes, however, cloud cover would certainly limit it.
Another approach to disaggregation was central to the Hydros mission design. In
addition to the passive microwave instrument, Hydros would have included a radar
that shared the same antenna. This instrument would provide a 3 km soil moisture
product with possibly limited accuracy. However, the real role of this sensor was
to serve as a disaggregation and data fusion tool that would be combined with the
coarser resolution passive product. It is fairly well established that a radar responds
to soil moisture in much the same way as a passive microwave sensor for smooth
bare soils. However, roughness and vegetation effects can be more difficult to account for using radar, which leads to a higher uncertainty in radar only soil moisture retrieval algorithms. A potential disaggregation scheme is described in Narayan
et al. (2006).
Not only would the radar serve in disaggregation it would provide the option
of using data fusion of passive and active microwave in soil moisture retrieval.
The concept of data fusion of passive and active could be explored using another
L band mission being implemented by NASA called Aquarius (Koblinsky et al.,
2003). Aquarius would provide (2010) both active and passive coarse spatial resolution measurements using a push broom approach. It would provide some useful
information for algorithm science but the technology does not provide a pathway to
high spatial resolution.
It was anticipated that with Hydros on its successor a highly reliable 10 km soil
moisture product could result from the integration of passive and active microwave
remote sensing. A 10 km soil moisture product would open the doors to a much
wider range of applications and the demonstration of the approach could lead to
even higher resolutions in the future. There are likely to be limits on the differences
in scale that can be used when disaggregating. There has to be some overlap in the
dominant processes that control variability at the two resolutions. If a passive technology is demonstrated that can lead to future missions with higher spatial resolutions, then radars with higher resolutions can also be consider that could eventually
result in products <1 km. Schemes for disaggregation that could be compatible with
SMOS, as well as other missions, are described in Merlin et al. (2005) and Pellenq
et al. (2003).
2.5 Summary
Land applications, in particular soil moisture retrieval, have been hampered by the
lack of low frequency passive microwave observations and the coarse spatial resolution of existing sensors. The next decade will see several exploratory missions using
17
new technologies, and innovative approaches to integrating passive microwave observations with active microwave measurements that could lead the way to an order
of magnitude improvement in spatial resolution.
References
Basist A, Grody NC, Peterson TC, Williams CN (1998) Using the Special Sensor Microwave/
Imager to monitor surface temperatures, wetness, and snow cover. J. Appl. Met. 37:888911
Becker F, Choudhury, BJ (1988) Relative sensitivity of NDVI and microwave polarization difference index (MPDI) for vegetation and desertification monitoring. Remote Sens. Environ.
24:297311
Bindlish R, Jackson TJ, Wood E, Gao H, Starks P, Bosch D, Lakshmi V (2003) Soil moisture estimates from TRMM Microwave Imager observations over the Southern United States. Remote
Sens. Environ. 85:507515
Chauhan NS, Miller S, Ardanuy P (2003) Spaceborne soil moisture estimation at high resolution:
a microwave-optical/IR synergistic approach. Int. J. Remote Sens. 24:45994622
Entekhabi D, Njoku E, Houser P, Spencer M, Doiron T, Belair S, Crow W, Jackson T, Kerr Y,
Kimball J, Koster R, McDonald K, ONeill P, Pultz T, Running S, Shi JC, Wood E, van
Zyl J (2004) The hydrosphere state (Hydros) mission concept: an Earth System Pathfinder
for global mapping of soil moisture and land freeze/thaw. IEEE Trans. Geosci. Remote Sens.
42:21842195
Gaiser PW, St. Germain KM, Twarog EM, Poe GA, Purdy W, Richardson D, Grossman W,
Jones WL, Spencer D, Golba G, Cleveland J, Choy L, Bevilacqua RM, Chang PS (2004) The
WindSat spaceborne polarimetric microwave radiometer: sensor description and early orbit
performance. IEEE Trans. Geosci. Remote Sens. 42:23472361
Jackson TJ (1997) Soil moisture estimation using special Satellite Microwave/Imager satellite data
over a grassland region. Water Resour. Res. 33:14751484
Jackson TJ, Schmugge TJ (1991) Vegetation effects on the microwave emission from soils. Remote
Jackson TJ, Levine DM, Swift CT, Schmugge TJ, Schiebe FR (1995) Large-area mMapping of
soil-moisture using the ESTAR passive microwave radiometer in Washita92. Remote Sens.
Environ. 54:2737
Kerr YH, Waldteufel P, Wigneron JP, Font J, Berger M (2001) Soil moisture retrieval from
space: the Soil Moisture Ocean Salinity (SMOS) mission. IEEE Trans. Geosci. Remote Sens.
39:17291735
Koblinsky CJ, Hildebrand P, LeVine D, Pellerano F, Chao Y, Wilson W, Yueh S, Lagerloef G
(2003) Sea surface salinity from space: science goals and measurement approach. Radio Sci.
38(4):8064, doi:10.1029/2001RS002584
Le Vine DM, Griffis A, Swift CT, Jackson TJ (1994) ESTAR: a synthetic microwave radiometer
for remote sensing applications. Proc. IEEE. 82:17871801
Li L, Njoku EG, Im E, Chang P, St. Germain K (2004) A preliminary survey of radiofrequency interference over the U.S. in Aqua AMSR-E data. IEEE Trans. Geosci. Remote Sens.
42:380390
Merlin O, Chehbouni AG, Kerr YH, Njoku EG, Entekhabi D (2005) A combined modeling and multi-spectral/multi-resolution remote sensing approach for disaggregation of surface soil moisture: application to SMOS configuration. IEEE Trans. Geosci. Remote Sens.
43:20362050
Narayan U, Lakshmi V, Jackson TJ (2006) High-resolution change estimation of soil moisture
using L-band radiometer and radar observations made during the SMEX02 experiments. IEEE
Trans. Geosci. Remote Sens. 44:15451554
18
T.J. Jackson
Narvekar PS, Jackson TJ, Bindlish R, Li L, Heygster G, Gaiser P (2007) Observations of land
surface passive polarimetry with the WindSat Instrument. IEEE Trans. Geosci. Remote Sens.
45:20192028
Neale CMU, McFarland MJ, Chang K (1990) Land-surface type classification using microwave
brightness temperatures from the Special Sensor Microwave/Imager. IEEE Trans. Geosci. Remote Sens. 28:829838
Njoku E, Koike T, Jackson T, Paloscia S (2000) Retrieval of soil moisture from AMSR data. In:
P. Pampaloni, S. Paloscia (eds), Microwave Radiometry and Remote Sensing of the Earths
Surface and Atmosphere. VSP, Zeist, The Netherlands, pp 525533
Njoku EG, Jackson TJ, Lakshmi V, Chan TK, Ngheim SV (2003) Soil moisture retrieval from
AMSR-E. IEEE Trans. Geosci. Remote Sens. 41:215229
Njoku EG, Ashcroft P, Chan T, Li L (2005) Global survey and statistics of radio frequency interference in AMSR-E land observations. IEEE Trans. Geosci. Remote Sens. 43:938947
Owe M, van de Griend AA, Chang AT (1992) Surface moisture and satellite microwave observations in semiarid Southern Africa. Water Resour Res. 28:829839
Owe M, de Jeu R, Walker J (2001) A methodology for surface soil moisture and vegetation optical
depth retrieval using the microwave polarization difference index. IEEE Trans. Geosci. Remote
Sens. 39:16431654
Paloscia S, Macelloni G, Santi E, Koike T (2001) A multifrequency algorithm for the retrieval
of soil moisture on a large scale using microwave data from SMMR and SSM/I. IEEE Trans.
Geosci. Remote Sens. 39:16551661
Pellenq J, Kalma JD, Boulet G, Saulnier G, Wooldridge S, Kerr Y, Chehbouni A (2003) A scheme
for disaggregation of soil moisture along topography and soil depth. J. Hydrology. 276:112127
Ulaby F, Moore R, Fung A (1982) Microwave Remote Sensing: Active and Passive, vol. II,
Addison-Wesley, Reading, MA
Wigneron JP, Waldteufel P, Chanzy A, Calvet JC, Kerr Y (2000) Two-dimensional microwave
interferometer retrieval capabilities over land surfaces (SMOS mission). Remote Sens. Environ.
73:270282
Chapter 3
Active Microwave Remote Sensing Systems

and Applications to Snow Monitoring
Jiancheng Shi
Abstract This chapter describes research activities made over terrestrial snow
covered areas with active microwave instruments, especially with Synthetic
Aperture Radar (SAR). Even though an exhaustive review of all studies could
not be considered in the framework of the book, the chapter consists of many significant developments which have been made recently in the space-borne and air-borne
based active microwave remote sensing of snow properties including techniques of
mapping snow, inferring snow wetness and snow water equivalence (SWE). The
abilities and existing problems with the current sensors will be discussed.
3.1 Introduction
Seasonal snow cover is one of the most important components in predicting global
water- and energy-cycle consequences of Earth-system variability and change. Seasonal snow cover and its subsequent melt can dominate local-to-regional climate and
hydrology. It is the major source of fresh water over wide areas of the mid-latitudes.
Understanding, characterizing, and predicting snow-related processes across spatial
scales in coupled atmospheric and hydrologic models requires improved capability for accurately monitoring spatial and temporal distributions of seasonal snow
properties on land, especially snow water equivalence (SWE) and snow wetness.
In situ measurements provide direct characterization, but at limited spatial and temporal extent and resolution, and frequently must be acquired under challenging or
dangerous field conditions. Furthermore, because of the high spatial and temporal
variability of snow cover, snow properties derived from in situ measurements often
do not provide the reliable spatial and temporal characterization of distributed snow
properties across model domains at the required range of resolutions.
Jiancheng Shi
Institute for Computational Earth System Science, University of California, Santa Barbara, USA
shi@icess.ucsb.edu
19
20
J. Shi
Because visible and near-infrared radiation does not penetrate snow, and because the optical properties of ice and water are similar, snow reflectance in this
part of the electromagnetic spectrum is not sensitive to snow depth (except for very
shallow snow) or free liquid water in the snowpack. For large, flat regions, passive
microwave data at 18 and 37 GHz allow sequential mapping of SWE distribution. In
dry snow, the profile of the snow grain sizes also influences the microwave remote
sensing signal. A more serious problem for measurement of SWE is that the snow
depths are often too large for existing passive microwave remote sensing retrieval
algorithms to provide accurate estimates. The signal becomes asymptotic and insensitive to snow water equivalence at and above about 0.5 m at 37 GHz (a commonly
used frequency). Moreover, the spatial resolution of the current passive microwave
satellite sensors is too coarse to provide useful information at regional and drainage
basin scales for hydrological investigations and applications. Active microwave sensors (radars), on the other hand, are sensitive to many snow parameters such as snow
density, depth, grain size, free liquid water content, and snow-pack structures that
are useful for hydrologic applications.
Active microwave sensors, especially Synthetic Aperture Radar (SAR), have
been used to estimate snow properties, such as snow wetness and snow water equivalency (SWE), and to discriminate snow with other surfaces. They acquire image
measurements of the Earth day or night, in all weather, through cloud cover, smoke
and haze, and with high resolution especially useful in monitoring geophysical properties in alpine regions. This chapter summarizes the recent progresses in the technical developments for estimation of snow and vegetation properties using active
microwave instruments, mainly synthetic aperture radar (SAR).
3.1.1 Current Available Satellite SAR Data

There are several Spaceborne SAR instruments available, in past, current and near
future, from different space agencies over the world. The major sensor parameters
are summarized in Table 3.1.
ERS-1 and ERS-2 SAR instrument is one of the sensors on the first and second
European Remote Sensing Satellites that were launched in July 1991 and 1995 by
the European Space Agency. They are circling the Earth every 100 min. In 3 days,
it can cover the entire planet with exact repeat coverage at 35 days. They have the
polar orbit at 780 and 785 km for ERS-1 and ERS-2, respectively. Both ERS-1 and
ERS-2 SAR transmit and receive the microwave signals with vertical polarization
at frequency 5.3 GHz and a fixed incidence angle of 23 . Its primarily applications
oriented towards ocean and ice monitoring including those of sea state, sea surface winds, sea surface temperature, ocean circulation and sea and ice level. Due
to its all-weather and high resolution microwave imaging capability, its applications
have been found over land from the monitoring of crops, tropical deforestation to
flooding monitoring. The commonly used products of ERS-1/2 SAR in geophysical
3 Active Microwave Remote Sensing Systems and Applications to Snow Monitoring
21
Table 3.1 Characteristics of current available spaceborne SAR

Sensor
Frequency
in GHz
Polarization
ERS-1/2
ASAR
5.3
5.33
RADARSAT-1
5.3
SIR-C/X-SAR
JERS-1
PALSAR
1.25 and
5.3/9.6
1.27
1.27
RADARSAT-2
5.3
TerraSAR-X
9.6
Fixed at 23
Varying with
mode
HH
Varying with
mode
Fully
Varying with
polarimetric/VV data takes
HH
Fixed at 35
Fully polarimetric, Varying with
dual-polarization, mode
and HH
Fully polarimetric, Varying with
dual-polarization, mode
and HH
VV or dualVarying with
polarization
mode
VV
Dual-polarization
Incidence
in degree
Pixel
resolution
in m
Available time
frame
3.812.5
3.8150
Since 1991/1995
Since 2002
10100
Since 1997
627
April and
October, 1994
19941997
Since December
2005
18
10100
3100
2007
116
2007
applications are the single look complex product (SLC), the precision image (PRI),
and the geocoded product that produced from PRI data after Earth ellipsoid and
terrain corrections. The more detailed information about ERS-1 and ERS-2 SAR
instrument and data products can be found at http://www.earth.esa.int/ers.
In March 2002, the European Space Agency launched Envisat, an advanced
polar-orbiting Earth observation satellite which provides measurements of the
atmosphere, ocean, land, and ice. An Advanced Synthetic Aperture Radar (ASAR),
operating at C-band, ASAR ensures continuity with the image mode (SAR) and the
wave mode of the ERS-1/2 AMI. It enhanced capabilities in comparison with ERS
SAR include coverage (a ScanSAR mode), range of incidence angles, polarization (dual-polarizations), and modes of operation (with nine different modes). The
resulting improvements in image and wave mode beam elevation steerage allow
the selection of different swaths, providing the swath coverage of over 400 km wide
using ScanSAR techniques. In alternating polarization mode, transmit and receive
polarization can be selected allowing scenes to be imaged simultaneously in two
polarizations.
RADARSAT-1, launched in November 1995, is an operational radar satellite system developed by Canadian Space Agency (CSA) to monitor environmental change
and the planets natural resources. RADARSAT-1 has the sun-synchronous orbit at
798 km altitude with 6 p.m. ascending node and 6 a.m. descending node. It transmits and receives the microwave signals with horizontal polarization (HH) at C-band
(5.3 GHz). It has seven beam modes, 35 beam positions and ScanSAR modes for a
wide range of imaging options, varying resolutions from 8 to 100 m, swath widths of
22
J. Shi
50500 km, and incidence angles from 10 to 59 , depending on the mode selection.
The more detailed information about RADARSAT-1 instrument and data products
can be found at http://www.space.gc.ca/asc/eng/satellites/radarsat1.
ALOS PALSAR is the Phased Array type L-band Synthetic Aperture Radar
(PALSAR) on the Advanced Land Observing Satellite (ALOS) that has been
launched on January 24, 2006 and is a follow on the Japanese Earth Resources
Satellite-1 (JERS-1). It provides higher performance than the JERS-1s SAR for
land observations. PALSAR has the ScanSAR mode with the swath width of 250
350 km depending on the number of scans in addition to the fine resolution in a
conventional mode. This swath is three to five times wider than conventional SAR
images. It has also different polarization selections from single, dual, to fully polarimetric observations and improved orbit controlling capability that allows the more
frequent repeat-pass interferometer applications.
SIR-C/X-SAR is the Shuttle Imaging Radar-C and X-Band Synthetic Aperture
Radar. It was a cooperative experimental SAR mission between the National Aeronautics and Space Administration (NASA), the German Space Agency (DARA),
and the Italian Space Agency (ASI) (Evans et al., 1997). The SIR-C/X-SAR system was flown onboard NASAs Space Shuttle on two 10-day missions in April and
October 1994. SIR-C provides radar polarimetric digital images simultaneously at
two wavelengths, L-band (24 cm), and C-band (5.6 cm). These polarimetric data
allows derivation of the complete scattering matrix on a pixel-by-pixel basis and
more detailed information about the geometric structure and dielectric property of a
target (Van Zyl et al., 1987; Evans et al., 1988). X-band (3 cm) with VV polarization
results in a three-frequency capability. The resolution of this system ranged between
10 and 40 m, and it collected image swaths between 15 and 90 km wide.
In addition, there have been two new SAR satellites launched in 2007.
RADARSAT-2 is the follow on RADARSAT-1 system co-funded by CSA and
MacDonald Dettwiler (MDA). It has been launched in 2007 with 7 years mission duration and will support all RADARSAT-1 imaging modes. It has the same
orbit, repeat cycle and ground track as RADARSAT-1. Its improvements in comparison with RADARSAT-1 include the higher resolution (up to 3 m), routine
left-looking and right-looking mode for more frequent revisit, selective polarizations with fully-polarimetric imaging modes, on-board GPS receivers to improve
the real-time position knowledge, and higher downlink power. The more detailed
information about RADARSAT-2 instrument and data products can be found at
http://www.radarsat2.info/rs2 satellite.
TerraSAR-X is a new German radar satellite with 514 km sun-synchronous orbit
at 98 Inclination and has been launched in June, 2007 with a 5 year lifetime. It carries a high frequency X-band (9.65 GHz) SAR sensor that can be operated in three
different basic modes the ScanSAR, Stripmap and Spotlight mode at a varying
geometrical resolution between 1 and 16 m. TerraSAR-X will provide single or dual
polarization data. On an experimental basis additionally quad polarization and along
track interferometry are possible.
23
3.1.2 SAR Measurement Properties

Generally, SAR provides three types of measurements that can be used in study
geophysical properties depending on the sensor properties: (1) backscattering measurements at the given frequency and polarization such as ERS1/2, ASAR, and
RADARSAT-1; (2) the measurements of the polarization properties by a fully
polarimetric instrument (SIR-C, PALSAR, RADARSAT-2, and TerraSAR-X with
selected mode); and (3) interferometry from the repeat-pass measurements.
Backscattering SAR imagery provides high resolution digital images but with
only fixed polarization state of the antenna. With only one or a few intensity measurements per pixel, the applications to monitor snow properties have to rely mainly
on the radiometric properties such as snow classification and estimations of snow
wetness and SWE.
Rather than just measuring amplitude, an imaging radar polarimeter measures the
amplitude and relative phase for every polarization state. These complex measurements lead to nine independent real elements of the Stokes matrix, which describes
how the scattering mechanisms in each pixel transform the illuminating electromagnetic wave back to the receiving antenna. The polarization feature derived from the
Stokes matrix, is the radar cross section as a function of the antenna polarization
state and is useful in interpreting the scattering mechanisms within a resolution cell.
The first element of the matrix is the total power or span. The nine cross products
of the pixel scattering matrix can be obtained by the linear combinations from the
span and the other Stokes matrix elements. In addition to the polarization signature, many other features describing the scattering mechanisms within a pixel can
be derived from the Stokes matrix. These include: the pixel intensity synthesized
from the polarization signature for fixed antenna polarization states, the polarization
phase difference of the scattering matrix elements, the coefficient of variation, the
enhancement factor, the scattering mechanism and the degree of polarization. Thus,
polarimetric measurements provide much more information per pixel than the single
fixed antenna SAR imagery, and have been shown to be effective in classification
of terrain. These measurements have great opportunity to minimize the topographic
effects on radar images.
In addition to single, multi-polarization, and fully polarimetric radar observations, the repeat-pass interferometric radar measurement also provides the useful
information in analyses of land surface geophysical properties. Its techniques for
topographic mapping of surfaces promise the high-resolution digital elevation models; but they also permits inference of changes in the surface over the orbit repeat
cycle from the correlation properties of the radar echoes. Measurements of interferometer correlation describe processes occurring on the time scales of the orbit
repeat time and size scales on the order of a radar wavelength, such as vegetation
growth, glacier motion, permafrost freezing and thawing, and soil moisture induced
effects.
Furthermore, SAR images have two distinct special characteristics: image
speckle and a more complicated geometric mapping in contrast to images from
optical sensors. The SAR image generation involves a coherent processing carried
24
J. Shi
out on the received signal: fading causes on SAR imagery a grainy appearance
referred to as speckle (Ulaby et al., 1982, 1986). This is because a SAR resolution cell, that contains many different scatterers, is very large when compared
to the wavelength of the illuminating electromagnetic wave. The returned radar
echo results from the coherent summation of all the returns from the amplitudes
and phases of the single scatterers. Commonly, it requires processing a multi-look
image or applying a speckle filter to reduce its effects and to improve the image
quality (Ulaby et al., 1982, 1986; Lee et al., 1998). Furthermore, SAR mapping is
mainly an integration of backscattered signals having the same Doppler frequency
as along-track measurements and the same distance as across-track measurements.
The characteristic radar measurements are range the distance from the sensor to
an object point, and time the position of the sensor along its flight path where the
data are collected. They define the two-dimensional SAR image space. The total
received power from a resolution cell is proportional to the radar cross section, and
the rest of factors in the radar equation are generally assumed constant. The radar
cross section is usually a function of polarization, frequency, viewing geometry,
and illuminated area, in addition to factors such as the dielectric properties and
geometric structure of the targets. During the data processing, the backscattering
coefficient o is usually presented in terms of a normalized radar cross section to
eliminate the dependence on illumination area.
(3.1)
A
where A is illumination area. For a flat area, both illumination area and incidence
angle of a pixel can be reasonably estimated so that the backscattering coefficient
can be obtained.
Topographic effects on radar images can be considered as two aspects: (1) the
effects on received radar power, which result from a great variation in illuminated
area and incidence angle for a pixel resolution. It can be described as the variation of
the received power from an inclined surface compared to the received power from
a horizontal surface. This variation is a function of the relative orientation of the
surfaces with respect to the illuminating source and their position relative to the sensor. (2) The geometric distortions on image coordinates that are mainly dependent
upon the distance between an imaged pixel and the sensor. The effects of rugged
terrain on both the received power result from the change in illuminating area and
the geometric distortion can be corrected for a resolution cell if a digital elevation
model (DEM) with the comparable resolution and the sensor location information
are available. They are available only with the geocoded data products.
o =
3.2 Snow Mapping with SAR

For climatological and hydrological investigations, the area of snow cover is one
of the important parameters. For instance, the snow line on glaciers is an important quantity for hydrological applications and mass balance studies. Visible and
25
near-infrared sensors have been used extensively to measure these quantities, but
are hampered by cloud cover, which can be pervasive in some regions. In particular, snow cover must be measured on a timely basis to be useful for operational
hydrology, and the opportunities for obtaining suitable data from these sensors can
be infrequent. Microwave remote sensing is a methodology that is less influenced
by cloud conditions, depending on frequency. Especially, Synthetic Aperture Radar
(SAR) provides high resolution measurements that are comparable to the scales of
the topographical variation in mountain areas and more suitable for mapping snow
cover than passive microwave instruments.
In the microwave part of the electromagnetic spectrum, ice is almost transparent, and the radar penetration depth, depending on the frequency, can reach tens
of meters for dry snow. The major scattering source is the snowground interface, and it is difficult to discriminate dry snow cover from bare surfaces or short
vegetation with backscattering measurements from a single-polarization radar measurements. Earlier investigators found no significant difference in backscattering
from dry snow and snow-free surfaces at either C-band (Rott et al., 1993) or X-band
(Fily et al., 1995).
However, there is a large dielectric contrast between the solid and liquid phases
of water at microwave frequencies. As snow starts melting, even a small amount of
liquid water reduces the penetration depth of the radar signal, and thereby changes
the dominant scattering source from the snowground interface to the snow volume
and the airsnow interface. The backscattering coefficients decrease substantially at
C-band and X-band (Stiles and Ulaby, 1980; Nagler and Rott, 2000; Shi and Dozier,
1995). As the snow continues melting and the surface ages, backscattering increases
and surface scattering at the airsnow interface becomes the dominate scattering
source, as Rott et al. (1993) observed in ERS-1 data over the Otztal,
Austria, a
snow-covered alpine glacier region. Surface scattering also dominates the signals
from bare rock, soil, and glaciers, but these surfaces are usually rougher than wet
snow and therefore have stronger backscattering signals. However, the backscattering from wet snow cover may have the similar intensities as that from very smooth
bare soils, which can result in a difficulty for separation, as Haefner et al. (1993)
showed in ERS-1 data from the Swiss Alps.
Since the radiometric quality of SAR images in an alpine region is dependent
on flight and imaging parameters (e.g., flight altitude, radar elevation angle) and the
topography of the imaged area, the representation of the target materials is likely
to be inaccurate. This variation in radar backscatter that is unrelated to the surface
cover type is particularly evident for high relief surfaces where a large variation of
slope and aspect creates a great variation of local incidence angles and illuminated
areas. For example, rock surfaces with greater incidence angle could have lower
power return than that from snow or glacier covered areas with smaller incidence
angle. The effect of topography, therefore, is another major problem we are facing
in using radar to map snow covered areas, especially in alpine regions.
The study (Van Zyl et al., 1993) indicates that the topographic effects on radiometric properties measured from spaceborne SAR can be explained by variation in
imaged pixel area and in local incidence angle. When topographic information of
26
J. Shi
the study area is available, both spaceborne and airborne SAR image data can be
radiometrically calibrated. The remaining problem for target discrimination is the
effect of local incidence angle on the received power. If topographic information is
not available, we need to consider the effects of variations in both local incidence
angle and imaged pixel area for satellite SAR measurements.
Therefore, in attempting to use SAR data to map seasonal snow cover over remote and inaccessible areas, we are faced with two major problems:
Compensation must be made for the effects of rugged terrain.
Snow must be distinguished from other surface covers.
The currently available techniques for mapping snow cover with SAR imagery can
be summarized mainly into three catalogues based on the usage of the SAR measurement properties.
3.2.1 Multi-temporal Single-polarization Techniques

Previously, a single-polarization SAR imager provides high-resolution images but
with only one available measurement per pixel at a fixed polarization state. With
only one intensity measurement per pixel, we have to rely mainly on radiometric
properties to distinguish snow covered area from other targets, such as bare ground
or vegetation. A practical technique to mapping wet snow cover, that has been developed recently, is the multi-temporal methods. Nagler and Rott (2000) used four
ERS-1 repeat-pass images, two from ascending and two from descending orbits to
reduce the effect of layover. They compared backscattering ratios of an image with
wet snow and a reference image (either snow-free or dry snow) to generate a map
of wet snow cover, based on change detection. They found that wet snow cover
can be identified by using the time series measurements. The radar observations
have shown that the backscattering from wet snow cover could be reduced 34 dB
at C-band in comparison with eight dry snow cover before melting or bare surface
after melting. In addition, they also found that the requirement for accurate local
incidence angle was relaxed when they used multi-temporal ERS-1 data. (Rott and
Nagler, 1993) since the classification algorithm used the ratio measurements of wet
snow cover image to the reference image in which the part of terrain effects pixel
illumination area has been cancelled out if two temporal images are co-registrated
accurately.
For wet snow mapping in forested regions, Luojus et al. (2006) demonstrated a
two step technique to map wet snow cover fraction using multi-temporal ERS-2
measurements for boreal forest region in Finland during the snow melting season. The first step is the forest canopy compensation. This is done by nonlinearly fitting ERS-2 measurements with a semi-empirical forest backscattering model
(Pulliainen et al., 2003) with the forest stem volume information to estimate the
backscattering signals at ground surface and the volume backscattering signals, the
two-way transmittivity of the forest canopy. The second step is the employment
of linear interpolation algorithm that uses the reference images and the estimated
27
backscattering signals at ground surface to calculate the fraction of wet snow covered area SCA.
o o
surf
ground,ref
(3.2)
SCA = 100% o
o
snow,ref ground,ref
where osurf is the estimated backscattering coefficient at the ground surface,
oground,ref is the reference signal from the snow-free ground, and osnow,ref is the reference signal from the wet snow ground. One reference image describes the signal
at the fully wet snow cover situation and the other describes the snow-free at the
end of the snow melting season. In addition to two reference images, this method
requires the knowledge of the forest stem volume distribution of the study area and
suitable to sub image scale applications (many pixels averaged together). This is because the uncertainty produced by radar speckle and the use of forest compensation
in which the requested forest stem information is more reliable at coarse resolution
than that at ERS-2s pixel resolution.
3.2.2 Multi-frequency and Multi-polarization Techniques

Shi and Dozier (1997) evaluated the characteristics of the backscattering, polarization, and frequency ratios of the targets of the study site near Mammoth Mountain
in the Sierra Nevada, U.S using multi-frequency and polarization SIR-C/X-SARs
measurements. They developed two type of supervised classifiers based on classification tree technique. The first type of the classifier was developed by using
intensity measurements, polarization properties, and frequency ratios. It can map
dry snow and discriminate dry from wet snow, but it requires topographic information for radiometric terrain correction and to reduce effects of local incidence angle.
It is about 79% as accurate as a TM binary classification, but it suffers the same
shortcoming it underestimates total snow cover in regions of mixed pixels, especially forested regions. Its performance on the two data-takes where the snow was
dry showed that only a few pixels were misclassified as wet snow.
The second type of classifier was developed based on polarization properties
and backscattering ratios between different frequencies. Since these measurements
can be obtained correctly without radiometric terrain calibration, the classifier does
not require topographic information and can be used to map wet snow. Similarly,
Shi et al. (1994) also developed a method with multi-polarization C-band airborne
SAR to map wet snow and glacier ice without a DEM, using only measurements
of the polarization properties. Its accuracy is 77% when compared with TM binary
classification, but both underestimate total snow cover.
3.2.3 Repeat-pass Interferometric Technique

Except the technique of the first type of classifier as in the multi-frequency and polarization techniques, all above methods are restricted to mapping wet snow-cover
28
J. Shi
since it is difficult to discriminate dry snow cover with bare ground and short vegetation. In the study of using SIR-C/X-SAR data to map snow cover (Shi and Dozier,
1997), it was found that wet snow cover had very similar backscattering intensity
and polarization characteristics to smooth bare surface at C-band and X-band. For
instance, the backscattering from wet snow-cover is very similar to smooth dry soil,
alluvial surfaces, and relatively rough water surfaces. At a large drainage basin or
regional scale, where many different targets are within a scene, those techniques
might not be reliable. For similar reasons, change detection measurement might be
also unreliable since the similar change in backscattering could be caused by different natural environment changes. In order to develop a large-scale snow mapping
technique other measurement are required to discriminate between snow and other
targets.
Interferometric radar techniques for topographic mapping of surfaces promise the
high-resolution digital elevation models; but they also permits inference of changes
in the surface over the orbit repeat cycle from the correlation properties of the radar
echoes. Measurements of interferometer correlation describe processes occurring on
the time scales of the orbit repeat time and size scales on the order of a radar wavelength, such as vegetation growth, glacier motion, permafrost freezing and thawing,
and soil moisture induced effects. The coherence measurement between two repeatpasses, therefore, provides a useful measurement in addition to backscattering intensities in each scene and their changes between two passes, and makes it possible
to develop an algorithm for mapping both dry and wet snow covers over large area.
Strozzi et al. (1999) demonstrated that coherence measurements could provide
the separation between wet snow cover and bare ground in the cases where the
backscatter discrimination failed from analyses of ERS tandem data in Switzerland.
The low coherence observed over wet snow cover is mainly caused by the rapid
change in scattering properties and geometry as the result of wet snow metamorphism due to the movement of free liquid water content, ice grain growth,
displacements of adjacent scatterers, and formation of density heterogeneities
(layering, ice-lenses, etc.), which all result in a significant decorrelation. On the
other hand, the high coherence is regularly observed over no-forested snow free
areas. For forested areas, it can be easily separated with wet snow cover due to their
huge difference in backscattering intensity even if its coherence is generally low.
They provide the physical bases for separating wet snow with other surfaces. For using C-band measurements, however, it requires the short temporal scale (a few days)
between the two repeat-pass measurements in order to avoid the significant temporal
decorrelation in other surface targets such as bare or short vegetation surfaces.
Shi et al. (1997) evaluated the L-band coherence measurements between two
repeat-pass SIR-C image data from its first mission in April (with snow) and second mission in October (without snow), 1994. This measurement indicates that the
ground is completely undisturbed between viewings the signals will be highly correlated. Otherwise, the decorrelation will occur. They evaluated the coherence measurements of L-band VV polarization for five land cover types snow, lake, bare
surface, short vegetation, and forest. It was found that lake, snow and some forest
areas had very low coherence between two data-takes. For the lake, the decorrelation
between two data-takes is mainly due to the low S/N ratio and the changes of the
29
lake surface roughness characteristics, because of change in wind conditions (speed

and direction). For the dry snow case, the dominant scattering at L-band is from the
snow-ground interface. In addition to the change of the dielectric contrast from the
air-ground to snow-ground, existing dry snow cover will result in large decorrelation. This is due to change in local incidence angle when radar signal passes through
snow layer will cause a spatial baseline decorrelation. The radar echoes will be also
expected to be close to complete decorrelation when measuring correlation between
wet snow-cover and bare ground image passes. This is because the radar signal in
snow-covered pass can only penetrate a few centimeters so that the radar senses two
different targets. The coherence measurements from forest can have large dynamic
range with very low values similar to those from snow and lake, especially from
dense forest. On the other hand, the coherence measurements from the bare surface are significantly higher than those from lake. For the bare surface, a change of
soil moisture will result in a decorrelation. However, the amount of decorrelation
is expected to be smaller because radar senses a same target with a same scattering
mechanism (only change in magnitude). The short vegetation (mainly sagebrush and
grass in this study area) has very similar coherence measurement to the bare surface
mainly due to the dominant scattering source is from the ground surface at L-band.
Therefore, the coherence measurements between one snow covered scene and one
without snow provide a very good separation between snow cover and bare surface
as well as short vegetation. These two targets are most difficult to discriminate with
snow cover. Thus, the correlation measurement provides a significant information
to map snow covered area.
A pixel-based decision tree classifier (DTC) was developed based on above characteristics. While the coherence measurements to discriminate forest, open water,
and snow with short vegetation and bare ground, snow can be separated with forest
and open water by using backscattering measurements. In order to verify the classification results, a cloud-free Landsat Thematic Mapper (TM) scene on April 14,
1994 (the SIR-C/X-SAR data-take on April 13, 1994) was acquired. A snow map
was generated by the TM data and then projected (or co-registered) to the slant
range presentation of the SAR images by using the Space Shuttle ephemeris data
and ground control points. In the classification of the TM scene, there are only four
target categories: lake, snow, forest, and range land that includes both bare ground
and short vegetation. Figure 3.1 shows the SAR classification map (right) and the
Fig. 3.1 Comparison of SAR (right) and TM (left) derived classification maps. Black Forest and
lake, gray bare surface and short vegetation, white snow
30
J. Shi
TM classification map (left). The comparison estimation of this two results indicated that at 86% accuracy can be obtained for snow cover area under consideration
of the TM classification map as the ground truth.
3.3 Estimates of Snow Wetness from C-Band SAR

3.3.1 Relationships Between Snow Wetness and C-Band SAR
Measurements
Monitoring spatial and temporal changes of liquid water content in snow is important for hydrological modeling, because the presence of liquid water shows that a
particular area of the basin can contribute immediately to runoff. When snow starts
melting, there is a great change in scattering mechanism in comparison with the
dry snow cover. The radar can not see through snow-pack since the penetration
depth is about a radar wavelength (a few centimeters) for typical wet snow cover at
C-band.
Radar backscattering measurements from wet snow are affected by two sets of
parameters: (1) sensor parameters, which include the frequency, polarization, and
viewing geometry, and (2) snow parameters, which include snow density, liquid
water content, particle sizes and shapes of ice and water inclusions, type of the correlation function, and surface roughness. It has been understood that the backscattering from wet snow cover is mainly controlled by snow volume backscattering and
the surface backscattering at airsnow interface.
The volume scattering is inversely correlated to snow wetness. The liquid water
content mainly causes an increase of snow permittivity because of the high
dielectric contrast between ice and water. This results in a significant (1) decrease in transmission at the airsnow interface and (2) high dielectric loss, which
greatly increases the absorption coefficient.
The surface scattering is proportional to snow wetness. As the liquid water content increases, the reflectivity at airsnow interface increases greatly so does the
surface backscattering.
The effects of snow wetness on two major scattering signals are in opposite directions and the actual relationships between radar measurements and snow wetness
depend on which scattering component is dominant scattering source. If the volume scattering is dominant scattering source, the radar measurements will show a
negative correlation to snow wetness. However, the radar measurements will have a
positive correlation to snow wetness if the surface scattering is dominant scattering
source. The relationship between backscattering and snow wetness, however, can
be either positive or negative, depending on snow characteristics and surface roughness and on incidence angle. Generally, as snow wetness increases the backscattering decreases rapidly for low wetness (<3%) conditions because volume scattering
albedo decreases. When wetness is low, the dielectric contrast between air and snow
31
is small and volume scattering dominates, so backscattering is not sensitive to surface roughness. As snow wetness further increases, backscattering becomes sensitive to surface roughness. This is because the surface scattering component becomes
dominating, resulted from rapidly increasing surface scattering component and decreasing volume scattering component. This complexity of the relationship between
the backscattering and snow wetness makes it unrealistic to develop an empirical
relation between the radar signal and field measurements.
3.3.2 Snow Wetness Algorithm Using C-Band Polarimetric SAR

Measurements
An algorithm has been developed for snow wetness retrieval using C-band polarimetric SAR imagery (Shi and Dozier, 1995). This algorithm is developed using
a database, which covers the most possible wet snow physical properties including the wide ranges of snow wetness, density, particle size and surface roughness,
simulated by the first-order scattering model with both surface and volume scattering components. The major developments in this algorithm were: (1) a simplified
surface backscattering model, that describes the relationships between the different
polarization measurements for the conditions of most seasonal wet snow covers,
to minimize the surface roughness effects with multi-polarization measurements;
and (2) the property of the volume scattering ratio in co-polarizations which is only
a function of snow permittivity and incidence angle to minimize the volume scattering albedo effects on estimation of snow wetness. The more details on how the
algorithm was developed can be found in (Shi and Dozier, 1995). The final inverse
model for estimating snow wetness by using three C-band measurements tvv , thh ,
and tvvhh is given as
M1 [avx Re[vv hh
] (avhx DRS DT S ) + bvx M2 ]

bvx M2
= M2 avx Re[vv hh
]+
DTV |vv |2
avhx DRS DT S
(3.3)
with
M1 = tvvhh DTV (i , s ) tvv ,
DT S =
M2 = tvv + thh DT S (i , s ) tvvhh
DTV (i , s ) + DT H (i , s )
|vv (i , s )|2 + |hh (i , s )|2
, DRS =
( , )]
DTV (i , s ) DT H (i , s )
Re[vv (i , s )hh
i s
where tvvhh = Re[Stvv Sthh ] is the real part of the cross product of VV and HH
complex scattering elements, vv and hh are the polarization amplitudes (Ulaby
et al., 1982). avx (i ), avhx (i ), and bvhx (i ) are coefficients derived from statistical
analyses and depend only on incidence angle. They are given in (Shi and Dozier,
1995). DTV (i , s ) and DT H (i , s ) denote the volume backscattering ratios for
Re[VVHH ] to VV and Re[VVHH ] to HH which are only depending on the local
32
J. Shi
incidence angle and the dielectric constant of a wet snowpack. This concept is from
the first-order volume backscattering model for an inhomogeneous dielectric half
space medium:

3
vpp = T2pp exp 2s2 (k1 cos(i ) k2 cos(i ))2
4
(3.4)
Tvv and Thh are power transmission coefficients for a plane interface for vertical and
horizontal polarization. s is the standard deviation of the random surface height.
The loss factor exp[2s2 (k1 cos(i ) k2 cos(i ))2 ] is the rough surface effect on
the power transmission coefficient (Ulaby et al., 1986). is the volume scattering
albedo, which depends on snow density, wetness, particle size, size variation and
shape. Under the assumption of spherical grains or randomly orientated particles,
the volume scattering albedo is independent of the polarization. Therefore, the ratio
for the first-order volume backscattering signals of VV and HH polarizations can be
represented as a function of the local incidence angle and dielectric constant:
DTH (i , s ) =
Re [Tvvhh (i , s )]2
vvvhh
,
=
2 ( , )
hh
v
Thh
i s
DTV (i , s ) =
Re [Tvvhh (i , s )]2
vvvhh
=
vvh
Tvv2 (i , s )
v
The algorithm derived above requires no information about the volume scattering
albedo or the surface roughness parameter. With known local incidence angle, it
involves only the calculation of snowpack permittivity, which can be directly related to snow wetness. This algorithm is applicable to the situations of incidence
angle from 25 to 70 , and the snow surface roughness rms height <0.7 cm and
correlation length <25 cm.
Figure 3.2 shows the comparison between the field snow wetness measurements
(x-axis) and the derived snow wetness from C-band polarimetric SAR data of SIR-C
(black squares) and AIRSAR (white triangles) at the study site around Mammoth
Mountain, California during the first SIR-C/X-SAR mission in April, 1994. Those
included eight snow pits and two transects. The SAR-inferred snow wetness values were obtained from an average value of 3 3 windows around the snow pit
locations. Snow wetness measurements from transects were also averaged to 100 m
scales and the inferred snow wetness from SAR was determined from a mean value
of 4 2 windows along the transect measurements. Most SAR-derived snow wetness agreed well with estimates of snow wetness. The standard deviation of absolute
error was about 1.3% by volume which gives 2.5% error bars at 95% confidence interval for absolute error. The algorithm performed well on both local and regional
scales and provided a quantitative estimate of spatial distribution of snow wetness
at the top snow layer.
As we noticed that the above algorithm requires the C-band polarimetric SAR
measurements and can not be directly applied to ASAR image data since it does
not provide the measurement of Re[Svv Shh ]. In addition, the pair of co-polarization
measurements VV and HH have little difference at small incidence angle, which
results in VV and HH measurements that are almost identical regardless of snow
wetness and surface roughness conditions. Therefore, the measurements of VV and
33
10
Estimated
8
6
4
Sirc
Airsar
2
0
10
Measured snow wetness in %

Fig. 3.2 Comparison between the field measurements and the derived snow wetness using C-band
SIR-Cs and NASA/JPL AIRSARs polarimetric SAR data that are represented as black squares
and white triangles in the plot
HH only provide the opportunity for estimation of snow wetness where the pixels
with the moderate to large incidence. Further studies are needed to modify the above
algorithm for using only two (ASAR) polarization measurements.
3.4 Estimates of Snow Water Equivalence from SAR

Snow water equivalence (SWE), the product of snow density and depth, is the most
important parameter for hydrological study because it represents the amount of
water potentially available for runoff. Measurement of the amount of water stored
in the snowpack and forecasting the rate of melt are thus essential for management
of water supply and flood control systems. Because of rough, irregular topography,
these attributes exhibit large spatial variability over alpine drainage basins, making
it impractical to gather enough in situ measurements. Snow density and depth are
generally not highly correlated, so they must be considered as independent variables in field surveys of the spatial distribution of snow and in the study of snows
microwave backscattering properties.
In studies of active microwave remote sensing from SAR on retrieval SWE,
Bernier et al. (1998) related the multi-temporal C-band backscattering measurements over frozen agriculture fields to thermal resistance in the snowpack. The relationship between thermal resistance and SWE was found and used to estimate SWE.
However, it is commonly known that dry snow is almost transparent at C-band. The
snowpack only contributes a small amount of signal so that the sensitivity of C-band
measurements to SWE is poor. The more studies may be needed to demonstrate the
physics of this technique and its capability to be applied to other areas. This chapter
34
J. Shi
mainly summarizes the other two types of techniques that employ different SAR
measurement properties and with strong physical principles:
1. SAR backscattering technique with Multi-frequency (L, C, and X bands) and
dual-polarization (VV and HH) measurements (Shi and Dozier, 2000a, b).
2. Repeat-pass SAR interferometric technique: The differential phase shift measurements from repeat-pass SAR interferometry can be directly related to SWE
(Rott et al., 2004) or its changes (Guneriussen et al., 2001).
3.4.1 Estimation of SWE with Multi-frequency and Polarization

SAR Backscattering Measurements
During the past years, theoretical modeling of snowpack microwave backscatter has
achieved advanced significantly with respect to understanding snowpack extinction
properties and the interactions of the microwave signal with the snowpack volume and surfaces. The dense media theory considers a snowpack as consisting
of random, discrete, spherical particles of a single size and permittivity (Tsang
et al., 1985), accounting for coherent scattering (near-field effect) within the snowpack, and has been extended to consider multi-size and multi-permittivity randomly
distributed spherical particle systems and systems where the particles are not randomly distributed but may have a tendency to form clusters and bonds (Tsang,
1992; Ding et al., 1994). This approach is appropriate for application to the snowpack microstructure because snow metamorphic processes can give rise to clusters
or aggregations of ice grains. There has also been significant improvement in the
modeling of surface scattering characteristics. The Integral Equation Model (IEM)
(Fung, 1994) and the more recent Advanced Integral Equation Model (AIEM) allow
characterization of a wider range of the surface roughness conditions than past models. Validations of AIEM (Wu et al., 2001; Li et al., 2002; Chen et al., 2003) have
demonstrated significant improvement in modeling surface scattering and emission
for microwave remote sensing of land surfaces. These efforts have established a
fundamentally-improved understanding of the effects of snow physical parameters
and underlying surface dielectric and roughness properties on the microwave measurements of snow-covered terrain, making it possible to characterize the microwave
backscatter behaviors more accurately.
Radar backscattering coefficient measurements at a given incidence angle i
over seasonal snow covered terrain can, generally, be expressed as a four component model:
t
a
v
gv
g
pq
( f ) = pq
( f ) + pq
( f ) + pq
( f ) + Tp Tq L p Lq pq
(f)
(3.5)
The total backscattering signals consist of the surface backscattering from the interfaces at the airsnow interface and at the snowground interface, direct volume
backscattering from snow-pack, the interaction term between snow volume and the
snowground interface, and are represented by the superscript, t, a, g, v, and gv. The
35
f is the radar frequency. The subscript p and q represents polarization status of the
observed radar signals. T is the power transmittivity at the airsnow interface. g pq
is the surface backscattering signal at snowground interface. Lp = exp[p / cos r ]
is the snow-pack attenuation factor. = e d is the optical thickness, which is the
product of snow extinction coefficient e and snow depth d. To carry out a forward simulation of a dry snow covered terrain Eq. (3.5) requires a total of 13 surface
and snow parameters. The six snowpack parameters include snow depth, density, ice
particle size, size variation, stickiness, and temperature. In addition to the dielectric
constant of the ground, the six surface roughness parameters include RMS height,
correlation length, and the two-parameter correlation functions at the airsnow and
the snowground interfaces. The importance of each scattering component depends
on the sensors frequency, polarization, incidence geometry, and snow properties.
Based on the frequency dependence of SAR measurements to snow and ground
properties, Shi and Dozier (2000a, b) developed a multi-frequency (L, C, and X
bands) and dual-polarization (VV and HH) technique to estimate SWE and applied
SIR-C/X-SAR image data. This technique uses L-band measurements to estimate
snow density and the underground dielectric and roughness properties. The relationship between underground backscattering signals at C-band and X-band can
be estimated with the dielectric and roughness properties estimated from L-band
measurements. Then, using C-band and X-band measurements with the minimized
effects of the underlying backscattering signals estimate snow depth and ice particle
size. This technique requires all 3 SIR-C/X-SAR frequency measurements.
3.4.1.1 Estimation of Snow Density with L-Band Dual-polarization SAR

Effects of Snow on L-Band SAR Measurements
Volume scattering and extinction from dry snow are all very small at L-band. At microwave frequencies, the absorption coefficient (the imaginary part of the dielectric
constant) of ice is small, and snow grains are also small compared to an incident
L-band wavelength (24 cm). During the past two decades, little attention has been
paid to microwave interactions of snowpack at L-band (1.25 GHz) frequencies or
lower, while much work has been done at C-band (5.5 GHz) or higher frequencies
(Rott and Matzler, 1987; Ulaby et al., 1982; Stiles and Ulaby, 1980; Ulaby and
Stiles, 1980; Ulaby et al., 1984; Kendra et al., 1998). Assuming there are no large
structures including the melting/frozen draining fingers and channels in the snowpack which may occur during snow melting season, snow grains can not generate
significant volume scattering at L-band since snow grains are much smaller than
incident L-band wavelengths. Under this condition, we can simplify the backscattering model by considering a dry and homogeneous snowpack over a bare soil or
rock surface and Eq. (3.5) becomes
t
g
qp
(k0 , i ) = Tq (i ) Tp (i ) qp
(k1 , r )
(3.6)
36
J. Shi
i , r , and k1 represent the radar wave incidence angle at the airsnow interface, cosine propagation angle in snowpack, and radar wave propagation number in snowpack for the coherent component, respectively. It indicates that the L-band radar
measurements over season dry snow cover are affected by two sets of properties:
(1) snow properties through the power transmittivity at the airsnow interface T
which is a function of snow density, polarization, and incidence angle and (2) surface backscattering properties at snowground interface through surface parameters
including the dielectric constant and surface roughness (roughness is generally described by an autocorrelation function and standard deviation of surface roughness
height) for a random rough surface with no orientation of features.
When the electromagnetic wave passes through the snowpack, versus directly
striking the ground, the following differences occur:
Because of refraction within the snow, the incidence angle at the snowground
interface is smaller.
The incident wavelength at the snowground interface is shorter because the
snow is dielectrically thicker than air.
The snow layer reduces the dielectric contrast at the snowground interface,
which in turn reduces the reflectivity at snowground interface.
The power loss at the airsnow interface reduces the total energy incident on the
snowground interface.
The first two factors result in a change of the sensor observing parameters. Note that
the dielectric contrast g /s should be used instead of g and that k1 should be used
g
at the snowground interface.
instead of k0 when calculating the backscattering qp
The shift in wavenumber k1 = k0 s (or wavelength 0 = 1 s ) is a function of

snow density. For the range of snow densities considered 100 to 550 kg m3
the L-band propagation wavelength in snow ranges from 21 to 16 cm, compared to
24 cm in air. Because the surface roughness effect depends on its size relative to
the incident radar wavelength, the shortening of the incident wavelength for higher
snow densities will result in the soil surface appearing rougher than it would if the
snow were absent. This causes an increase in the surface backscattering signal and
is especially strong for a nearly smooth surface. However, this effect caused by
the wavelength shift becomes smaller when the surface is rougher or the incident
frequency is higher.
Snells law specifies the change in incidence angle: sin(i ) = sin(r ) s . The
incidence angle at the snowground interface depends only on the incidence angle
at the snow surface and the dielectric constant of snow, not on its thickness, thus
it is a function of snow density. For a given incidence angle at the snow surface,
a greater snow density causes a greater change in the incidence angle at the snow
ground interface. For a given snow density, however, a larger incidence angle at the
snow surface results in a greater change in the refractive angle in the snow layer.
Therefore, a greater increase in the backscattered power at larger incidence angles
is expected than that at smaller incidence angles.
Furthermore, the effects of snow density at different co-polarizations HH and VV
shows a smaller increase in VV polarization than in HH polarization for the same
37
snow density and ground surface properties. The backscattering of VV polarization

as a function of incidence angle declines more slowly than that of HH polarization
when the surfaces are not too rough. Therefore, the changes in VV polarization are
smaller than that of HH polarization. However, for a very rough surface, the angular
dependence is smaller than that for a smooth surface, and the difference between
VV and HH polarization is smaller.
While a dry snowpack does not absorb or scatter the radar signal at low frequencies, it nevertheless affects the magnitude of the backscattering from the underlying
rock or soil and the relationship between HH and VV polarization. The magnitude of
the effect depends on the radar incidence angle, snow density, roughness and dielectric properties of the soil. Snow is more likely to enhance the backscattering magnitude of a smooth soil than a rough soil. These factors enable development of an
algorithm for inferring snow density using L-band SAR measurements. At C-band
or higher frequencies, however, snow density affects the magnitude of the volume
scattering and the surface scattering properties at the snowground interface.
Snow Density Algorithm with L-band Dual-polarization SAR Measurements

Based on the above understanding of snow density effects on radar backscattering
at L-band, Shi and Dozier (2000a) develop an algorithm to estimate snow density
by characterizing the dependence of the surface backscattering on both the incidence angle and the wavelength. It was done by establishing a backscattering g hh
and g vv database using the IEM model (Fung, 1994) over a wide range of incidence angles, dielectric and roughness conditions, and incidence wavenumbers,
corresponding to a range of snow densities from 100 to 550 kg m3 . Then, the relationship of HH and VV backscattering signatures with the wide range of surface
dielectric and roughness conditions at each incidence angle and wavenumber were
characterized by using regression analysis to find coefficients to parameterize this
relationship:

g
g

g
g
g
+ c(r , k1) log10 hh
hh
+ vv
+ vv
log10
= a(r , k1) + b(r , k1) log10 hh

+ d(r , k1 ) log10
g
hh
g
vv

+ e(r , k1 ) log10
g
hh
g
vv
2
(3.7)
It represents a relationship between the surface backscattering coefficients g hh and

g vv at a given incidence angle and wavenumber for a wide range of random rough
surfaces. The form of the relationship minimizes its sensitivity to the surface dielectric and roughness properties, while maximizing its sensitivity to the incidence
angle and wavenumber. The coefficients a, b, c, d and e in Eq. (3.7) depend only on
incidence angle and wavenumber at the ground surface. They are given in (Shi and
Dozier, 2000a).
38
J. Shi
By placing Eq. (3.6) in Eq. (3.7), the algorithm for estimation of snow density by
t and t SAR measurements can be derived:
using only hh
vv

t
t
hh
vv
+
log10
Thh (i , s ) Tvv (i , s )

t
t
hh
vv
+
2 ( , )
Tvv2 (i , s )
Thh
i s

t 2

t
Tvv (i , s )
+ d(r , k1 ) log10 hh
+c(r , k1 ) log10 2 hh
t T 2 ( , )
Thh (i , s )
vv
hh i s
= a(r , k1 ) + b(r , k1 ) log10

+e(r , k1 ) log10
t T 2 ( , )
hh
vv i s
t T 2 ( , )
vv
hh i s
2
(3.8)
In Eq. (3.8), Tpp depends on the polarization pp, the incidence angle i at the air
snow interface, and the dielectric constant of snowpack s . s is the only unknown;
i can be calculated from a combination of the spaceborne orbital data and a digital
elevation model. Therefore, for a given L-band SAR measurements of t hh and
t vv , we can estimate s numerically by varying the coefficients of a, b, c, d, and e
to find the root of Eq. (3.8). It does not require a priori knowledge of the dielectric
and roughness properties of the soil under the snow. Furthermore, snow density
can be estimated from Looyengas semiempirical dielectric formula (Looyenga,
1965), which provides a good fit to Polder and van Santens physical formula
(Matzler, 1996).
s = 1.0 + 1.5995s + 1.861s3
(3.9)
Three spatial distributed snow density maps were derived by this technique using three SIR-C/X-SARs L-band data-takes from its first mission in April 1994
at Mammoth Mountain study site. The three data-takes were acquired in the early
morning around 6 a.m. while the snow pit measurements taken at 11 a.m. showed
no signs of liquid water in the snow, even at 2,850 m elevation. Thus the imaged
snowpacks were dry. Using the Space Shuttle orbital geometry data and a digital
elevation model, the terrain radiometric calibration factor and local incidence angle
images corresponding to the SIR-C image data were derived. The terrain radiometric correction factor is sin 0 / sin i , where 0 is the incidence angle used in the
initial SAR data processing under a flat surface assumption and i is the actual local incidence angle (Van Zyl et al., 1993). To reduce the effect of image speckle
on estimation of snow density, the Stokes matrixes were first averaged to azimuth
(25.1 m) and slant range (13.3 m) pixel spacing to form multi-look imagery. Then,
the terrain correction was made. The snow-free pixels were also masked out based
on the snow classification results derived from SIR-C/X-SARs image data (Shi and
Dozier, 1997).
Figure 3.3 shows the comparison of the field measurements (x-axis) with SIRCs L-band image data derived snow density (y-axis). The RMSEs are 42 kg m3
39
Estimated
0.4
0.3
0.2
0.2
0.3
0.4
Measured snow density

Fig. 3.3 Compares of the field measurements (x-axis) with SIR-C/X-SARs L-band image data
(y-axis) derived snow density
and 13% for absolute and relative errors, respectively. It should be noticed that the
estimated snow density should represent the mean value from the snowpacks top
and bottom layers. This is because the backscattering signal is mainly controlled by
two factors: (1) the snow density at the top layer near the surface mainly affects the
power transmissivity at the airsnow interface; (2) snow density at the bottom layer
controls the incidence angle and wavenumber at the snowground interface.
It should be noticed that the snow density algorithm described above is based on
a radiative transfer model concept Eq. (3.6) with an assumption of the wavelength
shift when radar wave propagates in snowpack without regarding how thick of snow
is. It is understandable if snow depth is too thin, it may not result in the notable wavelength shift. Unfortunately, the current used technique does not allow the evaluation
of this issue. The studies of the well controlled laboratory experiment or Monte
Carlo Model simulations are needed for further quantitative identification on what
fraction of snow thickness to radar wavelength will result in the wavelength shift.
3.4.1.2 Technique on Estimation of Snow Water Equivalency

with C- and X-Bands
Relationships Between Snow Water Equivalence and SAR Measurements
Field experiments using ground scatterometer data have shown different relationships between radar backscattering and SWE. For example, Ulaby and Stiles
(1980) showed that backscattering at 8.2 and 17.0 GHz had a positive relation
with SWE. Similarly, a positive relationship was also observed by an experiment
over a smooth subsurface at 5.3 and 9.5 GHz (Kendra et al., 1998). However, this
40
J. Shi
positive relationship existed only over a frozen subsurface, with no correlation over
the unfrozen subsurface (Bernier and Fortin, 1998). On the other hand, negative
relationships have been observed at similar frequencies, 5.3 and 9.6 GHz (Strozzi,
1999). In addition, Rott and Matzler (1987) observed no significant difference between snow-free and dry snow covered regions at 10.4 GHz. Each field experiment
represented particular snow and ground conditions. The existence of both positive
and negative relationships between radar backscattering and snow water equivalence indicates that this relationship is quite complex. The confusion may result
from varying combinations of snow and ground properties, because backscattered
power received by the radar over dry snow depends not only on the total snow mass
but also on the snows density, grain size, structure, and stratification, along with the
dielectric and roughness properties of the underlying surface. Understanding this
relationship is essential to the development of a reliable algorithm for estimating
snow water equivalence.
In order to understand the general relationship between the radar measurements
at the frequency range between C-band and Ku-band and snow depth or SWE for
dry snow cover, for simplicity, we deploy a simple backscattering model cloud
model in which the surface scattering term from airsnow interface (the first term
in Eq. (3.5)) and the interaction term between the ground surface and the snow
volume (the third term in Eq. (3.5)) are ignored. Two remaining scattering terms
the direct volume and ground surface backscattering components (the second and
fourth terms in Eq. (3.5)) are dominant backscattering signals. The former can be
described as a function of the optical thickness ( a product of the snow extinction
coefficient e and snow depth d) and the volume scattering albedo ( a function
of the snow density, particle size, size variation, stickness, and temperature). It is
given in terms of the first-order backscattering model as

3
2 (f)
v
pp
(f, i ) = T2pp (i , s )
(f) cos(r ) 1 exp
(3.10)
4
cos(r )
In the direct volume backscattering signals, the is positively related to SWE. The
underground scattering component, however, is the product of the backscattering
signals from underground surface and snow-pack attenuation properties through the
optical thickness , which is negatively related to this scattering component. In other
words, the effects of SWE in these two major scattering components are in an opposite way. It results in the major confusion in describing the relationship between
SWE and radar observations.
To further demonstrate the general relationship, we take the first-order derivative of cloud model with respect to snow depth d that represents the slope of the
backscattering curves in response to snow depth; it can be written as

t ( )
pp
3
i
2 2 e
g0
= Tpp
exp(2 e d/r )
r pp
(3.11)
d
r
4
The ground surface backscattering signal has a great impact on the relationship
between the backscattering and snow depth. The sign of the bracketed term in
41
Eq. (3.11) determines whether the relationship between the measured backscattering
signal and snow depth is positive or negative. If g0 pp > 3/4 r , the slope is negative. Physically, this means that the attenuated subsurface scattering signal, after it
passes through the snowpack, is larger than the backscattering signal generated by
the snowpack. It is possible for the bracketed term to be zero, in which case there is
no correlation between backscattering and snow depth. In this case, the attenuated
amount of the subsurface scattering signal is the same as that of the backscattering
signal generated by snowpack. Even for the same snowpack, a positive relationship
to snow depth may be observed over a smooth surface but a negative relationship
will be observed over a rough surface.
Moreover, the sensitivity of the backscattering signal to snow depth depends not
only on the snow properties, but also on the incidence angle, and the magnitude of
the subsurface backscattering. From Eq. (3.11), we can also see that as the snow
depth increases, the change in backscattering will decrease. Therefore, we expect
that the change in the backscattering measurements is more sensitive to shallow
snow than to thick snow. The extinction coefficient and volume scattering albedo
are positively correlated to the sensitivity of backscattering to snow depth. For a
given snow depth, the larger these parameters are, the greater changes will be expected. Similarly, the sensitivity is also proportional to the angle of incidence, which
indicates that a large incidence should be more effective than a small incidence for
the purpose of monitoring snow depth. Note that the parameters, such as the ground
dielectric and roughness properties in the groundsurface and snowvolume interaction terms, also affect the relationships. The combinations of all parameters control
the overall result.
Therefore, the backscattering signal from a seasonal natural snow cover and the
signals relationship to snow depth is affected by three sets of parameters:
1. Sensor parameters, which include the frequency, polarization, and incidence
geometry
2. Snowpack parameters including snow density, particle size and size variation,
free liquid water content, characteristics of particle spatial distribution (stickiness), and stratification
3. Subsurface parameters that include the dielectric and roughness properties at the
snowground interface
These complex relationships make it implausible to characterize the parameters
from the limited field experiment measurements and to derive an empirical model
for estimating snow depth or water equivalence from SAR measurements. To estimate snow depth and thereby water equivalence, we must separate the varying
backscattering signals of the subsurface or minimize the effect of the backscattering signal generated by the snowground interface. It is clear that estimating the
snow depth requires a physically based inversion model that considers all important
scattering terms and a technique to separate the backscattering components of the
snowpack itself from those of the airsnow and snowground interfaces.
42
J. Shi
Snow Depth Algorithm with C-band and X-band SAR Measurements

As shown in (Shi and Dozier, 2000a), the snow density s , ground dielectric constant g , and surface RMS height s can be estimated by L-band SAR measurements.
From these parameters, we can obtain s , r , k1 , T pp , and the reflectivity Rpp at
the snowground interface. The remaining unknowns in Eq. (3.5) are the surface
g
backscattering components pp ( f ) from the snowground interface and a pp ( f )
from the airsnow interface, the extinction coefficients e ( f ), the volume scattering albedo ( f ), and the snow depth d. Through analyses of the simulated data,
the techniques for estimating snow depth and ice particle size using SIR-C C-band
and X-SAR measurements were developed. They were based on the following approaches (Shi and Dozier, 2000b):
Reducing the effect of surface backscattering signal from airsnow interface. In
considerations of each backscattering component from dry snow cover in Eq. (3.5)
for the SWE retrieval, the backscattering signals from the airsnow interface are
generally considered as the noise since they are typically small in comparison
with the other scattering contributions and have no information on SWE. Its effect
on retrieval snow properties can be reduced by using the signal generated by the
estimated snow density and typical roughness parameters.
Developing semi-empirical models to characterize the snowground interaction terms. These models represent the snowground interaction components more
realistically than formulas developed under the assumption of independent scattering. Natural surfaces in alpine regions are quite rough. Therefore, a significant
contribution to the snowground interactions from non-coherent components is expected. The semi-empirical models are developed explicitly in terms of snowpack
volume scattering albedo, optical thickness, ground reflectivity, and surface RMS
height. They are easy to implement in the inverse algorithm.
Developing semi-empirical model to characterize the relationships between the
ground surface backscattering components at C-band and X-band. With L-band
SAR estimates of the ground dielectric constant and RMS height, the relationships
between the ground surface backscatterings at C-band and X-band can be well characterized. Thus, the unknowns in the ground surface backscattering formulation can
be reduced to one.
Parameterizing the relationships between snowpack extinction properties at
C-band and X-band. Because the extinction properties are highly correlated, we
developed analytical forms for the extinction relationships at C-band and X-band.
Using these relationships, the number of unknowns in backscattering components
from the snowpack can be reduced to two: the volume scattering albedo and the
optical thickness .
From the developments in the above, the number of unknowns has been reduced
to three in Eq. (3.5), that is , and the surface component: svv (X). Thus, these
three unknowns can be solved numerically at each pixel by using Eq. (3.5) with three
SAR measurements: t vv (C), t hh (C), and t vv (X).
To estimate snow depth d, the separation between the extinction coefficient e
and d in the estimated is needed. In other words, we need to determine the absolute
43
value of e from the estimated parameters, i.e., ( f ), ( f ) and snow density s at

each pixel. This can be done by first estimating the absorption coefficient a (X):

(X) [1 (X)]
a (X) = 1.334 + 1.2182 log(Vs ) 3.4217 log
(3.12)
(C) [1 (C)]
where Vs is the volume fraction of ice and can be derived from the estimated snow
density at each pixel by using L-band measurements. [1 ] = a d = a is the
absorption part of the optical thickness at each frequency after removing scatter (X)]
a (X)
ing effects. (X)[1
(C)[1 (C)] = a (C) gives snowpack temperature information (Shi and
Dozier, 2000b). Then, snow depth can be estimated by
d=
(X) [1 (X)]
a (X)
(3.13)
Figure 3.4 shows the comparison of the field measurements with the depths estimated from the SIR-C/X-SAR image data at Mammoth Mountain study site. The
algorithm inferred the overall trend of the snow depths, with an RMSE of 34 cm.
During the SIR-C/X-SAR overflights, 19 snow pits and the grain size measurements with detailed vertical profiles at depth intervals of 510 cm were obtained in
the field measurements. Each individual ice particle size was measured by taking
average values from three cross diameters with direction differences of 60 . There
were more than 200 grains measured in each snow pit. By taking one step further,
a formula was also developed for estimating the optically equivalent particle size,
defined as a particle size at which the extinction properties from a natural snow
volume equal those obtained from an ideal snow volume with a uniform distribution of ice particles (Shi et al., 1993). This optically equivalent particle size can be
300
Estimated
250
200
150
100
50
50
100
150
200
250
300
Measured snow depth in cm

Fig. 3.4 Compares of the field measurements (x-axis) with SIR-C/X-SARs C- and X-band image
data (y-axis) derived snow depth
44
J. Shi
2.0
Estimated
1.5
1.0
0.5
0.5
1.0
1.5
2.0
Measured grain size in mm

Fig. 3.5 Compares of the field measurements of snow particle radius with SIR-C/X-SARs C- and
X-band image data derived grain size
represented by the weighted mean size with respect to the scattering properties for
a natural snowpack and can be estimated by

rs =
0.01s (X)
2Vs S f (2.8332 + 6.6143Vs )
1
3
(3.14)
where rs is in cm. s (X) can be obtained from the estimated s (X) and (X) at
each pixel.
Figure 3.5 compares the snow particle radius between the ground measurements
and those estimated from the SIR-C/X-SAR image data. The estimated snow particle sizes were obtained from a mean value of 3 3 window at corresponding snow
pit locations on the SAR images. The algorithm performed quite well. It inferred
the overall trend of the snow particle sizes, with an RMSE of 0.27 mm.
The current technique described above for estimation of SWE requires five measurements: L-VV and L-HH to estimate snow density and ground dielectric and
roughness properties, plus C-VV, C-HH, and X-VV to estimate snow depth and
grain size. The sensitivity analysis indicated that the C-band SAR measurements
were affected mainly by the ground surface properties. The parts of the signal that
comes from a typical snowpack at C-band are about 30% and 15% for HH and
VV polarization, respectively. The C-band measurements are expected mainly sensitive to under-ground surface condition. Estimation of snow depth using C-band
SAR measurements, therefore, requires an accurate technique to estimate the ground
backscattering component. At X-band it about 60% of the signal comes from the
snowpack itself. Thus, we expect that the measurement is much more sensitive to
snowpack and that the requirement for estimation of the ground backscattering component is less severe for radar measurements at X-band or higher.
45
3.4.2 Estimations of SWE and Its Change with SAR Interferometry

From last section, it can be seen that retrieving SWE from SAR backscattering measurements requires very accurate inversion models and is a very complex process.
The other promising technique for SWE estimation is to use the repeat-pass SAR
interferometric measurements. It has been shown that the interferometric phase shift
resulting from differences in the propagation due to constant changes in snow properties offers a direct method for estimating SWE (Rott et al., 2004) or relative change
in SWE (Guneriussen et al., 2001). This technique has been applied to low frequency
interferometric SAR measurements at C-band (Guneriussen et al., 2001; Rott et al.,
2004) and L-band (Rott et al., 2004).
Considering a layer of dry snow, the radar return is dominated by scattering signal
from snow/ground interface at low frequencies (C-band or lower). The repeat-pass
interferometric phase consists of following contributions (Rott et al., 2004):
= f lat + topo + atm + snow + noise
(3.15)
flat and topo are the phase differences due to changes of the relative distance between satellite and target for flat earth and for topography, respectively. atm results
from changes in atmospheric propagation, and noise is phase noise. snow is the
two-way propagation difference in the snow-pack relative to air and is resulted from
the refraction of radar wave in dry snowpack. When snowpack volume scattering
is neglected, the snow phase term snow for an uniform layer of snow with the
depth d can be written as (Guneriussen et al., 2001):

snow = 2ki d cos i s sin2 i
(3.16)
For ERS SAR with an incidence angle i = 23 , the phase shift due to a change of
SWE can be approximated by a linear relation (Guneriussen et al., 2001):
snow = 2ki 0.87 SW E
(3.17)
It indicates that at the ERS wavelength one fringe is equivalent to 32.5 mm SWE,
and for L-band ( = 24 cm, i = 23 ) one fringe corresponds to SWE = 138 mm.
The coherence is determined by several factors and given by Hoen and Zebker
(2000):
total = thermal sur f ace volume temporal
(3.18)
thermal depends on the signal-to-noise ratio. Its contribution to decorrelation of dry

snow covered ground is usually small since the signals from dry snow ground covered are generally strong. For a dry snow cover, the wave number shifts in slant
range and in vertical direction have to be considered, which affects mainly on
the terms of sur f ace and volume . Their calculations are given by Hoen and Zebker
(2000). The temporal decorrelation is a major effect on in coherence of repeat pass
SAR data over dry snow covered areas. It is mainly resulted from changes of the
46
J. Shi
snowpack properties due to snow fall or snow drift. They result in changes of the
snowpack structure and the roughness of the snow-surface at sub-pixel scale, and
consequently change the propagation path in the snowpack of the radar return from
each surface element. Rott et al. (2004) shows the term of the temporal correlation
for dry snow cover can be expressed as

temporal = exp 2 ki2 z2 cos i s sin2 i
(3.19)
where z is the standard deviation of the geometric path length through the snowpack. It takes into account of differential phase delay due to non-uniform snow
accumulation or erosion.
Guneriussen et al. (2001) applied this technique using ERS-1/2 tandem data and
found that the repeat-pass interferometric phase measurements were very sensitive
to small changes in snow properties that could introduce a significant error in DEM
estimation even in the case of high degree of coherence. To avoid the temporal
decorrelation effects, Rott et al. (2004) used ERS repeat pass data without snowfall between data acquisitions and derived SWE map over their study site using this
technique with the required corrections for all other phase components in Eq. (3.14)
and the reference points with zero and known snow accumulation. Both studied have
shown that the interferometric phase shift from repeat-pass SAR measurements provides a physically based means for mapping SWE of dry snow. Temporal decorrelation due to differential phase shifts at sub-pixel scale caused by snow fall or wind
re-distribution is the major limitation for application of this method. In C-band SAR
data, these effects often result to complete decorrelation within a few days. L-band
is preferable for this technique on mapping SWE than C-band because it is less
affected by temporal decorrelation, better coherence and larger measurement range.
3.5 Need for Future Spaceborne System for Monitoring Snow

Properties
The current available spaceborne C-band imaging radar systems are either single polarization (ERS-2 with VV polarization and RADARSAT-1 with HH polarization) or
dual-polarization ASAR (HV/VV or VH/HH). For monitoring snow properties with
backscattering measurements, these current systems are only capable to identify the
extent of wet snow. They are not capable of measuring SWE or of quantifying snow
wetness. For interferometric measurements, the current and near future satellite systems, however, have the repeat orbit at least 24 days or longer. The presented SWE
estimation technique in Section 3.4.2 requires the acquisitions within a few days
and is only applicable to ERS 1/2 tandem data. For using interferometric technique
to estimate SWE, the low frequency SAR instruments (L-band) is required with a
quick repeat time interval. The recent launched PALSAR system has 48 days repeat pass and does not match this requirement. However, it provides the unique
tool for mapping both dry and wet snow cover with the combined interferometric
47
coherence and backscattering measurements as presented in Section 3.2.3 but do

not offer capabilities to estimate SWE and snow wetness over the range of Earths
environments.
Taking into account the capabilities and deficiencies of present remotely sensed
snow cover information, its is obvious that improvements are necessary in particular
regarding the observation of extent and physical properties of snowpacks SWE and
wetness. Sufficiently high spatial resolution is needed to account for the small scale
heterogeneity of snowpack related to topography and land cover.
For using backscattering signals, two critical capabilities of the radar instrument
are required for monitoring snow water equivalence (SWE) globally. First, the radar
signal can penetrate the natural snowpacks so that the snow depth information can be
obtained from the measurements. Secondly, the radar signal from snowpack itself
has to be large enough to provide the first-order information for snowpack itself.
At the frequencies at or lower than C-band, snowpack does not generate the significant volume scattering signals and the attenuation of the ground signals also
weak (Shi and Dozier, 2000b). At Ku-band (1319 GHz), the penetration depths
are ranged from 0.5 (extreme large grain) to 10 m (extreme small grain). The studies have shown the one-way penetration depth of dry snow from experimental data
which at 17 GHz ranges from about 4 m for coarse grained Antarctic firn (Rott et al.,
1993) to 6 m for fine-grained Alpine winter snow (Matzler, 1987). Secondly, the
direct volume scattering component contributes about 60% radar measurement signals in the co-polarization for a typical dry snow condition at X-band 9.25 GHz
(Shi and Dozier, 2000b). The numerical simulations (Shi et al., 2003; Shi, 2004)
have indicated that the direct volume contribution will be significantly increased in
both co-polarized and cross-polarized signals at Ku-band 17 GHz. The direct volume scattering becomes the dominant scattering source. These numbers indicate that
17 GHz is a good choice for sensing volume properties of dry snow and firn, because
at this frequency the snow-pack provides a clear signal but is optically transparent
enough to achieve sufficient penetration. Recent studies of the global snow cover
with NSCAT, operating at 14 GHz also demonstrated the sensitivity of Ku-band to
sense dry and wet snow cover (Nghiem and Tsai, 2001). Therefore, we expect that
X and Ku-band instruments have capability to penetrate the most of natural dry
snow cover globally and provide much more sensitive to snowpack itself than those
by C-band. For future satellite mission on monitoring SWE over range of Earths
environments, the optimal sensor configuration and corresponding algorithm development are needed.
References
Bernier M, Fortin JP (1998) The potential of time series of C-band SAR data to monitor dry and
shallow snow cover. IEEE Trans. Geosci. Remote Sens. 36:226243
Chen KS, Wu TD, Tsang L, Li Q, Shi J, Fung AK (2003) The emission of rough surfaces calculated
by the integral equation method with a comparison to a three-dimensional moment method
simulations. IEEE Trans. Geosci. Remote Sens. 41(1):90101
48
J. Shi
Ding K, Zurk LM, Tsang L (1994) Pair distribution functions and attenuation rates for sticky
particles in dense media. J. Electromagnet. Waves Appl. 8:15851604
Evans D, Farr TG, van Zyl JJ, Zebker HA (1988) Radar polarimetry: analysis tools and applications. IEEE Trans. Geosci. Remote Sens. 26:774789
Evans D, Plaut J, Stofan E (1997) Overview of the spaceborne imaging radar-c/x-band synthetic
aperture radar (SIR-C/X-SAR) missions. Remote Sens. Environ. 59(2):135140
Fung AK (1994) Microwave scattering and emission models and their applications. Artech House,
Norwood, MA
Guneriussen T, Hgda KA, Johnsen H, Lauknes I (2001) InSAR for estimation of changes in snow
water equivalent of dry snow. IEEE Trans. Geosci. Remote Sens. 39(10):21012108
Haefner H, Holecz F, Meier E, Nusch D, Piesbergen J (1993) Capabilities and limitations of ERS-1
SAR data for snow cover determination in mountainous regions. Proc. Second ERS-1 Symposium, ESA PS, pp 353359
Hoen EW, Zebker HA (2000) Penetration depths inferred from interferometric volume decorrelation observed over the Greenland ice sheet. IEEE Trans. Geosci. Remote Sens.
38(6):25712583
Kendra JR, Sarabandi K, Ulaby FT (1998) Radar measurements of snow: experiment and analysis.
IEEE Trans. Geosci. Remote Sens. 36:864879
Lee JS, Papathanassiou KP, Ainsworth TL, Grunes MR, Reigber A (1998) A new technique
for noise filtering of SAR interferometric phase images. IEEE Trans. Geosci. Remote Sens.
36(5):14561465
Li Q, Shi J, Chen KS (2002) A generalized power law spectrum and its applications to the
backscattering of soil surface on the integral equation model. IEEE Trans. Geosci. Remote
Sens. 40(2):271280
Looyenga H (1965) Dielectric constant of heterogeneous mixtures. Physica 21:401406
Luojus K, Pulliainen J, Metsamaki S, Hallikainen M (2006) Accuracy assessment of SAR databased snow-covered area estimation method. IEEE Trans. Geosci. Remote Sens. 44(2):277287
Matzler C (1987) Applications of the interaction of microwaves with the natural snow cover. Remote Sens. Rev. 2:259387
Matzler C (1996) Microwave permittivity of dry snow. IEEE Trans. Geosci. Remote Sens.
34:573581
Nagler T, Rott H (2000) Retrieval of wet snow by means of multitemporal SAR data. IEEE Trans.
Geosci. Remote Sens. 38(2):754765
Nghiem SV, Tsai WY (2001) Global snow cover monitoring with spaceborne Ku-band Scatterometer. IEEE Trans. Geosci. Remote Sens. 39(10):21182134
Pulliainen J, Engdahl M, Hallikainen M (2003) Feasibility of multitemporal interferometric
SAR data for stand-level estimation of boreal forest stem volume. Remote Sens. Environ.
85:397409
Rott H, Matzler C (1987) Possibilities and limits of synthetic aperture radar for snow and glacier
surveying. Annals Glaciology 9:195199
Rott H, Nagler T (1993) Capabilities of ERS-1 SAR for snow and glacier monitoring in alpine
areas. Proc. Second ERS-1 Symposium, pp 16
Rott H, Matzler C, Strobl D, Bruzzi S, Lenhart KB (1988) Study on SAR land applications for
snow and glacier monitoring. Technical Report 6618/85/F/FL(SC), European Space Agency
Rott H, Nagler T, Scheiber R (2004) Snow mass retrieval by means of SAR interferometry Proc.
FRINGE 2003 Workshop, ESA SP-550, article-46, Frascati, Italy
Shi J (2004) Estimation of snow water equivalence with two Ku-band dual polarization radar. Proc.
IGRASS04, IEEE No. 04CH37612C
Shi J, Dozier J (1995) Inferring snow wetness using C-band data from SIR-Cs polarimetric synthetic aperture radar. IEEE Trans. Geosci. Remote Sens. 33(4):905914
Shi J, Dozier J (1997) Mapping seasonal snow with SIR-C/X-SAR in mountainous areas. Remote
Sens. Environ. 59(2):294307
Shi J, Dozier J (2000a) Estimation of snow water equivalence using SIR-C/X-SAR, part I: inferring
snow density and subsurface properties. IEEE Trans. Geosci. Remote Sens. 38(6):24652474
49
Shi J, Dozier J (2000b) Estimation of snow water equivalence using SIR-C/X-SAR, part II: inferring snow depth and particle size. IEEE Trans. Geosci. Remote Sens. 38(6):24752488
Shi J, Jiang L (2004) Numerical simulations on estimation of snow wetness with dual-frequency
and polarization radar. Proc. SPIE, Honolulu 5654:149156
Shi J, Davis RE, Dozier J (1993) Stereological determination of dry snow parameters for discrete
microwave modeling. Annals Glaciology 17:295299
Shi J, Dozier J, Rott H (1994) Snow mapping in alpine regions with synthetic aperture radar. IEEE
Trans. Geosci. Remote Sens. 32(1):152158
Shi J, Hensley S, Dozier J (1997) Mapping snow cover with repeat pass synthetic aperture radar.
Proc. IGARSS 97, IEEE No. 97CH36042, vol II, pp 628630
Shi J, Yueh S, Cline D (2003) On estimation of snow water equivalence using L-band and Ku-band
radar. Proc. IGRASS03, IEEE No. 03CH37477C
Stiles WH, Ulaby FT (1980) The active and passive microwave response to snow parameters, 1,
wetness. J. Geophys. Res. 85:10371044
Strozzi T, Matzler C (1998) Backscattering measurements of Alpine snowcovers at 5.3 and 35 GHz.
IEEE Trans. Geosci. Remote Sens. 36(3):838848
Strozzi T, Wiesmann A, Matzler C (1997) Active microwave signatures of snowcovers at 5.3 and
35 GHz. Radio Sci. 32(29):479495
Strozzi T, Wegmuller U, Matzler C (1999) Mapping wet snow covers with SAR interferometry.
Int. J. Remote Sens. 12(20):23952403
Tsang L (1992) Dense media radiative transfer theory for dense discrete random media with particles of multiple sizes and permittivities. Prog. Electromagnetic Res. 6(5):181225
Tsang L, Kong JA, Shin RT (1985) Theory of microwave remote sensing. Wiley, New York
Ulaby FT, Stiles WH (1980) The active and passive microwave response to snow parameters: 2,
water equivalent of dry snow. J. Geophys. Res. 85:10451049
Ulaby FT, Stiles WH (1981) The active and passive microwave response to snow parameters: 2,
water equivalent of dry snow. J. Geophys. Res. 85(C2):10451049
Ulaby FT, Moore RK, Fung AK (1982) Microwave remote sensing: active and passive, 2, radar
remote sensing and surface scattering and emission theory. Addison-Wesley, Reading, MA
Ulaby FT, Stiles WH, Abdelrazik M (1984) Snowcover influence on backscattering from terrain.
IEEE Trans. Geosci. Remote Sens. GE-22:126132
Ulaby FT, Moore RJ, Fung AK (1986) Microwave remote sensing, active and passive. vol III.
Artech House, Durham, MA.
Van Zyl JJ, Zebker HA, Elachi C (1987) Image radar polarization signatures: theory and observations. Radio Sci. 22(4):529543
Van Zyl JJ, Chapman BD, Dubois P, Shi J (1993) The effect of topography on SAR calibration.
Wu TD, Chen KS, Shi J, Fung AK (2001) A transition model for the reflection coefficient in surface
scattering. IEEE Trans. Geosci. Remote Sens. 39(9):20402050
Chapter 4
Multi-angular Thermal Infrared Observations

of Terrestrial Vegetation
Massimo Menenti, Li Jia, and Zhao-Liang Li
Abstract This chapter reviews the experimental evidence on the anisotropy of

emittance by the soilvegetation system and describes the interpretation of this signal in terms of the thermal heterogeneity and geometry of the canopy space. Observations of the dependence of exitance on view angle by means of ground-based
goniometers, airborne and space-borne imaging radiometers are reviewed first to
conclude that under most conditions a two-components, i.e., soil and foliage, model
of observed Top Of Canopy (TOC) brightness temperature is adequate to interpret observations. Particularly, airborne observations by means of the Airborne
Multi-angle TIR/VNIR Imaging System (AMTIS) and space-borne observations
by means of the Along Track Scanning Radiometers (ATSR-s) are described and
examples presented. Modeling approaches to describe radiative transfer in the soil
vegetationatmosphere system, with emphasis on the thermal infrared region, are
reviewed. Given the dependence of observed TOC brightness temperature on leaflevel radiation and heat balance, energy and water transfer in the soilvegetation
atmosphere system must be included to construct a realistic model of exitance by
soilvegetation systems. A detailed modeling approach of radiation, heat and water transfer is first described then applied to generate realistic, multi-angular image
data of terrestrial landscapes. Finally, a generic algorithm to retrieve soil and foliage
component temperatures from Top Of Atmosphere (TOA) radiometric data is described. Column water vapor and aerosols optical depth are estimated first, to obtain
TOC radiometric data from the TOA multi-angular and multi-spectral observations.
M. Menenti
TRIO/LSIIT, University Louis Pasteur (ULP), Strasbourg, France
Istituto per i Sistemi Agricoli e Forestali del Mediterraneo (ISAFOM), Naples, Italy
m.menenti@isafom.cnr.it
L. Jia
Alterra, Wageningen University and Research Centre, The Netherlands
Z.-L. Li
TRIO/LSIIT, University Louis Pasteur (ULP), Strasbourg, France
Institute of Geographic Sciences and Natural Resources Research, Beijing, China
51
52
M. Menenti et al.
Then vegetation fractional cover and soil and foliage component temperatures are
determined by inverting a simple two-components mixture model. The accuracy of
all elements of the algorithm is evaluated by using a combination of actual measurements and synthetic radiometric data. Although applicable to multi-angular radiometric data irrespective of spatial resolution, the approach presented would be
particularly relevant if space-borne observations with a footprint of 100 100 m or
better would be available. Observing systems, presently at the design stage, with
this capability are briefly described.
4.1 Introduction
4.1.1 VegetationAtmosphere Exchanges of Energy and Water
The exchange of energy between the land surface and the atmosphere and within
terrestrial vegetation canopies is a significant determinant of processes in the atmospheric boundary layer and in terrestrial ecosystems. In these processes it is
crucial to determine accurately the partitioning of available energy into sensible
heat flux density H (heating or cooling of the surface) and latent heat flux density
E (evaporation from surface) over a wide range of spatial and temporal scales.
Observation and modeling of turbulent heat fluxes at the land surface has been a
very active area of research at least since the work of Bowen (1926) on the relative
magnitude of heat transfer over dry and wet surfaces (Monteith, 1965; Feddes, 1971;
Verma et al., 1976; Hall et al., 1979; Price, 1982; De Bruin and Jacobs, 1989;
Beljaars and Holtslag, 1991; Lhomme et al., 1994; Sellers et al., 1995, 1996). Most
conventional techniques that employ point measurements to estimate the components of energy balance are representative only of local scales and cannot be extended to large areas because of the heterogeneity of the land surface, of the dynamic
nature and of the spatial distribution of heat transfer. Remote sensing is one of the
few techniques to provide representative measurements, e.g., surface temperature
and albedo, at regional and global scales.
Methods using remote sensing techniques to estimate heat exchange at the landatmosphere interface fall into two main categories: (1) use surface radiometric temperature Trad to calculate H then obtain E as the residual of the energy balance
equation (Blad and Rosenberg, 1976; Seguin et al., 1989; Hatfield et al., 1984);
(2) use Trad to estimate the Crop Water Stress Index or the evaporative fraction (the
ratio of evapotranspiration to the available energy) (Jackson et al., 1981; Menenti
and Choudhury, 1993; Moran et al., 1994). The former category can be further
subdivided into single-source, dual-source and multisource models corresponding
with a single-, dual- or multi-layer schematization of the surface respectively. Successful estimations of heat fluxes have been achieved over horizontal homogeneous
surfaces, such as a surface fully covered by vegetation, open water and bare soil
(Jackson, 1985; Huband and Monteith, 1986; Choudhury et al., 1986). Large deviations from these conditions occur at partial canopies which are geometrically and
4 Multi-angular Thermal Infrared Observations of Terrestrial Vegetation
53
thermally heterogeneous. Recent years have seen increasing evidence of specific

difficulties inherent to the heterogeneous nature of terrestrial vegetation. For instance, in many semi-arid environments where the surfaces are partially covered by
vegetation, both the soil surface and cooler foliage determine the heat exchanges.
This leads to the challenge of relating the separate contributions from these elements
to the turbulent transport of heat across the land-atmosphere interface.
Exchange of water and CO2 between land surface and atmosphere determines
to a significant extent the dynamics of the Convective Boundary Layer (CBL)
(Mahfouf et al., 1987; Segal et al., 1988; Hatfield, 1991; Xue and Shukla, 1993).
Over homogeneous land surfaces the controlling factor is the partition of net radiation into sensible, latent and soil heat flux. The partition of net radiation is
determined by the presence and functioning of vegetation and by available soil moisture. Heterogeneous land surfaces compound the complexity of these processes,
since the spatial pattern of land surface properties determines CBL motion at small
length scales (Hatfield, 1991; Wang and Mitsuta, 1992). These studies brought the
attention of a wide scientific community to the need for significant improvements in
models of such land surface processes and of the interactions of land surfaces with
the atmosphere (e.g., Avissar and Pielke, 1989; Pielke et al., 1991). High resolution
atmospheric models may be used to interpret observations in complex landscapes as
for example done by Peng and Cheng (1993) and Yan (1999), in the framework of
the Hei He basin International Field Experiment (HEIFE).
Observations of the anisotropic emittance of land cover provide unique access to the thermal heterogeneity of soilvegetation systems. All land surfaces
are anisotropic and in the thermal-infrared domain, the directional variation of
emitted fluxes (described by the so-called brightness temperature) is mainly determined by the distribution of temperature and emissivity between the elements
of the canopy, and by the structure of the vegetation (see Balick et al., 1987;
Kimes and Kirchner, 1983). Similar to the solar domain, the distribution of shadowed and illuminated parts, as well as the amount of soil and vegetation observable
from a particular direction, are the main drivers of the anisotropy models that have
been developed to describe the directional variations in thermal infrared spectral
domain (e.g., Norman and Chen, 1990; Ottermann, 1990; Smith and Goltz, 1995;
Francois et al., 1997).
Determination of the soil and foliage component temperature requires inverse
modeling of observed emittance, which is a challenge (Jia et al., 2003; Jia, 2004;
Francois et al., 1997; Li et al., 2001; Francois, 2002; Liu, 2002), given the strength
of the signals and the accuracy which can be achieved. On the other hand it provides
the only opportunity to overcome a major shortcoming of current parameterizations
of heat fluxes at the landatmosphere interface (Jia, 2004).
Observations of foliage and soil temperature provide also a reliable indicator
of crop water stress, either directly as soil-foliage temperature difference or indirectly through modeling of soil evaporation and plant transpiration (Francois, 2002;
Luquet et al., 2003).
54
M. Menenti et al.
4.1.2 Photosynthesis: Light Use Efficiency of Sun-lit

and Shadowed Leaves
Observations of leaf and soil temperature are also quite relevant for understanding
and modeling carbon exchange between terrestrial vegetation and the atmosphere.
The rate of photosynthesis depends on many factors, including carbon dioxide
concentration in the atmosphere, leaf temperature, or mineral deficiencies (in particular nitrogen) in the soil. The nitrogen content of leaves is strongly related to their
chlorophyll content (Field and Mooney, 1986).
Foliage temperature. Autotrophic respiration is the process by which some of the
chemical energy stored by photosynthesis is used by the plants themselves to grow
and develop. This process is critical to the carbon cycle because it results in the rapid
release of a large fraction of the carbon initially stored through photosynthesis back
to the atmosphere. Autotrophic respiration depends on foliage temperature, growth
rates and total biomass, as well as on the biochemical composition of the products
formed in the plants (Medlyn et al., 2002).
Soil temperature. Heterotrophic respiration is the process by which some of the
carbon stored in organic soil components is released. The soil carbon reservoir can
be very large compared to the above-ground biomass. Understanding the fluxes of
carbon to (senescence, mortality) and from (respiration or mineralization) this soil
reservoir becomes a major issue when closing the carbon cycle at the local scale.
Heterotrophic respiration is very dependent on soil temperature, and the availability
of water and nutrients, particularly nitrogen. Apart from nitrogen fertilization or
deposition, symbiotic fixation of atmospheric nitrogen, and leakage or volatilization,
the nitrogen cycle is intimately linked to the carbon cycle within the soil via the
biotic activity. Evaluation of heterotrophic respiration is a major challenge in the
description and modeling of NEP (Valentini et al., 2000; Matteucci et al., 2000).
Due to the strong temperature dependence (Fig. 4.1) of both leaf photosynthesis and soil respiration, observations of foliage and soil temperature are useful to
Soil respiration [umol CO2 m2 s1 ]
An(Normalisedto 1 at 25C)
1.6
An
1.4
1.2
1
Betula pendula
0.8
Quercus petraea
0.6
Acer pseudoplatanus
0.4
0.2
Gossypium hirsutum
0
0
10
15
20
25
30
Leaf temperature ( C)
35
40
6
5
Col., Q10=2.21
4
3
2
Wal., Q10=2.16
1
Mdm., Q10=2.09
10
15
20
Soil temperature at 5 cm depth [C]
Fig. 4.1 Carbon fluxes and component soil and foliage temperatures: (left) relative net
photosynthesis, An (normalized to the observed value at 25 C), vs. foliage temperature
(Medlyn et al., 2002); (right) soil respiration vs. soil temperature. (After Matteucci et al., 2000.)
55
understand and model these processes. As regards field experiments, the use of
very high resolution TIR images is a widely used solution to obtain direct measurements of foliage and soil temperature (see Section 4.5). As regards observations from space, the only feasible solution is multi-angular observations of Top Of
Canopy (TOC) brightness temperature and subsequent inverse modeling to determine foliage and soil temperature from observed anisotropy of emittance. A minimum of two spectral bands is required to establish the brightness temperature of the
surface, taking atmospheric influences into account.
4.2 Nature of Anisotropic Emittance of Land Surface

4.2.1 Anisotropic Emissivity of Leaves and Soils
The observed anisotropy of land emittance is (see Section 4.3) caused by a combination of two different processes: the inherent anisotropic emissivity of terrestrial
materials and the thermal heterogeneity of complex, three-dimensional (3D) land
targets. The latter is due to the interaction of radiative and convective energy transfer with the 3D structure of land targets These processes are briefly reviewed in this
section.
The accuracy of radiometric measurements of land surface temperature depends
significantly on accurate knowledge of land emissivity. A 1% uncertainty on surface
emissivity can cause about 0.6 error on land surface in temperature (Becker, 1987).
An angular variation of emissivity has been observed by a number of scientists either
in the field or in the laboratory (Becker et al., 1985; Labed and Stoll, 1991; Xu
et al., 2000). For example, when sea surface wind produces wavelets, an angular
variation of surface emissivity is observed (Masuda et al., 1988; Francois and Ottlè,
1994), and an angular variation of the effective surface emissivity has also been
observed with satellite data (Nerry et al., 1998; Petitcolin et al., 2002).
Many efforts have been devoted to measure the directional emissivity of soil,
leaves and other natural surfaces. Commonly, emissivity of natural surfaces decreases with increased zenith angle of observation. Becker et al. (1985) measured
in the thermal infrared band the bi-directional reflectivity (BDR) of different types
of bare soils including quartz sands, agricultural soil, and well-calibrated powders,
a large variability of the BDR for different samples with no apparent systematic behavior apart from the backscattering peak was observed and the roughness has large
impact on the BDR which means that the same material with different grain sizes
may exhibit different angular distributions.
As an example, Fig. 4.2 gives the measured angular variation of emissivity for
water, sand, clay, slime, and gravel (Sobrino and Cuenca, 1999).
56
M. Menenti et al.
Fig. 4.2 Angular variation of surface emissivity for several natural surfaces. (After Sobrino and
Cuenca, 1999, Fig. 3.4.)
Fig. 4.3 Anisotropy change resulting from changing the single-scattering reflectance from 0.3 to
0.2 in the Hapke model. The plot is in the 30 azimuth plane with the angle of incidence at 32 .
Positive zenith values represent backscattering. (After Snyder et al., 1997.)
To describe the anisotropic emissivity of sands and soils, Snyder et al. (1997)
defined a so-called anisotropy factor, a:
a=
f
1
(4.1)
where f is the bi-directional reflectance distribution function (BRDF) of the surface,

with the same meaning of the bidirectional reflectivity, the emissivity.
To illustrate the magnitude of anisotropy, a simulation with Hapke BRDF model
was done by varying single-scattering reflectance and zenith angle (Fig. 4.3).
57
This figure shows that the difference between the two curves for different singlescattering reflectance is relatively small with respect to the overall variation. The
simulation by Snyder et al. (1997) demonstrates that for typical materials, over a realistic range of azimuth and incident and reflected zenith angles, the upper bounds of
the impact of this difference is 4.8% (RMS) and 9.5% (maximum error) respectively.
Efforts to evaluate the anisotropy of foliage emissivity have been limited, because
it is very close to unity and has small angular variation. Therefore, for most studies, leaves are assumed to be Lambertian. The ray-tracing method has been applied
to estimate the directional reflectance of canopy in VIS/NIR. An extension of this
method to the TIR region may be useful to simulate the angular variation of canopy
thermal infrared reflectance.
4.2.2 The 3D Structure of Vegetation Canopies

The architecture of most vegetation canopies leads to a complex three-dimensional
distribution of absorbed radiant energy and, therefore, of the local balance of energy
within the canopy space (Fig. 4.4). On the one hand, within the canopy space the
surface temperature of foliage and soil varies significantly. On the other hand, the
20
Frequency (%)
11 April, 13:00, s50

15
10
5
0
10
15
20
25
30
35
Brightness temperature (C)

20
Frequency (%)
11 April, 10:30, s50
15
10
5
0
10
15
20
25
30
35
Brightness temperature (C)
Fig. 4.4 Thermal anisotropy of soilvegetation system resulting from canopy geometry and variability of absorbed solar radiation by different components: wheat leaves (blue) are significantly
colder than soil (red). (After Jia, 2004.)
58
M. Menenti et al.
nadir view
off - nadir view

v
cool vegetation
warm soil
Fig. 4.5 Illustration of observed TIR radiance as a function of canopy geometry, foliage and soil
component temperatures and the zenith view angle v . The light gray bar indicates the fraction
of foliage in the instantaneous field of view (IFOV) of the sensor, the dark gray bar indicates the
fraction of soil in the IFOV. (Jia 2004.)
vertical distribution of foliage temperature is also variable with the solar elevation,
the density of leaves and the angle distribution of leaves. The thermal heterogeneity
within a vegetation canopy leads to the fact that Trad measured by thermal infrared
(TIR) sensors is a function of canopy geometry, vertical distribution of foliage temperature Tf , soil temperature Ts , sensor view angle (v , v ) and incoming radiation
(Kimes, 1980; Francois et al., 1997) (Fig. 4.5).
4.2.3 Radiation and Convection in the Canopy Space

For incomplete canopies, a frequent case in nature, the interaction between the
canopy and the atmosphere becomes complex due to the canopy geometry in terms
of the size and spacing between plants, the leaf density and the leaf angle distribution. Figure 4.6 illustrates how the elements of a sparse canopy are interacting
with their environment. The complex canopy geometry determines the distribution
of absorbed solar radiation in the canopy, thereafter inducing spatial variability of
sources and sinks of heat and water vapor in the canopy space. A large spacing
between plants or lower leaf density, for instance, makes the exposed soil to play
an important role in the landatmosphere interaction. Canopy geometry has also
influence on the airflow in the canopy space and the boundary layer resistance of
leaves and soil, thus changing the source/sink strength. The interaction between
thermodynamic and dynamic processes will lead to thermal heterogeneity, which
will in turn give rise to the anisotropy in the exitance of canopy. (See, e.g., Albertson
et al., 2001; Wilson et al., 2002)
59
Horizontal
Divergence
Canopy Height
NET
SENSIBLE
RADIATION
HEAT
WATER
VAPOR
Water Vapor
Water Vapor
Sensible Heat
Horizontal
Divergence
dz
dy
dx
Storage terms in volume:

Temp. change of crop
Temp. change of moist air
Absolute humidity change
Sensible Heat
Soil surface
SOIL HEAT
Horizontal
Divergence
Fig. 4.6 Schematic illustration of the 3D structure of a vegetation canopy and of the interactions
between canopy elements and the canopy environment.
4.2.4 Thermal Heterogeneity of Vegetation Canopies

The radiative and thermal state of elements in a soil-vegetation canopy system is
strongly dependent on its geometric structure (radiative transfer) and on its environmental situation (convection and conduction processes). The interactions between
the land surface and the atmosphere consist of the interactions between foliage and
soil surface, foliage and air in the canopy space, soil surface and air in the canopy
space, and between the soil surface and deeper soil layers, i.e., root zone. A realistic
model requires describing the processes involved at each spatial point. However, it
may not be possible or necessary to do so. Adequate simplification is necessary to
redefine the canopy, which should retain the dominant aspects of 3D radiative, heat
and mass transfer (see Section 4.3).
Commonly, the soil temperature is much higher than foliage temperature because
of the different thermal properties between soil and foliage. Thus, to simulate energy
exchange between canopy and natural environment, vegetative canopy is commonly
represented by two-source model: the sensible heat flux between soil and air, and
the one between foliage and air. As the penetration of downwelling solar shortwave radiation and long-wave sky radiation, the foliage temperature changes with
the depth of canopy. The temperature profile of vegetative canopy with the height
of canopy has been observed by several scientists, e.g., by Paw et al. (1989) who
measured the variation of leaf temperature with height within the canopy space by
means of a Teletemp Infrared Thermometers (Fig. 4.7).
The temperature profile of canopy layers result from the energy balance for each
layer. As seen from Fig. 4.7 both foliage at top layers and at bottom layer have
60
M. Menenti et al.
HEIGHT(cm)
200
150
110
50
HEIGHT(cm)
0
150
200
50
0
19
20
21
22
23
24
25
26
27
28
29
30
31
TEMPERATURE(C)
Fig. 4.7 Leaf temperature profile in the canopy, at 1,2001,300 h, for all days. The top graph
shows each individual leaf temperature measurement. The bottom figure shows a 10-point moving
average of leaf temperature. (After Paw et al., 1989, Fig. 7.)
higher temperature. The higher temperature of top layer is due to the relatively large
incident short-wave radiation which may decrease with depth within the canopy.
The bottom layer of vegetative canopy will receive more long-wave radiation than
the middle canopy layer from the soil which temperature is usually higher than
foliage, with convective cooling being less efficient than at the top of the canopy.
Lei (2004) made a detailed temperature measurements of sunlit and shaded
leaves of Blackbrush and sunlit soil from March through December 2001 at the
Clark Mountain (roughly 35.7 N and 115.5 W; elevation 1.475 m). Figure 4.8
shows the temperature difference between sunlit and shaded foliage and the
temperature difference between sunlit soil and foliage.
4.3 Observed Magnitude of Anisotropy

4.3.1 Ground Observations of Tb ( ) at TOC
The anisotropy in canopy exitance implies that brightness temperature (Tb0 ) at the
Top Of the Canopy (TOC) changes with view zenith angle v and azimuth angle v
as shown by field measurements over a range of canopies, especially over sparse
Soil-leaf temperature difference (C)
Leaf temperature difference (C)
4
Spring
Summer
Winter
9
12 15
Hour of day
18
61
20
15
Spring
Summer
Winter
10
9
12 15
Hour of day
18
Fig. 4.8 Temperature difference between: sunlit and shaded leaves (left); sunlit soil surface and
sunlit leaves (right); Clark Mountain of southeastern California. (After Lei, 2004, Figs. 4 and 5.)
Fig. 4.9 Observation of TOC brightness temperature Tb0 at different view angles. The circles represent the footprints of IFOV at the Top Of Canopy (TOC) and at the bottom of the canopy for
different view angles. The components in the volume between TOC and the bottom are observed
by a radiometer located above the canopy. (Jia, 2004.)
canopies. The dependence of observed Top Of Canopy (TOC) brightness temperature Tb0 (v , v ) on the view angle is best documented by ground measurements
with a goniometer-mounted radiometer (Fig. 4.9, After Jia, 2004). Tb0 (v , v ) is the
temperature measured by a radiometer at zenith angle v and azimuth angle v and
is simply derived from the radiance R measured by a radiometer by inverting the
Plancks function. Therefore, the observation of Tb0 (v , v ) and the observation of
TIR radiance are equivalent. In this chapter, for simplification the symbol R denotes only TIR radiance and the subscript TIR will be neglected. TIR radiance of
62
M. Menenti et al.
a target is usually referred to as exitance, i.e., the sum of emitted and reflected
TIR radiance by the target concerned. Nielsen et al. (1984) have shown that it is
common to have large (up to 20K or more) differences between sun-lit soil and
shadowed leaf surfaces, particularly when the top soil is dry. [127] found differences between bare soil and air temperature as large as 27 C. For a soybean canopy
with 35% ground cover, the soil temperature exceeded the canopy temperature by
11 C and was 15 C higher than air temperature (Kimes 1980). Usually, Tb0 (v , v )
is measured by a radiometer in a specific spectral range (centered at some wavelength) and in a particular direction (v , v ), within an instantaneous field-of-view
(IFOV) IFOV . The portions of canopy components with different surface temperatures in the IFOV will change with the view angle (Fig. 4.9). As a consequence,
strong anisotropy in exitance, i.e., a significant variation in Tb0 (v , v ) with the direction of observation, can be observed over thermally heterogeneous systems like
sparse canopies. For instance, Qualls and Yates (2001) observed in a cotton field that
the difference in Tb0 (v , v ) between the 0 (mixture of vegetation and soil) and the
80 (vegetation only) zenith view angles was 16.2 C at noon, while the difference
was only 0.9 C in the early morning. Lagouarde et al. (1995) observed a difference
of up to 3.5K for a corn canopy and 1.5K for grass (20 cm high) with a view zenith
angle between 0 and 60 around solar noon.
A goniometer (Fig. 4.10a) was designed specifically for canopy directional
Tb0 (v , v ) measurements (Zhang et al., 2002; Jia, 2004; Li et al., 2004). Two arms
are connected perpendicularly to each other with the longer one being fixed onto a
(a)
(b)
30
330
300
60
90
270
(c)
330
0 60
30
40
Target
300
60
20
270
90
120
240
210
180
150
Fig. 4.10 Goniometric system used to measure Tb0 (v , v ): (a) overview of the goniometer system
and of the footprint of the TIR radiometer; (b) zenithal and azimuth sampling scheme; (c) detail of
the azimuth sampling scheme. (Liu, 2002; Jia, 2004.)
63
circular track, which is set up on the ground, and the shorter one is kept horizontal
on top of which radiometers can be mounted. The longer arm can move along
the track to change the azimuth position. At each specific azimuth position v the
longer arm sways over a range of zenith angles v (maximum 60 ). Such movements are designed and performed to measure the TIR radiance of the same target
on the ground (Fig. 4.10b) within a desired range of azimuth and zenith angles.
The diameter of the footprint, however, increases with increasing v because of
slant viewing. A 10 interval is normally used for the zenith angle change, and 15
interval for azimuth angle change (Fig. 4.10c).
Two radiometers were used (Jia 2004) to measure Tb0 (v , v ). One radiometer
was set up on the top of the short horizontal arm to measure the radiance of the
canopy at each azimuth and zenith angle. Distance to the target was the same at
any position of the arm so that the radiometer footprint included the same target at
all positions as shown in Fig. 4.10b. The other radiometer was mounted on a mast
observing continuously the canopy at nadir. The second radiometer provided the
continuous measurements needed to correct for the temporal change in the measurements of Tb0 (v , v ) during a complete goniometer scan. The latter usually took
about 20 min during which the surface temperature may change significantly due
to the variation of the solar radiation and windspeed. Due to technical problems,
different radiometers were used during the experiment (Table 4.1).
A thermal camera (AGEMA THV 900 LW), mounted on top of another goniometer, was used to obtain images of surface temperature of the wheat crop Tb0 (v , v )
for prescribed azimuth and zenith directions. The AGEMA thermal camera has a
scanning HgCdTe detector and a Stirling cooler with the single channel covering
the spectral range between 812 m, the frame rate is 15 Hz for 136 272 pixels,
and the nominal sensitivity is 80 mK at 30 C. The camera was equipped with a lens
having a FOV of 5 10 .
Figure 4.11 shows the change in Tb0 from nadir to off-nadir view zenith angles
at each view azimuth angle at different hours during the two selected days: (a) 11
April and (b) 21 April. Only the measurements across two perpendicular planes are
shown: one in the NS direction along the canopy rows and one in the EW direction
across the row. At each azimuth direction, measurements were made twice, e.g., at
both 0 and 90 azimuth the first observation started from v = +60 through nadir
to v = 60 (denoted as go in Fig. 4.11) and the second observation went back
Table 4.1 The characteristics of the radiometers used to measure Tb0 (v , v ) during the field
campaign of QRSLSP (Jia 2004)
Instrument
Wavelength (m)
Radiometer 1
Radiometer 2
Radiometer 3
Radiometer (single channel)
Raytek radiometer
811, 10.414
811, 10.414
811, 10.414
814
814
FOV ( )
4.7
8.6
8.6
15
7
64
M. Menenti et al.
(a)
0-180 go
0-180 back
90-270 go
90-270 back
0
60
40
20
20
40
11 April
11 : 00
0
60
60
40
20
40
0
60
40
20
20
40
0
60
60
40
20
20
40
0
40
20
20
40
40
60
0-180 go
0-180 back
90-270 go
90-270 back
0
40
20
20
40
60
Zenith view angle ()
0-180 go
0-180 back
90-270 go
90-270 back
0
60
40
20
20
40
60

4
20
4
11 April
14:00
60
11 April
13:00
60
Tb0 (nadir) - Tb0 (off-nadir) (C)
0-180 go
0-180 back
90-270 go
90-270 back
20
60
60
40
0-180 go
0-180 back
90-270 go
90-270 back
11 April
13:30
4
11 April
12:30
60
0-180 go
0-180 back
90-270 go
90-270 back
60
0-180 go
0-180 back
90-270 go
90-270 back
11 April
12 : 00
Tb0 (nadir) Tb0 (offnadir) (C)
20
11 April
11:30
11 April
14:30
0-180 go
0-180 back
90-270 go
90-270 back
0
60
40
20
20
40
60
4
0-180 go
0-180 back
90-270 go
90-270 back
11 April
15:00
b0
T (nadir) - T (off-nadir) (C)
0-180 go
0-180 back
90-270 go
90-270 back
11 April
16:00
2
1
0
b0
0-180 go
0-180 back
90-270 go
90-270 back
Tb0 (nadir) Tb0 (offnadir) (C)
11 April
10:30
Tb0 (nadir) - Tb0 (offnadir) (C )
1
60
40
20
20
40
60
60
40
20
20
40
60
Fig. 4.11 Measurements of the directional variation of TOC brightness temperature difference Tb0
(nadir) Tb0 (off nadir) with zenith view angle: (a) on 11 April and (b) on 21 April at different
times of the day at the QRSLSP site (Liu et al. 2002). Tb0 (nadir) is Tb0 at v = 0 , Tb0 (off nadir)
is Tb0 measured at v = 0 . The positive zenith view angle correspond to the azimuth 0 and 90
planes, the negative zenith view angle correspond to the 180 and 270 azimuth planes. (Jia 2004.)
from view zenith v = 60 through nadir to v = +60 (denoted as back in

Fig. 4.9). All the measurements shown in Fig. 4.11 have been corrected for temporal
change taking into account the measurements provided by the second nadir looking
radiometer.
Changes in Tb0 with v are significant and show different trends at different hours
during a day. Around noon, the observed near-nadir Tb0 is always higher than that
at off-nadir positions. On the contrary, in the early morning and the late afternoon,
65
near-nadir Tb0 tends to be lower than the off-nadir values. Maximum difference
between near-nadir Tb0 and those at off-nadir is about 2.8 C on 11 April, while
4.4 C is observed on 21 April, both around noon time.
The row structure of the winter wheat also plays an important role in determining
the angular change of Tb0 . Such structure effects are evident when comparing the
shape of the curves in the along-row direction (the plane from 0 to 180 ) and the
curves in the across-row direction (the plane from 90 to 270 ) in Fig. 4.11. The latter shows a steeper slope, particularly in the position opposite to the sun, e.g., at
270 plane (negative zenith view angle) at 10:30, 11:00 and 11:30 on 11 April 2001
when the sun was located between 90 and 180 planes, and asymmetric than the
former.
For the proper interpretation of multi-angular measurements the geometry of observations needs to be taken into account, because of significant differences in footprint size, position and shape at different view angles. The change in the diameter of
the radiometer footprint when the radiometer observes the target at different zenith
view angles implies significant differences in the canopy elements captured by observations.
4.3.2 Airborne Observations

To our knowledge the only airborne imaging system having the capability to perform
multi-angular thermal infrared observations is the Airborne Multi-angle TIR/VNIR
Imaging System (AMTIS) developed by Wang et al. (2002). This is a prototype three-channel multi-angle imaging system. It provides high-resolution data
in visible/near infrared and thermal infrared spectral bands for use in deriving bidirectional reflectance factors. The precision of the viewing angle is determined by
the pointing precision of the AMTIS and the knowledge of the attitude of the airplane (pitch, roll and heading angles).
The AMTIS consists of one visible CCD camera, one near-infrared CCD camera, one thermal video system (TVS), camera bench, swing assembly, motor driver
and controller, exposing synchronization signal generator, data grabber, real-time
display and recorder. Two CCD cameras and the TVS are mounted on the camera
bench. A stepper motor rotates the camera bench.
The instantaneous field of view (IFOV) of each pixel has been fixed at 0.3 mrad
for the VNIR bands, 1 mrad for the TIR band. The total field of view of each camera
is about 20 . The altitude range is roughly from 500 m to 10,000 m which thus translates into a spatial resolution ranging from 0.15 m to 3 m for VNIR, 0.5 m to 10 m for
TIR. Normally, AMTIS is operated at an altitude of 3,000 m with a ground speed
of 250 km/h, and swing along-track through discrete nine programmable viewing
angles within a range of 45 . A step motor through a gearbox drives the cameras bench. The accuracy of the pitch angle is 0.3 . The maximum swing speed is
200 steps/s, or 60 /s. Normally, the swing angular velocity is about 10 per second
when AMTIS takes photograph, then swings back at the angular velocity of about
30 per second.
66
M. Menenti et al.
4.3.3 Space Observations of Tb ( ) at TOA and TOC

The measurements of thermal emission at top of atmosphere include the contributions of atmosphere (through atmospheric upwelling radiation and scattering of
surface thermal emission) and those of the surface.
To perform nearly simultaneous observations of Tb0 (v , v ) from space at multiple view angles two technical solutions may be used: (a) multiple optics or (b)
accurate along-track pointing. The Along Track Scanning Radiometer 1 (ATSR-1)
flown on ERS-1 was based on solution (a). ATSR-1 was followed by ATSR-2 in
1995 and the Advanced ATSR (AATSR) on ENVISAT in 2001. The characteristics
of these systems (Mutlow et al., 1994) are summarized below.
All ATSR sensors acquire dual-view angle data (approximately 0 and 53 at
surface) in four channels for ATSR-1 and seven channels for ATSR-2 and AATSR.
The nominal noise equivalent temperature difference (NET) of ATSR-2 for IRT
channels is 0.04K. The use of the along track scanning technique (Fig. 4.13) makes it
possible to observe the same point on the earths surface at two view angles through
two different atmospheric path within a short period of time. The first view is at a
view angle of 55 (approximately 53 at the earth surface) along the direction of the
orbit track when the satellite is flying toward the target point, which is referred to as
forward observation in this Chapter. Within 21/2 min the nadir (0 ) view observation
is made over the same point, which will be referred to as nadir observation later on.
Figure 4.12 shows the viewing geometry of ATSR-2 observations. The swath width
of ATSR-2 is 500 km, which provides 555 pixels across the nadir (0 zenith angle)
swath and 371 pixels across the forward (55 zenith angle) swath. The nominal pixel
size of ATSR-2 is 1 1 km at the centre of the nadir swath and 1.5 2 km at the
centre of the forward swath.
The Along-Track Scanning Radiometer (ATSR) instruments are imaging radiometers which are currently the only observing system able to provide from
Frequency (%)
0
6
10
12
Tb(0) - Tb(45)
Fig. 4.12 Difference of surface brightness temperature Tb between nadir view and forward view
(45 ); observations by AMTIS in Shunyi of Beijing on 11 April 2001.
67
ATSR
Instrument
Sub-satellite Track
55
Nadir view swath

(555 nadir pixels
1 km resolution)
Flight direction
Forward view swath
(371 along track pixels
1.5 km x 2 km resolution)
Fig. 4.13 Illustration of Along Track Scanning Radiometer (ATSR) observation. (After Mutlow
et al., 1999.)
space quasi-simultaneous bi-angular radiance measurements of the earths surface

in the TIR and SWIR spectrum regions (in addition to VIS/NIR channels). ATSR1 onboard the first European Remote Sensing satellite (ERS-1) was launched in
July 1991 and operated until June 1996. ATSR-1 had four channels one short
wave infrared (SWIR) channel located at 1.6 m and three TIR channels centred at
3.7 m, 11 m and 12 m. ATSR-1 was designed particularly for providing data over
the sea. ATSR-2 onboard the ERS-2 satellite was launched in April 1995 and is currently providing data both over land and over sea. In addition to one SWIR channel
and three TIR channels as on ATSR-1, ATSR-2 and AATSR have three narrowband visible-near infrared channels in the blue, green and red spectrum located
at 0.55 m, 0.67 m and 0.87 m respectively for vegetation monitoring. Table 4.2
gives information on channel spectral characteristics. (see the World Wide Web site
at www.atsr.rl.ac.uk/software.html for details).
The ATSR-2 data analysed below pertain to images acquired on June 6th 1999
over Spain at 10:30 a.m. local solar time. The major part of this image is over a cultivated zone, the rest being bare soils. This image is considered cloud-free, inasmuch
as ATSR-2 data do not indicate the presence of clouds.
The four images have been corrected for atmospheric effects: channel 1 (12 m)
nadir and forward views; channel 2 (11 m) nadir and forward views. Since ATSR
has a limited dynamical range of brightness temperatures, and radiometric saturation
is supposed to occur at 312 K, all pixels having Tb 312 K have been masked. Atmospheric corrections have been performed using the split-window method given
68
M. Menenti et al.
Table 4.2 Central wavelength and bandwidth of ATSR-2 spectral channels

Channel No.
Central wavelength
(m)
Full width at half maximum

(m)
12.0
11.0
3.7
1.6
0.87
0.65
0.55
11.6012.50
10.5211.33
3.473.90
1.5751.642
0.8530.875
0.6470.669
0.5430.565
1
2
3
4
5
6
7
TOC(nd)-TOA11(nd)
45
TOC(nd)-TOA12(nd)
40
TOC(fw)-TOA11(fw)
TOC(fw)-TOA12(fw)
Frequency (%)
35
30
25
20
15
10
5
0
2
10
12
14
Tb(TOC) - Tb(TOA)
Fig. 4.14 Atmospheric impact on surface brightness temperature for two TIR channels at both
nadir and forward views. Data are from ATSR-2 on June 6th 1999 over Spain. (nd = nadir;
fw = forward.)
by Eq. (4.7) (Li et al., 2003). The difference in brightness temperature at a given
wavelength and a given view angle between the TOC and TOA levels (Fig. 4.14)
indicates the magnitude and the dependence on view angle of atmospheric effects.
The temperature at TOC is always higher than the temperature at the TOA level and
this difference is larger at the 45 view angle because of the larger atmospheric optical depth. For a given view angle, the difference is larger for channel 1 (12 m) than
for channel 2 (11 m) due to the stronger atmospheric absorption in channel 1. The
surface anisotropy signature is observable after atmospheric correction (Fig. 4.15).
Most pixels show a lower temperature in the forward image than in the nadir image. The histogram peaks around 3.5 K at TOC level and around 5.0 K and 5.2 K at
TOA level for channels 2 and 1, respectively. This figure shows also that the angular
variation observed at TOA comes from both the angular variation at TOC and the
69
TOC(nd) - TOC(fw)
25
TOA11(nd) - TOA11(fw)
TOA12(nd) - TOA12(fw)
Frequency (%)
20
15
10
10
15
Tn - Tf
Fig. 4.15 Difference in surface brightness temperature between nadir and forward view at TOC
and TOA. Data are from ATSR-2 on June 6th 1999 over Spain
angular dependency of the atmospheric effects, and the effects of atmosphere generally increases the TOA anisotropy signature as compared to the angular variation
observed at TOC.
The experimental evidence reviewed in this section leads to conclude that the
thermal heterogeneity of soilvegetation systems is very significant (see, e.g., Figs.
4.4, 4.8 and 4.11). On the other hand, the same observations suggest that soil
foliage temperature differences are much larger than the difference between sunlit
and shadow elements, both foliage and soil. In other words, available data suggest
that a 2-components conceptual model of soilcanopy systems may be adequate in
most cases, with 4 components (sunlit, shadow; foliage, soil) necessary under extreme radiative forcing. This conclusion leads to the modeling approach presented
below.
4.4 Modeling the Anisotropic Exitance of Soil-Canopy Systems

Several models have been proposed and developed to describe and handle the
anistropic exitance of soilcanopy systems. They can be loosely classified in three
categories:
1. Simple geometric (deterministic) model of the system: this approach applies
to structured vegetation (row crops, tree lines, patches) inasmuch geometry is
known and the system can be described with a small number of known parameters (e.g., dimensions, soil emissivity, vegetation emissivity, ...). (Sutherland and
70
M. Menenti et al.
Bartholic, 1977; Kimes et al., 1981; Kimes, 1983; Sobrino and Caselles, 1990;
Caselles et al., 1992). Attempts to incorporate a coupling with the atmospheric radiation (down welling) have been scarce (Colton 1996). Such an approach is quite
useful for sensitivity studies or to assess the feasibility of the retrieval coupled to
the atmospheric correction requirements. Except when the geometry is accurately
known, or can be well inferred from other measurements (also from satellite), usefulness of this type of model is however limited since it cannot deal with the physical
processes within the system, and model inversion is very sensitive to uncertainties
in parameters.
2. Radiative transfer within the canopy: this approach applies to dense (or
less dense) systems that can be described statistically and using biome characteristics. Models in this domain solve radiative transfer in the canopy with
atmosphere and soil as boundary conditions, assuming plant type, distribution, plant architecture, LAI (total, horizontally/vertically projected), LIDF, etc.
(Ottermann, 1990; Ottermann et al., 1992, 1995, 1999; Balick et al., 1987; McGuire
et al., 1989; Kimes, 1981; Norman and Chen, 1990; Smith and Goltz, 1995; Smith
et al., 1996). Soil temperature, leaf temperature, temperature gradient within the
canopy may either be assumed, or simultaneously solved for.
Observed TIR anisotropy may reveal whether there exists a temperature gradient
within the canopy or not. However, interpretation of directional radiance implies
that all parameters of the system are known or can be accurately retrieved from
other measurements (from satellite in the visible, NIR and SWIR domains).
Since the fluxes within the canopy are coupled to the flux above the canopy,
the micrometeorological parameters have to be known. It turns out that the surface
TIR directional effect is quite sensitive to ambient conditions. Thus, for a given
biome, the TIR emitted radiance may reverse the sign of its angular variation with
zenith angle (i.e., decrease or increase with increasing zenith angle), or even show
no variation at all.
This category of models may not be best for inverse modeling of observations.
Nevertheless, such models are extremely useful for: (1) for evaluating the order of
magnitude of the angular effect that can be expected and (2) for comparing what is
observed with outputs of models. It is worth noting that the modeled anisotropy is in
no case large, no more than a few K only if large (>60 ), zenith angle can be used.
The Geometric-Optics method was initially proposed to model the radiative transfer
through conifer canopy in near infrared and visible domains (Li and Strahler 1985).
After taking into account the component emittances at the pixel scale, the geometricoptics method has been extended to the thermal infrared domain to model the
angular variation of thermal emission from vegetative canopies at the local scale
(Su et al. 2003). The core idea behind geometric optics method is that, a vegetation
canopy is assumed to be represented by different component elements with different
shapes (such as cone, elliptical, or sphere), the parameters for these shapes (such
as height, radius, and apex angle) are specified, the spatial distribution for these
canopies in pixel plane is previously determined (random, regular, or row), and
the position of solar and sensor is used as input. After accounting for the mutual
shading between canopies, the fractions of four components (sunlit foliage, sunlit
71
soil, shaded foliage and shaded soil) are computed through geometrical optical rules.
Then the contributions of each component to the reflected or emitted radiance are
combined to compute the reflected or emitted radiation at top of canopy level or
pixel level. Multi-scattering between different components is also included.
Some improvements on the original geometric optic method have been implemented to take into account the spatial heterogeneity including clumping index of
vegetative canopy which describe the non-random distribution of canopy elements
including leaves, branches, and the non-random spatial distribution of vegetative
canopies in pixel scale (Chen and Leblanc 1997).
A limitation of the geometric optic method is that the mutual shading, gap distribution inside vegetative canopy, multi-reflection, and scattering cannot be accounted
for. To overcome this shortcoming, a so-called hybrid geometric-optical radiative
transfer method was proposed to improve the original geometric-optical method. By
using this geometric-optical radiative transfer method, Yu et al. (2004) simulated the
directional brightness temperature over a maize canopy with a row structure. The
gap probability between rows was computed with geometric-optical rules, while
the gap inside rows was calculated with radiative transfer theory. Row and canopy
structure parameters including row width, canopy height and width between rows
are needed to initialize their model. The comparison between the simulations with
their model and in situ measurements showed that, the angular variation of brightness temperature can be precisely captured with this type of model.
Radiosity (Gerstl et al., 1991; Gerstl and Borel, 1992) is a computational algorithm to describe the scattering of light between ideally diffuse (Lambertian) surfaces. Although this method has been devoted to modeling bi-directional reflectance
distribution function (BRDF) in VIS/NIR optical remote sensing for a long time
(Borel et al., 1991), there are few reports on its application to thermal infrared.
3. Inhomogeneous thick vegetation layer: that can be statistically described by an
angle dependent gap fraction or gap frequency (Nilson, 1971). This approach
represents an intermediate situation between 1 and 2. It allows the directional TIR
radiance to be described as a linear combination of the foliage and soil radiance contributions, weighted by the gap fraction. Inverting directional TIR radiance may be
used to retrieve vegetation temperature and soil temperature, assuming, for instance,
an angle dependent canopy emissivity. A detailed and comprehensive discussion of
direct and inverse modeling is found in (Francois et al., 1997).
The advantage of this type of models is that the gap fraction can be phenomenologically correlated with measurements in the visible-NIR domain through appropriate vegetation indices (Baret et al., 1995). Hence, combination of measurements
in the solar reflected domain and in the thermal infrared domain may be of great
value, if not yet enough, to solve the problem (i.e., retrieve soil and foliage temperatures).
A further step can be made by exploiting detailed BRDF (including hot spot
as much as possible) measurements from satellites in the visible NIR domain to
assess, e.g., LIDF type, that would help improving the gap fraction estimate.
72
M. Menenti et al.
4.4.1 Detailed 3D Models of Radiative Transfer in Vegetation

Canopies
Realistic simulation of the canopy structure and remote sensing scenario has been
developed to model the radiative transfer of downwelling solar radiance and atmospheric radiance reflected by vegetative canopy and soil (Kimes and Kirchner,
1982) and of thermal emission from leaves and soil by Guillevic et al. (2002). The
so-called DART (Discrete Anisotropic Radiative Transfer) model has been developed (Guillevic et al., 2002; Gastellu-Etchegorry et al., 2004; CESBIO, 2006) to
simulate pixel scene and radiation energy budget in this scene. The whole scene
is specified by the locations of different objects (including tree, lake, soil, lake, and
building) and combined by a 3-D matrix of parallelepiped cells of various sizes. The
atmosphere part is simulated with air cells, while the terrestrial part is made
up of air cells. Cells are associated with 4 types of elements (soil, vegetation, wall,
water), with or without relief (DEM, Digital Elevation Model). More details of this
model can be found in Guillevic et al. (2002) and Gastellu-Etchegorry et al. (2004).
A realistic representation (Fig. 4.16) of canopy geometry and of the spatial heterogeneity of radiative and convective processes can be constructed using a discrete
3D grid (Welles and Norman, 1991). The grid points are the basic unit in the 3D
model dealing with the radiation, heat and water vapor fluxes calculations. The
components of radiative environment, such as the direct beam radiation from the
sun, the diffuse radiation from the sky and soil, and the diffuse radiation scattered
by foliage at any grid point are calculated. This description of geometry may also
be used to model convection of heat and water vapor. One should note that since we
have assumed wind speed, air temperature and vapor pressure are homogeneously
distributed along the horizontal directions, the grid points at the same vertical level
have identical values of these variables wherever they are located in the horizontal directions.
4.4.2 Simpler Models: Linear vs. Non-linear, 2 vs. 4 Components

Mixture Models
Typically, the pixel or IFOV (Instantaneous Field Of View) of an imaging radiometer
includes both foliage and soil where soil is commonly much warmer than foliage.
The experimental evidence reviewed in Section 4.3 suggests that the thermal heterogeneity of a soil-canopy system may be represented under most environmental
conditions using foliage and soil components only. Under extreme radiative forcing
the difference between sunlit and shaded canopy elements becomes larger. Simple
linear and non-linear models have been proposed to simulate radiance or brightness
temperature measured at TOC.
A 2-component mixture model can be used to simulate the directional distribution of thermal radiation as (Ottermann et al., 1992, 1995):
73
Reference height
in atmospheric surface layer
TOC
F
Canopy space
R, u , S, C
Soil surface
Root zone
Fig. 4.16 Simplified schematic illustration of interactions between points (either containing foliage
or soil) in a three-dimensional canopy (soil + vegetation) and in the atmosphere just above the
TOC, with all possible physical, chemical and biophysical processes included. TOC is represented
by green plane while soil surface is represented by orange plane. The green blocks represent
sub-canopies. Points are symbolized by symbol interaction between points are represented by
lines. The symbol
implies that the interactions are 3D (vertical and horizontal exchanges).
Each point is characterized by absorbed radiation flux density R, windspeed u, concentration C
of a scalar (i.e., temperature, moisture, CO2 , etc.), source/sink S of a scalar. F represents the flux
density of a scalar between points. (Jia, 2004.)
B (Tb0 ) = f v B (Tv ) + (1 f )g B (Tg )
(4.2)
where, Tb0 is the brightness temperature at TOC. v and g are emissivities of foliage and soil, respectively. f is the fractional coverage of vegetation. Tv and Tg
are temperatures of foliage and soil, respectively. B denotes the Planck function.
Previous study showed that the directional thermal radiation is not sensitive to the
uncertainty of soil and leaf emissivities (Francois et al., 1997). This model implies
that both the multiple reflection between vegetation and soil, and the temperature
differences between sunlit and shaded canopy elements are neglected.
It should be noted that Eq. (4.2) is a linear representation of directional thermal
radiation at top of canopy without taking into account the atmospheric downwelling
radiation and the temperature differences between sunlit and shaded elements inside
canopy. As indicated, under extreme radiative forcing, the sunlit elements are much
warmer than the shaded ones. The latter implies that accurate modeling of TOC
emittance requires taking into account leaf level radiation balance and 3D canopy
74
M. Menenti et al.
structure, either by explicit modeling of leaf level radiative processes or by parameterization. Thus, the sunlit and shaded elements need to be treated differently as,
4
B (Tb0 ) = fi i B (Ti )
(4.3)
i=1
Where: fi , i , and Ti are the fraction, emissivity, temperature of component element

i, respectively. Totally, four elements, namely sunlit foliage, shaded foliage, sunlit
soil and shaded soil are included. The equations for a four component system are
derived in a similar way as shown here for two components (Jia, 2004).
The Eqs. (4.2) and (4.3) do not account for the multiple interactions between
canopy elements and soil which is the so-called cavity effect (Sutherland and
Bartholic, 1977). To simulate such a cavity effect, a non-linear item is introduced as
(Li et al., 2001; Menenti et al., 2001):
B(Tb0 ( )) = P( )s B(Ts ) + [1 P( )] v B(Tv ) + (1 Ph )v B(Tv )P( )(1 s )
+ (1 v )v B(Tv ) [1 P( )] + (1 v )s B(Ts ) [1 P( )]
+(1 c )Ratm
(4.4)
where P( ) is the ground fractional cover viewed at angle P( ) = 1 f ( ), s and
Ts are the soil emissivity and temperature, v and Tv are leaf emissivity and temperature, Ph is the hemispheric gap frequency defined as the ratio of the radiation
travelling through the canopy and reaching the soil to the incident radiation into
the canopy over the hemisphere, and are respectively the probability of the
radiation emitted by a leaf and reflected by other leaves in the canopy and the probability of the radiation emitted by soil and reflected by the leaves above it, c is the
canopy emissivity and Ratm is the downward hemispheric atmospheric radiance divided by . The first term represents the proportion of the soil radiation that reaches
the top of the canopy in the direction . The second term is the upward emitted
radiation from the vegetation in the direction. The third term represents the downward radiation emitted by the vegetation and reflected by the soil and subsequently
traveling upwards through the vegetation in the view direction (vegetationsoil interaction). The fourth term is the contribution of the radiation emitted by the leaves
towards other leaves into the canopy and reflected by these leaves towards outside
the canopy in the view direction (vegetationvegetation interaction). The fifth term
is the contribution of the radiation emitted by soil towards the leaves and reflected by
these leaves towards outside the canopy in the view direction (soilvegetation interaction), The last term is the downward hemispheric atmospheric radiance reflected
by the canopy system. If we define the effective emissivity of soil and vegetation,
Es and Ev , as:
Es = s + (1 Ph )v (1 s )B(Tv )/B(Ts )
(4.5a)
Ev = v + (1 v )v + (1 v )s B(Ts )/B(Tv )
(4.5b)
Eq. (4.4) is reduced to Eq. (4.2) with effective emissivity instead of actual emissivity.
75
4.5 Modeling of Observations and Sensor Design

4.5.1 Modeling Approach
The relationship between canopy geometry, leaf and soil properties, radiative and
convective processes is rather complex (see previous Sections) leading to significant
thermal heterogeneity, while radiometric measurements capture the overall effect of
such heterogeneity only.
To understand the nature and information conveyed by radiometric measurements it is useful, therefore, to apply the modeling approach outlined in Section 4.4
to model observations and, particularly, their dependence on canopy (e.g., LAI) and
soil (e.g., water content) properties Jia et al., 2005). Simulation of radiometric data
or full images can be done considering realistic canopy and environmental conditions by combining a soilvegetationatmosphere transfer (SVAT) model and atmospheric radiative transfer (RT) model (Jia, 2004).
Some TOC biophysical variables needed as input to the SVAT model are simulated using the RT models PROSPECT and GeoSAIL (Verhoef and Bach, 2003).
The TOC Tb () was simulated using the SVAT model Cupid. Radiative transfer
in the atmosphere is simulated using the RT model MODTRAN 4.0 taking into account the sensor specifications, particularly spectral coverage, spectral sampling and
channel spectral response. The simulated images of directional radiance in two TIR
channels (11 and 12 m) at top-of-atmosphere (TOA) are obtained by adding the
atmospheric effects to the top-of-canopy radiance (or images).
The SVAT model Cupid is a one-dimensional, multi-layer model that simulates various plant-environment interactions (Norman, 1979; Norman and Campbell, 1983; Wilson et al., 2003). The essential processes simulated in the Cupid
model are divided as above-ground processes and below-ground processes stratified
by canopy layers (see Fig. 4.16).
Above-ground processes are dominated by vegetation including the transport of
energy, mass and momentum between plants and their environment. Above-ground
canopy is stratified by horizontal layers by identical increment of leaf area index
and each horizontal vegetation layer is stratified by leaf angle class. Interactions
between leaves in each individual leaf angle class in each horizontal layer and their
local environment are first formulated. Collective effect of all the leaves in each horizontal layer is integrated to obtain the response of the layer. Canopy-level responses
are simulated by numerical integration over all canopy layers where soil layers are
also included.
The below-ground processes describe the transport of heat and mass between the
roots and their soil environment and between the soil layers defined by identical
increment of depth. Unlike many other SVAT models that take soil surface as lower
boundary, Cupid has been developed to have plant root zone as the lower boundary.
The soil lower boundary conditions consist of soil temperature and soil water content near the bottom of the root zone. All the processes occurring at the soil surface
layer are simulated rather than input.
76
M. Menenti et al.
Leaf optical (reflectance, transmittance and emissivity) and physiological

properties (photosynthetic rate, respiration rate and stomatal conductance) and
leaf position and orientation arrangement determine energy and mass exchanges
at leaf level, while canopy structure (leaf density distribution, total leaf area, and
canopy height) determines the distribution of absorbed solar radiation in canopy
layers thereafter determines mass and heat transfer between layers.
The forcing of all the processes described in Cupid model is ambient atmospheric
conditions above canopy and soil boundary conditions at the bottom of the root
zone. Ambient atmospheric conditions are defined by solar radiation, air temperature, humidity, wind speed, and precipitation measured at some reference height
above the canopy. Equations describing heat and mass throughout the entire canopy
(leaf energy budget for all leaves and the vertical flux-gradient equations of soil
and vegetation layers) are solved simultaneously. Among others canopy layer temperature profile (including soil surface layer) and thermal radiation flux profile are
obtained from the solution of Cupid which can then be used to simulate canopy directional radiance in the concerned wavelength. This is done by assuming that the
contributions of various leaf layers and of the soil layer are appropriately weighted
by the fraction of each layer viewed by the radiometer in a particular view direction.
An overview of the simulation procedure is given in Fig. 4.17.
4.5.2 Generation of Synthetic Multi-angular Images

The at-sensor radiance is a combination of the surface-emitted radiance and the atmospheric contributions. The anisotropy of at-sensor radiance is due to atmospheric
scattering, absorption and emission in addition to the inherent anisotropy of TOC
radiance. The most significant interaction of TIR radiance with the atmosphere is
attributed to atmospheric absorption primarily due to ozone and water vapor in the
atmosphere. The optical path length is greater in the off-nadir views where the water vapor path increases, thus contributing higher upwelling atmospheric radiance,
while atmospheric transmittance is smaller.
To demonstrate the approach described above a simulated data has been generated (Fig. 4.17; Jia et al., 2005) for three European sites with a heterogeneous
and detailed spatial structure to represent major, rather different biomes, such as
evergreen forest, deciduous forest, savannah, semi-grasslands and agricultural land.
Result presented here relate to two sites briefly described below:
1. Alpine Foreland site. The land surface is classified by 22 land use classes with
grassland, mixed forest, shrubs, coniferous forest, deciduous forest, and various
crops as dominant land cover types. Four dates were selected for image simulation: 22/042009, 17/06/2009, 19/07/2008, 18/09/2008. The central coordinates
of these images are about 47.9 N and 11.1 E.
77
Broadband emissivity
of leaf and soil
Other inputs
Cupid model
Sensor geometry
Atmospheric
profiles
- canopy temperature profile

(including soil surface layer)
- thermal radiation flux profile
MODTRAN
Canopy RT
model
(,, ), Ratm(,, ), Ratm()

c( ,, )
Trad( ,, )
Tb0( ,, )TOC
Atmospheric RT
model
Sensor channel response

function
Tb(11, , ) TOA
Tb(12, , ) TOA
Spatial model
SPECTRA images of
Tb(11, q, j) TOA
Tb(12, q, j) TOA
Fig. 4.17 The major steps and applied models in the TIR image simulation. (Jia et al., 2005.)
2. Boreal forest site. Boreal forest site is located in Sodankyla of Finland with typical boreal forest covered mainly by deciduous forest (ca. 42%) and coniferous
forest (ca. 40%) and dark litter (ca. 12%). The soil in Sodankyla site is mainly
sandy type.
After obtaining the atmospheric path radiance and transmittance using MODTRAN
4.0 in combining the atmospheric profiles at the two sites, the TOA TIR images
were generated by applying these variables to the TOC TIR images. Figure 4.18
shows the images at the Sodankyla site as an example.
78
M. Menenti et al.
Fig. 4.18 Synthetic multi-angular images of TOC Tb ( ); Boreal forest at Sodankyla, Finland:
= 0 (left) and = +60 (right).
4.6 Retrieval Algorithms

4.6.1 Introduction
An operational algorithm was described by Jia et al. (2003) to retrieve soil and foliage component temperatures over heterogeneous land surface based on the analysis of bi-angular multi-spectral observations made by ATSR-2. This algorithm is a
good illustration of the synergistic use of multiangular observations in the VNIRSWIR region and in the TIR region to retrieve simultaneously both land surface
and atmospheric variables (see also Verhoef and Menenti, 1998). This algorithm is
described in this section. On the basis of the radiative transfer theory in a canopy,
a model is developed to infer the two component temperatures using six channels
of ATSR-2. Four visible, near-infrared and short wave infrared channels are used
to estimate the fractional vegetation cover within a pixel. A split-window method
is developed to eliminate the atmospheric effects on the two thermal channels. An
advanced method using all four visible, near-infrared and short wave channel measurements at two view angles is developed to perform atmospheric corrections in
those channels allowing simultaneous retrieval of aerosol optical depth and land
surface bi-directional reflectance (Fig. 4.19).
This general approach has been applied to image data acquired with the airborne
AMTIS (see Section 4.3). The algorithm was applied to retrieve foliage and soil
temperatures of a winter wheat canopy (Liu, 2002). since detailed ground measurements of component temperatures and of directional emittance were only available
for this land cover type. Validation could, therefore, be done only for retrievals over
winter wheat. All other land cover types were left out of consideration (light blue
area in Fig. 4.20).
79
(TOA) ATSR data :Tb (nadir), Tb (forward), (nadir),
Cloud
Retrieval of Aerosol Optical Depth
Retrieval of Water Vapor
Atmospheric correction
for VIS / NIR / SWIR channels
Fractional vegetation cover

ff (nadir), ff (forward)
Atmospheric correction
for TIR channels
(BOA) Surface brightness temperature

Tb0 (nadir), Tb0 (forward)
Spatial Smoothing
Inversion model
Component temperatures
Tf , Ts
Fig. 4.19 Scheme of the operational algorithm for retrieval of Tf and Ts from ATSR multi-spectral
and dual-angular measurements. Tb is the brightness surface temperature at TOA measured by TIR
channels of ATSR-2; is reflectance at TOA measured by the VIS/NIR/SWIR channels of ATSR2. (After Jia et al., 2003.)
4.6.2 Retrieval of Tb ( ) at TOC (Top of Canopy) from Tb ( )

at TOA (Top of Atmosphere)
Since the satellite measures the TOA brightness temperature (T ), and the inversion
of separate soil and vegetation temperature model needs the TOC brightness temperature (Tb0 ), atmospheric corrections have to be performed. Moreover, since directional ground radiance (equivalent to ground brightness temperature) is likely to
result from radiometric 3-D heterogeneity of the surface or the surface cover, neither
a kinetic surface temperature nor an emissivity can be simply and uniquely defined
80
M. Menenti et al.
Fig. 4.20 Foliage (left) and soil (right) component temperatures determined from AMTIS multiangular measurements of exitance at 4,200 m height; Shunyi experiment, China, April 2001. (After
Liu, 2002.)
(Becker and Li, 1995) or even a value or angular behaviour assumed. Any separation
between temperature and emissivity would rely on a priori assumptions or definitions of such variables. Thus, the first step in looking for directional effects is to
consider the ground radiance as a whole. The consequence of that is that accurate
atmospheric corrections have to be applied to TOA radiances (or brightness temperatures) in a given channel, preferably less affected by atmospheric perturbations,
or, to say things differently, a common Split-Window method (SW) is not adequate
(nor is a double angle method with (A)ATSR nadir/forward views). Indeed, a SW
or similar regression algorithm for atmospheric corrections is designed to give the
surface kinetic temperature, based on several assumptions regarding spectral and
angular emissivity.
Single channel atmospheric correction is just inverting the radiative transfer
equation integrated over the channel (i) bandwidth:
Bi (Ti ) = Bi (Tb0i )i + Ratm

i
(4.6)
in order to get the ground radiance Bi (Tb0i ), where Tb0i is the ground brightness
temperature, Ti is the TOA measured brightness temperature, i is the atmospheric
transmission and Ratm
is the atmospheric upward radiance. All quantities refer to
i
a particular view direction, defined by zenith angle at ground level. The accuracy
on Bi (Tb0i ()) is determined by that on the atmospheric quantities, which in turn
depends on the goodness of the radiative transfer code and the description of the
vertical structure of the atmosphere. The second condition is by far the most important source of error. There is no simple parameterisation allowing for determination
. Atmospheric PTU profiles are needed.
of both i and Ratm
i
Alternatively, if the ground brightness temperature is, or can be assumed,
independent of the channels used to measure it, the Split Window (SW )
(Sobrino et al., 1994; Becker and Li, 1995) method can be used to get Tb0 . Following
the procedure developed by Becker and Li (1990, 1995) a general SW algorithm
is derived for ATSR-2 nadir and forward views using simulated radiometric data
(Li et al., 2003):
Tb0 ( ) = [a( ) + b( )W ] + [c( ) + d( )W ]T11 ( )

+[e( ) + f ( )W ][T11 ( ) T12 ( )]
81
(4.7)
where is view angle, W is the total column water vapor in the atmosphere. For
the large range of surface parameters and atmospheric conditions (W 4.5 g/cm2 ,
air temperature at surface Ta , 272 K Ta 311 K and 5 K Tg Ta 15 K), the
coefficients a- f have been generated for ATSR-2 nadir and forward views. They
are a = 4.89, b = 3.74, c = 1.0205, d = 0.0151, e = 0.916, f = 0.509 the rms
residual retrieval error = 0.10K for the nadir image, and a = 14.41, b = 8.51, c =
1.0582, d = 0.0343, e = 0.565, f = 0.857 = 0.24K for the forward view.
4.6.3 Retrieval of Water Vapor from ATSR-2 Split-window Channel

Data over Land
Water vapor content in the atmosphere is required to improve the accuracy of the
remotely sensed surface parameters (Sobrino et al., 1994, Francois and Ottle, 1996).
Nowadays, a number of different satellite approaches have been proposed and developed over the past two decades to measure atmospheric water vapor. According
to the wavelength used, these approaches may be grouped into three categories:
Near-infrared techniques (Frouin et al., 1990, Kaufman and Gao, 1992); Passive
microwave techniques (Prabhakara et al., 1985, Alishouse et al., 1990, Schulz et al.,
1993); and Thermal infrared techniques (Chesters et al., 1983, Susskind et al., 1984,
Kleespies and McMillin, 1990, Jedlovec, 1990, Iwasaki, 1994, Ottle et al., 1997,
Sobrino et al., 1999).
Because the near-infrared technique is based on detecting the absorption by water
vapor of the reflected solar radiation as it is transferred down to the surface and up
through the atmosphere, use of this technique needs to have at least one channel in
the water absorption band (0.94 m), and one nearby channel in the atmospheric
windows (0.86 m, 1.05 m and 1.24 m). Since ATSR-2 (Along-Track Scanner
Radiometer) on board ERS-2 (European Remote Sensing) has only four channels in
the visible and near infrared domain (0.55 m, 0.65 m, 0.87 m, 1.60 m) and three
channels in the thermal infrared domain (3.7 m, 11 m and 12 m), no channel in
the water absorption band is available, the near-infrared technique cannot be applied
to ATSR2 data, the only one applicable technique is thermal infrared technique.
Up to now, there have been several attempts to derive water vapor using two splitwindow channels (11 m and 12 m). For instance, Kleespies and McMillin (1990)
proposed a method based on the ratio of split-window channel brightness temperature differences assuming that the atmosphere and surface emissivities in the
split-window channels are invariant. Jedlovec (1990) proposed an extension of this
concept and showed that the water vapor content can be derived using the ratio of
the spatial variance of the channel brightness temperature. On the basis of these
methods, Iwasaki (1994) developed a new algorithm to reduce the non-linear effect of air temperature and unresolved cloud effect on the estimation of water vapor
82
M. Menenti et al.
content using the split-window data. Sobrino et al. (1994) improved Jedlovec (1990)
method by the use of Split-Window Covariance-Variance Ratio (SWCVR). It has
been shown that all these split-window methods are sensitive to instrument noise
and are difficult to be applied to satellite data such as AVHRR in an operational
manner (Sobrino et al., 1994, 1999). Under the condition that the atmosphere and
directional surface emissivity are constant or the effects of their spatial variations
are not larger than the combined effects of both instrument noise over the N neighboring pixels, Li et al. (2003) presented a new algorithm to determine quantitatively
column water vapor content (W) directly from (A)ATSR SplitWindow radiance
measurements using the following formulae:
(a) For ATSR2 nadir view:
W = 13.73 13.662 j /i
(4.8)
where the subscripts i, j denote respectively channel 11 m and channel 12 m

of ATSR.
(b) For ATSR2 forward view (
= 53 ):
W = 10.02 9.971 j /i
with
(4.9)
j
i
= R ji with R ji =
i
j
(Ti,k T i )(T j,k T j )
k=1
(4.10)
(Ti,k T i )2
k=1
in which the subscript k denotes pixel k, Ti,k and T j,k are the brightness temperatures of pixel k in channels i and j measured at satellite level, respectively, T i
and T j are the mean (or the median) brightness temperatures of the N neighboring pixels considered, respectively.
This method was developed and applied to several ATSR2 data sets. The water
vapor contents retrieved using ATSR2 data from SGP97 (USA), Barrax (Spain) and
Cabauw (The Netherlands) are in good agreement with those measured by the quasisimultaneous radiosonde. The mean and the standard deviation of their difference
are respectively 0.04 and 0.22 g/cm2 . It is shown that water vapor content derived
from ATSR2 data using the proposed algorithm is accurate enough in most cases for
surface temperature determination with split-window technique using ATSR2 data
and for atmospheric corrections in visible and near-infrared channels of ATSR2.
A more detailed description of this method and its applicability can be found in Li
et al. (2003).
4.6.4 Retrieval of Aerosol Optical Depth from ATSR-2 Data

for Atmospheric Corrections
Atmospheric perturbations (mainly due to absorption and scattering processes)
are responsible for substantial modifications of the surface spectral reflectance
83
measured by satellite instruments. It is therefore necessary to correct the atmospheric effects to retrieve the surface reflectance. Methods of atmospheric
corrections are generally concerned with the estimation of the atmospheric effects
associated with molecular absorption, molecular and aerosol scattering. Current
methods for the estimation of the atmospheric effects employ a radiative transfer
model (Vermote et al., 1997, Beck et al., 1999) whose inputs are generally the
vertically integrated gaseous contents, aerosol optical properties and geometric
conditions.
If i is the reflectance measured in channel i at the top of atmosphere (TOA),
from radiative transfer theory, the surface reflectance in channel i, i , can be expressed as (Rahman and Dedieu, 1994; Vermote et al., 1997)
iac (s , , )
1 + Si iac (s , , )
i (s , , ) =
(4.11)
with
i (s , , )
iac =
tgi (s , )i (s , , )
a
ti (s )ti ( )
where s and are solar and viewing zenith angles, respectively. is the relative
azimuth between sun and satellite direction. Si is the spherical albedo of atmosphere
in channel i. tgi is the total gaseous transmission in channel i associated with gaseous
absorption along the suntargetsensor atmospheric path. ia (s , , ) is the atmospheric reflectance. ti (s ) and ti ( ) are the total atmospheric scattering transmittance along the sun-target and target-sensor atmospheric paths, respectively.
In general, the independent measurements of atmospheric composition and
aerosol optical properties are not available; it is therefore desirable to derive them
directly from satellite data. The most important gases in atmospheric corrections in
visible and near infrared channels are water vapor and ozone. Water vapor content
in the atmosphere may be derived from the two split-window channel measurements as shown above, and ozone content is taken from climatological data. As
for the determination of the aerosol optical properties, if the surface reflectance
may be considered isotropic, then the difference in surface reflectance retrieved
from multi-angle directions using Eq. (4.1) may be used to derive the atmospheric
optical thickness if aerosol type is assumed. However, most land surfaces are far
from Lambertian (Hapke, 1981) With multi-angle measurements, it is imperative to consider non-Lambertian reflectances. Several multi-look aerosol retrieval
schemes for ATSR-2 have been proposed (Flowerdew and Haigh, 1997; Mackay
and Steven, 1998; North et al., 1999). The iteration of a two step-process proposed
by North et al. (1999) can be used. The first step is to derive using Eq. (4.1) eight
land surface reflectances i (s , v , ) from the TOA reflectance made at four
channels (0.55 m, 0.65 m, 0.87 m, 1.60 m) and two view angles (nadir and forward views),
an initial estimate of the atmospheric aerosol and optical depth
agiven
. The second step is to fit land surface bi-directional reflectance
at 550 nm 550
84
M. Menenti et al.
model to eight retrieved surface reflectances by the minimization of the error metric
function
2
E=
[i (s , v , ) im (s , v , )]2
(4.12)
v =1 i=1
where im (s , v , ) is the land reflectance in channel i predicted by the reflectance

model.
Since there are maximum eight land surface reflectance measurements, land surface bi-directional model must have maximum seven free model parameters so
that there is at least 1 of freedom available
a for atmospheric parameter retrieval,
. Considering the land surface to be
for instance, the aerosol optical depth 550
composed of opaque facets, each with Lambertian reflectance i , and separating
parameters relating to the wavelength invariant three-dimensional structure of the
surface from wavelength dependent parameters describing the component spectra,
North et al. (1999) developed a seven free parameter (P nadir, P forward, and
i (i = 1, 4)) model as
i (s , v , ) = (1 Di )P(s , v , )i +
i
[Di + (1 )i (1 Di )]
1 (1 )i
(4.13)
where Di is the incident diffuse fraction which excludes the radiation scattered close
to the solar beam direction. Di can be estimated by radiative transfer model for solar
direction and aerosol optical depth. P(s , v , ) is the geometric parameter dependent only on view and illumination directions, denotes the mean hemispherically
integrated probability of escape of light without further interaction, after a scattering
event at the land surface.
In case where there are no four channels available, an alternative scheme can be
used to retrieve the aerosol optical depth by assuming that the functional shape of
the bidirectional effects is invariant with respect to the wavelength within the visible
and near-infrared region (Flowerdew and Haigh, 1997; Mackay and Steven, 1998),
namely:
i (s , 1 , ) j (s , 1 , )
=
(4.14)
i (s , 2 , ) j (s , 2 , )
This relationship gives a constraint for atmospheric correction by forcing the retrieved bidirectional reflectance to have a consistent angular variation, even thought
the magnitude of the reflectance may vary greatly.
The aerosol optical thickness is therefore obtained through the minimization of
the error metric function

n n
i (s , 1 , ) j (s , 1 , ) 2
(4.15)
E =
j (s , 2 , )
i=1 j>i i (s , 2 , )
where n is the total number of channels available, i and j are channel numbers. The
details of these methods can be found in North et al. (1999).
85
4.6.5 Retrieval of Vegetation Fractional Cover

To retrieve soil and foliage temperatures from Eqs. (4.2) or (4.4), the viewing angledependent vegetation fraction needs to be determined independently from additional
observations, such as the TOC reflectances in the VNIR and SWIR spectral channels
of ATSR-2 and AATSR. The approach presented here is an example, while for a
review on retrieval of fractional cover see, e.g., Weiss et al. (1999). The approach to
estimate the fractional vegetation cover uses visible and near infrared data.
A stepwise multiple linear regression were applied in this study to estimate the
fractional vegetation cover ff (v ) using TOC reflectances i (s , v , ). The stepwise multiple linear regression is written
n
ff (v ) = a0 (s , v ) + ai (s , v )i (s , v , )
(4.16)
i=1
where n is the number of channels used.

The model OSCAR (Verhoef, 1998) has been applied to generate surface reflectances for the viewing and illumination conditions applying to the ATSRobservations used in this study and to an ensemble of canopy and atmospheric conditions. This synthetic data base, where f f is known for any given canopy condition,
has been used to determine the coefficients ai in Eq. 4.16. A different set of coefficients is obtained for each viewing geometry (see also Jia, 2004).
4.7 Conclusions
Vegetation fractional cover. The regression model Eq. (4.16) works rather well when
vegetation is green, i.e., corn and alfalfa, while slightly larger errors appear for
vegetation with some fraction of brown leaves, like barley at this time of the year
(DOY = 179). Table 4.3 gives the summary of the model performance represented
by RMSE and Absolute Difference (AD) for each case and for the total data set.
Table 4.3 The performance of the regression equation for estimating fractional vegetation cover
Eq. (4.15) as presented by RMSE and AD at nadir and forward 53 for each vegetation type. Total
indicates the overall performance over the entire simulation dataset
Vegetation
Corn
Barely
Alfalfa
Total
v = 0
v = 53
RMSE
AD
RMSE
AD
0.0339
0.1086
0.0327
0.0456
0.025
0.109
0.030
0.046
0.1038
0.0745
0.0515
0.0785
0.103
0.073
0.049
0.076
86
M. Menenti et al.
Table 4.4 RMSE between retrieval and simulation of Tf and Ts for corn, barley and alfalfa crops
RMSE of Tf
RMSE of Ts
Corn
Barley
Alfalfa
0.381905
0.669429
1.164064
0.867213
0.590805
0.772624
Retrieval of soil and foliage component temperatures. The validation of the retrieved component temperatures using ATSR-2 bi-angular measurements is challenging due to the difficulty of obtaining observations of temperatures of soil and
foliage in situ at ATSR-2 spatial resolution (i.e., 1.5 2.5 km for the forward view.
As done to evaluate the algorithm for ff (v ), the same method, i.e., using synthetic radiometric data generated with detailed modeling of radiative transfer in the
soilvegetationatmosphere system and inversion using a simplified algorithm (suitable for operational processing of actual data), can be used to evaluate the retrievals
of Tf . The reference Tf and Ts were simulated using the complete model CUPID
and the TOC radiance was calculated (for each image pixel) with the procedure
outlined in Section 4.5. The Tb0 (v , v ) at v = 0 and v = 53 obtained in this
way were then treated as observations in Eq. 4.4 to retrieve Tf and Ts . The comparisons between the retrievals and the simulations of Tf and Ts for the three crops are
encouraging (Table 4.4).
4.8 Conclusions: Limitations of Current Systems

and Perspectives
Over the last 25 years measurements of emittance by goniometer-mounted radiometers has been the main source of observations to document and understand the
anisotropic emittance of vegetation canopies. This experimental body of knowledge led to the development of detailed models of the complex processes which
determine the significant thermal heterogeneity of the canopy space. The latter is of
particular relevance towards a better understanding of the relation between canopy
architecture and foliageatmosphere exchanges of radiation, water and carbon.
On the other hand, the capability of goniometer-mounted radiometers is severely
limited both as regards the size of targets and sampling of the spatial variability.
Actual exploitation of the unique information conveyed by anisotropic emittance requires imaging radiometers to sample sufficiently large terrestrial targets at a spatial
resolution sufficiently large to integrate emitted radiance over an area representative
of the target, but sufficiently small to provide independent observations of different
land cover types.
The ATSR series of instruments is the only current observing system providing
directional (at two view angles) observations of emittance and these observations
have been used to demonstrate the relevance of directional measurements in the
87
TIR region. (see, e.g., Jia et al., 2001; Menenti et al., 2001; Hurk et al., 2002). On
the other hand, the large footprint of (A)ATSR and even more the large difference in
footprint between the nadir and forward view, make validation of data products very
challenging and restrict applications to large scale modeling. For example, the difference between foliage and soil temperature might be used as a measure of drought
or crop water stress, but given the low resolution is difficult to relate to the spatial
scale of crop and water management practices.
The European Space Agency had been studying a high spatial resolution mission
for multi-angular land observations (SPECTRA, see Menenti et al., 2005), but this
project has been abandoned. The Italian (ASI) and Canadian (CSA) Space Agencies
are developing an Earth Observation mission with similar characteristics (Galeazzi
et al.,2006) with expected launch in 2010. The latter would provide simultaneous
VNIR SWIR hyper-spectral and multi-spectral observations at high spatial resolution and is being designed specifically for land observations.
In general terms, multi-angular measurements of TIR emittance over land provide access to radiative and convective processes in the canopy space which concur
to determine the response of terrestrial vegetation to environmental forcing.
Acknowledgements Work by the Authors summarized in this chapter has been funded by
the European Space Agency (Contracts 13177/98NL/GD; 15164/01/NL/SF; 17169/03/NL/GS;
17179/03/NL/GS), the European Commission (Contracts FP6 GMES no. 502057, FP5 MIND
EVK2-CT-2002-00158), the Netherlands Users Support Program (GO-2 Contract SRON EO049.) and the Italian Space Agency (ASI Contract ASI I/R/073/01). The Shunyi field campaign
was part of the Quantitative of Remote Sensing theory and application for Land Surface Parameters (QRSLSP) project funded by Chinas Special Funds for Major State Basic Research
(project No. G2000077900, led by Prof. X. Li). Dr. Qiang Liu is thanked for the data processing
of goniometer and AMTIS measurements.
References
Albertson J, Katul G, Wiberg P (2001) Relative importance of local and regional controls on coupled water, carbon, and energy fluxes. Adv. Water Resour. 24:11031118
Alishouse JC, Snyder SA, Vongsathorn J, Ferraro RR (1990) Determination of oceanic total precipitable water from the SSM/I. IEEE Trans. Geosci. Remote Sens. 28:811816
Avissar R, Pielke RA (1989) A parameterization of heterogeneous land-surface for atmospheric
numerical models and its impact on regional meteorology. Month. Weather Rev. 117:2113
2136
Balick LK, Hutchison BA, Smith JA, McGuire MJ (1987) Directional thermal exitance distributions of a deciduous forest in summer. IEEE Trans. Geosci. Remote Sens. 25:410412
Baret F, Clevers JPW, Stevens MD (1995) The robustness of canopy gap fraction estimates from
the red and near infrared reflectance: a comparison between approaches for sugar beet canopies.
Remote Sens. Environ. 54:141151
Beck A, Anderson GP, Acharya, PK, et al. (1999) MODTRAN4 Users Manual, Air Force Research Laboratory, Hanscom AFB, MA
Becker F, Li Z-L (1990) Towards a local split window method over land surfaces. Int. J. Remote
Sens. 11(3):369393
88
M. Menenti et al.
Becker F, Li Z-L (1995) Surface temperature and emissivity at various scales: definition, measurement and related problems. Remote Sens. Rev. 12:225253
Becker F (1987) The impact of spectral emissivity on the measurement of land surface temperature
from a satellite. Int. J. Remote Sens. 8:15091522
Becker F, Ramanantsizehena P, Stoll M-P (1985) Angular variation of the bidirectional reflectance
of bare soils in the thermal infrared band. Appl. Opt. 24(3):363375
Beljaars ACM, Holtslag AAM (1991) Flux parameterization over land surface for atmospheric
models. J. Appl. Meteorol. 30:327341
Blad BL, Rosenberg NJ (1976) Evaluation of resistance and mass transport evapotranspiration
models requiring canopy temperature data. Agron. J. 68:764769
Borel CC, Gerstl SA, Powers BJ (1991) The radiosity method in optical remote sensing of structured 3-D surfaces. Remote Sens. Environ. 36:1344
Bowen IS (1926) The ratio of heat losses by conduction and by evaporation from any water surface.
Phys. Rev. 27:779787
Caselles V, Sobrino JA, Coll C (1992) A physical model for interpreting the land surface temperature obtained by remote sensors over incomplete canopies. Remote Sens. Environ. 39:203211
CESBIO, 2006, DART User Manual v1.1.
Chen JM, Leblanc S (1997) A 4-scale bidirectional reflection model based on canopy architecture.
Chesters D, Uccellini LW, Robinson P (1983) Low-level water vapor fields from the VISSR atmospheric sounder (VAS) split-window channels. J. Appl. Meteorol. 22:725743
Choudhury BJ, Reginato RJ, Idso SB (1986) An analysis of infrared temperature observations over
wheat and calculation of the latent heat flux. Agric. Forest Meteorol. 37:7588
Colton AL (1996) Effective thermal parameters for a heterogeneous land surface. Remote Sens.
Environ. 57:143160
De Bruin HAR, Jacobs CMJ (1989) Forest and regional scale processes. Phil. Trans. R. Met Soc.
B (324):393406
Feddes RA (1971) Water, heat and crop growth. Ph.D. thesis, Wageningen, Agricultural University,
The Netherlands. H. Veenman, Zonen, Wageningen, The Netherlands, 184pp
Field CB, Mooney HA (1986) The photosynthesis nitrogen relationship in wild plants. In: TJ
Givnish (ed) On the economy of plant form and function. Cambridge University Press, New
York, pp 2555
Flowerdew RJ, Haigh JD (1997) Retrieving land surface reflectances using the ATSR-2: a theoretical study. J. Geophys. Res. 102(D14):1716317171
Francois C (2002) The potential of directional radiometric temperatures for monitoring soil and
leaf temperature and soil moisture status. Remote Sens. Environ. 80:122133
Francois C, Ottlè C (1994) Estimation of the angular variation of the sea surface emissivity with
the ATSR/ERS-1 data, Remote Sensing of Environment, 48:302308
Francois C, Ottle C (1996) Atmospheric corrections in the thermal infrared: global and water vapor dependent split-window algorithms-applications to ATSR and AVHRR data. IEEE Trans.
Francois C, Ottle C, Prevot L (1997) Analytical parameterization of canopy directional emissivity
and directional radiance in thermal infrared. Application on the retrieval of soil and foliage
temperatures using two directional measurements. Int. J. Remote Sens. 18(12):25872612
Frouin R, Deschamps P-Y, Lecomte P (1990) Determination from space of atmospheric total water
vapor amounts by differential absorption near 940 nm: theory and airborne verification. J. Appl.
Meteorol. 29:448460
Galeazzi C, Gamache M, Hollinger A, De Cosmo V, Menenti M, Staenz K, Beck J, Bergeron
M, Ananasso C, Loizzo R, Maszkiewicz M (2006) ASI-CSA joint development phase for a
hyperspectral mission. In: I. Reusen and J. Cools (eds.) Imaging Spectroscopy: Innovation in
Environmental Research. EARSEL, Paris:111
Gastellu-Etchegorry JP, Martin E, Gascon F (2004) DART: a 3D model for simulating satellite
images and studying surface radiation budget. Int. J. Remote Sens. 25(2):7396
89
Gerstl SAW, Borel CC (1992) Principles of the radiosity method versus radiative transfer for
canopy reflectance modelling. IEEE Trans. Geosci. Remote Sens. 30(2):271275
Gerstl SAW, Borel CC, Powers BJ (1991) The radiosity method in optical remote sensing of structured 3D surfaces. Remote Sens. Environ. 36:1344
Guillevic P, Gastellu-Etchegorry JP, Demarty J, Prevot L (2002)Thermal infrared radiative transfer within three-dimensional vegetation. J. Geophys. Res. (Atmos.) 108(D8):ACL 6-1, CiteID
4248, doi:10.1029/2002JD002247
Hall AE, Canell GH, Lawton HW (1979) Agriculture in semi-arid environments. Springer, Berlin
Heidelberg, New York, 340pp
Hapke B (1981) Bidirectional reflectance spectroscopy. 1. Theory. J. Geophys. Res. 86:30393054
Hatfield JL, Reginato RJ, Idso SB (1984) Evaluation of canopy temperature evapotranspiration
models over various crops. Agric. Meteorol. 32:4153
Hatfield JL (1991) Development of agricultural crops and practices adapted to arid land conditions,
1991. Proc. SCS Multiconference on AI simulation: Agricultural Crops and Practices Adapted
to Arid Lands. University of New Mexico Press, Albuquerque, pp 3960
Huband NDS, Monteith JL (1986) Radiative surface temperature and energy balance of a wheat
canopy. Bound.-Lay. Meteorol. 36:117
Hurk BJJM, Jia L, Jacobs C, Menenti M, Li Z-L (2002) Assimilation of land surface temperature
data from ATSR in an NWP environment a case study. Int. J. Remote Sens. 23(24):51935210
Iwasaki H (1994) Estimation of precipitable water over land using the split-window data from the
NOAA satellite. J. Meteorol. Soc. Jpn. 72(2):223233
Jedlovec GJ (1990) Precipitable water estimation from high-resolution split window radiance measurements. J. Appl. Meteorol. 29:863876
Jia L, Menenti M, Su ZB, Djepa V, Li Z-L, Wang J (2000) Modeling sensible heat flux using
estimates of soil and vegetation temperatures: the HEIFE and IMGRASS experiments. In:
M Beniston, M Verstraete (eds), Satellite remote sensing and climate simulations. Kluwer,
Dordrecht, The Netherlands
Jia L, Menenti M, Su ZB, Djepa V, Li Z-L, Wang J (2001) Modeling sensible heat flux using
estimates of soil and foliage temperatures: the HEIFE and IMGRASS experiments, in Remote
sensing and climate modeling: Synergies and Limitations, M Beniston and M Verstraete (eds)
Kluwer, Dordrecht, 2349
Jia L, Li Z-L, Menenti M, Su Z-B, Verhoef W, Wan Z-M (2003) A practical algorithm to infer
soil and foliage component temperatures from bi-angular ATSR-2 data. Int. J. Remote Sens.
24(23):47394760
Jia L (2004) Modeling heat exchanges at the land-atmosphere interface using multi-angular thermal
infrared measurements. Wageningen University, The Netherlands, ISBN 90-8504-041-8, 199pp
Jia L, Li Z-L, Menenti M (2005) Simulation of directional thermal infrared radiance images using
soil-vegetation-atmosphere heat and radiation transfer models. Proc. 9th Int. Symp. on Physical
Measurements and Signatures in Remote Sens. (ISPMSRS), ISSN 16821750 ISPRS, XXXVI
(7/W20):331333
Jackson RD, Idso SB (1975) Surface albedo and desertification. Science 189:10121013
Jackson RD, Idso SB, Reginato RJ, Pinter PJ (1981) Canopy temperature as a crop water stress
indicator. Water Resour. Res. 17(4):11331138
Jackson RD (1985) Evaluating evapotranspiration at local and regional scales. Proc. IEEE
73:10861095
Kaufman YJ, Gao B-C (1992) Remote sensing of water vapor in the near IR from EOS/MODIS.
Kimes DS (1980) Effects of vegetation canopy structure on remotely sensed canopy temperatures.
Kimes DS (1981) Remote sensing of temperature profiles in vegetation canopies using multiple
view angles and inversion techniques. IEEE Trans. Geosci. Remote Sens. 19(2):8590
Kimes DS (1983) Remote sensing of row crop structure and component temperatures using directional radiometric temperatures and inversion techniques. Remote Sens. Environ. 13:3355
90
M. Menenti et al.
Kimes DS, Smith JA, Link LE (1981) Thermal IR exitance model of a plant canopy. Appl. Opt.
20(4):623632
Kimes DS, Kirchner JA (1982) Radiative transfer model for heterogeneous 3-D scenes. Appl. Opt.
21(22):41194129
Kimes DS, Kirchner JA (1983) Directional radiometric measurements of row-crop temperatures.
Int. J. Remote Sens. 4(2):299311
Kleespies TJ, McMillin LM (1990) Retrieval of precipitable water from observations in the split
window over varying surface temperature. J. Appl. Meteorol. 29:851862
Labed J, Stoll MP (1991) Spatial variability of land surface emissivity in the thermal infrared band:
spectral signature and effective surface temperature. Remote Sens. Environ. 38:117
Lagouarde JP, Kerr YH, Brunet Y (1995) An experimental study of angular effects on surface
temperature for various plant canopies and bare soils. Agric. Forest Meteorol. 77:167190
Lei SA (2004) Leaf temperatures of Blackbrush: relationships with air and soil surface temperatures. USDA Forest Service Proc. RMRS-P-31
Li Z-L, Jia L, Su Z, Wan Z, Zhang R-H (2003) A new approach for retrieving precipitable water
vapour from ATSR2 split-window channel data over land area. Int. J. Remote Sens. 24:5095
5117
Li X-W, Strahler AH (1985) Geometric-optical modeling of a conifer forest canopy. IEEE Trans.
Geosci. Remote Sens. GE-23(5):705721
Li Z-L, Stoll MP, Zhang RH, Jia L, Su Z (2001) On the separate retrieval of soil and vegetation
temperatures from ATSR2 data. Science in China, Series D 44(2):97111
Li Z-L, Zhang R-H, Sun X-M, Su H-B, Tang X-Z, Zhu Z-L, Sobrino JA (2004) Experiment system for the study of the directional thermal emission of natural surfaces. Int. J. Remote Sens.
25(1):195204
Liu Q (2002) Study on Component Temperature Inversion Algorithm and the Scale Structure for
Remote Sensing Pixel. Ph.D. dissertation of Chinese Academy of Sciences, 111p
Liu QH, Li XW, Chen LF (2002). Field campaign for quantitative remote sensing. In: Beijing.
Proc. IEEE 2002 Int. Geosci. Remote Sens. Symp. (IGARSS02) vol VI, pp 31333135
Lhomme J-P, Monteny B, Chehbouni A, Troufleau D (1994) Determination of sensible heat flux
over Sahelian fallow savannah using infra-red thermometry. Agric. Forest Meteorol. 68:93105
Luquet D, Begue A, Vidal A, Clouvel A, Dauzat J, Olioso A, Gu X-F, Tao Y (2003) Using multidirectional thermography to characterize water status of cotton. Remote Sens. Environ. 84:411
421
Mackay G, Steven MD (1998) An atmospheric correction procedure for the ATSR-2 visible and
near infrared land surface data. Int. J. Remote Sens. 19(15):29492968
Mahfouf JF, Richard E, Mascart P (1987) The influence of soil and vegetation on the development
of mesoscale circulations. J. Clim. Appl. Meteor. 26:14831495
Masuda K, Takashima T, Takayama Y (1988) Emissivity of pure and sea waters for the model sea
surface in the infrared window regions. Remote Sens. Environ. 313329
Matteucci G, Dore S, Stivanello S, Rebmann C, Buchmann N (2000) Soil respiration in beech
and spruce forests in Europe: trends, controlling factors, annual budgets and implications for
the ecosystem carbon balance. In: ED Schulze (ed), Carbon and Nitrogen Cycling in European
Forest Ecosystems, vol 142. Springer, Berlin, pp 217236
McGuire MJ, Smith JA, Balick LK, Hutchison BA (1989) Modeling directional thermal radiance
from a forest canopy. Remote Sens. Environ. 27:169186
Medlyn BE, Dreyer E, Ellsworth DS, Forstreuter M, Harley PC, Kirshbaum MUF, Le Roux X,
Montpied P, Strassemeyer J, Walcroft A, Wang K, Loustau D (2002) Temperature response of
parameters of a biochemically-based model of photosynthesis. II. A review of experimental
data. Plant, Cell Environ. 25:11671179
Menenti M, Choudhury BJ (1993) Parameterization of land surface evapotranspiration using a
location-dependent potential evapotranspiration and surface temperature range. In: HJ Bolle, et
al. (eds), Exchange processes at the land surface for a range of space and time scales, IAHS,
Oxfordshire, vol 212, pp 561568
91
Menenti M, Jia L, Li Z-L, Djepa V, Wang J, Stoll MP, Su ZB, Rast M (2001) Estimation of soil and
vegetation temperatures from directional thermal infrared observations: the HEIHE, SGP97,
IMGRASS experiments. J. Geophys. Res. 106:1199712010
Menenti M, Rast M, Baret F, Hurk B, Knorr W, Mauser W, Miller J, Moreno J, Schaepman M,
Verstraete M (2005) Understanding vegetation response to climate variability from space: recent advances towards the SPECTRA Mission. Riv. Ital. Teleril. Special Issue Hyperspectral
Observation of Terrestrial Environments. 33/34:193206
Monteith JL (1965) Evaporation and environment. In: GE Fogg (ed), The state and movement of
water in living organisms, Sympos. Soc. Exp. Biol. Academic, New York, 19:205234
Moran MS, Clarke TR, Inoue Y, Vidal A (1994) Estimating crop water deficit using the relation between surface air temperature and spectral vegetation index. Remote Sens. Environ. 49(3):246
263
Mutlow CT, Zavody AM, Barton IJ, Llewellyn-Jones DT (1994) Sea surface temperature measurements by the along-track scanning radiometer on the ERS 1 satellite: early results. J. Geophys.
Res. 99(C11):2257522588
Mutlow C, Murray J, Bailey P, Birks A, Smith D (1999) ATSR-1/2 User Guide, http://www.atsr.rl.
ac.uk/documentation/docs/userguide/atsr user guide rev 3.pdf
Nerry F, Labed J, Stoll M-P (1988) Emissivity signatures in the thermal IR band for remote sensing:
calibration procedure and method of measurement. Appl. Opt. 27(4):758764
Nielsen DC, Clawson KL, Blad BL (1984) Effect of solar azimuth and infrared thermometer view
direction on measured soybean canopy temperature. Agron. J. 76:607610
Nilson T (1971) A theoretical analysis of the frequency of gaps in plant stands. Agric. Meteorol.
8:2538
Norman JM (1979) Modeling the complete crop canopy. In: BJ Barfield, JF Gerber (eds) Modification of the Aerial Environment of Plants. ASAE Monogr. Am. Soc. Agric. Engr., St. Joseph,
MI, pp 249277
Norman JM, Campbell G (1983) Application of a plant-environment model to problems in irrigation. In: D Hillel (ed) Advanced in irrigation, vol 2. Academic, New York, pp 155188
Norman JM, Chen JL (1990) Thermal emissivity and infrared temperatures dependence on plant
canopy architecture and view angle. In: Proc. Int. Geosci. Remote Sens. Symp. (IGARSS90,
IEEE Geosci. Remote Sens. Soc., New York), pp 17471750
North PRJ, Briggs SA, Plummer SE, Settle JJ (1999) Retrieval of land surface bi-directional reflectance and aerosol opacity from ATSR-2 multiangle imagery. IEEE Trans. Geosci. Remote
Sens. 37(1):526537
Otterman J (1990) Inferring parameters for canopies non uniform in azimuth by model inversion.
Otterman J, Brakke TW, Susskind J (1992) A model for inferring canopy and underlying soil
temperatures from multi-directional measurements. Bound.-Lay. Meteorol. 61:8197
Otterman J, Susskind J, Brakke T, Kimes D, Pielke R, Lee TJ (1995) Inferring the thermal-infrared
hemispheric emission from a sparsely-vegetation surface by directional measurement. Bound.Lay. Meteorol. 74:163180
Ottermann J, Brakke TW, Fuchs M, Lakshmi V, Cadeddu M (1999) Longwave emission from a
plant/soil surface as a function of the view direction: dependence on the canopy architecture.
Int. J. Remote Sens. 20(11):21952201
Ottle C, Outalha S, Francois C, Le Maguer S (1997) Estimation of total atmospheric water vapor
content from split-window radiance measurements. Remote Sens. Environ. 61:410418
Paw U KT, Ustin SL, Zhang CA (1989) Anisotropy of thermal infrared exitance in sunflower
canopies. Agric. Forest Meteorol. 48:4558
Peng XD, Cheng LS (1993) Numerical simulations of the PBL mean structure and fluxed in HEIHE
region. Proc. Int. Symp. HEIFE. Kyoto University, Kyoto, Japan, pp 437442
Petitcolin F, Nerry F, Stoll MP (2002) Mapping temperature independent spectral indice of emissivity and directional emissivity in AVHRR channels 4 and 5. Int. J. Remote Sens. 23(17):3473
3491
92
M. Menenti et al.
Pielke RA, Dalu G, Snook JS, et al. (1991) Nonlinear influence of mesoscale land use on weather
and climate. J. Climate 4:10531069
Prabhakara C, Chang HD, Chang ATC (1985) Remote sensing of precipitable water over the oceans
from NIMBUS 7 microwave measurements. J. Appl. Meteorol. 21:5968
Price JC (1982) On the use of satellite data to infer surface fluxes at meteorological scales. J. Appl.
Meteorol. 21:11111122
Qualls RJ, Yates DN (2001) Directional radiometric temperature profiles within a grass canopy.
Adv. Water Resour. 24:145155
Rahman H, Dedieu G (1994) SMAC: a simplified method for the atmospheric correction of satellite
measurements in the solar spectrum. Int. J. Remote Sens. 15:123143
Schulz J, Schluessel P, Grassl H (1993) Water vapor in the atmospheric boundary layer over oceans
from SSM/I measurements. Int. J. Remote Sens. 14:27732789
Segal R, Avissar M, McCumber C, Pielke RA (1988) Evaluations of vegetation effects on the
generation and modification of mesoscale circulations. J. Atmos. Sci. 45:22682292
Seguin B, Assad E, Freteaud JP, Imbernon J, Kerr YH, Lagouarde JP (1989) Use of meteorological
satellite for water balance monitoring in Sahelian regions. Int. J. Remote Sens. 10:10011017
Sellers PJ, Heiser MD, Hall FG, Goetz SJ, Strebel DE, Verma SB, Desjardins RL, Schuepp PM,
McPherson JI (1995) Effects of spatial variability in topography, vegetation cover and soil
moisture on area-averaged surface fluxes: A case study using the FIFE 1989 data. J. Geophys.
Res. Vol. 100 (D12): 2560725630
Sellers PJ, Randall DA, Collatz GJ, Berry JA, Field CB, Dazlich DA, Zhang C, Colleli GD,
Bounoua L (1996) A revised land surface parameterization (SIB2) for atmospheric GCMS.
Part 1: model formulation. J. Climate 9(4):676705
Smith JA, Goltz SM (1995) A thermal exitance and energy balance model for forest canopies.
Smith JA, Chauchan NS, Ballard J (1996) Remote sensing of land surface temperature: the directional viewing effect. IEEE Trans. Geosci. Remote Sens. 13(4):21462148
Snyder WC, Wan Z-M, Zhang Y, Feng Y-Z (1997) Thermal Infrared (314 m) bidirectional
reflectance measurements of sands and soils. Remote Sens. Environ. 60(1):101109
Sobrino JA, Caselles V (1990) Thermal infrared model for interpreting the directional radiometric
temperature of a vegetative surface. Remote Sens. Environ. 33:193199
Sobrino J, Li Z-L, Stoll MP, and Becker F, 1994. Improvements in the split-window technique
for land surface temperature determination. IEEE Trans. Geosc. and Rem. Sens. Vol. 32(2):
243253
Sobrino J, Li Z-L, Stoll MP, Becker F (1994) Improvements in the split-window technique for land
surface temperature determination. IEEE Trans. Geosci. Remote Sens. 32(2):243253
Sobrino JA, Raissouni N, Simarro J, Nerry F, Petitcolin F (1999) Atmospheric water vapor content
over land surfaces derived from AVHRR data: application to the Iberian Peninsula. IEEE Trans.
Sobrino JC, Cuenca J (1999) Angular variation of thermal infrared emissivity for some natural
surfaces from experimental measurements. Appl. Opt. 38(18):39313936
Su H-B, Zhang R-H, Tang X-Z, Sun X-M, Zhu Z-L, Li Z-L (2003) An alternative method to
compute the component fractions in the geometrical optical model: visual computing method.
Susskind J, Rosenfield J, Renter D, et al. (1984) Remote sensing of weather and climate parameters
from HIRS2/MSU on TIROS-N. J. Geophys. Res. 89:46774697
Sutherland RA, Bartholic JF (1977) Significance of vegetation in interpreting thermal radiation
from a terrestrial surface. J. Appl. Meteorol. 18:759763
Valentini R, Dolman AJ, Ciais P, Schulze ED, Freibauer A, Schimel DS, Heimann M (2000) Accounting for carbon sinks in the biosphere, European perspective. CARBOEUROPE European
Office. Max-Planck Institute for Biogeochemistry, P.O. Box 100164. 07700 Jena, Germany, 17p
Verhoef W, Menenti M (1998) Final Report of Spatial and Spectral Scales of Spaceborne Imaging
Spectro-radiometers (SASSSIS). Contract Report 12072/96/NL/CN, ESA. Final Report, NLRCR-98213, National Aerospace Lab. NLR, Amsterdam, The Netherlands, 314pp
93
Verhoef W (1998) Theory of radiative transfer models applied in optical remote sensing of vegetation canopies. Ph.D. thesis, Wageningen University, The Netherlands, 311pp
Verhoef W, Bach H (2003) Simulation of hyperspectral and directional radiance images using coupled biophysical and atmospheric radiative transfer models. Remote Sens. Environ. 87:2341
Verma SB, Rosenberg NJ, Blad BL, Baradas NW (1976) Resistance-energy balance method for
predicting evapotranspiration, determination of boundary layer resistance, and evaluation of
error effects. Agron. J. 68:776782
Vermote EF, Tanre D, Deuze JL, Herman M, Morcrette JJ (1997) Second simulation of the satellite
signal in the solar spectrum, 6S: an overview. IEEE Trans. Geosci. Remote Sens. 35:675686
Wang J, Mitsuta Y (1992) An observation study of turbulent study of turbulent structure and transfer characteristics in Heihe oasis. J. Meteor. Soc. Jpn. 70:11471154
Wang J, Yan F, Xiao J, Wei T, Wang J (2002) Development of an airborne multi-angle TIR/VNIR
imaging system. IEEE 2002 Int. Geosci. Remote Sens. Symp. (IGARSS02), vol IV, pp 2245
2248
Weiss M, Baret F, Myneni R, Pragnère A, Knyazikhin Y (1999) Investigation of a model inversion
technique for the estimation of crop characteristics from spectral and directional reflectance
data. Agronomie 20:322
Welles JM, Norman JM (1991) Photon transport in discontinuous canopies: a weighted random
approach. In: RB Myneni, J Ross (eds), Photon-vegetation interactions. applications in optical
remote sensing and plant ecology. Springer, Heidelberg, Germany, pp 389414
Wilson K, Goldstein A, Falge E, Aubinet M, Baldocchi D, Berbigier P, Bernhofer C, Ceulemans R,
Dolman H, Field C, Grelle A, Ibrom A, Law BE, Kowalski A, Meyers T, Moncrieff J,
Monson R, Oechel W, Tenhunen J, Valentini R,Verma S (2002) Energy balance closure at
FLUXNET sites. Agric. Forest Meteorol. 113(14):223243
Wilson TB, Norman JM, Bland WL, Kucharik CJ (2003) Evaluation of the importance of Lagrangian canopy turbulence formulations in a soilplantatmosphere model. Agric. Forest Meteorol. 115:5169
Xu X, Chen L, Zhuang J (2000) The passive measurements of objects directional emissivity in
laboratory. Science in China 43:5561
Xue Y, Shukla J (1993) The influence of land surface properties on Sahel climate. Part I: desertification. J. Climate 6:22322245
Yan Y (1999) Numerical simulation of land surface process on heterogeneous surface. Ph.D. dissertation. Lanzhou Institute of Plateau Atmospheric Physics. Chinese Academy of Sciences,
Lanzhou, China, 90pp
Yu T, Gu X-F, Tian G-L, Legrand M, Baret F, Hanocq J-F, Bosseno R, Zhang Y (2004) Modeling
directional brightness temperature over a maize canopy in row structure. IEEE Trans. Geosci.
Remote Sens. 42(10):22902304
Zhang R-H, Sun X-M, Su H-B, Zhu Z-L, Tang X-Z (2002) A new design for measuring directional radiant temperature and data analysis. Proc. IEEE 2002 Int. Geosci. Remote Sens. Symp.
(IGARSS02), vol V, pp 27682770
Chapter 5
Terrestrial Applications of Multiangle Remote

Sensing
Mark J. Chopping
Abstract Multiangle remote sensing is an emerging technology that enables

important applications of terrestrial (land) remote sensing, in ecology and land
cover mapping as well as in a variety of disciplines in the Earth Sciences. Advances
have been realized in three major areas: the measurement or characterization of
canopy structure and surface roughness; the separation of the contributions of the
upper canopy and the background in forest and shrub-dominated environments; and
improvements in the accuracy of classifications of land cover in environments as
dissimilar as deserts and ice sheets. The focus of the chapter is on land surface applications of solar wavelength multiangular data acquired from the air and space; it
avoids discussion of methods for retrieval of essentially radiometric quantities such
as shortwave albedo, bidirectional and hemispheric reflectance factors (Chapter 9),
retrieval of land surface temperature (Chapter 4), theoretical radiative transfer
modeling studies, and those based uniquely on field measurements. The chapter
introduces existing instruments providing multiangle data; a review of work performed under the broad headings Empirical and Synergistic Approaches, Radiative
Transfer, The RAMI Exercise, Canopy Openness, Clumping Index, Structural Scattering Index, Geometric-Optical and Hybrid Models, Direction and Wavelength,
The Background in Canopy Reflectance Modeling, Land Cover and Community
Type Mapping, Snow and Ice, and Dust Emissions; and finally, brief discussions
under the headings Near-Simultaneous and Accumulated Sampling, Angular Sampling, and Scale and Multiangle Observation. The emphasis is on the added value
that existing solar wavelength multiangular data can provide to applications. There
has been a wide array of approaches, all of which have resulted in studies demonstrating progress and many of which show the advantages possible over the use
of purely nadir-spectral techniques, particularly for accessing measures of canopy
structure.
Mark J. Chopping
Department of Earth and Environmental Studies, Montclair State University, USA
chopping@pegasus.montclair.edu
95
96
M.J. Chopping
5.1 Introduction
The exploitation of data obtained by observing the Earths surface at different
viewing and illumination angles in the solar wavelengths is a rapidly growing field.
An excellent summary of work in this field is given in Liang et al. (2000) with a
recent update by Diner et al. (2005). Multiangle remote sensing seeks to exploit
the reflectance anisotropy of the planets terrestrial surface that is described by the
bidirectional reflectance distribution function (BRDF) that quantifies the angular
distribution of spectral radiance that is scattered by a surface (Nicodemus et al.,
1977). The BRDF is thus a fundamental description of the terrestrial surface from
the photons point of view. A key concept is that multiangle measurement provides
a unique way to infer surface information within the observed element (pixel) and
is thus useful for mapping land surface parameters over large areas. Remote sensing
instruments acquiring imagery at varying angles explicitly are designed to collect
multiple looks over as short a timeframe as possible and in the case of those flown
on satellites taking into account the constraints imposed by the orbits of their
platforms. Multiangle data sets may also be constructed by accumulation of observations over periods from days to weeks, following the longstanding practice of
filtering images from many swaths to construct quasi-cloud-free composites. A limiting factor of this method is the temporal displacement of the measurements, since
either the atmosphere or the surface, or both, may vary over short timescales (hours
to days). The BRDF gives the reflectance of a target as a function of illumination
geometry and viewing geometry; it depends on wavelength and is determined by
the structural and optical properties of the surface, such as shadow-casting, multiple
scattering, mutual shadowing, transmission, reflection, and absorption by surface elements, facet orientation distribution and facet (e.g., leaf, soil particle) density. The
BRDF of a surface (unit: sr1 ) is defined for all illumination and viewing directions
over the upper hemisphere and is not itself measurable as the light incident on the
surface is partly diffuse and the measurements involve finite solid angles (Liang,
2004; Bacour and Breon, 2005). Researchers have sometimes ignored the correct nomenclature and taken multiangle observations to provide a (usually sparse)
sampling that is approximately bidirectional, accepting the assumption that the
size/distance ratio of the illuminating source and the observing sensor is zero. However there are now efforts under way to ensure that the terminology in use is more
accurate and consistent (Martonchik et al., 2000; Schaepman-Strub et al., 2006).
The radiometric data products available from multiangle instruments include
bidirectional reflectance factor (BRF; BRDF normalized by the equivalent reflectance of a Lambertian, or diffuse scattering surface) and/or hemisphericaldirectional reflectance factor (HDRF; the single integral of BRDF on the incoming
directions; i.e., including direct and diffuse illumination), directional-hemispherical
reflectance (black-sky albedo; single integral of BRDF on the outgoing directions)
and bihemispherical reflectance (white-sky albedo; the double integral of BRDF
over all viewing and illumination directions). These quantities must be derived via
modeling since infinitesimal elements of solid angle do not include measurable
amounts of radiant flux, and small light sources and unlimited sensor fields-of-view
5 Terrestrial Applications of Multiangle Remote Sensing
97
do not exist; all measurable quantities of reflectance are therefore performed in the
conical or hemispherical domain (Schaepman-Strub et al., 2006). The most accurate
description of the spectro-directional radiance measurements made by remote sensing instruments is therefore hemispherical-conical reflectance factor, even though
solar illumination is usually highly directional, especially under clear skies.
The most important physical phenomena underlying the BRDF are surface scattering (geometric effects, shadowing and shadow-hiding; Camacho-de Coca et al.,
2004); and volume scattering by facets (leaves, soil particles). It is an intrinsic surface property. Multiangle remote sensing seeks to exploit fundamental aspects of the
remote sensing problem and leads to important improvements in the accuracy with
which vegetation and other surface parameters can be retrieved. The information returned is unique in that the data are sensitive to both the optical (spectral reflectance,
transmittance, and absorption) and structural properties of surfaces (e.g., canopy architecture: horizontal and vertical heterogeneity), that is, from both the angular and
spectral domains (Diner et al., 1999, 2005). Since most multiangle work pursues
the one-dimensional (1-D) approach that seeks to derive sub-pixel information via
variation in radiance with solar and/or viewing angles and in which each pixel is
treated independently, the range of scales over which multiangle sensing can play a
role is large (the term pixel is used here since in remote sensing it is commonly understood to correspond to the area under the sensor ground-projected field-of-view,
even though this usage is incorrect: pixel is a contraction of picture element).
Much work has gone into developing algorithms for retrieving well-understood
physical parameters with straightforward interpretations via 1-D model inversions
(Chopping et al., 2003; Widlowski et al., 2004; Koetz et al., 2005a), although researchers must grapple with the problem of non-uniqueness of solutions (different
combinations of model parameters or state variables can result in similar patterns
of observations), which can make this difficult (Weiss et al., 2000; Gobron et al.,
2000; Combal et al., 2002). Nevertheless, using simulations of top-of-atmosphere
radiance observations in 201 spectral bands and seven look angles, Verhoef (2005)
demonstrated that:
Multi-angular hyperspectral measurements could in some cases provide dramatically improved retrievals of canopy structure, leaf properties, soil moisture, and
radiative variables (e.g., fraction of photosynthetically-active radiation, albedo)
than monodirectional measurements.
The information content of the spectral and angular domains is highly complementary, i.e., the additional information from multiangle remote sensing is not
redundant.
For remotely estimating canopy architecture variables, multiangle information is
more important than hyperspectral information (Fig. 5.1).
Although the results above were obtained in the absence of real-world constraints,
it can be asserted from first principles that canopy leaf foliar chemistry and ecosystem function are better assessed using hyperspectral methods, while canopy physical structure (in the sense of physiognomy, or architecture) is better assessed using
multiangle methods (Schaepman, 2006). These two approaches are not mutually
Retrieval accuracy (dB)
98
M.J. Chopping
30
20
Hyperspectral Monodir.
Hyperspectral Multidir.
Multispectral Multidir.
Multispectral Monodir.
10
s
en
fd e
ar
fb
ta
ze
%
Cv
ss
Di
fB
t
ho Fb
D
LI a
DF
LI
I
LA
b
N_
_b
Cs _b
m
Cd _b
Cw _b
b
Ca
g
N_
_g
Cs _g
m
Cd _g
Cw _g
b
Ca
%
SM
do
be
Al R
PA
fA ver
o
fC
Geo-biophysical variable
Fig. 5.1 Relative retrieval accuracies expressed in dB above the a priori uncertainty levels, for low
sensor noise (0.1 W/m2 /m/sr) and four mission types. The parameters estimated were fractional
cover (fCover); fraction of absorbed PAR (fAPAR); albedo; soil moisture (SM%); green leaf chlorophyll (Cab g), water (Cw g), dry matter (Cdm g), senescence pigments (Cs g) and the N parameter
representing leaf mesophyll structure (CN g); brown leaf parameters (the same but identified with
the suffix b), canopy leaf area index (LAI), leaf inclination distribution function (LIDF in 2 parameters); a hot spot parameter (hot), the fraction of brown leaves (fB), dissociation factor (Diss),
crown cover fraction (Cv%), and tree shape factor (zeta); and the fractions of dense vegetation
and bare soil in the immediate environment of the observation (fDENS and fBAR), accounting for
adjacency effects. (Courtesy of Wout Verhoef).
exclusive but to date no studies have demonstrated the joint benefits of a spectrodirectional system with regular and broad geographic coverage; however, case studies over limited regions have demonstrated the considerable potential afforded by
exploiting both domains simultaneously (Garca-Haro et al., 2006).
5.2 Major Multiangle Instruments

Under a restricted definition for Earth Observation, including only instruments that
provide near-simultaneous acquisition of >2 angular looks of the same area on
the Earths surface, and excluding the thermal or longer wavelengths, only three
such instruments have ever been placed in orbit: the NASA Multiangle Imaging
Spectro-Radiometer (MISR) flown on the NASA Earth Observing System Terra;
the French Space Agencys (CNES) POLarization and Directionality of the Earths
Reflectance (POLDER), flown initially on the Japanese ADEOS satellite series and
currently on the CNES Parasol platform; and Sira (UK)s Compact High Resolution Imaging Spectrometer (CHRIS), flown on the European Space Agency (ESA)
Proba-1 satellite. Terra was launched in December 1999, Proba in 2001, and Parasol
in 2004. Unprecedented efforts have also been made in the last 15 years properly
to treat data from across-track scanning instruments such as the NOAA Advanced
Very High Resolution Radiometer (AVHRR) and the MODerate resolution Imaging
Spectroradiometer (MODIS) flying on NASAs Terra and Aqua EOS satellites as
99
intrinsically dependent on surface BRDF. Although for the latter instruments multiangular data sets are acquired sequentially rather than near-simultaneously, work
with these accumulating, or sequential systems is also making important contributions to multiangle remote sensing. For this reason, reference is made here to
MODIS as a multiangle instrument, even though it was not designed to acquire
multi-angle imagery and the variation in viewing and illumination angles is an accident of the across-track acquisition method. Indeed, under this paradigm many
other moderate resolution sensor systems, including those on geostationary platforms, could also be deemed to be multiangle. It is preferable to think of these
systems in this way rather than to insist on the narrower definition because angular
artifacts in data from these instruments might otherwise be overlooked.
This chapter will concentrate on applications using data from orbiting instruments providing more than two views; this excludes Japan Aerospace Exploration
Agencys ASTER on Terra, which exploits dual views via a geometric strategy
rather than the radiometric strategy necessary with data from sensors with a
larger instantaneous field-of-view (IFOV) and the Advanced Along-track Scanning
Radiometer (AATSR), an important European instrument whose dual views have
proven valuable for aerosol characterization (King et al., 1999). The geometric strategy involves real or apparent differences in the shape or location of observed objects,
resulting from changes in perspective (for targets with 3-D structure), stereoscopic
parallax (displacement dependent upon distance from the observer), or actual motion of the target during the time interval between views; the radiometric approach
refers to changes in the brightness, color, contrast, or other radiance-related quantities as a function of view and/or illumination angles (Diner et al., 2005, Table 5.1).
5.2.1 MISR on Terra

The Multi-angle Imaging SpectroRadiometer (MISR) was designed and built at
NASAs Jet Propulsion Laboratory (JPL), in Pasadena, California and flies on
NASAs Terra satellite. MISR measures Earths spectral radiance in four spectral
bands (blue, green, red and near-infrared), at each of nine look angles in the forward
and aft directions along the flight path (Fig. 5.2). One camera views at nadir (close
to 0 view zenith angle), while the others view at 26.1 , 45.6 , 60.0 , and
70.5 (Fig. 5.3). Spatial samples are acquired every 275 m. In the global mode, all
nine cameras report the red band and the nadir camera reports all other bands in at
this resolution; for the off-nadir cameras the other bands are downsampled to 1.1 km
via contiguous averaging of four cross-track by four along-track line samples within
the instrument before data transmission. Over a period of 7 min, a 400 km wide
swath of Earth comes into view at all nine angles. Global coverage is achieved every
9 days, with repeat coverage between 2 and 9 days depending on latitude (the repeat
time is shorter at higher latitudes). In local mode, all four spectral bands are retained
at the acquisition resolution of 275 m in all nine cameras. Special attention has
6 7 km nominal, 6 6
after processing,
2,400 km swath
Pole-to-pole
2, 400 1, 800 km
1 day
Pole-to-pole 400 km
9 days
Scene dimensions
Global coverage in
444.5, 444.9, 492.2,

564.5, 670.2, 763.3,
763.1, 907.7, 860.8 nm
446, 558, 672, 866 nm
275 m or 1.1 km,

depending on channel
Radiometric
Radiometric, geometric
August 1996, on ADEOS

1; December 02, on
ADEOS II; December 04,
on Parasol
Comprehensive (over
several days)
Wide Field-of-View
pushbroom radiometer
POLDER (ESA)
Spatial resolution
Multiangle strategies
employed
Spectral coverage
0 , 26 , 46 , 60 , 70
forward and backward of
nadir in along-track
direction
December 1999, on EOS

Terra
Launch Date
Angular coverage
Multiangle pushbroom
imager
Description
MISR (NASA) (CNES)
1834 m
depending on the
node selected,
13 km swath
13 13 km
approximately
NA
Radiometric,
geometric
Programmable up
to 63 bands,
0.41.0 m
Bandwidth:
<11 nm
Five looks in the

along-track
direction
Tiltable
pushbroom
spectrometer
October 2001, on
Proba-1
CHRIS/Proba
Table 5.1 Major satellite instruments with multiangle capability (Adapted from Diner et al., 2005.)
Pole-to-pole
2, 300 km
2 days
250 m 1 km at
nadir, depending
on channel
36 bands:
0.414.5 m
Radiometric
055
continuously
December 1999,
on EOS Terra
Across-track
imaging scanner
MODIS (NASA)
Pole-to-pole
limb-to-limb
1 day
20 km at nadir
Shortwave:
0.35 m Thermal
window: 812 m
Total: 0.3200 m
Radiometric
Limb to limb
December 1999,
on EOS Terra
Scanners
CERES (NASA)
Pole-to-pole
512 km
6 days
0.555, 0.659,
0.865, and 1.61
0.555, 0.659,
0.865, and
1.61 m (VIS-IR);
3.7, 10.8, and
123.7, 10.8, and
12 m (TIR)
1 km at nadir
Two looks, 55
forward and close
to nadir in the
along-track
direction.
Radiometric
March 1, 2002, on
ENVISAT
Circular scanner
AATSR (ESA)
100
M.J. Chopping
101
Fig. 5.2 Computer-generated image showing the Terra spacecraft and MISR instrument operation.
The direction of flight is toward the lower left. The locations imaged by the 9 MISR cameras, each
with 4 spectral bands, are illustrated here with translucent surfaces (Courtesy of Shigeru Suzuki
and Eric M. De Jong, Solar System Visualization Project. NASA/JPL-Caltech, image P-49081.)
Fig. 5.3 MISR images of Zambia and Botswana, Africa, acquired on August 25, 2000. The left
image is a true color composite from the red, green, and blue vertical-viewing (nadir) camera.
The middle image combines data from the green, red, and near-infrared bands. The right image
contains red band data only but as a composite of imagery from the nadir, 70.5 forward, and
70.5 aftward cameras. The color variations in the multi-angle composite arise because light is
reflected at different angles. (Courtesy of the MISR Science Team.)
been paid to providing highly accurate absolute calibration, using on-board hardware consisting of deployable solar diffuser plates and several types of photodiodes
(Diner et al., 1998; Bruegge et al., 2002). The land surface reflectance products currently generated at the NASA Langley Atmospheric Sciences Data Center include
the spectral hemisphericaldirectional reflectance factors (HDRF) at the nine MISR
view angles and the associated bihemispherical reflectances (BHR). The HDRF and
the BHR characterize the surface reflectance under direct and diffuse illumination.
102
M.J. Chopping
An algorithm for the generation of vegetation green leaf area index (LAI) and the
fraction of photosynthetically active radiation absorbed by vegetation (fPAR) from
MISR BHR and bidirectional reflectance factor (BRF) was implemented for operational processing (Knyazikhin et al., 1998c). Unlike single-angle approaches, the
MISR LAI/fPAR algorithm does not rely on prescribed vegetation type maps as it
uses the multiangular information to constrain the retrievals.
5.2.2 MODIS on Terra

The MODerate resolution Imaging Spectroradiometer (MODIS) instrument flying
on NASAs Earth Observing System (EOS) Terra and Aqua satellites provides 12
bit imagery in 36 spectral bands ranging in wavelength from 0.4 to 14.4 m. Two
bands (red and NIR) are imaged at a nominal resolution of 250 m at nadir, with five
bands at 500 m, and the remaining 29 bands at 1 km. MODIS is an across-track scanning radiometer providing observation along a scanline that observes the surface at
zenith angles of 55 . At the EOS orbit of 705 km, it achieves a 2,330 km swath and
provides global coverage every 12 days. MODIS-AM was placed into orbit with
the launch of Terra on December 18, 1999; the second instrument, MODIS-PM, is
flying on the Aqua spacecraft launched on May 4, 2002. Thanks to the very wide
swath, MODIS provides multiangular measurements of spectral radiance from the
same location on the surface only through the accumulation of looks over a period
of several days (sometimes called the sequential method); it is not capable of nearsimultaneous multiangle observation. While global coverage is provided in a period
of 48 h or less, cloud cover means than data from several overpasses are almost always required in order to map clear-to-surface surface-leaving spectral radiances.
The across-track scanning method provides angular observations closer to the solar principal plane (the plane intersected by the sensor, target, and the Sun) than
along-track instruments at mid-latitudes, although this is highly dependent on latitude. The major disadvantage of this method is that the surface may change during
the accumulation period; however the variation in the signal owing to accumulated
sampling is generally far lower than that owing to the BRDF. Since MODIS data are
highly dependent on viewing and illumination geometry and surface BRDF, major
efforts have been expended to properly account for BRDF effects, and to exploit the
directional signal to retrieve global land albedo (see Chapter 9).
5.2.3 CHRIS on Proba-1

The CHRIS sensor was developed by Sira Electro-Optics (UK). Since April 2006
Sira space and imaging operations have been a part of Surrey Satellite Technology Ltd., Guildford, UK, a spin-off company of the University of Surrey. CHRIS
produces imagery in up to 62 spectral channels in the range 4151,050 nm with a
103
Fig. 5.4 (a) Artists impression of the Proba satellite (b) CHRIS instrument multi-angle acquisition
via tilting of the platform (Courtesy of ESA and Mike Barnsley, University of Wales at Swansea.)
spectral resolution of 512 nm. It is remotely configurable in terms of both spectral

channels and spatial resolution (European Space Agency, 1999), with a nominal
ground resolution of between 17 and 34 m. The nominal swath is 13 km. CHRIS
was launched on the Proba-1 satellite (Verhaert, Belgium) on October 21, 2001
(Fig. 5.4). Proba-1 is a small platform, weighing approximately 100 kg and measuring approximately 60 60 80 cm. Thanks to its four reaction wheels the platform
is highly maneuverable: along-track pointing allows a given site to be imaged five
times during a single overpass, while across-track pointing ensures that the revisit
time for a site of interest is potentially less than a week (Barnsley et al., 2004). Proba
was not placed into its intended orbit because of an anomaly with the final stage of
the Indian Polar Satellite Launch Vehicle (PSLV), resulting in a much lower and
more elliptical orbit than intended. This caused a number of operational problems
that necessitated on-ground and on-board software changes and delayed the start of
the science program. It took more than 12 months to rewrite the software required
for precise orbital navigation, so that the first useable overlapping hyperspectral imagery was available to most users only in early 2003 (albeit often with four partially
overlapping looks, rather than the five intended).
Since then, further improvements have been made so that five overlapping images are generally available and the overlap area is larger. Both the satellite and the
CHRIS instrument continue to function well at the time of writing (April 2006),
although striping is evident in some images owing to detector anomalies; however,
third-party algorithms have been written to correct for this. The CHRIS/Proba system has several unique features: it employs tilting of the platform to gain a series
of angular views in the along-track direction; the platform nods in order to acquire
images; and the CHRIS sensor is programmable for spatial resolution and spectral coverage. There are well over 60 groups around the world engaged in research
with CHRIS in categories including calibration/validation, land and environment,
hydrology, atmosphere, agriculture, forestry, hazards, renewable resources, coastal
zones, topographic mapping, and geology (European Space Agency, 2006). It is
104
M.J. Chopping
worth noting that not all projects seek to exploit the multiangle feature of the instrument; many projects are focusing on the exploitation of the hyperspectral capability.
5.2.4 POLDER on ADEOS / Parasol

The Polarization and Directionality of the Earths Reflectance (POLDER) instrument developed by France was initially launched on the Japanese ADEOS-1
(ADvanced Earth Observation Satellite) platform in August 1996, with the satellite
failing in June 1997 (a solar panel failure terminated operations after 8 months of
continuous acquisition); a second identical instrument, was launched on ADEOS-2
in December 2002, with that satellite failing in October 2003 (another solar panel
failure terminated operations with 7 months of continuous acquisition). This version of POLDER provided a swath of 2, 400 km with a footprint of 6 7 km2 . The
instrument was a two-dimensional CCD detector array with wide field of view telecentric optics and a rotating wheel carrying spectral and polarization filters. There
were 15 filters and polarizers in the visible and the near infrared range providing
nine spectral bands, three of which (443, 670 and 865 nm) were associated with
polarized filters. Three bands were dedicated to the observation of ocean color (443,
490, 565 nm) but POLDER-1 was designed to provide dedicated measurements
of clouds and atmospheric aerosols. Its applications in terrestrial remote sensing
have therefore taken a secondary role, although they are not insignificant (Maignan
et al., 2004; Breon et al., 2002; Camacho-de Coca et al., 2002). While POLDER-1
acquired spectral radiance data at multiple angles for the same location on a nearsimultaneous basis in the along-track direction, its full angular sampling potential
was only realized after a number of successive orbits has been accomplished, typically within 15 days. This means that POLDER-1 was a hybrid system and has
some features in common with accumulating instruments.
The third POLDER instrument hereafter called POLDER-2 as it differs from
the first version, POLDER-1 is being flown on the CNES Parasol platform, which
was launched from Europes spaceport in Kourou, French Guiana in December 2004
(Fig. 5.5). Parasol is a micro-satellite that flies in formation with the NASA EOS
A Train. The POLDER-2 instrument has provided near-continuous acquisition
since March 2005. It observes in eight spectral bands from 440 (blue) to 910 nm
(NIR) using a bi-dimensional CCD matrix providing a spatial resolution of 6 km;
three of the bands are polarized: 440, 670, 865 nm. As with the POLDER instrument on the ADEOS-2 platform, the Parasol-borne POLDER-2 provides its most
dense angular sampling via accumulation over a period of a few days. Over successive days, the instrument is at different places in the sky, which provides another
set of angular measurements: after a few days, the directional space within 60
of view zenith angles is fully covered with a good sampling of the principal and
perpendicular planes, including the hot spot around which the solar and viewing
directions coincide.
(a)
105
(c)
50
45
[%]
40
35
30
Ref_443
Ref_565
Ref_670
25
20
80
60
40
20
20
Phase Angle
(b)
(d)
10
8
[%]
6
4
2
RefPol_443
RefPol_370
RefPol_865
0
2
80
60
40
20
20
Phase Angle
Fig. 5.5 POLDER-2 images over desert target in the Sahara (a) 3 color composites in total light
443670865 nm (b) 3 color composites in polarized light 443670865 nm (c) directional signatures in total light (d) directional signatures in polarized light (Courtesy of F-M Breon, Laboratoire
des Sciences du Climat et de lEnvironnement.)
5.3 Multiangle Applications: Vegetation Canopies

Satellite remote sensing would ideally be able to provide information on both the
structure and function of ecosystems and the types of plant communities from which
they are formed, at regular intervals, and for the entire terrestrial surface. While remote sensing of vegetation function is an established fundamental and very important application of remote sensing spectral measures have proved very effective in
mapping vegetation productivity less has been achieved with respect to remotely
measuring vegetation structure. The optical properties of vegetation have allowed
the exploitation of reflected light in the solar spectral wavelengths via spectral vegetation indices, spectral unmixing techniques and in the temporal domain, phenology
metrics, that have proved useful in obtaining global maps of vegetation parameters,
such as the MODIS Vegetation Continuous Fields (VCF) % Tree Cover product
(Hansen et al., 2002). In spite of these successes, it is difficult to access canopy
structure information using single-angle sensing since structural surface properties
are largely confounded in nadir-spectral measures.
106
M.J. Chopping
Canopy structure is defined as the vertical and spatial distribution, orientation

and density of foliage and its supporting structures. Vegetation structure determines
the pattern of light attenuation and the distribution of photosynthesis, respiration,
transpiration, and nutrient cycling in the canopy (Widlowski et al., 2004). Net primary production of woody vegetation is always dependent on vegetation structure,
since the maximum rate of carbon uptake is constrained by total leaf area, and the
rate of carbon loss is constrained by total woody biomass (Widlowski et al., 2004).
Measures of canopy structure are required in order to estimate biomass, to assess
geographically shifting abundances of woody plant material (shrubs and trees), to
estimate carbon emissions to the atmosphere from wildfires, and to examine where
forest regrowth and secondary succession are occurring, and at what rates (Liu and
Kafatos, 2007). Canopy architecture and spatial heterogeneity are important not
only in forests but also affect the available carbon storage, vegetation dynamics and
productivity of grasslands undergoing woody encroachment, shrublands, and savannas: different degrees of structural heterogeneity lead to a wide range of estimates
for these environments (Asner, 2000; Chopping et al., 2003). It is also necessary for
mapping fire fuel loads; and for estimating surface roughness, with the implications
this has for fluxes between the surface and the atmosphere.
The surface BRDF is dependent on the structural as well as optical characteristics of surfaces and it should thus be possible to exploit angular measurements to
retrieve structural canopy parameters. Although this is not straightforward, in the
last few years multiangle reflectance data from satellites have been used to develop
a variety of approaches (Pinty et al., 2002; Bacour et al., 2002; Gao et al., 2003;
Hu et al., 2003; Widlowski et al., 2004; Jenkins et al., 2004; Nolin, 2004; Koetz
et al., 2005b; Chen et al., 2005; Strahler et al., 2005; Diner et al., 2005; Kimes et al.,
2006; Laurent et al., 2005; Bach et al., 2005; Chopping et al., 2006c; Disney et al.,
2006; Garca-Haro et al., 2006; Heiskanen, 2006; Chopping et al., 2007). The
approaches taken are very diverse and include 3-D radiative transfer modeling,
geometric-optical canopy reflectance modeling, exploitation of the structural
information available in the parameters returned via inversion of physical and
semi-empirical models, purely empirical methods, and heuristic and data mining
algorithms. All have demonstrated some degree of success.
5.3.1 Empirical and Synergistic Approaches

For some years Earth observation scientists have been considering the range and
importance of synergies that may be realized between multiangle instruments and
other types of sensor. The focus has been on synthetic aperture radar (SAR) and
more recently lidar, both active systems, although good synergies with high resolution stereoscopic imagers can also be found for some applications (e.g., cloud
top heights). Disney et al. (2006) demonstrated that a combined structural and radiometric modeling approach provides a flexible and powerful method for simulating the remotely sensed signal of a forest canopy in the solar and microwave
107
domains, useful for exploring the impact of canopy structure on the resulting signal and for combined retrievals of forest structural parameters. They developed a
detailed 3-D structural model of a conifer forest canopy in order to simulate the
observed reflectance (optical) and backscatter (microwave) signals and showed that
it is feasible to model forest canopy scattering using detailed 3-D models of tree
structure (including the location and orientation of individual needles). A structural
growth model of Scots pine was modified to simulate observed growth stages of
a Scots pine canopy from age 5 to 50 years. Model canopies were generated by
adding needles to tree structural models and these were used to drive optical and
microwave models of canopy scattering. Simulated canopy radiometric response
was compared with airborne hyperspectral reflectance data from the HyMAP instrument (Integrated Spectronics, Australia) and airborne synthetic aperture radar
backscatter data. Model simulations agreed well with observations, particularly at
solar wavelengths. The choice of needle shape and phyllotaxy was shown to have
a significant impact on multiple scattering behavior at the branch scale. In the microwave domain, simulated backscatter values agreed reasonably well with observations at L-band and less so at X-band. L-band simulated backscatter significantly
underestimated observed backscatter at younger canopy ages, probably as a result
of inappropriate modeling of the soil/understory background.
Successful and convincing empirical applications of multiangle reflectance data
have been rare until recently, when the exploration of synergies with lidar data became feasible. Researchers at the NASA Goddard Space Flight Center, Greenbelt,
Maryland, performed work in which forest structural data from the airborne Laser
Vegetation Imaging Sensor (LVIS) were used with multivariate regression and
neural networks to allow the estimation of forest vertical structure from multiangle
reflectance data from the Airborne Multi-angle Imaging SpectroRadiometer (AirMISR) (Kimes et al., 2006). The AirMISR data were acquired on 28 August 2003
over a 7,000 ha experimental forest near Howland, Maine, consisting of small plantations, multi-generation clearings and large natural forest stands (predominantly
boreal-northern hardwood transitional forest with spruce-hemlock-fir, aspen-birch,
and hemlock-hardwood associations). Twenty-eight AirMISR multiangle spectral
radiance values were selected from the 36 possible combinations (nine angles and
four spectral bands). Data acquired at nadir viewing and at 26.1 , 45.6 , and
60.0 in the flying direction were used and resampled into 15 15 m pixels. The
LVIS data were acquired 3 days later, with a total of 160, 000 LVIS shots collected
in the study area. The LVIS data set provided a relatively direct measure of forest
vertical structure at a fine scale (20 m diameter footprints). For each laser shot, the
following were recorded: location (long, lat), surface height (m), and the heights of
the 25%, 50%, 75% and 100% of waveform energy data (m). These data were gridded into 1515m pixels using inverse-distance weighted average of the four nearest
shots. The best model accurately predicted the maximum canopy height as measured
by LVIS using AirMISR data with a root mean square error (RMSE) of 0.92 m and a
R2 = 0.89 (Fig. 5.6). This result is all the more remarkable because these high accuracies were achieved over a study site with an elaborate patchwork of forest communities with exceptional diversity in forest structure. The study concluded that models
108
M.J. Chopping
Predicated LVIS H100 Values (m)
25
20
15
10
5
10
15
20
25
75
100
True LVIS H100 Values (m)

25
True/Predicted Forest Height (m)
b
Predicted Undisturbed Stand
True Undisturbed Stand
Predicted Selective Cut
True Selective Cut
20
15
10
0
0
25
50
LVIS Energy Quartile Heights (%)
Fig. 5.6 (a) LVIS forest canopy heights predicted by AirMISR versus heights from LVIS for
17 17 pixel windows. The model was applied to 1,000 random test points. (b) Predicted versus
true LVIS energy for canopy height (H100) and quartile heights (H75, H50, H25) for the logged
and unlogged areas. The predictive models used only AirMISR data (Reprinted from Kimes et al.,
2006. Copyright, Elsevier. With permission.)
using multiangular data are capable of accurately predicting the vertical structure of
forest canopies. Further work using data from the Geoscience Laser Altimeter System (GLAS) on the Ice, Cloud, and Land Elevation Satellite (IceSAT) platform to
train MISR-derived estimates of canopy and extend them continuously over large
areas has provided encouraging but somewhat uneven results. This is most likely
to be at least partly owing to the azimuthal orientation of MISR data away from
the principal plane that sometimes can be advantageous, e.g., for the retrieval of
albedo and fPAR (B. Pinty, personal communication) since the AirMISR data were
acquired close to this plane (K.J. Ranson, personal communication).
109
Fig. 5.7 Predicting stand basal area with AirMISR (non-linear multiple regression) using an artificial neural network with 5 nodes. Training R2 = 0.68, RMS Error = 8 m2 /ha (Courtesy of Rob
Braswell and Julian Jenkins, Institute for the Study of Earth, Oceans and Space (EOS), University
of New Hampshire.)
Another study using independent field biometric measurements to train multiangle data was carried out over the Bartlett Experimental Forest in New Hampshire.
An artificial neural network with five nodes was used to map basal stand area over
an area of 19.1 km2 . A high correlation was obtained between optical multiangle
reflectance signature measured with AirMISR and field observations of stand basal
area, with an RMSE of 8 m2 /ha (Jenkins et al., 2004; Fig. 5.7).
These studies show that multiangle data can be used to leverage lidar or surface assets to provide wall-to-wall coverage using purely empirical methods, rather
than explicit modeling of light scattering. Further advances might be expected using
lidar synergistically with multiangle data in modeling approaches as well as in an
empirical training mode (Hese et al., 2005; see Geometric-Optical and Hybrid Models, below).
5.3.2 Radiative Transfer

Three-dimensional radiative transfer (RT) modeling is used as part of MISR surface
product retrievals. A look-up table (LUT) is used to rapidly determine the matching
modeled reflectances and the associated values of leaf area index (LAI) and fraction
110
M.J. Chopping
of photosynthetically active radiation (fPAR): all necessary radiative transfer parameters are pre-computed and stored in a Canopy Architecture Radiative Transfer
(CART) file. The MISR LAI retrieval algorithm does not depend on a particular
canopy radiation model as the elements of the CART are components of various
forms of the energy conservation law. The MISR leaf area index (LAI) product has
been shown to be accurate to within 0.5 LAI in herbaceous vegetation and savannas
and is an overestimate by about 1.0 in broadleaf forests (Diner et al., 2005). When
using single angle data, uncertainty in biome classification can lead to large errors.
However, the use of multiangle data minimizes the impact of biome misidentification on LAI retrievals; that is, despite errors in biome classification, the use of
multiangle information leads to a similar error rate even when the biome is misclassified as is obtained with single angle approaches and a prescribed biome map.
The multiangle method therefore permits LAI retrieval without the use of a prescribed biome map. In a case study in Africa, 80% of the LAI values were retrieved
using an incorrect biome type but with a probability of about 70%, uncertainties
in LAI retrievals due to biome misclassification did not exceed uncertainties in the
observations (Hu et al., 2003).
A parameter closely related to LAI, the canopy recollision probability (defined
as the probability with which a photon scattered in the canopy interacts with a phytoelement again) has been shown to describe canopy spectral absorption and scattering; this is a new and developing area of multiangle imaging research. Knyazikhin
et al. (1998a, b) proposed that to a good approximation the amount of radiation absorbed by a canopy should depend only on the wavelength and a canopy
structural parameter (p), which is wavelength independent. Knowing the recollision probability value of a canopy, the scattering coefficient of the canopy at any
wavelength can be predicted from the leaf scattering coefficient at the same wavelength. Knyazikhin et al. (1998a, b) also introduced a similar parameter (pt ) relating
canopy transmittances at two different wavelengths to the leaf scattering coefficients
at these wavelengths. Given the absorption (p value) and transmission (pt value), total reflectance (the upward scattered part of the incident radiation) is also known. It
is recognized that LAI may not depend on only these parameters alone but may also
vary with leaf orientation and the degree of foliage clumping. A recent study by
Smolander and Stenberg (2005) confirmed that the spectral absorption and scattering of structurally simple uniform canopies can be well described by the canopy p
value, which furthermore showed close relationship with the LAI but insensitivity
to solar zenith angle.
5.3.3 The Radiation Transfer Model Intercomparison (RAMI)

Exercise
Radiative transfer model accuracy and robustness are key issues for improving
remotely sensed surface variables that may depend heavily on the angular distribution of samples. The Radiation Transfer Model Intercomparison (RAMI) exercise
111
(Pinty et al., 2001, 2004a) has been ongoing since 1999 with the most recent and
third phase completed in 2005. RAMI provides a framework within which spectral bidirectional reflectance models designed to simulate the transfer of radiation
within plant canopies and over bare soil surfaces can be benchmarked. This is useful because the interpretation of remote sensing data generated by Earth Observing
satellites hinges on the exploitation of such models: it is clearly very important that
they are reliable, accurate, and fit for purpose. RAMI approaches this endeavor by
providing a series of test cases of different degrees of complexity, broadly divided
into homogeneous and heterogeneous cases (Fig. 5.8).
Participants perform simulations based on these scenarios and deliver modeling results to the Institute for Environment and Sustainability at the European
Commission Joint Research Center in Ispra, Italy, that coordinates the exercise.
The subsequent publication of the findings provides an objective way to evaluate
the performance and limits of applicability of the models. In addition, new models
may be developed and existing ones improved by comparing their output with the
available simulation results on the RAMI web pages (http://rami-benchmark.jrc.it/).
The overall benefit to the remote sensing community is a demonstration of model
maturity and a better understanding of how the models may be used in the interpretation of remote sensing data. The latest results, that include 5 1-D models and
Fig. 5.8 Selected scenarios in the Phase 3 RAMI set of radiative transfer modeling exercises: (a)
homogeneous turbid cases in the solar domain: leaves treated as a turbid medium (b) homogeneous
discrete cases in the solar domain: leaves represented as a large number of non overlapping discshaped objects (c) heterogeneous turbid cases in the solar domain: large number of non overlapping
spherical objects with crowns represented as a turbid medium (d) discrete floating spheres in the
solar domain: large number of non overlapping spherical objects representing the individual plant
crowns that contain randomly distributed finite size disc-shaped scatterers, with the orientation of
the scatterers following a uniform distribution function (e) Heterogeneous Scene-Based Experiments: large number of disc-shaped scatterers contained within a series of non-overlapping spherical and cylindrical volumes representing identically sized plant crowns (f) Boreal Birch Stand
Scene: large number of randomly located non-overlapping tree-like entities of differing sizes with
individual objects represented by ellipsoidal crowns with randomly distributed finite sized foliage.
See: http://rami-benchmark.jrc.it/HTML/RAMI3/RAMI3.php
112
M.J. Chopping
13 3-D models, are reported in Widlowski et al. (2006). It should be noted that
model accuracy is an important but not the only limiting factor in the application
of reflectance models in retrieving surface information useful to applications using
multiangle remote sensing data: inversion algorithms (Chapters 68), model paradigm and suitability for purpose, noise and contamination in data sets, and angular
and spectral sampling also play important roles.
5.3.4 Canopy Openness

One approach to estimating canopy openness makes use of parametric fits of red
wavelength MISR or AirMISR data to the Rahman-Pinty-Verstraete (RPV) model
(Rahman et al., 1993) or the modified version of this (MRPV, Engelsen et al.,
1996). The MRPV model parameters from inversion against MISR data are included
in the operational MISR Land product at level 2. The RPV and MRPV models
are based on a consideration of the main aspects of BRDF shapes and are threeparameter models. The first parameter is a factor describing the overall amplitude
of reflectance; the second describes the steepness of the BRDF bowl or bell shape;
and the third controls the relative importance of back- and forward-scattering via
a Henyey-Greenstein function (for the RPV model) or an exponential function in
scattering angle (for the MRPV model). The equations for the MRPV model are
given in Eq. (5.1):
BRFRPV = s (s , v , ; 0 , b, k) = 0 M F H
(5.1)
M = [cos s cos v (cos s + cos v )]k1

F = exp(b cos(g))
H = 1 + [(1 )/(1 + G)]
(5.2)
(5.3)
(5.4)
where:
where is the average measured reflectance factor and

G = [tan2 s + tan2 v , 2 tan s tan v cos ]1/2
cos g = cos s cos v + sin s sin v cos
(5.5)
(5.6)
The first parameter, 0 , is a factor describing the overall amplitude of reflectance;

the second, b, controls the relative importance of back- and forward-scattering via a
Henyey-Greenstein function in the original RPV formulation, replaced by an exponential term in the MRPV model Eq. (5.2) and in the MISR BRDF/albedo product.
The rationale for this modification is that except for the hotspot term the coefficients
of the MRPV model can be rapidly solved by linear least-squares by taking logarithms of the expression; but note that a recently-developed algorithm allows efficient, reliable and accurate inversion of the original RPV model and additionally
113
Fig. 5.9 Typical BRF anisotropy in the red spectral domain for radiatively homogeneous (left) and
heterogeneous (right) vegetation canopies. The BRFs of the heterogeneous surface were generated
using a 3-D RT model and typified by a bell-shape (k > 1). The BRFs of the homogeneous surface
covers were generated using a 3D RT model and a 1D IPA approach, and are generally bowl-shaped
(k < 1) (Reproduced from Widlowski et al., 2001. Copyright, AGU.)
provides a measure of uncertainty on the retrieved parameters (Lavergne et al.,

2007). The modified Minnaert function parameter, k, is used to quantify the degree
by which the observed bidirectional reflectance factor data resemble a bowl- or bellshaped pattern in azimuthal planes (Fig. 5.9). The presence of a very sparse or very
dense and quasi-homogeneous vegetation layer within the sensor IFOV is characterized by a bowl-shaped reflectance anisotropy shape (k < 1), while vertically elongated foliage clumps of medium-to-high densities uniquely generate bell-shaped
shapes (k > 1) (Widlowski et al., 2001, 2004; Pinty et al., 2002; Fig. 5.9). This
broad relationship holds well for snow-covered areas, with homogeneous bare or
non-forested areas exhibiting a bowl-shaped pattern (k < 1), indicating that subpixel homogeneity (sparse or extremely dense vegetation cover) will result in k < 1.
A dependency on vegetation density was demonstrated for the first time in Gobron
et al. (2002).
The k parameter has a limited physical meaning (Nolin, 2004); in order to translate these relationships to measurable canopy structural parameters, it is possible
to interpret k values as a ratio between mean effective scene height and the ratio of
mean tree density and mean nearest-tree distance (Widlowski et al., 2004). Effective
scene height is defined as the mean height of all structures within the IFOV weighted
by their fractional surface coverage. The relationship can be seen in kred results
obtained by these authors using radiation transfer simulations performed with an
illumination zenith angle of 30 , a soil reflectance of 0.126, a bark reflectance
of 0.251 and a leaf reflectance (transmittance) of 0.018 (0.021) (Fig. 5.10). The
data points generated differ only because of their structural characteristics (stem
and foliage density). Researchers using multiangle data from the European CHRIS
instrument found that the results of RPV model inversion enabled discrimination
114
M.J. Chopping
Fig. 5.10 The organization of the kred values in a plane defined by the mean effective scene
height, i.e., the height (m) of all structures within the IFOV weighted by their fractional surface
coverage and on a log scale the ratio between average tree density (stems/m2 ) and the mean
nearest tree distance (m). The data points show the characterizations of 200 structurally diverse
boreal forest scenes (Reproduced from Widlowski et al., 2004. Copyright, Springer. With kind
permission of Springer Science and Business Media.)
between different surface types based on the their inherent reflectance anisotropy
and that the k parameter was successfully linked to lidar measurements representing
the 3-D structure of the canopy (Koetz et al., 2005a).
5.3.5 Clumping Index

Foliage clumping is an important forest structural canopy attribute: it affects both
the gap fraction for the same LAI, radiation interception and distribution within the
canopy, which in turn affects photosynthesis. It can be quantified in a Clumping
115
Legend
0.0 - 0.3
0.3 - 0.4
0.4 - 0.5
0.5 - 0.6
0.6 - 0.65
0.65 - 0.7
0.7 - 0.8
0.8 - 1.0
No Value
Water
Excluded
Fig. 5.11 Global clumping index map derived from POLDER 1 data using the normalized difference between interpolated hotspot and darkspot NIR reflectance and applied to vegetated land
cover. Clumping increases with decreasing values of the index (Reprinted from Chen et al., 2005.
Copyright, Elsevier. With permission.)
Index (CI) based on a vegetation dispersion parameter obtained via a modification

to Beers law. A physical modeling approach to obtaining CI has been taken by
two collaborating groups in Canada at the University of Toronto and the Canada
Centre for Remote Sensing, Ottawa (Chen et al., 2005). These authors have produced
the first global maps of foliage clumping with multiangle POLDER data, assisted
by a geometrical optical model (Fig. 5.11). They used the Normalized Difference
between Hotspot and Darkspot (NDHD), an angular index that can be obtained from
multiangle remote sensing that characterizes reflectance anisotropy and has been
successfully related to ground measurements of CI (Chen et al., 2003). Using the
relative magnitude of the darkspot to the hotspot minimizes the dependence of the
index on foliage optical properties that are important determinants of bidirectional
reflectance. This work built on previous research using data from POLDER (Lacaze
et al., 2002).
5.3.6 Structural Scattering Index

A semi-empirical approach applied to MODIS data is the use of the Structural Scattering Index (SSI) developed to exploit the kernel weights derived by inversion of
the LiSparse-RossThick kernel-driven BRDF model (Gao et al., 2003). The MODIS
BRDF/Albedo (MOD43) product is generated on a regular 16-day cycle using this
model. The model depends on three parameters (kernel weights) describing the interaction of light with the surface. These parameters are used in a forward version
of the model to reconstruct the surface anisotropic effects and to correct MODIS
reflectances to a common view geometry (the MOD43B4 Nadir BRDF-Adjusted
116
M.J. Chopping
Reflectances or NBAR product). The model can also be used to compute

integrated black-sky albedo at some solar zenith angle as well as white-sky albedo
MOD43B3. The model is a linear, semi-empirical, kernel-driven (LiSK) model formulated as a superposition of weighted kernels Eq. (5.7):
R(s , v , ) = fiso + ( fgeo kgeo (s , v , )) + ( fvol kvol (s , v , ))
(5.7)
where R(s , v , ) is the modeled bidirectional reflectance for the solar zenith,
view zenith, and relative azimuth angles s , v , and , respectively. The model kernels kgeo and kvol are analytical functions of the solar and viewing angles derived via
simplifying physical terms (Roujean et al., 1992; Wanner et al., 1995): kgeo accounts
for geometric scattering and shadowing, while kvol accounts for volume scattering
from a discrete medium of randomly located facets, assuming the single scattering approximation, an isotropic facet distribution function, and an optically thick
medium. The kernel weights fiso , fgeo and fvol are theoretically dependent on the
three dimensional structure and optical properties of a canopy and its background;
fiso should represent the bidirectional reflectance viewing a surface at nadir with the
overhead Sun, while fgeo and fvol measure the relative contributions of the geometric/shadowing and volume scattering components. Models of this kind are also often referred to generically as Li-Ross models, since the terms describing anisotropic
scattering are obtained from functions derived by Xiaowen Li and Juhan Ross (see
Wanner et al., 1995).
The SSI is calculated as ln( fvol nir / fgeo red ), where fvol nir is the volume scattering kernel weight in the NIR band and fgeo red is the geometric kernel weight
in the red band. It is based on the principle that surfaces with a higher (lower)
and a relatively homogeneous (heterogeneous) vegetation cover will exhibit BRDFs
that exhibit a behavior closer to (further from) that of an ideal turbid medium and
will therefore generally have a higher (lower) volume scattering kernel weights in
the near-infrared (NIR) wavelengths; while in the red wavelengths sparser, more
clumped vegetation will exhibit higher kernel weights (Fig. 5.12). Since SSI depends on the vegetation cover fraction as well as vertical structure, a relative structural scattering index has been defined that partially removes the effects of cover
by estimating a linear relationship between SSI and a spectral vegetation index.
This approach has been found to be useful in exploiting canopy structure as translated through a kernel-driven model and should provide superior results for mapping
land cover.
5.3.7 Geometric-optical and Hybrid Models

The use of discrete object models such as geometric-optical (GO) and hybrid
geometric-optical/radiative transfer (GORT) models represents another approach to
exploiting multiangle data (Strahler et al., 2005). Geometric-optical (GO) models
treat the surface as an assemblage of discrete, identical, and large objects defined by
117
Structural Scattering Index
5
4
3
2
1
0
1
0
0.2
0.4
0.6
0.8
Nadir View NDVI (SZN = 30)

All Samples
Open Shrubland (Jan-Mar)
Barren/Sparsely Veg (Nov-Mar)
Evergreen Needleleaf Forest

Open Shrubland (Apr-Oct)
Barren/Sparsely Veg (Apr-Oct)
Fig. 5.12 Structural Scattering Index (SSI) vs nadir view NDVI from MODIS. The solid trend line
represents the change of SSI as the function of nadir view NDVI. Samples above the trend line
show the stronger volumetric scattering; those below show the stronger geometric scattering. SSI
is also dependent on temporal changes in cover types (Reprinted from Gao et al., 2003. Copyright,
Elsevier. With permission.)
geometric primitives (e.g., spheroids, cones, or cylinders) placed in a Poisson distribution above an underlying surface (Li and Strahler, 1985, 1992) which is frequently
considered Lambertian (Scarth and Phinn, 2000). The remotely sensed observation
is modeled as a linear combination of contributions from viewed sunlit and shaded
crown and background components, with each contribution a product of component
reflectance and the fraction of the sensors field of view occupied by the component
Eq. (5.8):
(5.8)
R = C k C + G kG + T kT + Z kZ
where C, G, T , and Z are the (assumed Lambertian) spectral reflectances or component signatures of the sunlit crown and background and shaded crown and background, respectively; and the kn are the component fractions.
GO models vary considerably in complexity, with some including terms that
allow for volumetric scattering within crowns rather than simple signatures (Ni and
Li, 2000; Chen et al., 2003, 2005); and in the way the background contribution is
represented (Ni and Li, 2000; Chopping et al., 2005). They hold the potential for obtaining upper canopy information such as stand density, openness, mean object size
and canopy height, and crown morphology, which are useful in multiple disciplines
(ecological modeling, forestry, hydrology, radiation budget; Garca-Haro et al.,
2006). GO models have been used in the calculation of Clumping Index (see above);
118
M.J. Chopping
some results from applying a simple GO model to estimating woody plant cover in
semi-arid environments in the southwestern USA are provided in the sections entitled The Background in Canopy Reflectance Modeling.
Using lidar data obtained over forest in the boreal ecosystem-atmosphere study
(BOREAS) sites in central Canada, Ni-Meister (2005) showed that lidar waveforms
can be simulated accurately and with precision using a modified version of the
Geometric-Optical/Radiative Transfer (GORT) canopy reflectance model (Li et al.,
1995), driven with canopy structure inputs. This model is dependent on canopy
geometry parameters such as tree size, shape, and tree density; and on the spectral reflectance properties of leaves and the BRDF of the background. The GORT
model was designed for use with passive remote sensing data and is appropriate
for the interpretation of multiangle data. Since it can be used to relate canopy parameters in both the active and passive domains it can be used as a bridge between
large footprint lidar and moderate resolution multiangle data. This is a particularly
promising direction.
5.3.8 Direction and Wavelength

The future will certainly see greater use made of the directional signal in many
wavelengths as the advantages of spectro-directional remote sensing become more
apparent (Garca-Haro et al., 2006; Schaepman, 2006). However, most use is still
made of red and NIR band data, mainly because it is in these regions that the dominant signatures in both spectral and angular domains are see; partly because these
have long been used in remote sensing of vegetation; partly because MODIS and
MISR both provide moderate resolution data (250275 m) in these bands; and partly
because interactions with the canopy are better understood. As seen above, several groups working with MISR data have adopted modeling approaches in which
only the red band reflectance data are used. This is predicated on the following
arguments:
The single scattering regime is dominant in the red spectral domain.
It exploits the reflectance and absorption contrasts between vertically clumped
vegetation and the background.
The linear mixture assumption underlying geometric-optical models is more
valid for the red than near infra-red wavelengths.
BRDF model inversion experiments using numerical methods show that there are
generally fewer problems such as trapping at local minima in the red compared
to the near-infrared (Gemmell, 2000).
It keeps the modeling problem (relatively) simple.
On the other hand, there have been attempts to investigate and exploit the spectral
and directional domains simultaneously (Weiss et al., 2000; Gobron et al., 2000,
2002; Bacour et al., 2002; Casa and Jones, 2005; Bach et al., 2005; Garca-Haro
et al., 2006). Weiss et al. (2000) addressed the question of a potentially optimal set of
119
spectral bands and directions for retrieval of biophysical parameters using the Scattering by Abitrarily Inclined Leaves (SAIL) radiative transfer model and the PARCINOPY nested radiosity/turbid medium model code. These were used to simulate
multiangle remote sensing data in the principal and perpendicular planes (32 directions) for a solar zenith angle of 45 and for a wide range of canopy conditions.
They found that for estimating the primary variables LAI (leaf area index) and Cab
(chlorophyll a + b content), the optimal number of bands is approximately six and
the optimal number of directions for these bands is between four and seven, with six
directions or less required for reliable retrievals of the primary variables plus fractional vegetation cover and fraction of absorbed photosynthetically active radiation
(fAPAR). Additional directions were found to contain redundant information and
so their use would merely induce noise in the retrieval process. The ideal directions
would be located in the principal plane close to the hot spot direction and in the
forward scattering direction for large zenith angles. Note that these results assume
idealized conditions (e.g., no cloud cover).
Bacour et al. (2002) sought to utilize both the directional and the spectral information of the images acquired in 16 flights by the AirPOLDER instrument (eight
bands in the visible-NIR) over the Alpilles test site (JanuaryOctober 1997), as
part of the Alpilles-ReSeDA campaign. Estimation of biophysical variables was
executed by inversion of three one-dimensional radiative transfer models, SAIL,
KUUSK and IAPI, coupled with the PROSPECT leaf optical properties model. They
focused on the capability of model inversion to retrieve the leaf area index (LAI) of
wheat, maize, sunflower and alfalfa crops. A quasi-Newton inversion algorithm was
used to estimate the Cab , LAI, the mean leaf inclination angle (l ), the hot spot parameter (sl ) and a multiplicative soil parameter (asoil ). The three models were shown
to accurately estimate the LAI compared to planimeter measurements but tended
to underestimate LAI for values above 2.3. Compensation effects between LAI and
Cab emerged from the spatial analyses of these variables, although the uncertainties
were low. Casa and Jones (2005) investigated an unconventional approach for the
estimation of leaf area index (LAI) and leaf angle distribution (LAD), based inversion of a canopy ray tracing model against multiangle data from a ground-based
multispectral camera. The model was developed using the Persistence of Vision
Raytracer (POVRAY) and inversion was carried out using a look-up-table approach.
Tests using an extensive data set gathered on a potato crop during experimental trials carried out at Viterbo (Italy) over 3 years showed that LAI was estimated with a
RMSE varying from 0.29 to 0.75 in the different years.
Bach et al. (2005) used data from the CHRIS instrument on the Proba platform to
invert the four-stream Soil-Leaf-Canopy radiative transfer model SLC, an extension
of the canopy reflectance models SAILH and GeoSAIL, producing maps of LAI,
leaf chlorophyll, and leaf inclination distribution for agricultural fields in the Upper
Rhine Valley test site along the German/French border for July 2003 (Fig. 5.13).
The input parameters include structural and physiological information on the
canopy, soil optical properties and the observation geometry. The Hapke model is
used to model a non-Lambertian soil BRDF, including its variation with moisture
content. The canopy itself is modeled with a two-layer version of the model SAILH
120
M.J. Chopping
>2.0
2.0-2.5
2.5.3.0
3.0-3.5
3.5-4.0
4.0-4.5
4.5-5.0
5.0-5.5
5.5-6.0
>35
35-40
40-45
45-50
50-55
55-60
60-65
65-70
70-75
Angle LIDF a
56 -0.30
52 -0.20
49 -0.10
Fig. 5.13 LAI, chlorophyll per leaf area (g/cm2 ) and average leaf angle maps for a 3 4 km
area of maize, retrieved by adjusting the SLC model against data from CHRIS for the Upper Rhine
Valley test-site along the German/French border (July 2003)
and transmittance of green and brown leaves is calculated using the PROSPECT
sub-model. The leaf angle distribution is approximated by parameters that describe the average leaf slope and the bimodality of the distribution. The spatial
distributions of all three retrieved parameters delineated fields and the additional
information obtained from the directional data showed that canopy structure was
crop-specific but also changed with phenological development. The field pattern in
the average leaf slope map (Fig. 5.13c) values was thought to result from different
varieties of maize. When the average leaf angle was retrieved for the same site
2 weeks after the initial CHRIS acquisition, the overall average angle retrieved
changed by about 10 to a more vertical distribution. This can be interpreted as maturation of maize where leaves become more vertical with development. The study
showed that crop parameters retrieved from multiangle CHRIS data can provide
input parameters essential for crop growth models.
Garca-Haro et al. (2006) developed an approach called directional spectral
mixture analysis (DISMA) for retrieving vegetation parameters with the focus on
fractional cover and leaf area index. This seeks to combine a consideration of the
spectral signatures of soil and vegetation components with an analytical approximation of the radiative transfer equation, resulting in a fast, invertible model suitable for use with discontinuous canopies. Data from the AirPOLDER and HyMap
instruments were used to test the model and its inversion using a lookup table
(LUT). Retrievals of LAI corresponded well to ground measurements of LAI, with
an RMSE of 0.50.6 and an R2 of the fitting of around 0.92. The spatial distribution
matched that obtained by inversion of the physically-based New Advanced Discrete
Model (NADIM) radiative transfer code, also known as the Semi-discrete model
(Gobron et al., 1997) (Fig. 5.14).
(a)
121
(b)
0.0 0.5 1.0 1.5 2.0 2.5 3.0 >3.5
Fig. 5.14 (a) Green LAI obtained via DISMA using AirPOLDER (b) the corresponding results
from inversion of the physically-based NADIM radiative transfer code. The maps shows a 5 5 km
area 28 km from Albacete, Spain; 1% and 7% of the pixels exceed the saturation value of 3.5 (From
Garca-Haro et al., 2006. Copyright, IEEE. With permission.)
5.3.9 The Background in Canopy Reflectance Modeling

Many researchers have recognized that for successful modeling of heterogeneous,
clumped, or non-closed canopies it is important to account adequately for the
contribution of the background of soil and understory (Gemmell, 2000; Ni and
Li, 2000; Chen et al., 2005; Koetz et al., 2005b; Bach et al., 2005; Chopping
et al., 2006a). In all canopies except closed especially where the understory is
heterogeneous attempts to exploit canopy reflectance models can be confounded
by spatial variation in background reflectance magnitude and anisotropy. In the last
few years modelers have moved from the assumption of a Lambertian background,
through imposing a single, static background BRDF, to attempts to estimate a spatially dynamic background BRDF. Recent work has shown that the soil/understory
reflectance can be obtained for both coniferous and deciduous forests using MISR
data, with the retrieved values following seasonal trajectories similar to those of
adjacent grasslands, a partial validation of the approach (Canisius and Chen, 2007).
Recent progress in dynamically estimating the anisotropic soil-understory contribution for discontinuous open shrub canopies in desert grasslands exhibiting a
wide range of canopy-background configurations including young, small honey
mesquite shrubs over a dark grassland matrix and older, larger ones over bright,
sandy soils has been made using the simplified geometric-optical (SGM) model
incorporating a kernel weighting approach with MISR data. Estimating the background contribution from MISR-derived LiSparse-RossThin model kernel weights
while fixing number shrub density, shape and height at typical values and using
numerical optimization to retrieve mean shrub radius has allowed the mapping
of fractional shrub cover at landscape scales in the United States Department of
122
M.J. Chopping
Agriculture (USDA), Agricultural Research Service (ARS) Jornada Experimental

Range in southern New Mexico. A root mean square error of 0.03 in fractional
cover with respect to values estimated from high resolution IKONOS panchromatic
imagery was obtained (measured fractional cover range: 0.030.19); however only a
moderate proportion of the variation in the measured data was explained by the estimated data set, with an overall R2 of 0.19 (Chopping et al., 2006b). When the same
algorithm was used to predict fractional shrub cover for pastures a few kilometers
from the study area for which independent estimates were available (from image
segmentation on a 0.6 m QuickBird panchromatic image), the distributions exhibited an improved spatial correlation (Fig. 5.15) and an R2 of 0.47 was obtained with
two model variants (Fig. 5.16).
When the algorithm was applied with MISR data over much larger areas for
the Jornada Experimental Range (783 km2 ) and the Sevilleta National Wildlife
Refuge (1, 000 km2 ) and their environs the maps include trees on the San
Andres mountains, in other elevated areas (e.g., Summerford Mountain in the Jornada), and in the riparian environments of the Rio Grande (southwest quadrant of
the Jornada map and center of the Sevilleta map), in addition to shrubs. The resulting distributions compare well with those of trees in the MODIS Vegetation
Fig. 5.15 (a) QuickBird shrub map for pasture 12 in the Jornada, red = shrub, white = background
(b) shrub cover aggregated to 250 m cells; brighter = greater shrub cover (c) retrieved using MISR
red band data to invert the SGM GO model (d) MISR/SGM shrub cover map for pastures 8/9 in the
Jornada (e) the corresponding QuickBird map (From Chopping et al., 2006c. Copyright, American
Geophysical Union.)
0.30
y = 0.4358x + 0.0208
R2 = 0.4676
0.20
0.10
0.00
0.00
++
+ +
+
++
+
+ +
++
+ + + +
+
+ +
+
++ +
+
+++ + +++ +
+ ++ ++ +
++
+++ ++++
+ ++ ++
+ + + +
+ ++ ++
+
0.10
0.20
0.30
Fractional Shrub Cover (QB-NIH)
Fractional Shrub Cover (MISR/SGM)
Fractional Shrub Cover (MISR/SGM)
123
0.30
y = 0.5427x + 0.0777
R2 = 0.4824
0.20
0.10 +
+
++
+
0.00
0.00
++
+ +
+
+ +
+
+
++ ++
+
++ + +
+ +++
++ +
+
+
+
+
+
+ ++ +
+
+ +++ +
++
+ ++++
+++
+
+
+
+ +++
++ +
+ + ++ ++
++
+ ++
+
0.10
0.20
0.30
Fractional Shrub Cover (QB-NIH)
Fig. 5.16 Retrieved vs. measured shrub cover (a) SGM (b) SGM with sunlit crown unweighted
by the fraction viewed (i.e., kC = 1.00) (Reproduced from Chopping et al., 2006c. Copyright,
American Geophysical Union.)
Continuous Fields (VCF) percent tree cover maps (Fig. 5.17). Note that the MISRderived map includes all large woody plants: it is based on exploiting the canopy
structure information encapsulated in the multiangle reflectance data (Chopping,
2006). Subsequent inversions of the SGM that allowed the crown shape parameter to vary indicated that it is possible to obtain regional maps of canopy height
as well as crown cover, allowing estimates of aboveground woody biomass. Retrievals of cover, canopy height, and biomass showed good matches with US Forest Service maps, with coefficients of determination of 0.78, 0.69, and 0.81, and
absolute mean errors of 0.10, 2.2 m, and 10.1 Mg/ha, respectively, after filtering for
high model fitting error, the effects of topographic shading, and a small number of
outliers (Chopping et al., 2007; http://csam.montclair.edu/chopping/wood/).
5.3.10 Land Cover and Community Type Mapping

Many studies have shown that there is much potential for improving the accuracy of land cover classification if patterns in the angular domain as well as the
spectral domain can be exploited (Abuelgasim et al., 1996; Hyman and Barnsley,
1997; Bicheron et al., 1997; Sandmeier and Deering, 1999). Chopping et al. (2002)
showed that the kernel weights of Li-Ross LiSK models obtained by adjusting a
LiSparse-RossThin variant against accumulated multiangle data from the NOAA
AVHRR provide contingency tests for community types in semi-arid environments
in Inner Mongolia Autonomous Region (IMAR; 10 types) and New Mexico (NM;
19 types) that are superior to those obtained using maximum-value compositing
using the Normalized Difference Vegetation Index (NDVI) as the criterion (denominated MVC), and in particular perform better in the worst case: the kernel weights
124
M.J. Chopping
Fig. 5.17 Woody plant fractional crown cover map obtained by adjusting the SGM geometricoptical model against MISR red band data (a) in the Sevilleta National Wildlife Refuge in central
New Mexico (b) % tree cover from the corresponding MODIS VCF product (c) in the Jornada
Experimental Range in southern New Mexico, dotted line is the tree-shrub boundary from the VCF
map (d) % tree cover from the corresponding VCF map product. Solid lines indicate the boundaries
of the Sevilleta and fencelines in the Jornada. (Reproduced from Chopping et al., 2006c. Copyright
2006 American Geophysical Union.)
provided minimum reliabilities of 83% and 69% for IMAR and NM experiments,
respectively, compared to only 11% and 26%, respectively, with the MVC data set.
These performances were reflected in Kappa Index values of 0.93 and 0.91 for the
IMAR and NM spectro-directional data sets against 0.74 and 0.46 for the MVC data
sets, respectively (the Kappa Index is a means to test two data sets to determine the
extent to which their similarities or differences are due to chance). When directional
125
information was incorporated into the signatures, errors of omission or commission

fell from 24% and 51% obtained with MVC to only 6% and 8% for the
IMAR and NM experiments, respectively.
More robust cross-validated tests of classifications of multiangle remote sensing metrics from MISR and nadir-spectral data from the An (nadir) camera have
demonstrated that using multi-angular data and anisotropy patterns raises the overall classification accuracy importantly. Su et al. (2007) studied maximum likelihood
and support vector machine (SVM) algorithms for mapping community types in
the Jornada Experimental Range and the Sevilleta National Wildlife Refuge, New
Mexico (19 classes). Half of the samples were randomly selected as the training
set and the other half retained as the testing set. A total of 66 classifications were
performed with various combinations of data sets: the 0 , k, and b parameters of
the MRPV model (see Canopy Openness, above); the isotropic, geometric and volume scattering kernel weights of a Li-Ross BRDF model; the structural scattering
index; and MISR surface reflectance estimates. Using multiangle data raised the
overall classification accuracy from 45.4% obtained with nadir observations only
to 60.9%, and with surface anisotropy patterns derived from MRPV and LiSparseRossThin BRDF models (separately) an overall accuracy of 67.5% can be obtained
with a maximum likelihood classifier. Using a non-parametric SVM algorithm the
classification accuracy was increased to 76.7%. Note that the classes in these experiments are community types that often differ very subtly in terms of their spectral
signatures, rather than broad land cover types (Fig. 5.18).
Fig. 5.18 Community type maps for the Sevilleta National Wildlife Refuge. (a) LTER vegetation
map (b) MISR maximum-likelihood classification (c) MISR support vector machine classification
(Reproduced from Su et al., 2007. Copyright, Elsevier. With permission.)
126
M.J. Chopping
5.4 Other Applications

Most applications or prototype applications of multiangle remote sensing of land
have been concerned with vegetation canopies. However, there are other important
Earth Science applications that are often tangentially related to vegetation, including estimating land albedo and the Earths radiant energy budget (see Chapter 9);
snow and ice; and mapping dust emission sources to simulate the likelihood of large
emissions.
5.4.1 Snow and Ice

Multiangle remote sensing data are able to provide important additional information
on Earth surfaces covered with snow and ice. For example, differentiating between
clouds and snow or ice surfaces using spaceborne detectors is difficult because the
surface may often be as bright and as cold as the overlying clouds, and because polar atmospheric temperature inversions sometimes mean that clouds can be warmer
than the underlying surface. Mega-sastrugi ice fields in East Antarctica, with dunelike features as high as 4 m and separated by 25 km a result of unusual snow
accumulation and redistribution processes influenced by the prevailing winds and
climate conditions appear more like cloud formations: they exhibit a rippled appearance. However, the Angular Signature Cloud Mask (ASCM), a MISR product
(Di Girolamo and Wilson, 2003) is able to detect clouds over snow and ice as well as
over ocean and land (Fig. 5.19). MISR imagery indicated that these mega-sastrugi
were stationary surface features between 2002 and 2004.
(a)
708 backward viewing camera
(b)
confidence high low low high

cloud
surface
Fig. 5.19 (a) December 16, 2004 MISR image over Antarctica, showing sastrugi (b) colorcoded image showing the Angular Signature Cloud Mask results (Courtesy of the MISR Team,
NASA/JPL/Caltech, and L. Di Girolamo and M.J. Wilson, University of Illinois.)
127
Hendriks and Pellikka (2004) presented a study of the multiangular reflectance

of a glacier surface. They used multiple ALTM (Optech, Ontario, Canada) digital
metric camera frames acquired from the air on August 12, 2003, to examine the
angular spectral reflectance response of snow, firn, and ice surfaces on the Hintereisferner glacier in Austria. The camera was operated in color infrared (CIR) mode,
resulting in images in three spectral bands: green (510600 nm), red (600720 nm)
and near-infrared (720800 nm); the pixel size varied between 0.250.30 m as a result of topography. They found important differences in the angular signatures of the
three surface types and compared their results to MISR imagery (local mode: nominal 275 m footprints at all angles and in all four bands). Their results showed that
MISR data reveal the backscattering of dirty ice, firn and old snow; BRF increases
going from the nadir image in the backward direction and at the viewing camera
Ca (60 ) backwards, BRF is 30% higher relative to the value in the nadir viewing
camera (An) (Fig. 5.20).
Multiangle views provide unique sub-pixel resolution information about the ice
sheet surface that can be used to improve the characterization of climate and ice dynamics processes. Nolin et al. (2002) had previously shown that MISR data can
be used as a proxy for surface roughness. They developed a normalized difference angular index (NDAI) using a combination of forward and backward scattered radiation in the MISR red band at the 60 fore and aft viewing angles of the
MISR instrument. A positive (negative) NDAI value indicates that backward (forward) scattering exceeds forward (backward) scattering and that the surface is rough
0.52
0.47
0.42
Firn/Old snow
BRF
0.37
0.32
0.27
Dirty ice
0.22
0.17
0.12
Cf
Bf
Af
An
Aa
Ba
Ca
60
45.6
26.1
26.1
45.6
60
View angle (degrees)
Fig. 5.20 MISR BRFs for dirty ice and firn/old snow. The green, red and NIR bands are shown
by the green, red and blue lines, respectively, for the Hintereisferner glacier in Austria, August 14,
2003 (Courtesy of J.P.M. Hendriks and P. Pellikka.)
128
M.J. Chopping
Fig. 5.21 Ice sheet surface classes for 20002005 derived from ISODATA classification of summer
NIR albedo and NDAI from MISR images acquired in the vicinity of the Jakobshavn glacier and
its upland drainage basin. Contour lines of elevation are at 250 m intervals (Reprinted from Nolin
and Payne, 2006. Copyright, Elsevier. With permission.)
(smooth). The NDAI was shown to be highly correlated with surface roughness derived from an airborne laser altimeter. Nolin and Payne (2006) used the ISODATA
clustering technique with NIR albedo and NDAI values from surface hemisphericaldirectional reflectance for consecutive MISR summer images from 20002006 over
an area in the vicinity of the Jakobshaven glacier in Greenland. The classified maps
(Fig. 5.21) demonstrated good spatial and temporal consistency for seven ice sheet
classes for all 6 years; moreover, the classes roughly correspond with glacier facies
mapped previously by other researchers. The classes differ in albedo, roughness (including the presence of crevasses), wetness, and the age of the snow at the surface.
Nolin (2004) performed a study using data from MISR to demonstrate how
the angular pattern of reflectance from vegetation over snow can provide information on forest cover density. This is important in snow studies as vegetation
structure and density affect the dynamics of snow accumulation and ablation and
affects the ability to estimate snow-covered area accurately from satellite-based
sensors. The study area was located in north-central Colorado. MISR red band
level 1B2 (top-of-atmosphere radiometrically and geometrically calibrated spectral
radiances) data from 15 February 2002 were converted to top-of-atmosphere bidirectional reflectance factors no atmospheric correction was applied. The Rahman
PintyVerstraete (RPV) semi-empirical parametric model was successfully used to
simulate the angular patterns of reflectance. The models k parameter was used to
characterize the angular signatures of selected pixels. In the RPV model this parameter is used to quantify the degree by which the observed bi-directional reflectance
129
2
71%
1.5
93%
Normalized BRF
99%
0.5
0
0.5
1
1.5
2
2.5
80
60
40
20
20
40
80
60
MISR camera angle

Fig. 5.22 MISR angular signatures for three forest cover densities of firspruce forest. Note the
change in shape with increasing tree cover; this is captured in the RPV models k parameter
(Reprinted from Nolin, 2004. Copyright, Wiley. With permission.)
factor data resemble a bowl- or bell-shaped pattern. The results showed distinct patterns in the retrieved k parameter values, with a marked dependence on density and
cover type. Non-forested areas exhibited a bowl-shaped pattern (k < 1.0) of reflectance versus viewing angle. Low-density deciduous forests also exhibited this
bowl-shaped reflectance pattern, changing as density increases. Other forest cover
types show transitional patterns between bowl and bell shapes and distinct bellshaped patterns (k > 1.0) for higher densities (Fig. 5.22). The relationship between
k and density does not hold for forest cover densities that approach 100%. For a
density of 99%, the fir spruce forest cover type has a distinct bowl shape and a k
value of only 0.69. This is in agreement with previous work indicating that sub-pixel
homogeneity (whether because of sparse vegetation cover or extremely dense vegetation cover) will result in k < 1.0. This study indicated from a qualitative standpoint
that multiangle reflectance data captures information on forest cover density at the
sub-pixel scale.
5.4.2 Dust Emissions

Laurent et al. (2005) used POLDER multiangle data with a surface roughness
parameter estimated from the Roujean kernel-driven model (Roujean et al., 1992).
They used the geometric kernel weight normalized by the diffuse (isotropic) scattering kernel weight (i.e., geo/iso) to simulate dust emissions from deserts in East Asia
130
M.J. Chopping
(N. China and Mongolia). The approach was derived from work previously accomplished by Roujean et al. (1997) and employed an empirical relationship established
by Marticorena et al. (2004). The composite surface roughness map they developed
describes the spatial variability of the erosion threshold (10 m wind velocities) on
dust emission frequency. The map was found to be consistent with geomorphologic
interpretations from Landsat imagery and with soil properties described in the literature. The retrieved roughness lengths are in agreement with the roughness lengths
experimentally determined over similar surface types in other deserts of the world.
The authors computed dust emission frequencies for 3 years (199719981999)
by combining the 10 m erosion threshold wind velocities, the European Centre for
Medium-Range Weather Forecasts (ECMWF) surface wind fields, the snow depth
and the soil moisture computed using the Food and Agriculture Organization of
the United Nations (FAO) soil texture profiles, and ECMWF meteorological data.
The simulated frequencies of significant dust emissions were compared to the frequencies of occurrence of Total Ozone Mapping Spectrometer (TOMS) Absorbing
Aerosol Index (AAI) higher than 0.7. Both the location and the relative intensity of
the highest dust emission frequencies identified from the simulations were in agreement with the observations (Fig. 5.23).
1
Simulated dust event frequency
0.8
0.6
0.4
0.2
0
0
0.2
0.4
0.6
0.8
Frequency of TOMS AAI > 0.7
Fig. 5.23 Monthly frequencies of significant simulated dust emissions (flux > 1010 g cm2 s1 )
as a function of the monthly frequencies of TOMS AAI > 0.7 over the Taklimakan desert for the
3 years 1997, 1998, and 1999. Small dots represent individual data; circles represent the averaged
frequency of simulated dust emissions for classes (5% width) of frequency of TOMS Absorbing
Aerosol Index (AAI) > 0.7; the solid line represent the linear fit of the averaged data (without
accounting for the last class which is not representative) (Reproduced from Laurent et al., 2005.
Copyright, American Geophysical Union.)
131
5.5 Some Considerations in Multiangle Acquisition

5.5.1 Near-simultaneous and Accumulated Sampling
Near-simultaneous acquisition of multiple observations of the same area on the
Earths surface is possible by viewing in the along-track direction. From low Earth
orbit, multiple looks of the same area can be acquired over a time span of less than
10 minutes. This strategy may be contrasted with accumulated sampling, in which
repeat multiangular coverage of a particular area is obtained with a wide cross-track
swath and observations are accumulated during multiple orbits. Accumulation of
multiangle data in this manner takes many days. This may degrade the angular data
set to some degree, depending on how rapidly surface conditions change and to
some extent on changes in the atmosphere, since it is more difficult to correct extreme conditions. For along-track systems, there will be delay of only a few minutes
between the first and last acquisitions, whether the instrument employs multiple
sensors (MISR), is tilted (CHRIS/Proba), has a conical scan (AATSR), or has a
very wide field-of-view (POLDER). This lag is unavoidable unless a constellation
of geostationary platforms is available an unlikely scenario for the near future.
Most Earth observing instruments for which the effects of the surface BRDF have
been studied and to a far lesser extent exploited are across-track instruments
on polar-orbiting instruments such as the AVHRR and MODIS. These systems are
designed to view large swaths, covering the globe at least once per 24 h period.
They are intrinsically off-nadir viewing devices but to be used as multiangle instruments, observations of the same land areas must be accumulated from multiple
overpasses. In addition, sensors such as Spinning Enhanced Visible and Infra-Red
Imager (SEVIRI) on Meteosat Second Generation (MSG) geostationary satellites
provide angular sampling through diurnal variation in illumination angles. The motivation for BRDF studies with accumulating instruments has been twofold: the need
to adjust observations to a common Sun-target-sensor geometry for consistency (a
much-touted but actually rather rare quality in Earth Observation in general); and to
obtain more accurate estimates of terrestrial albedo than could possibly be obtained
from a single view, since sunlight is never scattered in a perfectly diffuse manner at
the surface.
In the case of the MODIS BRDF/Albedo algorithm, an additional product is
Nadir-equivalent BRDF Adjusted Reflectance (NBAR), defined as the best estimate
of reflectance at nadir viewing at the mean solar zenith angle of the observations
used to invert the Li-Ross BRDF model (in this case LiSparseMODIS-RossThick).
NBAR was quickly found to provide superior performance in applications such as
land cover mapping (Zhang et al., 2000). The possibility of obtaining additional
surface information by exploiting the directional signal encapsulated in these data
was something of an afterthought until the 2000s; however, in the last 5 years this
has changed, with several studies addressing the retrieval of vegetation structural
parameters such as LAI and fractional vegetation coverage (Camacho-de Coca et al.,
2002; Chen et al., 2005).
132
M.J. Chopping
Accumulating instruments have one major advantage over those providing nearsimultaneous multiangular acquisitions: they provide greater spatial and temporal
coverage. This is particularly important for global applications where cloud and
cloud shadow contamination present a problem for instruments with revisit periods
of less than a few days. However, near-simultaneous multiangle imaging provides
some important advantages over accumulated sampling:
It greatly helps to solve the aerosol scattering problem over land and especially
over bright surfaces including deserts where spectral information alone currently appears to be inappropriate.
Directional data sets acquired by accumulation may afford serious limitations for
applications under those conditions where either the atmosphere or the surface
or both change more rapidly than the accumulation period. This advantage is
more applicable in some circumstances than in others and is probably more dependent on the development of the surface rather than that of the atmosphere.
For example, changes owing to snowfall, snowmelt, flood, and fire may not be
resolvable by accumulated sampling as they happen much more rapidly than the
sampling rate. Near-instantaneous imaging is therefore the only way to assess
surface directional properties under these circumstances. In other cases, for example, where changes in surface conditions are slow and/or sparse and intermittent, accumulated sampling is more acceptable.
There is the potential to assess information on canopy structure and retrieve LAI
and fPAR with greater accuracy (Hu et al., 2003).
5.5.2 Angular Sampling

The angular sampling of both multiangle imaging and accumulating instruments
is determined by sensor design (field-of-view/swath), the platform on which the
sensor is flown, and by the constraints imposed by the orbit into which the platform
is injected. The first and last of these are straightforward. However, that the platform
itself can impact the angular sampling regime of a sensor has a meaning restricted
to satellites that are able to re-orient themselves in order to acquire multiple looks.
Currently this definition applies to only one system: CHRIS on the Proba-1 satellite.
In the future, increasing use of constellations of small, agile satellites such as Proba
might be expected to provide greater flexibility in angular sampling. One such system concept developed at NASA/GSFC, dubbed Leonardo-BRDF has not yet been
pursued (Wiscombe, 2000). The angular sampling provided depends on whether
the instrument is an across-track scanner (using a revolving or oscillating mirror;
e.g., AVHRR, MODIS, the forthcoming US National Polar-orbiting Operational
Environmental Satellite System (NPOESS) Visible/Infrared Imager/Radiometer
Suite, or VIIRS); an along-track scanner ((A)ATSR(2)); multiple along-track
moderate swath CCD arrays (MISR); or an along-track wide swath CCD array
(POLDER/Parasol). Important issues with respect to angular sampling include:
133
The range of viewing and illumination zenith angles

The azimuthal coverage of the samples (how well the hemisphere is covered,
particularly with respect to the principal and perpendicular planes)
The angular sampling density, particularly around the hot spot (rarely achieved)
The spatial resolution, determined by the size of the solid angle of the individual
observing element, which is somewhat large for POLDER and much smaller for
MISR and CHRIS
The interval over which the sampling is acquired (minutes for true multiangle
imagers such as MISR, CHRIS, and POLDER; up to weeks for accumulators)
The frequency with which a given sampling may be repeated
Since most instruments are flown on polar-orbiting platforms with sun-synchronous
orbits having orbital inclinations only somewhat greater than 90 and acquisitions
are made either side of local noon, across-track scanners tend to provide a sampling
that is closer to the principal plane where variation in radiance with viewing and
illumination zenith is greatest for mid-latitudes. Along-track imagers on platforms
with sun-synchronous orbits such as MISR provide a sampling that is often far from
the principal plane but approaches it with increasing latitude; and for any given
latitude the sampling remains consistent. Owing to the Proba-1 orbit and the narrow
field-of-view, CHRIS provides a less regular sampling from site to site and it is only
occasionally able to image a target area from directly overhead view; Proba-1 has
to be tilted at some usually small angle in the across-track direction so that the
target area is viewed. The platform acquires images of the target when the zenith
angle of the platform is one of the following: 55 , 36 and 0 . This means
that the angles at which images are acquired vary from site-to-site, depending on
their positions with respect to the orbital track. The pattern with which any site is
accessible to the platform varies at roughly 8-day intervals, but with some change
in sampling because the orbit does not repeat exactly (Barnsley et al., 2004). The
POLDER design allows the most comprehensive sampling of the Earths radiation
field from space to date. The very wide field-of-view (2, 400 1, 800 km2 ) allows
observation of the same target under many different angular configurations (between 10 and 15 observations for each passage of the satellite), with view zenith
angles up to 60 for the full azimuth range (Laurent et al., 2005).
5.5.3 Scale and Multiangle Observation

The spatial scale at which multiangle observations over land are made has an important impact on the kinds of applications in which they might be used, as well as
on data transmission rates and geographic coverage. Progress in remote sensing has
often been measured in terms of increasing the sampling rate in the spatial domain,
so that the smaller the ground sampling distance and the higher the spatial resolution,
the better (NASA, 2000). Indeed, this is often postulated to be the major limitation of
moderate resolution remote sensing devices: while wall-to-wall coverage is provided,
heterogeneity in surface features cannot be resolved, limiting many applications,
134
M.J. Chopping
including cover type classification and the assessment of small-scale but widespread
fragmentation in a variety of ecosystems (Phinn et al., 1996). However, this paradigm is flawed in certain respects: although high resolution imagers show the details
of land features, this comes at the price of high data volumes monetary expense,
and restricted coverage. Although multi-angle observation can only provide statistical metrics on sub-pixel structure, this is often sufficient to address the problem of
interest. Moreover, data volumes are smaller and global coverage is achievable. For
example, an hypothetical 10 m resolution single-angle imager generates ten times
more data than a 100 m multiangle imager having ten angles, for the same swath; it
can therefore cover a swath ten times wider for the same data volume.
Note too that with multiangle remote sensing a small ground-projected instantaneous field-of-view (GIFOV) can be a disadvantage; specifically, for the radiometric, or 1-D, approach that seeks to derive sub-pixel information via variation in
radiance with Sun and/or viewing angles and in which each pixel is treated independently. In this approach, if the GIFOV is small in relation to surface features
such as trees or shrubs, there is a higher probability that an observation will include a
sampling of elements that is unrepresentative of the surface as a whole; furthermore,
it may also include important contributions from partial features at the extremities. These may result in noise-like fluctuations and adjacency effects in the angular
signal that will impact on models that treat the surface as either a semi-infinite,
homogenous, planeparallel medium, or as a set of identical objects distributed spatially in a Poisson distribution (e.g., geometric-optical models). In both cases, an
unrepresentative sampling of features and/or a large relative contribution from partial features at the edge of the GIFOV will obscure the angular signal. If the GIFOV
is large in relation to surface features then the sampling of surface elements will
be more representative and the periphery will make a relatively small contribution.
Other considerations are that a large GIFOV may result in greater proportions of observations with unavoidable cloud and cloud shadow contamination and with mixtures of too many surface types. On the other hand, a very small GIFOV may also
suffer from multiple scattering to/from adjacent footprints (pixel cross talking).
Clearly, for any given landscape there is an optimum sampling resolution: if the
GIFOV is much larger than the typical length scales then spatial trends in BRDF and
parameters derived from model inversions will be less well-defined. There is also a
sampling resolution that is the global optimum; i.e., over all surface types of interest.
Following Pinty et al. (2002, 2004), Widlowski et al. (2005) addressed the question
of the degree to which 1-D radiative transfer models can explain as well as describe
the reflectance fields of 3-D forest targets of varying composition and complexity,
over a range of spatial sampling scales (sensor ground sampling resolutions). Explaining these reflectance fields requires that the models internal parameters (state
variables) match and are consistent with those of the 3-D target; it is not sufficient
that the modeled anisotropy matches that of the target. Both conditions must hold if
any model is to provide surface parameters on inversion which have meaning and
utility in remote sensing applications; if they are not, there is considerable doubt
on the reliability of surface parameters retrieved. The study tackled this problem
through addressing two important questions:
135
Can 1-D models represent the anisotropy of a reflectance field generated by a

more complex 3-D model at multiple spatial resolutions? and
Will the inversion of a 1-D model (at low spatial resolution where they are more
appropriate) lead to systematic (and significant) biases in the retrieval of surface
properties?
These questions were pursued by using data sets for a level Scots pine landscape
simulated by the Rayspread 3-D Monte Carlo raytracing model, using allometric
relationships for a range of canopy parameter configurations in the red and near
infra-red wavelengths. A 1-D model known as the Semi-discrete model (Gobron
et al., 1997) was used to generate a large series of lookup tables for the comparisons.
The study examined the impact of changing sensor GIFOV and foliage clumping on
the match between the 3-D and 1-D models, as well as differences in interception,
absorption and transmittance of solar radiation through and by the canopy arising
from the 1-D and 3-D modeling. The conclusions were that 1-D models can provide
good matches to reflectance fields generated by 3-D models in terms of both magnitude and directionality but that it does not follow that the 1-D models internal
parameters and radiation budget match those of the 3-D target. It was shown that
differences in the shape of the reflectance anisotropy between a 3-D target and its
1-D homologue featuring identical spectral and structural properties with the exception of foliage clumping lead to the retrieval of BRF-equivalent 1-D solutions
with parameter values that diverge from those of the 3-D target. Part of this problem
is understood as the problem of lack of uniqueness of solutions. A key finding is that
for conifer forest canopies, an observation spatial scale of less than 100 m is more
likely to introduce large discrepancies between the reflectance fields generated by
3-D and 1-D models for a wide range of conditions.
5.6 Conclusions
This chapter has reviewed recent work towards exploiting solar wavelength remote sensing data acquired at multiple viewing angles from the air and space
in applications in forestry, ecology, land cover mapping and land cover change,
agriculture, hydrology, and glaciology, with important implications for enhancing
ecological, crop growth, biogeochemical, hydrologic, and energy balance models.
Even when excluding the numerous studies concerned with goniometric measurements, multiangle observations in the thermal wavelengths, and model development
and simulation work, there has clearly been a great deal of activity over the last
few years. Advances have been seen in a diversity of approaches, with notable gains
for synergistic methods that seek to use multiangle data together with other kinds
of remote sensing data and/or high resolution inventory data. The synergistic use
of multiangle data with those from lidar instruments has provided an especially
promising direction for the future, especially if the capability of models such as
GORT to link multiangle and lidar data can be further developed. While canopy reflectance models have become more accurate, often incorporating a wider range of
136
M.J. Chopping
spectral measures, the hindrances to the use of geometric-optical models that have
prevented widespread adoption in applications are being removed through methods
for isolating the contribution of the soil/understory background from that of the upper canopy. Studies directed at assessing the gains from adopting the multiangle
approach over nadir-spectral sensing, as well as obtaining meaningful and consistent metrics that characterize surface and canopy conditions, have universally found
multiangle data to be valuable.
Although efforts to exploit the additional, unique information available through
multiangle remote sensing have been ongoing for many years, progress towards
applications is accelerating, with high quality data sets available from MISR and
MODIS on NASAs EOS satellites, a third POLDER in orbit on Parasol, innovative
new experimental systems such as CHRIS on Proba, greater demonstrated potential
for satisfying multiple user groups in diverse disciplines (e.g., atmospheric science,
ecology, glaciology, land management, and climate modeling), the realization of
important synergies with active instruments, and a greater number of researchers
engaged. This is borne out by the tenfold increase in the number of peer-reviewed
publications using data from MISR, POLDER or CHRIS in the 10 years to 2005
(Fig. 5.24). Canopy reflectance modeling work with a variety of model types and
improved reference data from high resolution imaging and lidar continues to shed
new light on the constraints to robust retrievals of biophysical parameters and on
important issues such as optimal scales of observation. Efforts continue to be made
to engage remote sensing scientists with user groups (e.g., Chopping and Diner,
2005), although this is an ongoing task. Continued improvements in knowledge and
understanding, together with greater experience with both near-simultaneous and
70
POLDER-1 + POLDER-2 + PARASOL
60
MISR
Number of publications
CHRIS
50
40
30
20
10
0
1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005
Year
Fig. 5.24 Growth in peer-reviewed publications focused on exploitation of POLDER, MISR, and
CHRIS, 19902005 (Courtesy of David J. Diner, MISR scientist, NASA/JPL.)
137
accumulating multiangle instruments, will drive the design of future Earth observation systems. The examples presented in this chapter provide a good, albeit not exhaustive, indication of the range and depth of work that has been accomplished over
recent years in just one area of multiangle remote sensing (terrestrial applications
in the solar wavelengths). They reflect the progress achieved in the ten years since
the NASA-sponsored Workshop on Multi-angular Remote Sensing for Environmental Applications (Privette et al., 1997) and point the way to further opportunities for
making better use of the unique information provided by multiangle data over land.
Acknowledgements Special thanks are owed to David J. Diner for his very valuable suggestions on the manuscript as well as the many others who helped me in compiling this review;
and in particular Daniel Kimes, Alan Strahler, Bernard Pinty, Jean-Louis Roujean, Francois-Marie
Breon, David Jupp, Jon Ranson, Jing Chen, John V. Martonchik, Anne Nolin, Rob Braswell, Julian
Jenkins, Mike Barnsley, Wout Verhoef, Wenge Ni-Meister, Andres Kuusk, Soeren Hese, Hamlyn
Jones, Raffaele Casa, F. Javier Garca Haro, Johan Hendriks, Petri Pellikka, Janne Heiskanen,
Larry Di Girolamo, Mathias Disney, Wolfgang Lucht, Sampo Smolander, Gunar Fedosejevs,
Mike Cutter, Gabriela Schaepman, and Michael Schaepman. I thank also the participants in the
NASA/MISR Workshop on Multiangle Remote Sensing in Ecological Modeling not mentioned
above. I must also acknowledge the many people not named here who sent me their recent and
often unpublished work, and especially those whose work I could not incorporate here: I am deeply
grateful. Any errors or omissions are uniquely mine.
Glossary
1-D
3-D
AAI
AATSR
ADEOS
ARS
ASCM
AVHRR
BHR
BOREAS
BRDF
BRF
Cab
CART
CCD
CHRIS
CI
CIR
DISMA
ECMWF
EOS
One-dimensional
Three-dimensional
Absorbing Aerosol Index
Advanced Along-Track Scanning Radiometer
Advanced Earth Observation Satellite (Japan)
Agricultural Research Service (USDA)
Angular Signature Cloud Mask
Advanced Very High Resolution Radiometer (NOAA)
Bihemispherical Reflectance
Boreal Ecosystem-Atmosphere Study (NASA)
Bidirectional Reflectance Distribution Function
Bidirectional Reflectance Factor
Chlorophyll a + b content (leaves in a given canopy)
Canopy Architecture Radiative Transfer (MISR LAI algorithm)
Charge Coupled Device
Compact High Resolution Imaging Spectrometer
Clumping Index
Color Infrared
Directional Spectral Mixture Analysis
European Center for Medium range Weather Forecasting
Earth Observing System (NASA)
138
FAO
fAPAR
GIFOV
GLAS
GO
GORT
HDRF
ICESat
IFOV
ISRO
LAD
LAI
Lidar
LiSK
LVIS
MISR
MODIS
MRPV
MSG
MVC
NADIM
NBAR
NDAI
NDVI
NIR
NDHD
NPOESS
POLDER
POVRAY
PSLV
RAMI
RMSE
RPV
RT
SAIL
SAR
SEVIRI
SGM
SSI
TOMS
UK
USDA
VCF
VIIRS
M.J. Chopping
Food and Agriculture Organization of the United Nations

fPAR Fraction of Absorbed Photosynthetically-Active Radiation
Ground-projected (sensor) Instantaneous Field-Of-View
Geoscience Laser Altimeter System
Geometric-Optical (model)
Geometic-Optical/Radiative Transfer (model)
Hemispherical-Directional Reflectance Factor
Ice, Cloud, and Land Elevation Satellite (NASA EOS)
(sensor) Instantaneous Field-Of-View
Indian Space Research Organization
Leaf Angle Distribution
Leaf Area Index (one-sided leaf area per unit ground area)
Light Detection and Ranging (sensor)
Linear, Semi-empirical, Kernel-driven
Laser Vegetation Imaging Sensor
Multiangle Imaging Spectro-Radiometer (NASA/JPL EOS)
MODerate resolution Imaging Spectroradiometer (NASA EOS)
Modified Rahman-Pinty-Verstraete (model)
Meteosat Second Generation
Maximum-Value Compositing (on NDVI)
New Advanced Discrete Model (model)
Nadir BRDF-Adjusted Reflectances (EOS MOD43B4 product)
Normalized Difference Angular Index
Normalized Difference Vegetation Index
Near Infra-Red (sometimes written near-infrared)
Normalized Difference between Hotspot and Darkspot
National Polar-orbiting Operational Environmental Satellite System
Polarization and Directionality of the Earths Reflectance
Persistence of Vision Raytracer
Polar Satellite Launch Vehicle (ISRO; India)
Radiation Transfer Model Intercomparison (RAMI) Exercise
Root Mean Square Error
Rahman-Pinty-Verstraete (model)
Radiative Transfer
Scattering by Abitrarily Inclined Leaves (model)
Synthetic Aperture Radar
Spinning Enhanced Visible and Infra-Red Imager
Simplified Geometric-optical Model
Structural Scattering Index
Total Ozone Mapping Spectrometer
United Kingdom of Great Britain and Northern Ireland
United States Department of Agriculture
Vegetation Continuous Fields (MODIS product)
Visible/Infrared Imager/Radiometer Suite (NPOESS)
139
References
Abuelgasim AA, Gopal S, Irons JR, Strahler AH (1996) Classification of ASAS multi-angle
and multispectral measurements using artificial neural networks. Remote Sens. Environ.
57(2):7987
Asner GP (2000) Contributions of multi-view angle remote sensing to land surface and biogeochemical research. Remote Sens. Rev. 18:137162
Bach H, Begiebing S, Verhoef W (2005) Analyses of hyperspectral and directional CHRIS data for
monitoring of agricultural crops using the canopy reflectance model SLC. Proc. 9th Int. Symp.
on Physical Measurements and Signature in Remote Sens. (ISPMSRS 2005), vol. I, Int. Soc.
Photogrammetry and Remote Sens., 5355
Bacour C, Breon F-M (2005) Variability of biome reflectance directional signatures as seen by
POLDER. Remote Sens. Environ. 98:8095
Bacour C, Jacquemoud S, Leroy M, Hautecoeur O, Weiss M, Prevot L, Bruguier N, Chauki H
(2002) Reliability of the estimation of vegetation characteristics by inversion of three canopy
reflectance models on airborne POLDER data. Agronomie 22:555565
Barnsley MJ, Settle JJ, Cutter MA, Lobb DR, Teston F (2004) The PROBA/CHRIS mission: a lowcost smallsat for hyperspectral multiangle observations of the Earth surface and atmosphere.
Bicheron P, Leroy M, Hautecoeur O, Breon F-M (1997) Enhanced discrimination of boreal forest
covers with directional reflectances from the airborne POLDER instrument. J. Geophys. Res.
102(D24):2951729528
Breon F-M, Maignan F, Leroy M, Grant I (2002) Analysis of hot spot directional signatures measured from space. J. Geophys. Res. 107(D16):10.1029/2001JD001094
Bruegge CJ, Chrien NL, Ando RR, Diner DJ, Abdou WA, Helmlinger MC, Pilorz SH, Thome
KJ (2002) Early validation of the Multi-angle Imaging SpectroRadiometer (MISR) radiometric
scale. IEEE Trans. Geosci. Remote Sens. 40(7):14771492
Camacho-de Coca F, Garca-Haro FJ, Martnez B, Melia J (2002) An operational strategy for
retrieval of vegetation products from SEVIRI and AVHRR-3 data. Proc. LSA SAF Training
Workshop, Ed. EUMETSAT, Darmstadt, Lisbon, July 810, pp 110
Camacho-de Coca F, Breon FM, Leroy M, Garcia-Haro FJ (2004), Airborne measurement of hot
spot reflectance signatures. Remote Sens. Environ. 90:6375
Canisius F, and Chen JM (2007) Retrieving vegetation background reflectance from Multi-angle
Imaging SpectroRadiometer (MISR) data. Remote Sens. Environ. 107(12):312321
Casa R, Jones HG (2005) LAI retrieval from multiangular image classification and inversion of a
ray tracing model. Remote Sens. Environ. 98:414428
Chen JM, Liu J, Leblanc SG, Lacaze R, Roujean J-L (2003) Multi-angular optical remote sensing
for assessing vegetation structure and carbon absorption. Remote Sens. Environ. 84:516525
Chen JM, Menges CH, Leblanc SG (2005) Global mapping of foliage clumping index using multiangular satellite data. Remote Sens. Environ. 97:447457
Chopping MJ (2006) Progress in retrieving canopy structure parameters from NASA multi-angle
remote sensing. Proc. Int. Geosci. Remote Sens. Symp. 2006 and 27th Canadian Remote Sens.
Symp., Denver, CO, July 31August 4, 2006
Chopping M, Diner D (2005) Workshop on Ecological Modeling using NASA Multi-angle Remote
Sensing. The Earth Observer, NASA/GSFC, NovemberDecember 2005, 17(6):3234
Chopping MJ, Rango A, Ritchie, JC (2002) Improved semi-arid community type differentiation
with the NOAA AVHRR via exploitation of the directional signal. IEEE Trans. Geosci. Remote
Sens. 40(5):11321149
Chopping MJ, Rango A, Havstad KM, Schiebe FR, Ritchie JC, Schmugge TJ, French A, Su L,
McKee L, Davis RM (2003) Canopy attributes of Chihuahuan Desert grassland and transition communities derived from multi-angular airborne imagery. Remote Sens. Environ.
85(3):339354
140
M.J. Chopping
Chopping M, Martonchik JVM, Rango A, Peters DPC, Su L, Laliberte A (2005) Geometric-optical

modeling of desert grassland canopy structure with MISR. Proc. 9th Int. Symp. on Physical
Measurements and Signature in Remote Sens. (ISPMSRS 2005), vol. I, Int. Soc. Photogrammetry and Remote Sens. 141143
Chopping M, Su L, Rango A, Martonchik JV, Peters DPC, Laliberte A (2006a) Remote sensing
of woody shrub cover in desert grasslands using MISR with a geometric-optical canopy reflectance model, ISPMSRS05. Special Issue of Remote Sens. Environ.
Chopping M, Su L, Laliberte A, Rango A, Peters DPC, Kollikathara N (2006b) Mapping shrub
abundance in desert grasslands using geometric-optical modeling and multiangle remote sensing with CHRIS/Proba. Remote Sens. Environ. 104(1):6273
Chopping, MJ, Su L, Laliberte A, Rango A, Peters DPC, Martonchik JV (2006c) Mapping woody
plant cover in desert grasslands using canopy reflectance modeling and MISR data. Geophys.
Res. Lett., 33, L17402, doi:10.1029/2006GL027148
Chopping M, Moisen G, Su L, Laliberte A, Rango A, Martonchik JV, and Peters DPC (2007) Large
area mapping of southwestern forest crown cover, canopy height, and biomass using the NASA
Multiangle Imaging SpectroRadiometer, Remote Sens. Environ, in press.
Combal B, Baret F, Weiss M (2002) Improving canopy variables estimation from remote sensing data by exploiting ancillary information. Case study on sugar beet canopies. Agronomie
22:205215
Di Girolamo L, Wilson MJ (2003) A first look at band-differenced angular signatures for cloud
detection from MISR. IEEE Trans. Geosci. Remote Sens. 41(7):17301734
Diner DJ, Beckert JC, Reilly TH, Bruegge CJ, Conel JE, Kahn RA, Martonchik JV, Ackerman TP,
Davies R, Gerstl SAW, Gordon HR, Muller J-P, Myneni RB, Sellers PJ, Pinty B, Verstraete MM
(1998) Multi-angle Imaging SpectroRadiometer (MISR) instrument description and experiment
overview. IEEE Trans. Geosci. Remote Sens. 36(4):10721087
Diner DJ, Asner GP, Davies R, Knyazikhin Y, Muller J-P, Nolin AW, Pinty B, Schaaf CB, Stroeve J
(1999) New directions in Earth observing: scientific applications of multi-angle remote sensing.
Bull. Am. Meteorol. Soc. 80(11):22092229
Diner DJ, Braswell BH, Davies R, Gobron N, Hu J, Jin Y, Kahn RA, Knyazikhin Y, Loeb N, Muller
J-P, Nolin AW, Pinty B, Schaaf CB, Seiz G, Stroeve J (2005) The value of multiangle measurements for retrieving structurally and radiatively consistent properties of clouds, aerosols, and
surfaces. Remote Sens. Environ. 97:495518
Disney M, Lewis P, Saich P (2006) 3D modelling of forest canopy structure for remote sensing
simulations in the optical and microwave domains. Remote Sens. Environ. 100:114132
Engelsen O, Pinty B, Verstraete MM, Martonchik JV (1996) Parametric bidirectional reflectance
factor models: evaluation improvements and applications. Report EUR16426EN European
Commission Joint Research Centre Space Applications Institute, Ispra, Italy, 120pp
European Space Agency (1999) Exploitation of CHRIS data from the Proba mission for science
and applications. Experimenters Handbook Issue 4: Baseline Programme, Scientific Campaign
Unit, ESA-ESTEC
European Space Agency (2006) Achievements of the Proba mission: projects and publications at:
http://earth.esa.int/pub/ESA DOC/PROBA/ProjectsandPublications.html
Gao F, Schaaf CB, Strahler AH, Jin Y, Li X (2003) Detecting vegetation structure using a kernelbased BRDF model. Remote Sens. Environ. 86:198205
Garca-Haro FJ, Camacho-de Coca F, Melia J (2006) A directional spectral mixture analysis
method. Application to multiangular airborne measurements. IEEE Trans. Geosci. Remote
Sens. 44:365377
Gemmell F (2000) Testing the utility of multi-angle spectral data for reducing the effects of
background spectral variations in forest reflectance model inversion. Remote Sens. Environ.
72:4663
Gobron N, Pinty B, Verstraete MM, Govaerts Y (1997) A semi-discrete model for the scattering of
light by vegetation. J. Geophys. Res. 102:94319446
141
Gobron N, Pinty B, Verstraete MM, Martonchik JV, Knyazikhin Y, Diner DJ (2000) Potential of
multiangular spectral measurements to characterize land surfaces: conceptual approach and
exploratory application. J. Geophys. Res. 105(D13):1753917549, 10.1029/2000JD900154
Gobron N, Pinty B, Verstraete MM, Widlowski J-L, Diner DJ (2002) Uniqueness of multiangular
measurements part 2: joint retrieval of vegetation structure and photosynthetic activity from
MISR. IEEE Trans. Geosci. Remote Sens. 40:15741592
Hansen MC, DeFries RS, Townshend JRG, Sohlberg RA, Dimiceli C, Carroll M (2002) Towards an
operational MODIS continuous field of percent tree cover algorithm: examples using AVHRR
and MODIS data. Remote Sens. Environ. 83:303319
Heiskanen J (2006) Tree cover and height estimation in the Fennoscandian tundrataiga transition
zone using multiangular MISR data. Remote Sens. Environ. doi:10.1016/j.rse.2006.03.015
Hendriks JPM, Pellikka P (2004) Multiangular reflectance of a glacier surface. Presentation at the
2004 Joint Assembly (CGU, AGU, SEG & EEGS), Montreal, Canada, May 1723, 2004
Hese S, Lucht W, Schmullius C, Barnsley M, Dubayah R, Knorr D, Neumann K, Riedel T, Schroter
K (2005) Global biomass mapping for an improved understanding of the CO2 balance the
Earth observation mission Carbon-3D. Remote Sens. Environ. 94:94104
Hu J, Tan B, Shabanov N, Crean KA, Martonchik JV, Diner DJ, Knyazikhin Y, Myneni RB (2003)
Performance of the MISR LAI and FPAR algorithm: a case study in Africa. Remote Sens.
Environ. 88:324340
Hyman AH, and Barnsley MJ (1997) On the potential for land cover mapping from multiple view
angle (MVA) remotely-sensed images. In: Observations and Interactions: Proc. 23rd Annu.
Conf. Exhibition Remote Sens. Soc., 135140
Jenkins J, Ollinger S, Braswell R, Martin M, Plourde L, Smith M-L, et al. (2004) Detecting patterns
of canopy structure and carbon uptake with multi-angle remote sensing. Presentation at MISR
Science Team Meeting, Pasadena, CA, December 9, 2004
Kimes DS, Ranson KJ, Sun G, Blair JB (2006) Predicting lidar measured forest vertical structure
from multi-angle spectral data. Remote Sens. Environ. 100(4):503511
King MD, Kaufman YJ, Tanre D, Nakajima T (1999) Remote sensing of tropospheric aerosols
from space: past, present, and future. Bull. Am. Meteorol. Soc. 80(11):22292260
Knyazikhin Y, Kranigk J, Myneni RB, Panferov O, Gravenhorst G (1998a). Influence of smallscale structure on radiative transfer and photosynthesis in vegetation cover. J. Geophys. Res.,
D: Atmospheres 103:61336144
Knyazikhin Y, Martonchik JV, Myneni RB, Diner DJ, Running SW (1998b) Synergistic algorithm
for estimating vegetation canopy leaf area index and fraction of absorbed photosynthetically
active radiation from MODIS and MISR data. J. Geophys. Res. 103(D24):3225732276
Knyazikhin Y, Martonchik JV, Diner DJ, Myneni RB, Verstraete M, Pinty B, Gobron N (1998c)
Estimation of vegetation canopy leaf area index and fraction of absorbed photosynthetically active radiation from atmosphere-corrected MISR data. J. Geophys. Res. 103(D24):3223932256
Koetz B, Kneubuhler M, Widlowski J-L, Morsdorf F, Schaepman M, Itten K (2005a) Assessment of canopy structure and heterogeneity from multi-angular CHRIS-PROBA data. 9th Int.
Symp.on Physical Measurements and Signatures in Remote Sens. (ISPMSRS), Beijing, China,
2005, 7678
Koetz B, Kneubuhler M, Schopfer J, Schaepman M, Itten K (2005b) Influence of changing background on CHRIS/Proba data over an heterogeneous canopy. Proc. EARSeL 4th Workshop on
Imaging Spectroscopy, Warsaw, Poland, 19
Lacaze R, Chen JM, Roujean J-L, Leblanc SG (2002) Retrieval of vegetation clumping index using
hot spot signatures measured by POLDER instrument. Remote Sens. Environ. 79:8495
Laurent B, Marticorena B, Bergametti G, Chazette P, Maignan F, Schmechtig C (2005) Simulation
of the mineral dust emission frequencies from desert areas of China and Mongolia using an
aerodynamic roughness length map derived from the POLDER/ADEOS 1 surface products. J.
Geophys. Res. 110(D18):D18S04, doi:10.1029/2004JD005013
Lavergne T, Kaminski T, Pinty B, Taberner M, Gobron N, Verstraete M, Vossbeck M, Widlowski
J-L, Giering R (2007) Application to MISR land products of an RPV model inversion package
using adjoint and Hessian codes. Remote Sens. Environ. 107(12):362375
142
M.J. Chopping
Li X, Strahler AH (1985) Geometric-optical modeling of a conifer forest canopy. IEEE Trans.

Li X, Strahler AH (1992) Geometric-optical bidirectional reflectance modeling of the discrete
crown vegetation canopy: effect of crown shape and mutual shadowing. IEEE Trans. Geosci.
Remote Sens. 30(2):276291
Li X, Strahler AH, Woodcock CE (1995) A hybrid geometric optical-radiative transfer approach
for modeling albedo and directional reflectance of discontinuous canopies. IEEE Trans. Geosci.
Remote Sens. 33:466480
Liang S (2004) Quantitative Remote Sensing of Land Surfaces. Wiley, Hoboken, NJ, 534 pp
Liang S, Strahler AH, Barnsley MJ, Borel CC, Gerstl SAW, Diner DJ, Prata AJ, Walthall CL (2000)
Multiangle remote sensing: past, present and future. In:, Liang S and Strahler AH (eds), The
Special Issue on Land Surface Bidirectional Reflectance Distribution Function (BRDF); Recent
Advances and Future Prospects, Remote Sens. Rev. 18:83102
Liu X, Kafatos M (2007) MISR multi-angular spectral remote sensing for temperate forest mapping
at 1.1-km resolution. International Journal of Remote Sensing 28(2):459464
Maignan F, Breon F-M, and Lacaze R (2004) Bidirectional reflectance of Earth targets: evaluation
of analytical models using a large set of spaceborne measurements with emphasis on the hot
spot. Remote Sens. Environ. 90:210220
Marticorena B, Chazette P, Bergametti G, Dulac F, Legrand M (2004) Mapping the aerodynamic
roughness length of desert surfaces from the POLDER/ADEOS bi-directional reflectance product. Int. J. Remote Sens. 25:603626
Martonchik JV, Bruegge CJ, Strahler A (2000) A review of reflectance nomenclature used in remote sensing. Remote Sens. Rev. 19:920
NASA (2000) NASA Earth Science Enterprise Research Strategy for 20002010
Ni W, Li X (2000) A coupled vegetation soil bidirectional reflectance model for a semi-arid
landscape. Remote Sens. Environ. 74:113124
Ni-Meister W (2005) 3D vegetation structure extraction from lidar remote sensing. Proc. 9th Int.
Symp. on Physical Measurements and Signature in Remote Sens. (ISPMSRS 2005), vol. I, Int.
Soc. Photogrammetry and Remote Sens. 4749
Nicodemus FE, Richmond JC, Hsia JJ, Ginsberg IW, Limperis T (1977) Geometrical Considerations and Nomenclature for Reflectance. In: National Bureau of Standards, US Department of
Commerce, Washington, DC
Nolin AW (2004) Towards retrieval of forest cover density over snow from the Multi-angle Imaging
SpectroRadiometer (MISR). Hydrol. Process. 18:36233636
Nolin AW, Payne M (2006) Classification of glacier zones in Western Greenland using albedo and
surface roughness from the Multi-angle Imaging SpectroRadiometer (MISR). Remote Sens.
Environ. 107(12):264275
Nolin AW, Fetterer FM, Scambos TA (2002) Surface roughness characterizations of sea ice and ice
sheets: case studies with MISR data. IEEE Trans. Geosci. Remote Sens. 40(7):16051615
Phinn S, Franklin J, Hope A, Stow D, Huenneke L (1996) Biomass distribution mapping using
airborne digital video imagery and spatial statistics in a semi-arid environment. J. Environ.
Manage. 47:139164
Pinty B, Gobron N, Widlowski J-L, Gerstl SAW, Verstraete MM, Antunes M, Bacour C, Gascon F,
Gastellu J-P, Goel N, Jacquemoud S, North P, Qin W, Thompson R (2001) Radiation Transfer
Model Intercomparison (RAMI) exercise. J. Geophys. Res. 106:1193711956
Pinty B, Widlowski J-L, Gobron N, Verstraete MM, Diner DJ (2002) Uniqueness of multiangular
measurements Part I: an indicator of subpixel surface heterogeneity from MISR. IEEE Trans.
Pinty B, Widlowski J-L, Taberner M, Gobron N, Verstraete MM, Disney M, Gascon F, Gastellu
J-P, Jiang L, Kuusk A, Lewis P, Li X, Ni-Meister W, Nilson T, North P, Qin W, Su L, Tang S,
Thompson R, Verhoef W, Wang H, Wang J, Yan G, Zang H (2004a) RAdiation transfer
Model Intercomparison (RAMI) exercise: results from the second phase. J. Geophys. Res. 109,
D06210 101029/2003JD004252
143
Pinty B, Gobron N, Widlowski J-L, Lavergne T, Verstraete MM (2004b) Synergy between 1-D
and 3-D radiation transfer models to retrieve vegetation canopy properties from remote sensing
data. J. Geophys. Res. 109, D21205
Privette JL, Deering D, Wickland D (1997) Report on the workshop on multiangular remote sensing for environmental applications. NASA Technical Memorandum 113202, 54 pp.
Rahman H, Pinty B, Verstraete MM (1993) Coupled surface-atmosphere reflectance (CSAR)
model. 2. Semiempirical surface model usable with NOAA Advanced Very High Resolution
Radiometer data. J. Geophys. Res. 98:2079120801
Roujean J-L, Leroy M, Deschamps P-Y (1992) A bidirectional reflectance model of the Earths
surface for the correction of remote sensing data. J. Geophys. Res. 97:2045520468
Roujean J-L, Tanre D, Breon F-M, Deuze J-L (1997) Retrieval of land surface parameters from airborne POLDER bidirectional reflectance distribution function during HAPEX-Sahel. J. Geophys. Res. 102:1120111218
Sandmeier S, Deering DW (1999) Structure analysis and classification of boreal forests using airborne hyperspectral BRDF data from ASAS. Remote Sens. Environ. 69(3):281295
Scarth P, Phinn S (2000) Determining forest structural attributes using an inverted geometricoptical model in mixed eucalypt forests, Southeast Queensland, Australia. Remote Sens. Environ. 71:141157
Schaepman M (2006) Spectrodirectional remote sensing: from pixels to processes. Int. J. Appl.
Earth Obs. Geoinformation 6(34):271282
Schaepman-Strub G, Schaepman M, Painter T, Dangel S, Martonchik J (2006) Reflectance
quantities in optical remote sensing definitions and case studies. Remote Sens. Environ.
103(1):2742
Smolander S, Stenberg P (2005) Simple parameterizations of the radiation budget of uniform
broadleaved and coniferous canopies. Remote Sens. Environ. 94(3):355363
Strahler AH, Jupp DLB, Woodcock CE, Li X (2005) The discrete-object scene model and its
application in remote sensing. Proc. 9th International Symp. on Physical Measurements and
Signatures in Remote Sens. 1:166168
Su L, Chopping MJ, Rango A, Martonchik JV, Peters DPC (2007) Support vector machines for
recognition of semi-arid vegetation types using MISR multi-angle imagery. Remote Sens.
Environ. 107(12):299311
Verhoef W (2005) Earth observation model sensitivity analysis to assess mission performances
in terms of geo-biophysical variable retrieval accuracies. Proc. 9th Int. Symp. on Physical
Measurements and Signature in Remote Sens. (ISPMSRS 2005), Beijing, October 1719, 2005,
vol. I, Int. Soc. Photogrammetry and Remote Sens. 324327
Wanner W, Li X, Strahler AH (1995) On the derivation of kernels for kernel-driven models of
bidirectional reflectance. J. Geophys. Res. 100:2107721090
Weiss M, Baret F, Myneni R, Pragnere A, Knyazikhin Y (2000) Investigation of a model inversion
technique for the estimation of crop characteristics from spectral and directional reflectance
data. Agronomie 20:322
Widlowski J-L, Pinty B, Gobron N, Verstraete MM, Davis AB (2001) Characterization of
surface heterogeneity detected at the MISR/TERRA subpixel scale. Geophys. Res. Lett.
28(24):46394642
Widlowski J-L, Pinty B, Gobron N, Verstraete M, Diner DJ, Davis B (2004) Canopy structure parameters derived from multi-angular remote sensing data for terrestrial carbon studies. Climatic
Change 67:403415
Widlowski J-L, Pinty B, Lavergne T, Verstraete MM, Gobron N (2005) Using 1-D models to
interpret the reflectance anisotropy of 3-D canopy targets: issues and caveats. IEEE Trans.
Geosci. Remote Sens. 43:20082017, doi10.1109/TGRS.2005.853718
Widlowski J-L, Taberner M, Pinty B, Bruniquel-Pinel V, Disney M, Fernandes R, GastelluEtchegorry J-P, Gobron N, Kuusk A, Lavergne T, Leblanc S, Lewis P, Martin E, Mottus M,
North PJR, Qin W, Robustelli M, Rochdi N, Ruiloba R, Soler C, Thompson R, Verhoef W,
Verstraete MM, Xie D (2006) The third RAdiation transfer Model Intercomparison
(RAMI) exercise: documenting progress in canopy reflectance modelling. J. Geophys. Res.
112(D9):D09111. doi:10.1029/2006JD007821
144
M.J. Chopping
Wiscombe W (2000) Formation Flying, Leonardo, Balloon-ERB and Triana: Forays to the Frontier,
NASA/NIAC workshop presentation, November 7, 2000, online at http://www.niac.usra.
edu/files/library/meetings/misc/nov00 wks/wwiscombe.pdf, last access 06/14/06
Zhang X, Schaaf CB, Gao F, Friedl MA, Strahler AH, Hodges JCF (2000) Mapping land cover and
green vegetation abundance using MODIS-like data: a case study of New England. Proc. Int.
Geosci. Remote Sens. Symp. 2000: vol 5, pp 20052007
Part II
Physical Modeling and Inversion

Algorithms
Chapter 6
Modeling the Spectral Signature of Forests:

Application of Remote Sensing Models
to Coniferous Canopies
Pauline Stenberg, Matti Mottus, and Miina Rautiainen
Abstract The vertical and horizontal structure of forest canopies is one of the most
important driving factors of various ecosystem processes and has received increasing attention during the past 20 years and served as an impetus for earth observation
missions. In the remote sensing community, the variables which describe canopy
structure are called biophysical variables, and are directly coupled with the fundamental physical problem behind remote sensing: radiative transfer in vegetation.
There are basically three different approaches to interpreting biophysical variables
from remotely sensed data: (1) empirical, (2) physically based, and (3) various combinations of them. The physical approach builds upon an understanding of the physical laws governing the transfer of solar radiation in vegetative canopies, and formulates it mathematically by canopy reflectance models which relate the spectral
signal to biophysical properties of the vegetation. In this chapter, we will first outline the basic principles and existing physically based model types for simulating
the spectral signature of forests. After this, the focus is on the specific issues related
to applying these models to the complex 3D structure of coniferous canopies.
6.1 Introduction
The assessment of many fundamental ecological questions at global scale is possible
only through remote sensing, since integrated analyses of the biosphere, atmosphere
and hydrosphere require simultaneous measurements over large areas.
Pauline Stenberg
Department of Forest Resource Management, University of Helsinki, Finland
Pauline.Stenberg@helsinki.fi
Miina Rautiainen
Department of Forest Resource Management, University of Helsinki, Finland
Matti Mottus
Tartu Observatory, Toravere, Tartumaa, Estonia
147
148
P. Stenberg et al.
The three-dimensional (3D) structure of forest canopies is amongst the most important driving factors of various physiological and ecological processes. By analyzing the 3D structure of canopies, it is possible to detect growth patterns and
phenological cycles of forests, as well as other changes such as insect or fire outbreaks and damages, illegal deforestation, and environmental stresses caused by
drought or pollution. In the remote sensing community, the variables which describe canopy structure are called biophysical variables. Biophysical variables are
coupled with the radiative transfer problem (Chandrasekhar, 1960) in vegetation
(Ross, 1981) and can also be defined as state variables of the radiative transfer problem, in other words, the smallest set of variables which are needed to fully describe
the physical state of the system at a given scale (Verstraete et al., 1996). However,
mere description of the structure of the canopy through biophysical variables is not
the only goal: the variables should be applicable for end-users in, for example, water
and carbon cycle or climate modeling.
It is obvious that measuring biophysical variables in situ is both laborious and
time-consuming in forests and impossible at global scale within a short time frame.
Space-born monitoring is thus required, and algorithms for interpreting these variables from remotely sensed data need to be developed. Most common remotely
sensed biophysical variables of forests include leaf area index (LAI), fraction of
photosynthetically active radiation (400700 nm) absorbed by vegetation (fPAR)
and fraction of canopy cover (fCover). LAI and fCover are geometrical variables
which are related to canopy gap fraction, i.e., the fraction of ground seen in a given
direction (Nilson, 1977) canopy gap fraction is in fact determined by LAI and
its spatial distribution and leaf inclination distribution. fPAR, on the other hand, is
an outcome of radiative transfer in vegetation, and the opposite of reflectance from
vegetation. In addition, there are biophysical variables which are not geometric, but
instead influence the spectral properties of scattering elements. An example of such
variables is the chlorophyll content of green leaves (which also is closely connected
to the nitrogen content of foliage). However, there is evidence that no clear distinction between foliar biochemistry and LAI can be made in practical remote sensing
(Yoder and Pettigrew-Crosby, 1995).
There are basically three different approaches: (1) statistical (empirical) and (2)
physically based, and (3) various combinations of them (e.g., neural networks), to
assess biophysical variables from spectral signals provided by optical satellite images. In the empirical approaches, commonly used in, for example, regional or national forest inventories, the vegetation characteristics of interest are estimated based
on statistical relationships (regressions) obtained by collecting training data on the
spectral signatures of a variety of objects. These methods are limited to a specified
viewing and illumination configuration, and require large sets of reliable ground
truth data. The physical approach, in contrast, builds upon an understanding of the
physical laws governing the transfer of solar radiation in vegetative canopies and
formulating it mathematically by reflectance models. Reflectance models, in turn,
relate biophysical properties of the vegetation, or sets of canopy and stand parameters, to the spectral signal. Physically based methods are, at least in theory, more
robust since they are not limited to a single configuration or vegetation biome type.
6 Modeling the Spectral Signature of Forests
149
Physically based methods have progressively become more and more attractive
for the assessment or monitoring of biophysical characteristics of vegetation. They
are the only methods which can fully take into use the versatile hyperspectral and
multiangular information provided by the modern satellite sensors, and they are
better suited for many current large scale applications than the empirical techniques.
Several aspects must be considered when physically based models are developed
for operational monitoring purposes. To begin with, the sensitivity of the model
to the biophysical variable to be retrieved should be maximized. Simultaneously,
undesired effects, for example from the atmosphere or terrain topography, should be
eliminated as efficiently as possible using other available models and data. Finally,
when considering the optimal model, the larger the area (e.g., the whole Earth), the
more the model should rely on theory and the less on local data bases for input as the
sets are usually not compatible from one region or country to another. In addition,
we should take into account technical questions such as computation time required
by the model.
Canopy reflectance models were originally developed for agricultural crops, followed by broad-leaved forests, and even though presently there exist a number
of reflectance models designed to be applicable also to conifers, some missing
difficult-to-model properties and insufficient empirical data on some key parameters have still limited their optimal use. Several recent studies (Rautiainen and
Stenberg, 2005a; Smolander and Stenberg, 2005) give quantitative support to the
hypothesis put forward a long time ago (Norman and Jarvis, 1975; J. Ross, 1981
2002, personal communication), that a major reason for the distinct spectral signature of coniferous forests lies in their hierarchical grouped structure, which governs
the processes of multiple scattering within the canopy, as well as canopy absorption.
Moreover, modeling tools are emerging by which grouping at different scales can be
accounted for in canopy reflectance calculations. The effect of grouping on canopy
PAR absorption and photosynthesis has been a well studied subject in forest production ecology (Oker-Blom, 1986; Leverenz and Hinckley, 1990; Wang and Jarvis,
1990; Oker-Blom et al., 1991; Nilson, 1992; Cescatti, 1997a) but only more recently,
after becoming aware of the specific problems related to interpretation of satellite
images over the boreal zone, has modeling the radiation regime of coniferous forests
received increasing attention in the remote sensing community. The boreal zone
spreads through Fennoscandia, Siberia, Alaska and Canada, and hosts a multitude
of coniferous tree species adapted to the cold and drought climate conditions of
the region. The forests are typically rather open with dense crowns (consisting of up
to a few million needles per crown) and an abundant green understory and moss or
lichen layer. The documented complex structure of the forests is further complicated
by the fact that acquiring ground observations from many parts of the boreal zone is
especially difficult due to the remoteness and climate of the region. For application
of forest reflectance models in the boreal zone, remaining problems of high priority
are thus how to model efficiently and sufficiently realistically (1) the hierarchic 3D
canopy structure and (2) the contribution to the remotely sensed spectral signal from
the background, typically composed of mixed green understory, and (3) how to separate the signals of the forest canopy layer and the understory from each other for correct interpretation of tree canopy biophysical variables from optical satellite images.
150
P. Stenberg et al.
In this chapter, we will first outline the basic principles and current status of
existing models for simulating the spectral signature of forests. After this, we will
focus on the specific issues related to applying these models in the boreal zone,
including how to collect the required input data. Finally, more practical issues, such
as application of ground truth data, model scaling and validation, will be addressed.
6.2 Modeling the Spectral Signature of Forests

The basic premises for optical remote sensing of vegetation are that the solar radiation received by a remotely located sensor (e.g., on a satellite) upon interaction
with the vegetation canopy carries in it the signature of the canopy, and that this
spectral signature can be deciphered to obtain the information of interest (Goel,
1989). Satellite-borne sensors measure the mean intensities of radiation at different wavelengths emanating from a target on Earth. Correct interpretation of these
measurements to yield the biophysical variables of interest requires an accurate
specification of the relationships between these variables and the canopy leaving
radiation field. This specification is quantified by canopy reflectance models. The
models are parameterized using mathematical descriptions of canopy structure together with optical properties of the plant elements and the underlying surface to
produce spectral signatures of canopy leaving radiation. In addition, the spectral
and angular properties of the incoming radiation and the receiving sensor need to be
specified (Table 6.1). Finally, the spectral signature depends on the resolution of the
measurement (i.e., pixel size), which needs to be considered in defining the spatial
scale of the model.
Table 6.1 Variables affecting the spectral signature of a forest (excluding terrain topography and
atmosphere)
Sensor
Illumination
Forest tree layer

canopy
Understory and soil
Zenith and azimuth

viewing angles
Angles of incidence
and azimuth
Wavelength band
(i.e., spectral
sensitivity)
Wavelength
Macroscale structure
(distribution, size
and shape of tree
crowns)
Microscale structure
(distribution, size
and shape of leaves,
needles, shoots and
branches)
Other structural
elements (e.g.,
distribution, size and
shape of tree trunks
and branches)
Spectral properties
of all the canopy
elements
Geometrical structure
(amount, distribution,
size and shape of
understory plants)
Spectral properties
of understory plants
Resolution
Soil optical
properties (influenced
by e.g., soil moisture
and texture)
151
The mathematical descriptions of the process of radiative transfer in vegetation

canopies cover a range of different approaches, starting from the traditional turbid
medium approach to geometric-optical and hybrid models, the main difference being the degree of detail at which canopy structure at different hierarchical levels is
described (Table 6.1).
The turbid medium approximation, introduced to modeling the radiation regime
inside vegetation by Ross (1981), is based on the classical radiative transfer theory
(Chandrasekhar, 1960). The radiative transfer theory has been successfully applied
to predicting the distribution of shortwave (optical and infrared) photons in several
media like atmospheres, stellar dust or water. Radiative transfer theory is, in principle, the law of conservation of energy written out for beams of radiation traveling
in all possible directions in a volume filled with optically active, i.e., absorbing and
scattering, material. It is thus a function of five coordinates: three spatial coordinates determining the elementary volume for which the energy conservation law is
applied, and two variables defining the direction of the radiation beam. In the classical radiative transfer theory, the fate of a photon is determined by the probability
of being absorbed or scattered while traveling a unit distance (these probabilities
are called the absorption and scattering cross-sections, respectively). Additionally,
interactions between different rays for each point are described by the probability
distribution of directions the photon can be scattered in, measured relative to the
photons traveling direction before scattering. This is called the photon scattering
phase function. The equation of radiative transfer merely states that the number of
photons exiting each elementary volume equals the number of photons entering the
volume minus the number of absorbed photons, written out in a correct mathematical form. More generally, a source term has to be added for radiating media, e.g., inside stars or for the thermal infrared radiation in atmosphere. In the optical region
under consideration in this chapter, no radiation is usually emitted by the medium.
The difference between the classical radiative transfer theory and the turbid
medium approach lies in the treatment of the directional properties of the medium:
in the classical radiative transfer theory, the scattering phase function and the absorption and scattering cross-sections are assumed to be independent of the direction
of photon travel. This assumption does not hold for the transfer of radiation inside
a vegetation canopy. Thus, in the turbid medium approximation, plant cover is described as consisting of geometric elements which usually have a preferred orientation described, in turn, by a directional attribute. Now, the absorption and scattering
cross-sections become functions of the direction of photon travel inside the canopy
as well as of the actual shape of the elementary units (phytoelements) constituting
the canopy. For example, in a canopy of preferably horizontal flat leaves, the probability of hitting a leaf while traveling a distance unit is much larger for a photon
moving in the vertical direction than for one moving in the horizontal direction.
Consequently, the interaction cross-section (sum of absorption and scattering crosssections) in such a canopy is larger in the vertical direction than in the horizontal.
Orientation of plant elements is described by statistical distributions. For example,
the orientation of leaves is described by the direction (inclination and azimuth) of
their normal and that of stems, branches and needles by the direction of their axis.
152
P. Stenberg et al.
A number of different distributions suitable for describing the distribution of leaf

normals of actual canopies are given in, e.g., Weiss et al. (2004). Cross-sections can be
calculated from these distributions by integrating over all possible orientation angles.
Other than implying that the scattering phase function and the cross-sections depend on direction within the canopy, the turbid media approach leaves the classical
radiative transfer equations intact: canopy elements are still of infinitesimal size
and do not cast shadows, i.e., the far-field approximation still holds. Although
anisotropy (or dependence on photon travel direction) makes finding a solution to
the radiative transfer equation more difficult and yields some traditional methods
unpractical, a plethora of techniques is still available for finding an approximate
solution. The basic equations for the radiation field can be found, for example, in
the book by Ross (1981) or various other sources (e.g., Myneni et al., 1989).
The crudest (and fastest) way of solving the radiative transfer problem for a
plane-parallel turbid media (or in slab geometry) is by using the two-stream
(Kubelka-Munk) equations. Only two directions, up and down, are used to describe
the radiation field inside a horizontally homogeneous and infinite medium and the
solutions can be found analytically. In its simplest form, the two stream model assumes isotropic distribution of both up- and downward traveling radiation, i.e., the
intensity does not depend on view angle inside both hemispheres. This condition
is clearly violated if direct solar radiation is present. To overcome the problem, the
contribution of the direct beam is separated, thus creating an additional source for
the remaining more isotropic diffuse radiation field. This technique is almost universally employed in all algorithms for solving the radiative transfer problem in
plant canopies. This is because besides its extremely anisotropic character, direct
illumination gives rise to the so-called hot-spot phenomenon (see discussion later in
this chapter) that needs to be accounted for if measurements are taken near the backscattering direction. The solutions of the two-stream equation are analytical and thus
extremely computer-efficient. It should also be noted that the two-stream equations
cease to be approximations for a (theoretical) canopy of infinitesimally small horizontal ideally scattering leaves but, instead, give the exact analytical solution as the
reflected and scattered radiation field inside such a canopy is isotropic.
A classical example of an application of the two-stream approximation is the
SAIL (Scattering and Arbitrarily Inclined Leaves) model (Verhoef, 1984, 1985).
Besides the two fluxes, the model calculates intensities for two other directions: the
solar direction (for reasons discussed above) and the view direction. This enables to
take into account anisotropy in the distribution of reflected radiation. The absorption
and scattering coefficients, and the scattering phase function are calculated from a
distribution function of leaf inclination angles. Despite its age, the model is still used
for remote sensing purposes, although sometimes modified to fit the particular needs
or encapsulated in other models that include a higher level description of canopy
structure (Kuusk, 1995; Kuusk, 2001; Andrieu et al., 1997). The SAIL model is
reasonably accurate, and can take the competition from much more sophisticated
3D models, in the case of relatively homogeneous canopy types (crops, grasslands)
that have a small number of structural parameters (Jacquemoud et al., 1995; Koetz
et al., 2005; Meroni et al., 2004; Andrieu et al., 1997; Weiss and Baret, 1999; ZarcoTejada et al., 2003).
153
If the media cannot be considered homogeneous in the horizontal directions, as is

the case for many natural vegetation canopies, the radiative transfer equation can be
solved in three dimensions (Myneni, 1990; Knyazikhin and Marshak, 1991). This
approach is useful for highly structured scenes where the two-stream model cannot
be adequately parameterized to include canopy heterogeneity. Once such a scene
is generated based on measured data or a forest structure model, the canopy is divided into cells and the equation (or, after discretization, a set of coupled integrodifferential equations) is solved numerically using the discrete ordinates iterative
approach on the grid (Myneni, 1990; Gastellu-Etchegorry et al., 1996; Knyazikhin
et al., 1998a). This method is based on the calculation of radiation intensities in
discretized directions. In each iteration, the contribution of radiation scattered to
every discrete direction from all other directions is calculated for each canopy cell
using a pre-calculated scattering matrix until the intensities converge to a solution.
Naturally, this method, besides its mathematical complexity (which is further increased by using various convergence acceleration techniques), also requires a realistic description of the scene and reasonable cell-size to exclude shadowing effects.
The method is computer-intensive but the results can be considered accurate as this
method is based on the physical principles of radiative transfer.
Another common method used for calculating radiation reflected by a canopy, for
which a complete 3D description of the scene is given, is the Monte Carlo method.
As the radiation field above plant canopy is composed of small contributions by
individual photons, the angular and spatial variation in the intensity of reflected
radiation can be viewed as a distribution function describing the possible exit directions and locations of light quanta. Monte Carlo methods (or ray tracing algorithms)
are based on random sampling of this distribution. If a sufficient number of photons
are traced, the counts of photons exiting in each direction are relatively close to the
actual directional distribution of reflected radiation, although some numerical noise
is inevitable. Although in studies of radiative transfer in canopies, the term Monte
Carlo is commonly used to denote tracing the trajectories of photons in a realistic detailed 3D canopy scene (Gerstl et al., 1986; Ross and Marshak, 1988; North,
1996; Chelle and Andrieu, 1998; Govaerts and Verstraete, 1998; Thompson and
Goel, 1998; Disney et al., 2006), Monte Carlo calculations can also be used to trace
photons in a 3D turbid media with varying optical properties, like the geometricoptical approximation described below (Gerard and North, 1997; Garca-Haro and
Sommer, 2002). An overview of the Monte Carlo modeling approach is given by
Disney et al. (2000).
If a 3D description of a scene is given, the reflected radiation field can also be
calculated using the radiosity principle (Borel et al., 1991; Gerstl and Borel, 1992;
Goel et al., 1991; Chelle and Andrieu, 1998; Garca-Haro et al., 1999; Qin et al.,
2002; Soler et al., 2003). Radiation field inside the canopy is calculated using view
factors for individual elements: if the illumination conditions of a particular canopy
element are given, its brightness as viewed from the positions of all other elements
can be calculated based on its geometry and scattering properties. The brightness of
an object is a well-defined radiometric quantity that can be used to calculate the partial flux contributed by the object. Thus, if the magnitude and angular distribution
154
P. Stenberg et al.
of radiation field incident on a canopy element can be calculated from brightness

of all other canopy elements and the intensity of the incident unscattered radiation,
the brightness of the element as viewed from any point inside or above the canopy
can be calculated. It is an iterative process complicated by the large number and
mutual shading of elements present in a natural canopy. From the computing perspective, the most time-consuming task is calculating these view factors or solving
the geometrical relations between each pair of foliage elements. Once these factors are known, calculations for different wavelengths for which only the scattering
properties of elements are different, but not the geometry of the canopy as a whole
are relatively fast.
In the classical radiative transfer approach, the optical properties of the medium
are assumed to change continuously (for the 3D case, this assumption is valid inside
each cell). This is naturally not the case for actual vegetation canopies. The Monte
Carlo and radiosity methods that describe the canopy as consisting of separate solid
objects remove this restriction and take into account the mutual shading of elements
(remove the far-field approximation). The assumption of discontinuities in canopy
optical properties, or the existence of voids between scatterers, can also be inserted
into the radiative transfer equation. The foliage are a density can be described using
the statistical distribution of an indicator function, in other words, a function of the
spatial coordinates that equals unity only if a scattering element is present at the
point described by the coordinates. This leads to the stochastic radiative transfer
equation in plant canopies (Menzhulin and Anisimov, 1991; Shabanov et al., 2000).
Besides the mean reflecting properties of a canopy, this theory allows to calculate
its statistical moments.
Often the terms deterministic (non-stochastic) and stochastic are used as synonyms to 1D and 3D models, respectively. However, with the exception of Monte
Carlo models, most of the existing canopy radiation/reflectance models, in fact use a
non-stochastic approach in computing the spectral signatures. That is, even though
the canopy structure may be described using statistical distributions, the canopy
radiation field is solved for a mean realization (with average characteristics) of
canopy structure. A truly stochastic approach, in contrast, would be to evaluate the
3D canopy radiation field for all possible statistical realizations of the canopy, and
then average the corresponding radiation fields to obtain the ensemble average signature (Shabanov et al., 2000). The stochastic and non-stochastic approaches can
result in different relationships between mean characteristics of canopy structure
and canopy-leaving radiation.
Sometimes, a more abstract description of the 3D structure of a natural forest
canopy is required. The exact locations of all canopy elements are rarely known
or reliably predicted for a forest canopy. Therefore, instead of describing the structure of a canopy by specifying the locations of the smallest canopy elements (e.g.,
leaves, needles or shoots), the canopy can be first divided into larger subunits. An
obvious choice for the crudest first-level division would be the tree crown. The
canopy can then be viewed as a congregation of geometrical tree crown envelopes
with gaps between them. The locations of the tree crowns are described by a statistical distribution, thus accounting for mutual shadowing and the distribution of
155
between-crown gaps. Depending on the accuracy by which the processes of radiative transfer are described in these models, the models are commonly categorized as
geometric-optical or hybrid models, the term hybrid being a shortcut for hybrid
radiative transfer/geometric-optical model.
In the simplest form, a geometric-optical canopy reflectance model considers
just the effect of the geometrical structure of a forest on the remotely sensed image
(Li and Strahler, 1985, 1986, 1992; Welles and Norman, 1991; Chen and Leblanc,
1997). The image is supposed to consist of (sub-pixel-sized) regions of different brightness: directly illuminated tree crowns and ground, and shadowed tree
crowns and ground. The fractions of these components are calculated using given
illumination and view angles and mutual shadowing derived from geometrical
considerations. These types of models can accurately account for only first-order
scattered radiation, i.e., photons that reach the sensor having interacted only once
with the geometric shapes comprising the canopy. Diffuse radiation is included
using correction factors. This makes it difficult to use these models for different
wavelengths where the optical properties of canopy elements are not similar. The
share of first-order scattering can be calculated straightforwardly: this reflectance
component, besides depending on canopy geometry, which is the same for all wavelengths, is a linear function of leaf single scattering albedo (leaf reflectance plus
transmittance). Diffuse photons, on the other hand, interact with leaves (or other
canopy elements) several times and the possible number of interaction is itself a
function of leaf albedo at the specified wavelength. This makes canopy diffuse
reflectance strongly nonlinear with leaf albedo, and results in highly wavelengthspecific diffuse radiation correction factors.
Including the diffuse radiation field into the geometric-optical model leads to
a hybrid radiative transfer/geometric-optical model (Nilson and Peterson, 1991;
Li and Strahler, 1995; Ni et al., 1999; Atzberger, 2000; Chen and Leblanc, 2001;
Huemmrich, 2001; Kuusk and Nilson, 2000; Peddle et al., 2004). The methods
to include diffuse radiation can range from exact solutions (similar to those used
in solving the 3D radiative transfer) to tracing the photons in the canopy comprised of crown envelopes seen as large chunks of turbid media (similar to the
Monte Carlo approach). However, to maintain the high efficiency achieved by delineating the canopy into abstract crowns, a simple and fast approximation is often
used. Now, the problem of finding reflected radiation intensity is divided into two
sub-problems: (1) finding the first-order scattering component using a geometrical
figure filled with absorbing and scattering foliage elements illuminated by a beam
of direct radiation, taking into account mutual shadowing by other semi-transparent
crowns, and (2) calculating the share of diffuse radiation using a simpler (e.g., twostream) solution of the radiative transfer equation. Naturally, in calculating both
components one has to consider the effect of a partially reflecting ground surface or
undergrowth. Joining the two sub-models is not an easy task: the multiple-scattering
component has to be parameterized to include the effect of canopy structure so that
conservation of energy and correct partitioning between first- and higher order
scattering are maintained.
156
P. Stenberg et al.
Hybrid models are an efficient and flexible tool for describing the radiation
regime of complex forests. The effect of different crown shapes can be included
by using various geometrical shapes: ellipsoids and cones of different proportions
are the most commonly used. Also, higher-order clumping can be introduced by
further dividing the canopy into abstract objects and adding their contributions to
total reflectance (e.g., Smolander and Stenberg, 2005). However, most models still
assume that the crown envelopes are uniformly filled with scatterers, although from
biological considerations and biometrical measurements it is known not to be true.
Using more realistic foliage distributions would considerably affect the transparency
of the crowns by introducing more clumping at both branch and whorl scale or
by locating photosynthesizing material close to the crown edge. While accounting
for higher-order structure usually leads to an increase in predicted transmittance if
the higher-order objects are distributed randomly, a regular distribution of foliage
clumps that has been suggested for natural canopies (Cescatti, 1997b) may diminish
or even reverse the effect.
The choice of the model to be used for solving a problem involving calculation
of canopy reflectance or radiation balance clearly depends on many factors. The
amount of required computer resources, manpower and time is clearly different for
the different approaches. Also, the optimal choice depends on the object under investigation: for relatively homogeneous canopies (crops, grasslands), a two-stream
model give very good results. Sometimes, if canopy reflectance calculations form
only a small contribution to a larger problem under investigation that depends on
many variables with possibly high uncertainties, the use of a simpler model is justified. This is clearly not the case for highly structured vegetation covers like shrublands or boreal coniferous forests where the canopy upper surface is not flat, and
mutual shading and between-crown transmittance between tree crowns have to be
taken into account. Also, the amount of a priori knowledge can be a limiting factor when constructing a detailed 3D description of a canopy. In this case, a less
detailed approach with a smaller number of input parameters might be preferred,
e.g., a hybrid geometric-optical/radiative transfer model.
Another characteristic of the reflecting properties of a medium that contains
finites objects filling a three-dimensional volume is the hot spot phenomenon. Hotspot is a bright area in a remotely sensed image opposite to the source of illumination caused by a lack of shadows in the exact backscattering direction. The width
of this reflectance peak (or the size of the brighter area in an image) depends on the
geometric properties of the reflecting medium. Based on their working principles,
geometric-optical models take this phenomenon into account, but only partly. When
looking from the direction of illumination, no shaded crowns can be seen. This results in an increase in the predicted reflectance. Yet, the hot-spot phenomenon is also
produced at a finer level (e.g., leaves), which in turn leads to a distinct anisotropy
of the brightness of tree crowns. In more sophisticated reflectance models, this leaflevel hot-spot is added to the wider hot spot created by crown structure. Also, a
hot-spot correction can be added to models based on the radiative transfer equation
(e.g., Gerstl et al., 1986; Marshak, 1989; Verstraete et al., 1990; Jupp and Strahler,
1991; Kuusk, 1991). The correction factor is semi-empirical due to the complex
157
structure of vegetation canopies. The phenomenon itself is difficult to measure accurately since it tends to be in the shadow of the sensor. Due to common view angle
configurations, it is also not registered by most remote sensing instruments.
Apart from the choice of radiation transfer model and availability of adequate
computing resources, quality of both structural and spectral input data has a large
impact on model performance. Today, good-quality data of either sorts are scarce as
they are difficult to measure and have a high natural variance. Moreover, as developing accurate radiative transfer models for vegetation canopies is a work in progress,
a ready-made solution may very likely not yet exist. However, the models presented
here (or, a quite up-to-date performance comparison can be found from RAMI website (http://rami-benchmark.jrc.it) (Pinty et al., 2004b)) cover a large number of approaches which, if adequate input is provided in terms of both measurement data
and dedication, can be applied to a wide variety of problems. More information on
the subject can be found, in addition to the works referred to in this section, from
previously published reviews and textbooks (e.g., Myneni et al., 1991; Liang and
Strahler, 2000).
Canopy reflectance models can be constructed in many ways; some can be run in
the forward mode, some both in the forward and inverse modes. In model inversion
(discussed thoroughly in another chapter of this volume), radiation measurements
are converted into variables of interest characterizing the target. In remote sensing of forests, invertible models can be used to infer biophysical variables from
reflectances (or back-scattering) registered by the remote air-or satellite-borne sensors. Commonly used methods for the inverse estimation of vegetation characteristics from satellite images include (1) comparing the observed signal to a database of
previously computed spectral signals for a wide selection of different canopies and
choosing the closest matches (look-up tables, LUT) (Knyazikhin et al., 1998b), or
(2) iteratively optimizing model input parameters to match the observed signal as
closely as possible using different optimization routines (Kuusk and Nilson, 2000).
Needless to say, the goodness of the estimates depends crucially on how realistic the
model is. Another basic requirement for successful estimation is of course that the
vegetation characteristic in question has a detectable influence on the spectral signal,
and that measurement errors can be corrected for. Even so, a remaining limitation is
the well known fact that the inversion problem is ill-posed: no unique solution exists
but different combinations of input parameters produce the same spectral signal. To
be able to solve the inverse problem, in other words, to reduce the array of possible
solutions to one solution, we need to acquire information from outside the problem
itself. Technical or mathematical advances do not remove the underlying ambiguity
of the ill-posed nature of the inversion problem in remote sensing. Another central
problem in inversion is scaling: the forest reflectance models assume that the given
forest structure continues infinitely. However, often in the case coarse and medium
resolution satellite images, the vegetation (forested area) may cover only part of
the pixel. At best, thus, the averaged solutions to the inversion problem may be
accurate at larger, often regional, scales but not at small scales.
158
P. Stenberg et al.
6.3 Structural and Spectral Characteristics of Coniferous

Forests
Northern hemisphere boreal forests, forming the largest unbroken, circumpolar forest zone in the world, are dominated by coniferous tree species. The radiation regime
of coniferous forests is known to differ in many respects from that of other vegetation, e.g., broadleaved forests and agricultural crops (Norman and Jarvis 1975;
Oker-Blom et al., 1991; Nilson, 1992). From the perspective of optical remote
sensing, a specific problem encountered in the boreal forest zone is the poor performance of commonly used spectral vegetation indices (e.g., the NDVI) for the
estimation of biophysical variables (such as LAI). One reason for the generally observed relative insensitivity of these indices to changes in LAI is apparently caused
by an abundant presence of mixed green understory in boreal coniferous forests
(Chen and Cihlar, 1996; Eklundh et al., 2001; Stenberg et al., 2004; Peltoniemi
et al., 2005; Rautiainen et al., 2007). The influence of understory on the spectral
signal naturally poses a problem also for the correct image interpretation using physically based models. In recent years, yet another phenomenon specific to coniferous
forests has become a subject of increasing scientific interest by the optical remote
sensing community. From empirical observations, it has long been known that coniferous forests have a lower reflectance, especially in the near-infrared (NIR), than
broadleaved forests, but the physical background to this behavior is still partly unexplored. In the following, we will discuss how some conifer-specific characteristics
affect the spectral signature of coniferous forests.
There are several structural and spectral attributes which are specific to coniferous forests and require modifications to the radiative transfer models formulated
for their broadleaved counterparts (Fig. 6.1). The non-flat, three-dimensional structure of conifer needles of varying, species specific shape, first of all, requires proper
Needles
Shoots
Tree crowns,
internal shoot
distribution patttern
Fig. 6.1 Hierarchic structure of coniferous crowns: A three-layer scheme
159
(re-)definition of some of the key variables and concepts of these models: e.g., the
leaf orientation, the mean projection of unit foliage area (G-function), the leaf
scattering phase function, and the leaf area index (Stenberg, 2006). These variables and distributions were all originally defined and derived under the assumption of flat leaves and, so, for example, equations given for the mean projection
of planar leaf area (e.g., Warren Wilson 1967) are not directly applicable to needles but must be derived with consideration given to the needle shape (Oker-Blom
and Kellomaki, 1982; Lang, 1991; Chen and Black, 1992). A more serious theoretical problem for modeling however arises from that in many coniferous species
needles are closely grouped together as shoots. The resulting small-scale variation
in needle area density cannot readily be represented in the formulation of the radiative transfer equation based on the concept of an elementary volume, which
should be small enough that essentially no mutual shading between the elements
exists but large enough for statistical laws to apply (Ross, 1981). It has been proposed, therefore, that the shoot should be treated as the basic structural element
of coniferous canopies an approach that has actually long been used in models
of canopy light interception and photosynthesis (e.g., Oker-Blom and Kellomaki,
1983; Nilson and Ross, 1997; Cescatti, 1997a), but has not yet been fully implemented in forest reflectance models due to the lack of data and models describing the scattering properties of shoots (but see Smolander and Stenberg, 2003).
More data are available on another key parameter entering the shoot based models,
namely the shoot silhouette to total area ratio (STAR) (Oker-Blom and Smolander,
1988), which is conceptually analogous to the G-function defined for flat leaves
(Nilson, 1971), but includes a clumping coefficient accounting for the mutual shading of needles in the shoot. The clumping or mutual shading of needles in shoots
acts to decrease the interaction cross section area of a given amount of total needle
area, i.e., the extinction coefficient, and this effect can in radiative transfer models
be parameterized (quantified) using the STAR. The decrease in shoot single scattering albedo, as compared to that of a single needle, has also been shown to be
closely related to STAR, which thus can be used to modify the scattering properties of an elementary volume containing shoots (Smolander and Stenberg, 2003).
However, theoretical models and, above all, empirical data on the shoot scattering
phase function for different species are still needed for correct parameterization of
coniferous canopy reflectance models. In some current models (Knyazikhin et al.,
1998b; Kuusk and Nilson, 2000), shoot level grouping is accounted for in quantifying (i.e., reducing) the extinction coefficient (interaction cross section area of the
elementary foliated volume), but its effect on the volume scattering phase function
has not yet been fully implemented. It should be noted that when shoots (instead
of single needles, or needle surface area elements) are treated as the basic elements,
the optical properties (transmittance and reflectance) of single needles no longer suffice to describe the scattering properties of the elementary volume in the radiative
transfer equation.
Also at higher levels of organization, coniferous forests display a distinct grouped
pattern (Fig. 6.1). The canopies are typically formed of dense, narrow and deep tree
crowns. Crown shape, volume, and density have a considerable effect on the total
160
P. Stenberg et al.
canopy reflectance, as well as on the bidirectional reflectance factor (BRF) of conifer

stands (Rautiainen et al., 2004). The same properties, together with the stand density
(number of trees per hectare), also determine how much of the background (forest
floor) is visible to the satellite instruments. Because northern coniferous forests are
often rather open, with an abundant green understory, the spectral signal is a complex mixture of (overstory) canopy and background reflectance.
6.4 Effect of Clumping on the Spectral Signal: p-Theory

as an Example
The high degree of grouping (or clumping) of foliage at different hierarchical scales
in coniferous forest stands (Fig. 6.1) is considered to be the most important reason behind the different spectral signature of these forests, in particular the low
reflectance in NIR, as compared to broadleaved forests (Nilson et al., 1999; Rautiainen and Stenberg, 2005a). Incorporating the effects of grouping in canopy reflectance models poses a true challenge as very detailed canopy descriptions cannot
readily be integrated into models operating at large spatial scales this is typically
the case in remote sensing of vegetation. Thus, especially for large scale applications, it is important to search for some key parameters that could capture the most
essential structural features of a forest stand. One such canopy structural parameter,
proposed to govern canopy absorption and scattering, is the spectrally invariant (i.e.,
wavelength independent) p-parameter introduced by Knyazikhin et al. (1998b).
The p-theory, in short, predicts that the amount of radiation scattered by a canopy
(bounded underneath by a black surface) should depend only on the wavelength and
the spectrally invariant parameter (p), which can be interpreted as the probability
that a photon scattered from a leaf (or needle) in the canopy will interact within
the canopy again the recollision probability (Smolander and Stenberg, 2005;
Rautiainen and Stenberg, 2005a). The usefulness of this parameter in practical applications depends on whether and how well p can be related to (or derived from)
other commonly available forest stand data. Nonetheless, it is a powerful modeling
tool because it links canopy absorption ( C ) and scattering (C ) at any wavelength
( ) to the phytoelement (leaf or needle) scattering coefficient (L ) at the considered wavelength, while simultaneously preserving the law of energy conservation
(Panferov et al., 2001; Wang et al., 2003). The relationship between canopy and leaf
scattering coefficients is described by the simple equation:
C ( ) =
L ( ) pL ( )
1 pL ( )
(6.1)
It follows from Eq. (6.1) that, at any given wavelength, the canopy spectral scattering
coefficient (C ) decreases in a nonlinear fashion with increasing canopy aggregation
or grouping (quantified by the p parameter). More importantly, in any given canopy
(fixed p), the relationship between C and L is also nonlinear, implying that canopy
161
structure (or grouping, specifically) does not only affect the magnitude of the canopy
leaving radiation but also modifies its spectral distribution.
Simulation studies and empirical measurements have provided support for the
validity and usefulness of the p-theory. Equation (6.1) was shown to precisely predict the absorption and scattering in structurally homogeneous canopies simulated
with a Monte Carlo model (Smolander and Stenberg, 2003; Smolander and Stenberg, 2005), and to hold true also in heterogeneous canopies (Mottus et al., 2007)
simulated using the hybrid FRT model (Kuusk and Nilson, 2000).
Another spectral invariant parameter (pt ) has been proposed to control canopy
transmission, i.e., the part of the scattered radiation that exits the canopy downwards (Knyazikhin et al., 1998a, b; Panferov et al., 2001; Shabanov et al., 2003).
This structural parameter so far lacks a clear physical interpretation, as has been
given for the other spectral invariant the recollision probability (p), but if it can be
formulated and shown to be valid (both theoretically and empirically), the two parameters offer a simple and effective tool for parameterization of the canopy radiation
budget. Namely, given the absorption (p value) and transmission (pt value), total reflectance (the upward scattered part of the incident radiation) is also known (because
they all sum up to one). Thus, the spectral invariants p and pt would allow calculation of all components, i.e., spectral absorptance, transmittance and reflectance
of the canopy shortwave radiation budget for any given wavelength knowing the
leaf (or needle) scattering coefficient at the same wavelength. If the relationships
between p and canopy structural parameters such as LAI are known, the spectral
signature of the canopy can be predicted in terms of LAI or, conversely, inverse
estimation LAI can be done based on measured canopy reflectance.
The relationship described by Eq. (6.1), linking together canopy scattering coefficients at a specific wavelength to the leaf albedo at the same wavelength, can
actually be applied at different hierarchical levels and provides a tool for scaling
grouping effects. In the simulation studies by Smolander and Stenberg (2003, 2005),
it was found that Eq. (6.1) could be used to scale from needle to shoot scattering by
replacing p by the recollision probability within a shoot (psh ). Moreover, it was
shown that the canopy level recollision probability (p) could be decomposed into
(1) the probability (psh ) that the new interaction occurs within the same shoot where
the first (former) interaction took place, and (2) the probability (pc ) that interaction
occurs with another shoot in the canopy, according to the formula:
p = psh + (1 psh )pc
(6.2)
Equation (6.2), which in a similar manner could be further developed (decomposed)

to account, e.g., for grouping at the crown level, shows how the whole canopy p
value is affected by grouping at different spatial scales. For a given LAI, the p
value increases with the degree of grouping present in the spatial dispersion of the
leaf area. Thus, for example, considering a broadleaved and coniferous canopy with
similar LAI and macro-scale structure (Table 6.1), the coniferous canopy would
have higher p or, equivalently, a smaller escape probability for the radiation incident
162
P. Stenberg et al.
on the canopy. This explains, at least partly, the observed lower canopy reflectance
(albedo) of coniferous forests as compared to broadleaved forests.
Two shortcomings of the p-theory for practical applications should however be
noted: (1) It describes only canopy scattering, i.e., the contribution from background
reflectance must be modelled separately and (2) it is not able to describe the angular
distribution of scattered radiation. To solve these problems, the p-theory should be
combined with other physically-based reflectance modeling concepts.
With increasing p, the scattered part of the radiation intercepted by the foliated
canopy (the crowns), and thus also canopy reflectance, decreases. However, the effect of groping on whole stand reflectance is not as straightforward. To demonstrate
the combined effect grouping has on canopy reflectance, on one hand, and on the
contribution of understory reflectance, on the other hand, we use the simple parameterization model, PARAS (Rautiainen and Stenberg, 2005a). In this semi-physical
model, forest BRF is calculated as a sum of the ground and canopy components:
BRF = cg f (1 )cg f (2 )ground + f (1 , 2 )i0 (2 )
L pL
1 pL
(6.3)
The parameters of Eq. (6.3) are defined as follows: 1 and 2 are the viewing and
illumination zenith angles, cgf denotes the canopy gap fraction in the directions of
view and illumination (Sun), ground is the BRF of the ground, f is the canopy scattering phase function, i0 (2 ) is canopy interceptance or the fraction of the incoming
radiation interacting with the canopy, and L is needle (or leaf) scattering coefficient. With increasing degree of grouping (larger p), canopy interceptance (i0 ) simultaneously decreases while the canopy gap fractions (cgf) increase and, thus, the
contribution from ground (understory) reflectance increases. Especially grouping at
larger scale (between crowns) may more importantly influence the total stand reflectance through its effect on increasing the contribution from the background than
through its effect on the canopy contribution.
Incorporating the effect of shoot scale clumping in the PARAS canopy reflectance model was shown to produce realistic reflectance values in the nearinfrared (NIR) of coniferous forests. This can be seen as a major improvement since
the low NIR reflectance observed in coniferous areas is one of the main anomalies
that models have not been able to account for. The results give support to the hypothesis that in coniferous canopies large part of the clumping occurs at the shoot
level and, thus, that the incorporation of shoot structural and spectral properties into
current forest reflectance models will significantly improve their performance.
6.5 Scaling of Canopy Reflectance

In the previous section, the term scaling was introduced to describe an application
of recollision probability at different canopy grouping levels, i.e., the dimensionless
scattering characteristics calculated at one level were applied at another, larger scale.
This approach is similar to using downscaled models for predicting the behavior
163
of the full-scale object. The procedure can be applied repeatedly as long as the
character of the process (in our case, interaction of photons with canopy elements)
remains unchanged. In optical remote sensing, the limit is determined by the ratio
of photon wavelength to the size of the scatterers, i.e., radiative transfer is still
determined by geometric optics, and refraction can be ignored.
In the current section, scaling has a different meaning that is more familiar to
the remote sensing community. Here, we use it to describe the spatial variation of
canopies that has to be taken into account when comparing sensors with different
view configurations (i.e., sensors that produce images with different pixel sizes).
Generally, models can produce only point estimates of canopy reflectance, although
most of them have a method to account for within-plot variability (statistical distribution of tree locations, etc.). This works well as long as the canopy is horizontally
homogeneous (which is what these models actually presume) and its structure does
not vary across a relatively large area, substantially exceeding the dimensions of
a pixel in the image produced by the sensor. Natural forest canopies, however, almost never possess such a property: they include clusters of high tree concentrations
and canopy openings, due to harvesting or windfall. As the reflectance process is
strongly nonlinear, this heterogeneity is difficult to take into account in the spatially
averaged pixels of remote sensing instruments.
The scale at which model results can be related to the reflected signal in a
straightforward way depends on the size of the structural units of the canopy. For
example, models that utilize an assumption of a statistical distribution of tree locations predict an average signal produced by such a canopy and cannot be applied
to model distinct patterns created by a particular configuration of individual trees
or the variance of intensity in the image of a single crown. They were not designed
for these purposes as overly detailed patterns are commonly not required for remote
sensing of larger areas. Another reason is that biophysical variables (LAI, fCover)
are defined for a larger canopy area and cannot be used to describe a single tree.
When moving to larger resolutions, the models may also fail due to the nonlinearity of the radiative transfer problem and a highly varying or discontinuous
distribution of tree locations. Canopy spectral properties are a function of spatial
resolution (Tian et al., 2000; Pinty et al., 2004a), and especially in regions with a
fragmented forest area, the signal will always be a mixture of reflectance from many
different ground (soil/understory) and forest canopy compartments. This may lead
to a situation where a model cannot be used as such (without alterations) to model
signals by sensors with different spatial resolutions (Tian et al., 2002).
The scaling problem affects the retrieval of different biophysical variables with
varying severity. While some variables can be considered almost scale-independent
(fCover, fPAR), predicting others (e.g., LAI) requires a careful consideration of spatial variation (Weiss et al., 2000). Algorithms for a correct treatment of the scaling
problem of vegetation reflectance modeling are scarce although a physically-based
theory for scaling was developed by Tian et al. (2002), for example. The incorporation of scaling algorithms directly in radiation transfer models is so far an almost
unexploited subject (Widen, 2004).
164
P. Stenberg et al.
6.6 Use of Stand Inventory Data as Model Input

Ground truth, or ground observations, is required for both developing and testing
physically based forest reflectance models and their inversion. There are two general paths for acquiring this information: by conducting measurements or by using
existing data bases (e.g., land cover maps, regional or national forest inventories,
spectral data banks of common plant species). However, a general problem with all
these data sources may be their age when compared to the more recently obtained
remotely sensed image.
Stand inventory data (which typically consists of information on tree species,
tree height and diameter, and stand density or basal area) form a common input
set required by most 3D forest reflectance models. However, stand inventories do
not provide all the parameters typically included in physically based reflectance
models, e.g., crown shape, leaf area density and orientation, nor do they include
the biophysical variables of interest in this chapter (LAI, fPAR, fCover). Testing
reflectance models should be done with a carefully measured, comprehensive input
data set which includes all the input required by the model and which does not rely
on other models (i.e., is not generated by other models). However, when operational
inversion for obtaining biophysical variables (LAI, fPAR, or fCover in this case)
is considered, such comprehensive data sets are not available, and other, let us say
less appropriate, sets need to be utilized. A typical source would be regional or
national forest inventory data bases which have recently become more suitable for
the purpose in the boreal zone of North America and Europe. Traditionally forest
mensuration science dealt with determining the volume of stands and logs and then
studying stand growth and yield. Nowadays, however, the scope has widened and
regional and national forest inventories have become more efficient environmental
monitoring tools and are conducted to ensure the sustainable use of forests. Thus,
they include more stand variables than tree height and diameter and form a better
interface for forest reflectance models. Nevertheless, if forest reflectance models are
run through a routine forest management data base (e.g., Rautiainen, 2005), there
are several structural input parameters, such as crown shape, woody area index or
shoot size, in the models which can hardly be obtained from these data bases. Thus,
alternative solutions for obtaining the structural information need to be considered.
A possibility worth exploring further would be to create realistic input data on
these additional structural parameters from basic stand variables through regression models and allometric relationships. Such models exist for relating different
biomass components of a tree or forest (e.g., Marklund, 1988) but parameters describing crown architecture have been of less interest in forest mensuration. In the
context of radiative transfer modeling, one of the most central parameters not available from routine stand inventory data sets is crown shape which here serves as an
example. Crown shape determines the limits of integration over the crown envelope
and thus the scattering volume. Therefore, crown shape needs to be predicted from
routine stand variables such as tree height and breast height diameter. This can be
done by, for example, developing a crown shape model based on extensive measurements and then using the model to predict crown shape for a stand where it has not
165
been measured (e.g., Baldwin and Peterson, 1997; Rautiainen and Stenberg, 2005b;
Mottus et al., 2006). Alternatively, crown shape can be assumed to have a constant
geometric shape (e.g., a cone or an ellipsoid) for a given tree species.
6.7 Concluding Remarks

Remote sensing mapping techniques have traditionally been statistically oriented
but physically based methods have progressively become more and more attractive because they are better suited for many current large scale applications related
to global mapping of vegetation (e.g., Knyazikhin et al., 1998b). Implementation of
physically based remote sensing methodology has become feasible thanks to the development of fast computer hardware and high resolution and multi-angular satellite
instruments. The rapid technological and methodological development of satellite
derived products offers a range of new possibilities for the mapping of terrestrial
biophysical parameters such as vegetation type, forest cover (fCover) and leaf area
index (LAI) from remotely sensed data (Myneni et al., 1997; Chen et al., 2002).
Remote sensing based methodologies are especially important, and in fact the only
feasible alternative, in regions where reliable field information is not available, is
difficult to reach, or is too expensive. This, on the other hand, makes the importance of validation an even more crucial part of the development process aiming at
producing trustworthy information righteously required by the end-users.
Several research networks operate today on the refinement of remote sensing
mapping methods and on validation of satellite derived biophysical products, such as
LAI, fCover and fPAR, that are presently routinely generated by a range of sensors
(Knyazikhin et al., 1998b; Morisette et al., 2006).
For critical assessment of the accuracy of remotely sensed estimates of biophysical variables, there should be available a set of statistically representative (sufficient)
and independent reference data with known accuracy. Also, special care should be
taken that similar (standardized) measurement methodologies and designs are used
to produce reference data of the variable of interest (Morisette et al., 2006). Global
validation networks have a very important mission in providing these validation data
needed for the further development of the remote sensing methodology.
Physically based remote sensing methods still have room for improvement, and
we have claimed in this chapter that there is especially a need to further develop reflectance models designed for coniferous forests, accounting for their complex 3D
canopy structure and the small-scale variation in needle area density. The widely discussed global trends of greening and vegetation status require accurate algorithms
for boreal forests. An efficient use of physically based forest reflectance models in
the boreal coniferous forest zone would require that the model input data include
variables such as crown shape and canopy cover, and spatial pattern of trees, to represent their clumped structure. In combination with standard forest inventory data,
these key structural parameters would form a very valuable data base. It seems reasonable to believe that with better representation of the 3D canopy structure, the
166
P. Stenberg et al.
accuracy of remotely sensed estimates of canopy biophysical characteristics would

improve considerably. Accordingly, future tasks of high priority include means to
create realistic input data on crown shape, tree pattern and canopy cover from basic
stand variables, for example through regression models and allometric relationships.
In boreal forests, another limitation to successful application of canopy reflectance
models is a typically large proportion of non-contrasting background reflectance in
the visible and near-infrared (NIR) part of the spectra due to abundant mixed green
understory. Stand inventory data do not include spectral properties of the forests,
and there is a need for more data on the angular reflectance properties of common
forest species, including not only understory vegetation but tree components, such
as needles, shoots and bark.
Use of the p-theory is for application in large scale remote sensing mapping
methods is another question of large current interest. Can a simple parameterization model using the spectral invariants be built to mimic canopy reflectance with
sufficient accuracy? It has already been shown, and put into practice in the MODIS
LAI/FPAR algorithm, that the spectral invariants provide a powerful calculation tool
in reflectance models and preserve the law of energy conservation. However, if one
could derive specific relationships between the spectral invariants (p and pt ) and
basic canopy structural parameters such as LAI and canopy cover, it would provide
an effective tool by which the effect of the clumping of foliage at different hierarchical levels of the forest stand structure could be incorporated in forest reflectance
models.
References
Andrieu B, Baret F, Jacquemoud S, Malthus T, Steven M (1997) Evaluation of an improved version
of SAIL model for simulating bidirectional reflectance of sugar beet canopies. Remote Sens.
Environ. 60:247257
Atzberger C (2000) Development of an invertible forest reflectance model: The INFOR-Model.
In: Buchroithner (ed), A decade of Trans-European Remote Sensing Cooperation, Proc. 20th
EARSeL Symposium, 1416 June 2000, Dresden, pp 3944
Baldwin VC, Peterson KD (1997) Predicting the crown shape of loblolly pine trees. Can. J. For.
Res. 27:102107
Borel CC, Gerstl SAW, Powers BJ (1991) The radiosity method in optical remote sensing of structured 3-D surfaces. Remote Sens. Environ. 36:1344
Cescatti A (1997a) Modelling the radiative transfer in discontinuous canopies of assymmetric
crowns. I. Model structure and algorithms. Ecol. Model. 101:263274
Cescatti A (1997b) Modelling the radiative transfer in discontinuous canopies of assymmetric
crowns. II. Model testing and application in a Norway spruce stand. Ecol. Model. 101:275
284
Chandrasekhar S (1960) Radiative Transfer. Oxford University Press, London, UK
Chelle M, Andrieu B (1998) The nested radiosity model for the distribution of light within plant
canopies. Ecol. Model. 111:7591
Chen J, Pavlic G, Brown L, Cihlar J, Leblanc SG, White HP, Hall RJ, Peddle DR, King DJ,
Trofymow JA, Swift E, van der Sanden J, Pellikka PKE (2002) Derivation and validation of
Canada-wide coarse-resolution leaf area index maps using high-resolution satellite imagery
and ground measurements. Remote Sens. Environ. 80:165184
167
Chen JM, Black TA (1992) Defining leaf area index for non-flat leaves. Plant Cell Environ. 15:
421429
Chen JM, Cihlar J (1996) Retrieving leaf area index of boreal conifer forests using Landsat TM
images. Remote Sens. Environ. 55:153162
Chen JM, Leblanc SG (1997) A four-scale bidirectional reflectance model based on canopy architecture. IEEE Trans. Geosci. Remote Sens. 35:13161337
Chen JM, Leblanc SG (2001) Multiple-scattering scheme useful for geometric optical modeling.
Disney MI, Lewis P, North PRJ (2000) Monte Carlo ray tracing in optical canopy reflectance
modelling. Remote Sens. Rev. 18:163196
Disney MI, Lewis P, Saich P (2006) 3D modelling of forest canopy structure for remote sensing
simulations in the optical and microwave domains. Remote Sens. Environ. 100:114132
Eklundh L, Harrie L, Kuusk A (2001) Investigating relationships between Landsat ETM + sensor
data and leaf area index in boreal conifer forest. Remote Sens. Environ. 78:239251
Garca-Haro FJ, Sommer S (2002) A fast canopy reflectance model to simulate realistic remote
sensing scenarios. Remote Sens. Environ. 81:205227
Garca-Haro FJ, Gilabert MA, Melia J (1999) A radiosity model for heterogeneous canopies in
remote sensing. J. Geophys. Res. 104:1215912175
Gastellu-Etchegorry J, Demarez V, Pinel, V, Zagolski F (1996) Modeling radiative transfer in heterogeneous 3-D vegetation canopies. Remote Sens. Environ. 58:131156
Gerard FF, North PRJ (1997) Analyzing the effect of structural variability and canopy gaps on
forest BRDF using a geometric-optical model. Remote Sens. Environ. 62:4662
Gerstl SAW, Borel CC (1992) Principles of the radiosity method versus radiative transfer for
canopy reflectance modeling. IEEE Trans. Geosci. Remote Sens. 30:271275
Gerstl SAW, Simmer C, Powers BJ (1986) The canopy hot-spot as crop identifier. In: MSJ Damen,
G Sicco Smit, HT Verstappen (eds), Remote Sensing for Resources Development and Environmental Management. Proc. 7th International Symposium, Enschede, 2529 August 1986. AA
Balkema, Rotterdam
Goel NS (1989) Inversion of canopy reflectance models for estimation of biophysical parameters
from reflectance data. In: G Asrar (ed), Theory and Applications of Optical Remote Sensing.
Wiley, New York, pp 205251
Goel NS, Rozehnal I, Thompson RL (1991) A computer-graphics based model for scattering from
objects of arbitrary shapes in the optical-region. Remote Sens. Environ. 36:73104
Govaerts YM, Verstraete MM (1998) Raytran: a Monte Carlo ray-tracing model to compute
light scattering in three-dimensional heterogeneous media. IEEE Trans. Geosci. Remote Sens.
36:493505
Huemmrich KF (2001) The GeoSail model: a simple addition to the SAIL model to describe discontinuous canopy reflectance. Remote Sens. Environ. 75:423431
Jacquemoud S, Baret F, Andrieu B, Danson FM, Jaggard K (1995) Extraction of vegetation biophysical parameters by inversion of the PROSPECT + SAIL models on sugar beet canopy
reflectance data. Application to TM and AVIRIS sensors. Remote Sens. Environ. 52:163172
Jupp DLB, Strahler AH (1991) A hotspot model for leaf canopies. Remote Sens. Environ. 38:
193210
Knyazikhin Y, Marshak A (1991) Fundamental equations of radiative transfer in leaf canopies, and
iterative methods for their solution. In: RB Myneni, J Ross (eds), PhotonVegetation Interactions, Springer, Berlin-Heidelberg, pp 943
Knyazikhin Y, Kranigk J, Myneni RB, Panfyorov O, Gravenhorst G (1998a) Influence of smallscale structure on radiative transfer and photosynthesis in vegetation cover. J. Geophys. Res.
103(DD6):61336144
active radiation from MODIS and MISR data. J. Geophys. Res. 103(D24):3225732276
Koetz B, Baret F, Poilve H, Hill J (2005) Use of coupled canopy structure dynamic and radiative transfer models to estimate biophysical canopy characteristics. Remote Sens. Environ. 95:
115124
168
P. Stenberg et al.
Kuusk A (1991) The hot spot effect in plant canopy reflectance. In: RB Myneni, J Ross (eds),
PhotonVegetation Interactions, Springer, Berlin-Heidelberg, pp 139159
Kuusk A (1995) A fast, invertible canopy reflectance model. Remote Sens. Environ. 51:342350
Kuusk A (2001) A two-layer canopy reflectance model. J. Quant. Spectrosc. Radiat. Transfer
71:19
Kuusk A, Nilson T (2000) A directional multispectral forest reflectance model. Remote Sens.
Environ. 72:244252
Lang ARG (1991) Application of some of Cauchys theorems to estimation of surface areas
of leaves, needles and branches of plants, and light transmittance. Agric. For. Meteorol. 55:
191212
Leverenz JW, Hinckley TM (1990) Shoot structure, leaf area index and productivity of evergreen
conifer stands. Tree Physiology. 6:135149
Li X, Strahler AH (1985) Geometric-optical modeling of a conifer forest canopy. IEEE Trans.
Geosci. Remote Sens. GE-23:705721
Li X, Strahler AH (1986) Geometric-optical bidirectional reflectance modeling of a conifer forest
canopy. IEEE Trans. Geosci. Remote Sens. GE-24:906919
Li X, Strahler AH (1995) A hybrid geometric-optical radiative transfer approach for modeling
albedo and directional reflectance of discontinuous canopies. IEEE Trans. Geosci. Remote
Sens. 33:466480
Liang S, Strahler AH (guest editors) 2000. Land surface bi-directional reflectance distribution function (BRDF): Recent advances and future prospects. Special issue of Remote Sensing Reviews.
18:83511
Marklund L (1988) Biomassafunktioner for tall, gran och bjork i Sverige. Report 4 (in Swedish).
Department of Forest Survey, Swedish University of Agricultural Sciences, Umea, Sweden.
73 pp
Marshak AL (1989) The effect of the hot spot on the transport equation in plant canopies. J. Quant.
Spectrosc. Radiat. Transfer L.2:615630
Menzhulin GV, Anisimov OA (1991) Principles of statistical phytoactinometry. In: RB Myneni, J
Ross, (eds), PhotonVegetation Interactions, Springer, Berlin-Heidelberg, pp 111137
Meroni M, Colombo R, Panigada C (2004) Inversion of a radiative transfer model with hyperspectral observations for LAI mapping in poplar plantations. Remote Sens. Environ. 92:195206
Morisette JT, Baret F, Privette JL, Myneni RB, Nickeson J, Garrigues S, Shabanov N, Weiss M,
Fernandes R, Leblanc S, Kalacska M, Sanchez-Azofeifa GA, Chubey M, Rivard B, Stenberg
P, Rautiainen M, Voipio P, Manninen T, Pilant A, Lewis T, Iiames J, Colombo R, Meroni M,
Busetto L, Cohen W, Turner D, Warner ED, Petersen GW, Seufert G, Cook R (2006) Validation
of global moderate resolution LAI Products: a framework proposed within the CEOS Land
Product Validation subgroup. IEEE Trans. Geosci. Remote Sens. 44(7):18041817
Mottus M, Stenberg P, Rautiainen M (2007) Photon recollision probability in heterogeneous
forest canopies: Compatibility with a hybrid GO model. J. Geophys. Res. 112:D03104,
doi:10.1029/2006JD007445
Mottus M, Sulev M, Lang M (2006) Estimation of crown volume for a geometric radiation model
from detailed measurements of tree structure. Ecological Modelling 198(34):506514.
Myneni RB (1990) Modeling radiative transfer and photosynthesis in three-dimensional vegetation
canopies. Agric. For. Meteorol. 55:323344
Myneni RB, Ross J, Asrar G (1989) A review on the theory of photon transport in leaf canopies.
Agric. For. Meteorol. 45:1153
Myneni RB, Ross J (eds) (1991) PhotonVegetation Interactions, Springer, Berlin-Heidelberg,
565 pp
Myneni RB, Keeling CD, Tucker CJ, Asrar G, Nemani RR (1997) Increased plant growth in the
northern high latitudes from 1981 to 1991 Nature. 386 (6626):698702
169
Ni W, Li X, Woodcock C, Caetano M, Strahler A (1999) An analytical hybrid GORT model for

bidirectional reflectance over discontinuous plant canopies. IEEE Trans. Geosci. Remote Sens.
37:987999
Nilson T (1971) A theoretical analysis of the frequency of gaps in plant stands. Agric. Meteorol.
8:2538
Nilson T (1977) A theory of radiation penetration into nonhomogeneous plant canopies. In: The
Penetration of Solar Radiation into Plant Canopies, Acad. Sci. ESSR Report, Tartu, 570 (in
Russian)
Nilson T (1992) Radiative transfer in nonhomogeneous plant canopies. In: G Stanhill (ed), Advances in Bioclimatology 1. Springer, Berlin-Heidelberg, pp 5988
Nilson T, Peterson U (1991) A forest canopy reflectance model and a test case. Remote Sens.
Environ. 37:131142
Nilson T, Ross J (1997) Modeling radiative transfer through forest canopies: implications for
canopy photosynthesis and remote sensing. In: HL Gholz, K Nakane, H Shimoda (eds),
The Use of Remote Sensing in the Modeling of Forest Productivity. Kluwer, Dordrecht, The
Netherlands, pp 2360
Nilson T, Anniste J, Lang M, Praks J (1999) Determination of needle area indices of coniferous forest canopies in the NOPEX region by ground-based optical measurements and satellite images.
Agric. For. Meteorol. 9899:449462
Norman J, Jarvis P (1975) Photosynthesis in Sitka spruce (Picea sitchensis (Bong.) Carr.). V. Radiation penetration theory and a test case. J. Appl. Ecol. 12(3):839877
North PRJ (1996) Three-dimensional forest light interaction model using a Monte Carlo method.
Oker-Blom P (1986) Photosynthetic radiation regime and canopy structure in modeled forest
stands. Acta For. Fenn. 197:144
Oker-Blom P, Kellomaki S (1982) Effect of angular distribution of foliage on light absorption and
photosynthesis in the plant canopy: theoretical computations. Agric. Meteorol. 26:105116
Oker-Blom P, Kellomaki S (1983) Effect of grouping of foliage on the within-stand and withincrown light regime: comparison of random and grouping canopy models. Agric. Meteorol.
28:143155
Oker-Blom P, Smolander H (1988) The ratio of shoot silhouette area to total needle area in Scots
pine. For. Sci. 34:894906
Oker-Blom P, Lappi J, Smolander H (1991) Radiation regime and photosynthesis of coniferous
stands. In: RB Myneni, J Ross (eds), Photon-Vegetation Interactions: Applications in Optical
Remote Sensing and Plant Ecology. Springer, New York, pp 469499
Panferov O, Knyazikhin Y, Myneni RB, Szarzynski J, Engwald S, Schnitzler KG, Gravenhorst G
(2001) The role of canopy structure in the spectral variation of transmission and absorption of
solar radiation in vegetation canopies. IEEE Trans. Geosci. Remote Sens. 39:241253
Peddle DR, Johnson RL, Cihlar J, Latifovic R (2004) Large area forest classification and biophysical parameter estimation using the 5-scale canopy reflectance model in multiple-forward-mode.
Peltoniemi JI, Kaasalainen S, Naranen J, Rautiainen M, Stenberg P, Smolander H, Smolander S,
Voipio P (2005) BRDF measurement of understorey vegetation in pine forests: dwarf shrubs,
lichen and moss. Remote Sens. Environ. 94:343354
Pinty B, Gobron N, Widlowski J-L, Lavergne T, Verstraete MM (2004a) Synergy between 1-D
and 3-D radiation transfer models to retrieve vegetation canopy properties from remote sensing
data. J. Geophys. Res. 109, D21205, doi:10.1029/2004JD005214
Pinty B, Widlowski JL, Taberner M, Gobron N, Verstraete MM, Disney M, Gascon F, Gastellu
JP, Jiang L, Kuusk A, Lewis P, Li X, Ni-Meister W, Nilson T, North P, Qin W, Su L, Tang
S, Thompson R, Verhoef W, Wang H, Wang J, Yan G, Zang H (2004b) Radiation transfer
model intercomparison (RAMI) exercise: results from the second phase. J. Geophys. Res. 109,
D06210, doi:10.1029/2003JD004252
Qin W, Gerstl SAW, Deering DW, Goel NS (2002) Characterizing leaf geometry for grass and crop
canopies from hotspot observations: a simulation study. Remote Sens. Environ. 880:100113
170
P. Stenberg et al.
Rautiainen M (2005) Retrieval of leaf area index for a coniferous forest by inverting a forest reflectance model. Remote Sens. Environ. 99:295303
Rautiainen M, Suomalainen J, Mottus M, Stenberg P, Voipio P, Peltoniemi J, Manninen, T (2007)
Coupling forest canopy and understory reflectance in the Arctic latitudes of Finland. Remote
Rautiainen M, Stenberg P (2005a) Application of photon recollision probability in coniferous
canopy reflectance simulations. Remote Sens. Environ. 96:98107
Rautiainen M, Stenberg P (2005b) Simplified tree crown model using standard forest mensuration
data for Scots pine. Agric. For. Meteorol. 128:123129
Rautiainen M, Stenberg P, Nilson T, Kuusk A (2004) The effect of crown shape on the reflectance
of coniferous stands. Remote Sens. Environ. 89:4152
Ross J (1981) The Radiation Regime and Architecture of Plant Stands. Junk Publishers, The
Hague, 391 pp
Ross KJ, Marshak AL (1988) Calculation of canopy bidirectional reflectance using the Monte
Carlo method. Remote Sens. Environ. 24:213225
Shabanov NV, Knyazikhin Y, Baret F, Myneni RB (2000) Stochastic modeling of radiation regime
in discontinuous vegetation canopies. Remote Sens. Environ. 74:125144
Shabanov NV, Wang Y, Buermann W, Dong J, Hoffman S, Smith GR, Tian Y, Knyazikhin Y, Myneni RB (2003) Effect of foliage spatial heterogeneity in the MODIS LAI and FPAR algorithm
over broadleaf forests. Remote Sens. Environ. 85:410423
Smolander S, Stenberg P (2003) A method to account for shoot scale clumping in coniferous
canopy reflectance models. Remote Sens. Environ. 88:363373
Smolander S, Stenberg P (2005) Simple parameterizations for the radiation budget of uniform
broadleaved and coniferous canopies. Remote Sens. Environ. 94:355363
Soler C, Sillion FX, Blaise F, Dereffye P (2003) An efficient instantiation algorithm for simulating
radiant energy transfer in plant models. ACM Trans. Graphics. 22:204233
Stenberg P (2006) A note on the G-function for needle leaf canopies. Agric. For. Meteorol.
Stenberg P, Rautiainen M, Manninen T, Voipio P, Smolander H (2004) Reduced simple ratio better
than NDVI for estimating LAI in Finnish pine and spruce stands. Silva Fennica. 38(1):314
Thompson RL, Goel NS (1998) Two models for rapidly calculating bidirectional reflectance: Photon spread (ps) model and statistical photon spread (sps) model. Remote Sensing Reviews
16:157207
Tian Y, Zhang Y, Knyazikhin Y, Myneni RB, Glassy JM, Dedieu G, Running SW (2000) Prototyping of MODIS LAI and FPAR algorithm with LASUR and LANDSAT data. IEEE Trans.
Tian Y, Wang Y, Zhang Y, Knyazikhin Y, Bogaert J, Myneni RB (2002) Radiative transfer based
scaling of LAI retrievals from reflectance data of different resolutions. Remote Sens. Environ.
84:143159
Verhoef W (1984) Light-scattering by leaf layers with application to canopy reflectance modeling
the SAIL model. Remote Sens. Environ. 16:125141
Verhoef W (1985) Earth observation modeling based on layer scattering matrices. Remote Sens.
Environ. 17:165178
Verstraete MM, Pinty B, Dickinson RE (1990) A physical model of the bidirectional reflectance of
vegetation canopies. Part 1: theory. J. Geophys. Res. 95:1176511775
Verstraete M, Pinty B, Myneni R (1996) Potential and limitations of information extraction on the
terrestrial biosphere from satellite remote sensing. Remote Sens. Environ. 58:201214
Wang Y, Buermann W, Stenberg P, Smolander H, Hame T, Tian Y, Hu J, Knyazikhin Y, Myneni
RB (2003) A new parameterization of canopy spectral response to incident solar radiation: case
study with hyperspectral data from pine dominant forest. Remote Sens. Environ. 85:304315
Wang YP, Jarvis PG (1990) Influence of crown structural properties on PAR absorption, photosynthesis, and transpiration in Sitka spruce: application of a model (MAESTRO). Tree Physiology.
7:297316
Warren Wilson J (1967) Stand structure and light penetration. III. Sunlit foliage area. J. Appl. Ecol.
4(1):159165
171
Weiss M, Baret F (1999) Evaluation of canopy biophysical variable retrieval performances from
the accumulation of large swath satellite data. Remote Sens. Environ. 70:293306
Weiss M, Baret F, Myneni RB, Pragnèrea A, Knyazikhin Y (2000) Investigation of a model inversion technique to estimate canopy biophysical variables from spectral and directional reflectance data. Agronomie. 20:322
Weiss M, Baret F, Smith GJ, Jonckheere I, Coppin P (2004) Review of methods for in situ leaf
area index (LAI) determination. Part II. Estimation of LAI, errors and sampling. Agric. For.
Meteorol. 121:3753
Welles JM, Norman JM (1991) Photon transport in discontinuous canopies: a weighted random approach. In: RB Myneni, J Ross (eds), Photon-Vegetation Interaction, Springer, BerlinHeidelberg, pp 389414
Widen N (2004) Assessing the accuracy of land surface characteristics estimated from multiangular remotely sensed data. Int. J. Remote Sens. 25:11051117
Yoder B, Pettigrew-Crosby R (1995) Predicting nitrogen and chlorophyll content and concentrations from reflectance spectra (4002500 nm) at leaf and canopy scales. Remote Sens. Environ.
53:199211
Zarco-Tejada PJ, Rueda CA, Ustin SL (2003) Water content estimation in vegetation with MODIS
reflectance data and model inversion methods. Remote Sens. Environ. 85:109124
Chapter 7
Estimating Canopy Characteristics

from Remote Sensing Observations:
Review of Methods and Associated Problems
Frederic Baret and Samuel Buis
Abstract This article describes the methods and problems associated to the estimation of canopy characteristics from remote sensing observations. It is illustrated
over the solar spectral domain, with emphasis on LAI estimation using currently
available algorithms developed for moderate resolution sensors. The principles of
algorithms are first presented, distinguishing between canopy biophysical and radiometric data driven approaches that may use either radiative transfer models or
experimental observations. Advantages and drawback are discussed with due attention to the operational character of the algorithms. Then the under-determination
and ill-posedness nature of the inverse problem is described and illustrated. Finally,
ways to improve the retrieval performances are presented, including the use of prior
information, the exploitation of spatial and temporal constraints, and the interest in
using holistic approaches based on the coupling of radiative transfer processes at
several scales or levels. A conclusion is eventually proposed, discussing the three
main components of retrieval approaches: retrieval techniques, radiative transfer
models, and the exploitation of observations and ancillary information.
7.1 Introduction
Many applications require an exhaustive description of the spatial domain of interest
that may cover a large range of scales: from the very local one corresponding to precision agriculture where cultural practices are adapted to the within field variability,
through environmental management generally approached at the landscape scale, up
to biogeochemical cycling and vegetation dynamics investigated at national, continental and global scales. Most of these applications are using our knowledge on the
F. Baret and S. Buis
UMR1114, INRA-CSE, 84 914 Avignon, France
baret@avignon.inra.fr
173
174
F. Baret, S. Buis
main physical, chemical and biological processes involved such as energy balance,
evapotranspiration, photosynthesis and respiration. This knowledge is encapsulated
into a variety of surface process models. However, to account for the spatial heterogeneity observed at all scales, dedicated imaging systems are required to get a
distributed description of surface characteristics within the domain of interest. By
its capacity to cover exhaustively large space areas, remote sensing provides a very
pertinent answer to those requirements. However, remote sensing observations sample the radiation field reflected or emitted by the surface, and thus do not provide
directly the biophysical characteristics required by the models for describing some
state variables of the surface. An intermediate step is therefore necessary to transform the remote sensing measurements into estimates of the surface biophysical
characteristics.
Many methods have been proposed to retrieve surface characteristics from remote sensing observations. They span from simple empirical ones with calibration
over experimental data sets, up to more complex ones based on the use of radiative transfer models. Radiative transfer models summarize our knowledge on the
physical processes involved in the photon transport within vegetation canopies or
atmosphere, and simulate the radiation field reflected or emitted by the surface for
given observational configuration, once the vegetation and the background as well
as possibly the atmosphere are specified. Retrieving canopy characteristics from the
radiation field as sampled by the sensor aboard satellite needs to invert the radiative transfer model.
This article aims at presenting the state of the art in the estimation of surface
characteristics from remote sensing observations. Although this is a very general
problem in remote sensing, it will be illustrated by examples taken in the solar domain (4002,500nm), with emphasis put on the current operational algorithms that
are mainly used for medium resolution sensors such as MODIS, MERIS, AVHRR,
VEGETATION, POLDER and SEAWIFS. Among the possible canopy characteristics accessible from remote sensing in the reflective solar domain, we will focus on
leaf area index (LAI), defined as half the developed area of green elements per unit
horizontal soil (Stenberg, 2006). As a matter of fact, LAI is one of the key canopy
state biophysical variables required by many process models to describe energy and
mass exchanges in the soil/plant/atmosphere system.
7.2 Principles of Biophysical Variable Retrieval Algorithms

Remote sensing data result from radiative transfer processes within canopies that
depend on canopy variables, and observational configuration (wavelength, view and
illumination directions). Canopy variables include the variables of interest for the
applications such as LAI, and the other variables that are not of direct use for the
applications but that influence the radiative transfer, such as soil background properties. The causal relationship between the variables of interest and remote sensing data corresponds to the forward (or direct) problem (Fig. 7.1). They could be
7 Estimating Canopy Characteristics from Remote Sensing Observations
175
Other variables
Radiative
transfer
Variables of
interest
Inverse problem
Observation
configuration
Forward problem
Remote sensing
data
Retrieval
algorithm
Prior information
Fig. 7.1 Forward (solid lines) and inverse (dashed lines) problems in remote sensing
either described through empirical relationships calibrated over experiments or using radiative transfer models based on a more or less close approximation of the
actual physical processes. Conversely, retrieving the variables of interest from remote sensing measurements corresponds to the inverse problem, i.e., developing algorithms to estimate the variables of interest from remote sensing data as observed
in a given configuration. Prior information on the type of surface and on the distribution of the variables of interest can also be included in the retrieval process to
improve the performances as we will see later.
A panoply of retrieval techniques currently used have been reviewed in the early
1990s by several authors (Asrar et al., 1989; Goel, 1989; Pinty and Verstraete, 1991)
and more recently by Kimes et al. (2000) and Liang (2004). They can be split into
two main approaches (Fig 7.2) depending if the emphasis is put on remote sensing data (radiometric data) or on the variables of interest to be estimated (canopy
biophysical variables).
7.2.1 Canopy Biophysical Variables-driven Approach

The approach requires first to calibrate the inverse model: a parametric model representing the inverse model is adjusted over a learning data set (Fig. 7.2, left). It
mainly consists in adjusting the parameters to fit a response surface between reflectance values and the corresponding canopy variables of interest (LAI in this example). Once calibrated, the parametric model is run to compute the variables of
interest from the observed reflectance values. The learning data set can be generated either using simulations of radiative transfer models, or based on concurrent
experimental measurements of the variables of interest and reflectance data.
176
F. Baret, S. Buis
Radiometric data
driven approach
Biophysical variables
driven approach
Learning data set
LAI
RT
process
Reflectance
LAI
LAI*
RT
process
Reflectance
Inverse
Model
Reflectance
Measurement
Parameters
Fig. 7.2 The two main approaches used to estimate canopy characteristics from remote sensing
data for LAI estimation. On the left side the approach focusing on the biophysical variables showing the calibration of the inverse model. Once the inverse model is calibrated it can be applied
using the measured reflectance as input. On the right side, the approach focusing on radiometric
data showing the solution search process leading to the estimated LAI value, LAI . represents
the cost function to be minimized over the biophysical variables (left) or over the radiometric
data (right)
7.2.1.1 Calibration over Experimental Data Sets

This was the first approach historically used, the reflectance in few bands being
generally combined into vegetation indices (VI) designed to minimise the influence
of confounding factors such as soil reflectance and atmospheric effects (Baret and
Guyot, 1991). The relationships between VIs and canopy variables are calibrated
over experimental observations (Asrar et al., 1984; Huete, 1988; Wiegand et al.,
1990; Wiegand et al., 1992; Richardson et al., 1992). Recently Chen et al. (2002)
used simple VIs to derive LAI estimates from AVHRR and VEGETATION across
Canada. This was extended at the global scale by Deng et al. (2006). In agreement
with several observations, these authors found that the relationships vary from one
cover type to another as illustrated by Fig. 7.3. The development of such empirical transfer functions is limited by the difficulty to get a training data base that
represents the whole range of possible conditions encountered over the targeted surfaces, i.e., combinations of geometrical configurations, type of vegetation and states
including variability in development stage and stress level, and type of background
and state (roughness, moisture). Measurement errors associated both to the variables
of interest and to radiometric data may also propagate into uncertainties and biases in the algorithm and should be explicitly accounted for Fernandes and Leblanc
(2005) and Huang et al. (2006). Further, since ground measurements having a footprint ranging from few meters to few decametres, specific sampling designs should
be developed to represent the sensor pixel. This task is obviously more difficult for
medium and coarse resolution sensors as outlined by Morisette et al. (2006). Higher
spatial resolution observations could be used to extend the local ground measurements to the actual pixel size of medium or coarse resolution sensors.
177
10
Coniferous
Deciduous
Mixed
Others
LAI
5
10
RSR VEGETATION
Fig. 7.3 Empirical relationships between LAI values as a function of the simple ratio vegetation
index (RSR) computed for VEGETATION data for four types of canopies (After Chen et al., 2002)
7.2.1.2 Calibration over Radiative Transfer Model Simulations

To avoid limitations associated to the empirical nature of the training data base, radiative transfer models could be used alternatively to generate a training data base.
Radiative transfer models can be used to create a data base covering a wide range
of situations and configurations. Several authors have therefore proposed replacing actual observations by numerical experiments based on radiative transfer model
simulations to calibrate empirical relationships (Sellers, 1985; Baret and Guyot,
1991; Rondeaux et al., 1996; Leprieur et al., 1994; Banari et al., 1996; Huete et al.,
1997; Verstraete and Pinty, 1996). Based on these principles, operational algorithms
developed for medium resolution sensors are currently used: MGVI for MERIS
(Gobron et al., 2000) further extended to other sensors, MODIS back-up algorithm
based on NDVI (Knyazikhin, 1999), POLDER algorithm based on DVI computed
from bidirectional reflectance factor (BRF) measurements normalized to a standard
geometrical configuration (Roujean and Lacaze, 2002). Nevertheless, although quite
often effective, VIs are intrinsically limited by the empiricism of their design and the
small number of bands concurrently used (generally 23). This might not be a major
problem for fAPAR and fCover variables that are relatively simple to estimate, but
would be more difficult for variables such as LAI or chlorophyll content (Cab ) showing higher level of non linearity with reflectance measurements (Weiss et al., 2000).
The efficient interpolation capacity of neural network (NNT) can be exploited to
adjust surface responses (Leshno et al., 1993). Several authors have proposed such
an approach since the beginning of the 1990s (Smith, 1992; Smith, 1993; Atkinson
and Tatnall, 1997; Kimes et al., 1998; Abuelgasim et al., 1998; Gong et al., 1999;
178
F. Baret, S. Buis
Danson et al., 2003). Neural networks were compared with a specific implementation of multiple regression, the projection pursuit regression, and were concluded
to achieve very similar performances (Fang and Liang, 2005). Baret et al. (1995)
demonstrated that NNT used with individual bands were performing better than
classical approaches based on vegetation indices especially when calibrated with
radiative transfer model simulations rather than with experimental observations.
Weiss et al. (2002a) validated such techniques over a range of crops for estimating
the main canopy biophysical parameters LAI and fCover from airborne POLDER
instrument. Recently, several authors developed operational products for medium
resolution sensors, starting from top of canopy level: Lacaze (2005) for POLDER,
Bacour et al. (2006) for MERIS, and Baret et al. (2007) for VEGETATION instruments. Baret et al. (2006b) proposed an operational algorithm from the MERIS top
of atmosphere data by coupling an atmospheric radiative transfer model to the surface one, exploiting explicitly 13 over the 15 bands of MERIS.
Several ways may be used to build a data set for training empirical relationships
depending on the performances targeted. Evaluation of the performances of an algorithm is generally achieved by computing the Root Mean Square Error (RMSE)
value over a test data base made of representative cases. Best performances will
therefore be obtained when the variables in the training data base are distributed
similarly to those in the testing one, i.e., close to the actual distribution of the variables: the coefficients of the empirical transfer function will be optimized for these
conditions, and uncertainties will be minimal for the most frequent cases. Although
achieving poorer performances in term of RMSE, a more even distribution of the
uncertainties may be alternatively obtained using uniform distributions of the variables. Note that, for a given number of cases simulated in the training data base, the
density of cases that populate the space of canopy realization may rapidly decrease
as a function of the number of required variables. Experimental plans may be used
in this situation as proposed by Bacour et al. (2002b), in order to focus on the first
order effects and interactions. Additionally, Baret et al. (2006b) proposed to steamline the data base in the reflectance space by retaining the cases that belong both to
the simulated and actual remote sensing measurements spaces (Fig. 7.4). This allows discarding cases that were simulated but not actually observed. Conversely, it
allows also identifying cases which are observed but not simulated. This is achieved
by first compiling a large data base of reflectance measurements that should be representative of the possible situations available. Then the reflectance mismatch is
Simulations
Cases not represented
in the measured database
Actual measurements
Cases not represented
in the simulated database
Training database: selection
of cases in the intersection space
Fig. 7.4 Streamlining the simulated training data set by comparison to actual measurements. The
intersection between the space of simulated radiometric data (in dark gray) with that of the actual
measurements (in light gray) is used as the training data base
179
computed for each case in the simulated data base: it is the minimum RMSE value
computed between the reflectance in the simulated data base and the ensemble of actual measurements. A threshold corresponding to the uncertainties in the radiometric
measurements is then used to decide whether a simulated case is rejected from the
training data base. Additional criterions could be used to streamline the training data
base, based on the expected consistency between several products such as LAI and
fAPAR as proposed by Bacour et al. (2006).
Although the use of radiative transfer models appears very appealing, this approach is however limited by several aspects. The first one is the capacity of the
models to get a faithful description of the radiative transfer in canopies. Up to now,
most radiative transfer models used are computer efficient ones allowing populating
large training data base within few hours/days with a single regular computer. They
generally correspond to simple description of canopy architecture which may not
represent the actual one, particularly regarding the clumped nature of many vegetation types. This leads to model uncertainties that may dominate all other sources of
uncertainties for some of the vegetation types. Recent advances in modeling more
complex canopy architecture (e.g., Soler et al., 2001; Lewis et al., 2004) offer great
potential for improvement. However, the second limitation will probably counterbalance these advancements: building a realistic training data set requires a fair description of the distribution and co-distribution of the corresponding architectural
variables to define the actual space of canopy realization. For the simplest radiative transfer models (e.g., Verhoef, 1984; Kuusk, 1995; Gobron et al., 1997) at least
three architectural variables are required (LAI, leaf angle distribution function and
size of the leaves relative to canopy height), the distribution of which being very
poorly known. This is even more difficult when using more complex and realistic
architectural description that requires more variables.
Note that in these approaches based on radiative transfer model simulations, radiometric measurements uncertainties have to be added to the simulations when
building up the training data base. This allows more robustness within the training
process and thus improved retrieval performances. Accounting for these uncertainties is also critical when large differences exist among bands used or when these
uncertainties are strongly correlated.
Canopy biophysical variables driven approaches present the advantage of being
very flexible. For example, estimates of biophysical variables from one sensor could
be used to constitute the training data base for another sensor. This could be applied
over high spatial resolution products that are aggregated to coarser spatial resolution to generate an appropriate training data base. This could also apply to generate
consistent products between sensors.
7.2.2 Radiometric Data-driven Approach

While the previous approach was focusing on minimizing the distance between the
variables retrieved from the inverse model and those from the training data set,
the alternative approach is based on finding the best match between the measured
180
F. Baret, S. Buis
reflectance values and those either simulated by a radiative transfer model or stored
within a database made of experimental observations. No proper calibration step
is required in this approach. However, several ingredients of these techniques are
difficult to evaluate (uncertainties, parameters of the search algorithm) and need
generally some tuning over a prototyping data set. The performances of the approach will both depend on the minimization algorithm itself and on the level of
ill-posedness of the inverse problem as a function of measurement configuration
and model and measurement uncertainties. Several minimization techniques have
been used: classical iterative optimization, simulated annealing (Bacour, 2001), genetic algorithms (Fang et al., 2003; Renders and Flasse, 1996), look up tables and
Monte Carlo Markov Chains (Zhang et al., 2005). However, classical iterative optimization techniques (OPT) and look up tables (LUT) have been the most widely
used and will be described with more details below.
7.2.2.1 Iterative Optimisation (OPT)

This classical technique consists in updating the values of the unknown input biophysical canopy radiative transfer model variables until the simulated reflectance
closely fit the corresponding measurements (Goel and Deering, 1985; Kuusk, 1991a
and 1991b; Goel, 1984a and b; Pinty et al., 1990; Jacquemoud et al., 1995; Privette
et al., 1996; Bicheron and Leroy, 1999; Combal et al., 2000; Bacour et al., 2002a;
Combal et al., 2002). A good review on optimization methods used in remote sensing for land applications can be found in Bacour (2001). The goodness of fit between
measured and simulated reflectance spectra is quantified by a cost function (J) that
may account explicitly for measurements and model uncertainties. The cost function
may be theoretically derived from the maximum likelihood (Tarentola, 1987). When
no prior information is available and when uncertainties associated to each configuration used are assumed independent and gaussian, J is assessed using norm L2,
i.e., sum over the N observational configurations of the square of the difference be weighed by the
tween the measured reflectance values (R) and those simulated (R),
variance (2 ) associated to both reflectance measurements and model uncertainties:
(Rn R n )2
n2
n=1
N
J=
(7.1)
However, because of the difficulty to provide an estimate of 2 , several approximations have been used as shown in Bacour (2001). It spans from the simple ones such
as norm L1 to norm L2 with no weighing of the configurations, up to more complex
based on some modeling of the variance term (Table 7.1).
The main limitation of OPT techniques is twofold. (1) Firstly, the algorithm
might converge to a local minimum of the cost function that could be far away
from the global one expected to correspond to the actual solution. This can be partly
avoided by using a range of initial solutions, coupled with constraints on the range
of variation of the variables to be estimated. The use of a priori information in the
181
Table 7.1 The cost functions (J) used in several studies dealing with radiative transfer model inversion for canopy biophysical variables retrieval. N is the number of configurations (bands and
directions); R n and Rn being respectively the simulated and measured reflectance values for configuration n. and are the zenith and relative azimuth view angles
Cost function
N
J=
References
Rn R n
Rn
Gao and Lesht, 1997)
n=1
N

J = RnRnRn
n=1
N
J=
n=1
N
Rn R n
Rn +R n
(Qiu et al., 1998)
2
(Gobron et al., 1997)
J = (Rn R n )2
(Goel and Thompson, 1984; Pinty et al., 1990;

Privette et al., 1996; Braswell et al., 1996; Jacquemoud et al., 2000; Combal et al., 2002)
n=1
N
J=
n=1
N
Rn R n
Rn
J = n
n=1
2
Rn R n
Rn
2
(Nilson and Kuusk, 1989; Kuusk, 1991a and b;

Bicheron and Leroy, 1999; Weiss et al., 2000)
; n =
cos( sin( ))+1

2
(Bacour et al., 2002a)
cost function generally improves the convexity of the error surface, which is critical
as we will see later (Combal et al., 2002). The descent algorithm may also limit
the trapping in a local minimum by reducing the rate of descent. However, a compromise has to be chosen between rapid convergence achieved with large descent
rate, and limiting the probability of falling in a local minimum achieved with a slow
descent rate. Further, the optimization algorithm may sometimes lack of robustness
due to numerical problems occurring generally with very small values of J. The
criterion used to stop the iterations is in addition not always easy to adjust, requiring some preliminary tests (Bonnans et al., 2006). (2) Secondly, the OPT algorithm
requires large computer resources because of its iterative nature. However, there
are ways to speed up the process by limiting the number of model runs for each
iteration using the adjoint model that provides an analytical expression of the gradient of the cost function (Lauvernet et al., 2007). Nevertheless, OPT techniques are
still difficult to use routinely and exhaustively over large images, although image
segmentation may help reducing significantly the number of pixels to process, the
optimization process being restricted to a limited set of representative pixels. Note
that these techniques allow getting some estimates of the uncertainties associated
to the solution under some assumptions. However, the distribution of the solution
will be here always unimodal, conversely to what could be achieved with the other
radiometric driven approaches.
The main advantage of iterative optimization methods is their flexibility, allowing retrieving canopy characteristics from several observational configurations. It is
even possible to invert radiative transfer models concurrently over several pixels.
This opens great potentials for exploiting additional temporal or spatial constraints
as we will see later.
182
F. Baret, S. Buis
7.2.2.2 Look Up Tables

This is conceptually the simplest technique, although its implementation is not trivial (Weiss et al., 2000). It is the basis of the MODIS and MISR LAI and fAPAR
products (Knyazikhin et al., 1999). Firstly a large data base (the Look Up Table,
LUT) is generated, consisting of sets of input variables of the canopy radiative transfer model used. Then, the corresponding reflectance values are simulated. The LUT
can alternatively be based on experimental observations, although this requires a
very good sampling of the space of canopy realization. Once the LUT has been
generated, finding the solution for a given set of reflectance measurements consists
in selecting the closest cases in the reflectance table according to a cost function,
and then extracting the corresponding set of canopy biophysical variables. Note that
the distribution of the solution could be obtained by accounting for the uncertainties associated to the reflectance values as discussed by Knyazikhin et al. (1998a
and b).
This technique overcomes some of the limitations of iterative optimization techniques. As a matter of fact, the search for the solution is global here, leading to
the true minimum if the space of canopy realisation is sufficiently well sampled.
Note that for generating the LUT, the space of canopy realization has to be sampled
to represent the surface response, i.e., with better sampling where the sensitivity
of reflectance to canopy characteristics is the higher (Weiss et al., 2000; Combal
et al., 2002). This is different from the sampling of the training data base required
in canopy biophysical variables driven approaches.
The implementation of a LUT technique in algorithmic operational chains is very
efficient because the radiative transfer model is run off-line. However, LUT techniques require a fixed number of inputs unless having very large tables that could
be more difficult to manipulate. In addition, the way the solution is defined is not
always based on solid theoretical background. The cases selected as possible solutions are either defined as a fraction of the initial population of cases (after tests
and trials) such as in Weiss et al. (2000) or Combal et al. (2002). It can be also
defined by a threshold corresponding to measurement and model uncertainties as in
Knyazikhin et al. (1998a and b).
7.2.2.3 Bayesian Methods: Importance Sampling and MCMC

Alternative methods are available which are based on statistical backgrounds: MonteCarlo Markov Chains (MCMC) and Importance Sampling (IS) (Makowski D.,
J. Hiller, et al., 2006). These two Bayesian methods approximate the posterior
distribution, i.e., the distribution of the variables when the reflectance measurement
is known. Although very little attention has been paid to these techniques at the
exception of Zhang et al. (2005) who used with success the MCMC MetropolisHastings algorithm with MODIS data. However, Metropolis-Hastings algorithm is
an iterative process that might not be well suited for operational applications at
183
large scale, similarly to OPT methods. Conversely, IS methods that do not require
multiple iterations might be efficient for this purpose and need to be properly evaluated for remote sensing applications.
7.3 The Under-determined and Ill-posed Nature of the Inverse

Problem in Remote Sensing
7.3.1 Under-Determination of the Inverse Problem
Estimating biophysical variables from remote sensing measurements is often an
under-determined problem: the number of unknowns is generally larger than the
number of independent radiometric information remotely sampled by sensors. In
the case of a simple canopy radiative transfer model such as SAIL (Verhoef, 2002),
canopy reflectance at the top of canopy ( toc ) for a given illumination and view
geometry ( s, v, ) is simulated (Eq. (7.2)) using three variables describing
canopy structure that do not depend on wavelength (LAI, average leaf angle (ALA)
and hot spot parameter (hot) as modelled by Kuusk (1995)), and leaf reflectance
(refl) and transmittance (tran) as well as soil reflectance (Rs) that obviously depend
on wavelength ( ).
toc (, s, v, )
= CAN(LAI, ALA, hot, re f l( ), tran( ), Rs(, s, v, ), s, v, ) (7.2)
Several studies report that canopy (and soil) bidirectional reflectance distribution
function (BRDF) could be decomposed using empirical or semi-empirical orthogonal functions with generally 24 kernels (Lucht, 1998; Breon et al., 2002; Weiss
et al., 2002). Therefore, 79 characteristics (3 canopy structure, 2 leaf properties
[refl, tran] input variables and the 24 terms describing soil BRDF, Rs(, s, v, )
have to be estimated out of a maximum of 4 independent information derived from
BRDF measurements in a single band. Retrieval of canopy characteristics from
BRDF measurements in a single band is therefore not possible without introducing
other information in the system, particularly when soil background plays a significant role, i.e., for low to medium LAI values.
Similar observations are made when considering the reflectance spectral variation: leaf spectral properties may be described by a dedicated model such as
PROSPECT (Jacquemoud and Baret, 1990) requiring at least 5 input variables: mesophyll structure parameter (N), chlorophyll (Cab ), dry matter (Cdm ), brown pigment
(Cbp ) and water (Cw ) contents:
[re f l( ), tran( )] = LEAF(N, Cab , Cdm , Cbp , Cw , )
(7.3)
184
F. Baret, S. Buis
Soil reflectance Rs( , s, v, ) may be described by a model such as that proposed by Jacquemoud et al. (1992) and derived from that of Hapke (1981). It
requires a single scattering albedo ( ) that varies with wavelength and soil composition, between 1 to 4 phase function coefficients (i ), and a roughness parameter
(r). According to Price (1990), soil spectral variation, may be approximated as a
linear combination of 210 end-members. This is assumed to apply similarly to the
spectral variation of the single scattering albedo with weigh w j and end members
j ( ):
( ) = w j j ( )
(7.4)
j
The whole soil spectral and directional reflectance field could subsequently be simulated with at least five parameters:
Rs( , s, v, ) = SOIL([w j ], [i ], r, , s, v, )
(7.5)
Consequently, the whole spectral and directional top of canopy reflectance field
could therefore be modelled by coupling together the soil, leaf and canopy reflectance models, which leads to at least 13 input variables. These 13 unknowns
have to be estimated from the information content in remote sensing measurements.
Most of currently available sensors for which operational biophysical products are
available have a relatively small number of configurations: from two for AVHRR
(red and near infrared bands), to 15 bands for MERIS (VIS and NIR) and MODIS
(VIS, NIR, SWIR) with several bands dedicated to particular atmosphere, cloud,
snow/ice, or ocean characteristics. In the case of multidirectional sensors, the
number of configurations may be larger as in the case of MISR (36 configurations = 9 cameras 4 bands), or POLDER (84 configurations = 14 directions 6
bands). However, the actual dimensionality of remote sensing measurements is
much smaller than the number of available configurations considering the relatively
high level redundancy between bands (Price, 1994; Price, 1990; Liu et al., 2002;
Green and Boardman, 2001) and directions (Zhang et al., 2002a and b; Weiss et al.,
2002b). Although further investigation is required to better quantify the actual
dimensionality of remote sensing observations, it is clear that retrieval of surface
characteristics from reflectance measurements is an under-determined problem in
many cases. Improving retrieval performances will require introducing ancillary
information and constraints in the system.
7.3.2 Evidence of the Ill-posed Problem

A problem is well posed if and only if its solution exists, is unique, and depends
continuously on the data (Garabedian, 1964). Several authors have reported that the
inverse problem in remote sensing is ill-posed (Knyazikhin et al., 1999; Combal
et al., 2001; Baret et al., 2000) because of its under-determination and uncertainties
attached to models and measurements. In addition, models may incorporate sets of

80
Cab (g.cm-2)
Reflectance
0.3
0.2
0.1
0
400
185
600
800
Wavelength (nm)
60
40
20
0.5
1
LAI
1.5
Fig. 7.5 Actual reflectance measurements (left plot, solid lines representing the mean and standard deviations) and the corresponding closer simulations achieved with a simple turbid medium
radiative transfer model (the series of dots). On the right, the input LAI and Cab (the + symbols)
variables used to simulate the reflectance spectra shown on the left plot. The actual LAI and Cab
measurements are displayed with their associated confidence interval (bold line corresponding to
1 standard deviation). Data acquired over a sugar beet experiment conducted in 1990
variables that appear always in combinations such as products between variables. In

these conditions, very similar reflectance spectra simulated by a radiative transfer
model (Fig. 7.5, left) may correspond to a wide range of solutions (Fig. 7.5, right).
In the case illustrated by Fig. 7.5, high correlation is found between LAI and those
leaf chlorophyll content estimated values. This compensation between variables was
sometimes termed ambiguity (Baret et al., 1999) or equi-finality (Shoshany,
1991; Teillet et al., 1997). This may also indicate that the product LAI Cab should
be used in place of individual estimates of LAI and Cab . Although not appearing
formally in the radiative transfer model, this product is physically meaningful from
the radiative transfer processes perspective and corresponds to the actual optical
thickness of the medium (Weiss et al., 2000).
Measurement and model uncertainties may also induce instability in the solution of the inverse problem. This is particularly true for well developed canopies,
where a small variation in the measured reflectance can translate into large variation of variables such as LAI, for which reflectance saturates, i.e., is very little
sensitive to LAI variation. A proper sensitivity analysis should help quantifying interactions between input variables. A complementary sensitivity analysis conducted
over the cost function could also help evaluating the identifiability of the solution,
i.e., if output variables could be accurately retrieved from a given set of observations
(Salteli, 2004).
Regularization techniques are thus necessary to obtain a stable and reliable solution of the ill-posed inverse problem. This could be achieved both by using prior
information on the distribution of the variables, and by exploiting some constraints
on the variables. These two issues will be investigated separately in the following.
186
F. Baret, S. Buis
7.4 Improving the Retrieval Performances

7.4.1 Using Prior Information
If no remote sensing measurement is available, the best estimates of the variables
would come from the prior information on their distribution (Fig. 7.6d), capitalizing, all the knowledge coming from bibliography, past experiments or experts.
Conversely, when a radiative transfer model is available along with remote sensing
measurements, the variables can be estimated by inverting the RT model without
using any prior information. This will be illustrated using a simple example: estimating LAI from NDVI vegetation index. In this case the RT model consists in an
analytical relationship as proposed by Baret and Guyot (1991):
NDV I = NDV I + (NDV Is -NDV I ) eKLAI
1
0.8
0.8
0.6
0.6
0.6
0.4
0.2
0
NDVI
0.8
NDVI
NDVI
0.4
0.2
Remote Sensing
measurements (NDVI)
0.05
0.02
PDF
4
LAI
Estimates from RS
measurements and RT model
without prior information
4
LAI
0.05
Prior information on LAI
0.4
0.2
Radiative Transfer
model: NDVI=f(LAI)+e
0
0
0.04
(7.6)
NDVI estimates
with PI
0.04
0.8
0.04
0.03
0.6
0.03
0.02
0.4
4
LAI
NDVI estimates
without PI
0.02
Estimates from RS
measurements and RT model
with prior information
0.2
0.01
0
PDF
NDVI
PDF
Prior Information
4
LAI
0.01
4
LAI
Fig. 7.6 Estimation of canopy variables by combining remote sensing measurements, radiative
transfer model and prior information. All these pieces of information are represented by their probability distribution function (PDF): (a) PDF of remote sensing measurements in the simple case of
NDVI; (b) PDF of RT model simulations (NDV I = f(LAI)) accounting for model uncertainties; (c)
PDF of LAI as retrieved from RT model and NDVI measurement and their associated uncertainties, without using prior information; (d) PDF of LAI used as prior information; (e) Computation
of LAI PDF as estimated from NDVI measurements and RT model, using prior information on LAI;
(f) PDF of the solution (posterior distribution) when using only prior information (idem as plot d),
using RT model and NDVI measurements and their associated uncertainties only, and using all the
information available (RT model and NDVI measurements and their associated uncertainties and
prior information). The three contour plots (b, c, e) are coded from white to black for zero to max
PDF values with the same gray scale. Very simple assumptions on uncertainties models and values
are used here just for illustration
187
with NDV Is and NDV I being respectively the bare soil and asymptotic values
of NDVI, and K an extinction coefficient (K = 0.8). However, uncertainties are
associated both with remote sensing measurements (Fig. 7.6a NDV Ir = (0.8, 0.1)
where (x, 2 ) means a Gaussian distribution with mean x and variance 2 ) and
the RT model (Fig. 7.6b RT model represented by Eq. (7.2) with a Gaussian noise
(0, 0.1)). Accounting for these uncertainties in the form of the corresponding
probability distribution function (PDF) allows deriving the PDF of the estimated
variable (Fig. 7.6c). The small sensitivity of NDVI to LAI as compared to measurement and model uncertainties induce a relatively broad PDF for the larger
LAI values (Figs. 7.6c, f). This corresponds to an ill-posed problem, where a wide
range of possible solutions match very similar measurements. The combination of
RT model, remote sensing measurements and prior information on the variables
(here LAI = (2, 1.5) allows getting more reliable solutions accounting for all the
sources of information available in an optimal way (Figs. 7.6e, f).
The example provided above for a measurement value of NDV I = 0.8 could be
extended to the whole range of NDVI values. It shows that the mode of the distribution of the solution corresponding to the maximum likelihood (maximum of
the PDF) strongly depends on the type of input information used (Fig. 7.7, left).
When only prior information is used, the mode stays constant and obviously independent from measurements. When RT model and measurements are used with their
uncertainties, the LAI mode is generally close to the values obtained without considering uncertainties, assuming perfect model and measurements. However, over
the saturation domain corresponding to NDVI values higher than 0.85, accounting
for the uncertainties provides lower modal values because of the non linearity of the
LAI
4
3
Mode
Standard deviation
RT model
2.5
RT model and
measurement
uncertainties
RT model and
measurement
uncertainties
RT model and
measurement
uncertainties and
prior information
LAI
1.5
Prior information
2
Prior information
1
0
0
RT model and
measurement
uncertainties and
prior information
0.5
0.5
NDVI
0.5
NDVI
Fig. 7.7 Mode (plot on the left) of the distribution of the solution (LAI) of the inverse problem
as a function of the measured value (NDVI). The mode corresponds to the maximum PDF value,
i.e., the maximum likelihood. Four estimates are displayed: using only prior information; using RT
model (LAI = RT 1 (NDV I)) assumed to be perfect with perfect measurements (no uncertainities
accounted for); using RT model and measurements with their associated uncertainities; using RT
model and measurements with their associated uncertainities and prior information. On the right,
the standard deviation of the distribution of the solution is also displayed for the several cases. The
case with perfect RT model and measurements is not displayed here because its standard deviation
is null by definition
188
F. Baret, S. Buis
model. When prior information is used in addition to RT model and remote sensing
measurements, differences of LAI mode are marginal over the domain where NDVI
is sensitive enough to LAI. Conversely, over the saturation domain, LAI modal values are always lower (closer to prior information value) than those observed when
not using prior information which would lead to a bias. However, the interest of
using the prior information is clearly demonstrated when considering the standard
deviation of the distribution of the solutions (Fig. 7.7, right).
Introducing prior information in the inversion process provides a very significant
reduction of the variability of the posterior distribution. This is obviously more important for the larger NDVI values corresponding to the saturation domain: in this
case, very large scattering of the retrieved LAI values is expected when no prior
information is used. Although the maximum likelihood is often used as the solution, the variability within the posterior distribution as represented by its standard
deviation appears to be very informative and useful.
The theory behind this Bayesian approach has been extensively described by
Tarantola (2005). When restricting the solution as that maximizing the likelihood,
i.e., corresponding to the maximum of the PDF, a general formulation of the cost
function may be derived under Gaussian distribution assumption:
+ (V Vp )t C1 (V Vp )
t W 1 (R R)
J = (R R)

Radiometricinformation
(7.7)
Prior information
where V is the vector of the input biophysical variables estimates, R corresponds to

the vector of remote sensing measurements of dimension N (the number of bands
and directions used), R is the vector of the simulated reflectance corresponding to
the solution V (the vector of canopy biophysical variables) and Vp the vector of
prior values of biophysical variables. Matrices W and C are the covariance matrices
characterizing respectively the radiometric and model uncertainties, and that of the
prior information. Note that the first part of this equation corresponds to the distance
between the measured and the simulated radiometric data. It simplifies into Eq. (7.1)
if the covariance terms of matrix W are assumed to be zero, i.e., measurement and
model uncertainties are independent between configurations. The second part of
Eq. (7.7) corresponds to the distance between the values of the estimated variables
and those of the prior information. Very few studies are currently based on this
formulation of the cost function where prior information is explicitly used (Combal
et al., 2002).
Implementing the cost function as expressed by Eq. (7.7) requires some reasonable estimates of covariance matrices W and C as well as prior values Vp . The terms
of W should reflect both measurement and model uncertainties. While some rough
estimates of measurement uncertainties could be derived from the sensor specification, model uncertainties are far more difficult to estimate. Further, they may depend significantly on the type of situation considered, such as low or high vegetation amount. Even more difficult to estimate, are the covariance terms in W : measurement and model uncertainties may have important structure that translates into
high covariance terms which are however very poorly known. When using simul-
189
taneously a large number of configurations as in the case of hyperspectral observations, these covariance terms will be very important to account for: they will allow
weighing properly the several configurations used. The difficulty to estimate the
covariance terms explains why a small number of configurations is often selected
when a larger number is available as in the case of hyperspectral and/or directional
observations.
Retrieval approaches should be used within well defined and if possible restricted
domains. Larger domains will generally degrade retrieval performances since the
prior information will be looser defined, similarly to the covariance matrices characterizing uncertainties. However, splitting the whole domain into a set of subdomains may introduce problems due to misclassification and attribution errors as
observed by Lotsch et al. (2003), and artefacts at the limit between classes translating into more chaotic spatial or temporal variation of the solution.
The way prior information is introduced in the inversion process depends on the
inversion technique used. The cost function represented by Eq. (7.7) is used within
iterative optimization and LUTs. Bayesian methods include the a priori distribution
through the use of the Bayes theorem to estimate the a posteriori distribution. For
biophysical variables driven approaches the training data base should reflect the
actual knowledge on the distribution of the variables. Note that the difficulty in
defining explicitly the covariance terms in the uncertainties on remote sensing inputs
(RT model and measurements) for the radiometric data driven approaches remains
in the biophysical variables driven approaches for the generation of the training data
base. However, implicit introduction of these terms may be achieved when using a
training data base made from actual satellite measurements as suggested by Bacour
et al. (2006).
7.4.2 Using Additional Constraints

7.4.2.1 Coupling Models
The radiative transfer in each element of the soil/leaf/canopy/atmosphere system is
strongly coupled to the radiative transfer in the whole system. The simple example
given previously to demonstrate the under-determined nature of the inverse problem
in remote sensing shows that top of canopy reflectance could be written as:
toc ( , s, v, ) = CAN(LAI, ALA, hot, LEAF(N, Cab , Cdm , Cbp , Cw ),

SOIL ([w j ], [i ], r, , s, v, ) , s, v, ) (7.8)
The same applies when retrieving some characteristics of the system from top of
atmosphere reflectance ( toa ) measurements as usually achieved by sensors aboard
satellite:
Patm , Cwv , C03 , , s, v, )
toa ( , s, v, ) = ATM( toc ( , s, v, ), 550 , A,
(7.9)
190
F. Baret, S. Buis
where ATM represents an atmospheric RT model such as 6S (Vermote et al., 1997)

being respectively the aerosol optical
or MODTRAN (Berk et al., 1998), 550 , A,
thickness at 550 nm and the Angstrom coefficient, Patm is the atmospheric pressure,
Cwv is the water vapor content and C03 the ozone content.
Retrieval of characteristics of some element of the system without solving (implicitly or explicitly) the whole system will therefore be sub-optimal as demonstrated below.
Let consider retrieving leaf biophysical properties [N, Cab , Cdm , Cbp , Cw ] from
top of canopy remote sensing observations in B wavebands using a decoupled system and an iterative optimization technique. For sake of simplicity, soil reflectance
will be assumed to be known. Estimates of leaf properties could be achieved in
two steps. First, estimate the variables [LAI, ALA, hot, re f l( ), tran( )] from the
reflectance in each of the B bands. A cost function accounting for the reflectance
in the B bands should be minimized with the constraint that [LAI, ALA, hot] does
not vary with wavelength. The number of unknowns in the system will therefore be
(3 + 2 B) corresponding to the 3 canopy structure variables and the 2 (reflectance
and transmittance) leaf optical properties time the B bands. The second step of the
process consists in estimating leaf biophysical properties [N, Cab , Cdm , Cbp , Cw ]
from the retrieved leaf reflectance and transmittance in the B bands. The variables
[N, Cab , Cdm , Cbp , Cw ] are tuned by minimizing a cost function accounting for
leaf reflectance and transmittance in the B bands. Obviously, increasing the number
of bands will not improve the underdetermined nature of the problem because the
number of unknowns in the first step of the process will grow twice faster. In addition, since no biophysical constraints are set on the spectral variation of leaf optical
properties, canopy structure variables derived from the first step may express larger
and unrealistic range of variation. The proper way to solve this type of problem is to
minimize a cost function accounting for canopy reflectance over the B wavebands
based on the coupled leaf and canopy models. In this case, the number of unknowns
will be eight (the three canopy structure variables and the five leaf characteristics)
which is independent from the number of wavebands used. This allows limiting the
under-determined nature of the problem by increasing the spectral sampling.
Most of the retrieval approaches from top of canopy radiometric observations are
now using implicitly or explicitly coupled models as shown in Table 7.2. However,
although offering great potentials as demonstrated recently (Baret, 2006b), the use
of coupled atmosphere/surface models is still not very well developed because each
sub-problem was handled by different communities.
7.4.2.2 Spatial Constraints

Up to now, most retrieval algorithms are applied to independent pixels, neglecting
the possible spatial structure as observed on most images. However, some authors attempted to exploit these very obvious patterns at high spatial resolution. The object
retrieval approach proposed by Atzberger (2004) is based on the use of covariance
between variables as observed over a limited cluster of pixels representing the same
191
Table 7.2 Synthesis of the several algorithms currently used operationally to retrieve canopy biophysical variables. 1: (Lacaze, 2004); 2: (Knyazikhin et al., 1999); 3: (Gobron et al., 1999); 4:
(Weiss et al., 2002; Baret et al., 2007); 5: (Chen et al., 2002; Deng et al., 2006); 6: (Bacour et al.,
2006)
#
Algorithm
RT models
Canopy
Atmosphere
Inversion
technique
uncertainties
prior
information
Kuusk
LAI, ALA, hot
TOC
NNT
measurements
Some variables fixed

Range of variation
Hapke
3 typical +
understorey
DISORD
6 biomes
TOC
LUT
measurements
prescribed at
20%
specific values for 6

biomes
PROSPECT
N, Cab,
(Cw,Cdm, Cs)
5 typical soil
unique BRDF
Gobron
LAI, ALA, hot
TOA
(MODTRAN)
Parametric
not specified
Range of variation
(uniform distribution)
PROSPECT
N, Cab,
Cw,Cdm, Cbp
brightness
parameter
&reference
spectra
SAIL
LAI, ALA, hot,
vCover
TOC
NNT
TOC
Parametric
2 versions:
- TOC version
- TOA version
(SMAC)
NNT
leaf
soil
POLDER
LAI, fAPAR
PROSPECT
N, Cab,
(Cw,Cdm, Cs)
PRICE
2 abundances
MODIS/MISR
LAI, fAPAR
prescribed for
each biome
MERIS
MGVI
fAPAR
VEGETATION
CYCLOPES
LAI, fAPAR,
fCover
VEGETATIONCanada-Global
LAI
Empirical relations for specific biomes

using TM sensor and the corresponding ground
measurements over some sites
Prescribed BRDF model
MERIS
LAI, fAPAR,
fCover, LAIxCab
PROSPECT
green/brown
separated
N, Cab, Cdm,
Cw, Cbp
(1)
(2)
brightness
parameter
&reference
spectra
SAIL,
LAI, ALA, hot,
vCover
model and
measurements approximation of actual
prescribed at 4%
distribution
(relative)
not specified
Specific relations for

each biome
model and
measurements approximation of actual
prescribed at 4%
distribution
(relative)
class of object such as an agricultural field. Results show quite significant improvement of the retrieval performances for LAI, Cab and Cw , presumably because of a
better handling of the possible compensation between LAI and ALA in the retrieval
process as suggested by Atzberger (2004) and outlined by Jacquemoud (1993).
Other approaches based on models with random effects (Faivre and Fischer,
1997) may be also very attractive, although rarely used within the land remote
sensing community. They allow characterizing a population by their two first statistical moments (mean and variance). In the case of remote sensing applications, this
could be applied over a cluster of P pixels belonging to the same class of surface as
in the object retrieval approach of Atzberger (2004). The inversion process could
be achieved by tuning both the mean and variance values of each input variable over
the P pixels using iterative optimization techniques. The individual values of each
pixel could be derived from the estimated mean and variance values of the variables and the departure between the actual radiometric measurements of the pixels
and the mean values over the object. The under-determination of the problem could
significantly decrease with this approach: the number of unknowns to estimate is
independent on the number of pixels considered in the cluster and is just twice the
number of variables to estimate (mean and variance).
Although quite promising, these methods need further evaluation, and probably
adaptation before being accepted and used by the remote sensing community. Note
that only statistical distributions are used for both methods presented, although additional geo-statistical constraints could be exploited particularly for the higher spatial
resolutions, based on variograms (Garrigues et al., 2006).
192
F. Baret, S. Buis
7.4.2.3 Temporal Constraints

The dynamics of canopies results from elementary processes under the control of
climate, soil and the genetic characteristics of the plants that incrementally change
canopy structure and optical properties of the elements. Very brutal and chaotic time
course are therefore not expected, at the exception of accidents such as fire, flooding,
harvesting, or lodging. The smooth character of the dynamics of canopy variables
may be exploited as additional constraint in the retrieval process. Kotz et al. (2005)
proposed using a semi-empirical model of canopy structure dynamics to improve
remote sensing estimates of LAI over maize crops. Results show a significant improvement of estimates, particularly for the larger LAI values where saturation of reflectance is known to be a problem. This approach requires a semi-empirical model
of canopy structure dynamics (here LAI) describing the whole growth cycle with
few parameters. In the case of the model used by Kotz et al. (2005) five parameters
are needed. In this case, the under-determined nature of the inverse problem will
decrease only if more than five dates of remote sensing observations are available
and well distributed over the growth cycle. However, because the parameters of the
model of LAI dynamics have some biological meaning, prior information on them
could be accumulated and efficiently exploited.
More recently, Lauvernet et al. (2007) proposed a multitemporal patch inversion scheme to account for both spatial and temporal constraints. Reflectance
data are here considered observed from top of atmosphere. Atmosphere/canopy/
leaf/soil RT models are thus coupled to simulate top of atmosphere reflectance from
the set of input variables as stated by Eqs. (7.8) and (7.9). Spatial and temporal
constraints are based on the assumption that the atmosphere is considered stable
over a limited area (typically few kilometres) but varies from date to date, and
that surface characteristics vary only marginally over a limited temporal window
(typically 7 days) but may strongly change from pixel to pixel. This has obviously
important consequences on the under-determined nature of the inverse problem as
demonstrated hereafter. The atmosphere characteristics [Patm , Cwv , C03 ], except the
are assumed to be known from independent observations
aerosol ones [550 , A],
such as meteorological estimation or dedicated sensors or algorithms. The observational configuration [, s, v, ] is also known at the time of image acquisition.
Soil reflectance was simply approximated as lambertian, with reflectance proportional to a reference soil spectra according to a brightness parameter Bs (Bacour
et al., 2006). The brightness parameter is assumed to vary both from date to date
and pixels to pixels, without any constraints. The forward model resulting from
nesting the RT models presented previously could be written as a function of the
with nA = 2
ten unknowns [N, Cab , Cdm , Cbp , Cw , LAI, ALA, hot, Bs, 550 , A]
atmosphere variables, nC = 8 canopy and leaf variables and ns = 1 soil variables.
Let consider d dates of observation available over a limited temporal window during
which the canopy variables are about constant, and a spatial window of p pixels
for which the atmosphere is considered homogeneous. The number of unknowns,
N(p, d) in the case of concurrent inversion of an ensemble of d dates and p pixels
using the spatial and temporal constraints described above is therefore:
193
N(p, d) = d nA + p nC + d p nS
(7.10)
Inverting the nested radiative transfer models concurrently over an ensemble of d

dates and p pixels will significantly reduce the total number of unknowns. (N(p, d))
as compared to p times d independent instantaneous pixel inversions (p.d.N(1, 1)).
Figure 7.8 shows that the number of unknowns to be estimated within the same
inversion process for p pixels and d dates as compared to p.d single pixel and
single date inversions (N(p, d)/(p.d.N(1, 1))) decreases significantly up to about
10 pixels. However, the main advantage over ensemble inversion is reached when
applying concurrently the inversion process to several dates. Using two dates and
more than 10 pixels allows dividing by almost 2 the number of unknowns. Note that
these results concern only the number of unknowns, and is therefore applicable to
any observational configuration characterized by a set of bands and directions.
Results on the performances achieved demonstrate the interest of the approach
for the estimation of most of the variables, particularly for the aerosol characteristics
and for LAI, LAI Cab and ALA canopy characteristics. However, again, this new
approach was only demonstrated over RT simulations, and its interest should be
verified over experiments with actual remote sensing data and the corresponding
ground truth.
1
0.9
d=1
N(p,d) / (p.d.N(1,1 ))
0.8
0.7
0.6
0.5
d=2
0.4
d=3
0.3
d=4
d=5
d=6
0.2
0.1
10
15
20
25
30
35
40
45
50
Number of pixels (p)
Fig. 7.8 Ratio, N(p, d)/(p.d.N(1, 1)), between the number of unknowns when applying the inversion process concurrently to p pixels and d dates with that of single pixel and date inversion as a
function of the number of dates and pixels considered
194
F. Baret, S. Buis
7.5 Conclusion
This overview of retrieval approaches is based on methods currently used, while
alternative ways to solve the problem and hopefully improve the accuracy and robustness of estimates were briefly introduced. Several ingredients of the algorithms
were identified apart from the retrieval techniques themselves: radiative transfer
models, observations, additional information and constraints. We will briefly summarize the conclusions for each of these ingredients in the following.
7.5.1 Retrieval Techniques

The several techniques investigated have been classified as radiometric variables or
biophysical variables driven approaches. However, both types of methods could be
either derived from actual measurements or based on radiative transfer model simulations. The best approaches are obviously the ones that will be trained over data sets
that are as close as possible to the evaluation data set. For this very reason, canopy
biophysical variables trained over empirical data sets would be ideal. In addition,
canopy biophysical variables driven approaches present the advantage of being very
computer efficient once trained, allowing easy implementation within operational
processing chains. However, because of the difficulty of getting a large enough
training data set representing the actual distribution of cases (observational configuration, type of canopies and state, background properties, eventually atmosphere
characteristics), training data base made of radiative transfer model simulations is
preferred. These hybrid techniques as termed by Liang (2004) require however the
radiative transfer models to be well adapted to the type of canopy they target, and
their adequacy to be quantified to properly input model uncertainties. In addition,
the structure of uncertainties on the radiometric variables and distribution and codistribution of the input biophysical variables should be also known. An alternative
approach currently not yet explored would consist in bridging the two retrieval approaches: actual sensor measurements are used to build the training data base allowing to keep all the structure of measurement uncertainties. This data base should be
representative of the cases investigated, which might be possible by specific spatial
and temporal sampling schemes as proposed by Baret et al. (2006a) in the case of
global observations. The corresponding best estimates of canopy biophysical variables could be derived from inversion methods such as iterative optimization techniques for which all the information available should be exploited: fusion of all
currently available sensors observations, prior information and spatial and temporal
constraints.
As a matter of fact, most radiometric variables driven approaches are very flexible and could easily ingest data from several sensors, bands and directions, at the
expense of computer requirements which make these methods more difficult for an
operational use. Conversely, canopy biophysical driven approaches are not as flexible as radiometric driven approaches: they are generally tuned for a limited set of
195
observational conditions: using other configurations would require a specific training or a dramatic enlargement of the training data base.
Retrieval methods will be more efficient when applied to a limited set of surface
types as compared to a very generic (global) solution. Approaches based on a classification would thus allow closer adaptation to each class of both the radiative transfer
model and prior information. However, attribution errors may significantly alter the
performances. Using a continuous classification (Hansen et al., 2002; Hansen and
DeFries, 2004; Schwartz and Zimmermann, 2005) will probably limit this source
of uncertainties and avoid getting artefacts when two consecutive pixels will jump
from one class to another.
Biophysical variables estimates are generally integrated within other process
models such as hydrology or biogeochemical cycling along with other ground observations. Quantification of the associated uncertainties is therefore required to properly merge these several sources of information. Current available products did not
provide quantitative evaluation of the confidence interval around the solution, but
are limited to qualitative indices. Bayesian approaches provide a direct access to the
distribution of the solution of the inverse problem and may be very useful for estimating the uncertainties. Current operational algorithms need further developments
to fully satisfy this important user requirement.
7.5.2 Radiative Transfer Models

Performances of methods based on radiative transfer models are largely depending
on the realism of the simulations. Radiative transfer models are based on a set of
assumptions, particularly regarding the description of canopy architecture. A more
realistic description of canopy architecture will require additional input variables
and will be probably more demanding in computer resources. Knowledge of prior
distribution and co-distribution of these additional canopy structure variables will
constitute a limitation. Further, using such more realistic radiative transfer model
requiring a larger number of unknowns will not necessarily improve the retrieval
performances because the under-determination of the problem will be even more
limiting. A compromise should therefore be found between the realism of the description of canopy structure, and its complexity.
Particular attention should be paid on the definition of the variables used in the
radiative transfer model that should match the one required for the application. For
example, the original LAI definition (Stenberg, 2006) may be altered depending
on the way and scale at which leaf clumping is accounted for (Chen and Leblanc,
1997). Great caution should be also paid when comparing retrievals with ground
measurements or inter-comparing several products.
As demonstrated here, holistic approaches based on the coupling of canopy, leaf
and soil models are optimal for best performances. Eventually, coupled surface and
atmosphere models would certainly help solving in an elegant way the retrieval of
surface variables from top of the atmosphere observations.
196
F. Baret, S. Buis
7.5.3 Observations and Ancillary Information

The observational configuration is an important element that drives the accuracy
of canopy biophysical variables estimation. It depends obviously on the variables
targeted. For the time being, sufficient maturity is achieved for the estimation of
LAI, fAPAR, the cover fraction, chlorophyll and water contents variables to implement operational algorithms for delivering the corresponding products to the user
community. The interest of multidirectional and hyperspectral observations is still
to be rigorously demonstrated for these variables by comparison over actual ground
measurements.
Frequent observations are required to monitor the dynamics of the vegetation that
conveys a large amount of information on the functioning of the surface. With the
hopefully venue of systems capable of high revisit frequency with high spatial resolution, new retrieval methods should be developed to exploit the temporal and spatial
dimensions in addition to the more classical spectral and in a lesser way directional
ones. This would allow benefiting from the spatial and temporal constraints and consequently reduce the number of unknowns to be retrieved. Ultimately, this approach
will converge towards direct assimilation of top of atmosphere radiances into surface process models. However, the research community is not mature enough on the
coupling between radiative transfer models and canopy process models. Radiative
transfer model inversion had still to mature and improve the accuracy of surface
variables estimation before jumping towards radiance data assimilation.
Knowledge and management of uncertainties is one of the critical issues for the
retrieval algorithms. If measurement uncertainties coming from the sensor are relatively well known, their structure (covariance between bands and directions for
example) is poorly documented. This is even worse when considering model uncertainties that may change dramatically from place (and time) to place (and time) with
presumably specific features (covariance between configurations).
The other critical issue is the lack of prior information on the distribution of
most land surface attributes. However, this could be accumulated from the numerous experiments organized in support of satellite images. A mechanism should thus
be developed to capitalize on the information gathered within the remote sensing
research community as well as other communities working with ecosystems. Note
that getting high spatial resolution data will considerably ease the characterization
of prior distribution of the variables, provided that each pixel could be properly
classified.
Any retrieval algorithm should be properly validated before delivering its products to the user community according to consensus protocols (Morisette et al.,
2006). This process will not only provide a way to characterize the associated uncertainties, it will be also critical for improving the algorithms. A short feedback loop
should therefore be set-up between algorithm prototyping and validation. When retrieval algorithms are based on radiative transfer modeling, this will implicitly merge
observations and model to improve robustness and accuracy of the products at the
expense of a decrease in the desired independency between the validation and calibration processes.
197
References
Abuelgasim AA, Gopal S, Strahler AH (1998) Forward and inverse modelling of canopy directional reflectance using a neural network. Int. J. Remote Sens. 19(3):453471
Asrar G, Fuchs M, Kanemasu ET, Hatfield JL (1984) Estimating absorbed photosynthetic radiation
and leaf area index from spectral reflectance in wheat. AJ, 76:300, 306
Asrar G, Myeni RB, Kanemasu ET (1989) Estimation of plant canopy attributes from spectral reflectance measurements. Theory and Application of Optical Remote Sensing. J. Wiley,
Indianapolis: 252296
Atkinson PM, Tatnall ARL (1997). Neural network in remote sensing. Int. J. Remote Sens.
18(4):699709
Atzberger C (2004) Object-based retrieval of biophysical canopy variables using artificial neural
nets and radiative transfer models. Remote Sens. Environ. 93:5367
Bacour C (2001) Contribtion a` la determination des paramètres biophysiques des couverts vegetaux
par inversion de modèles de reflectance: analyse de sensibilite et configurations optimales. PhD
Thesis, Universite Paris 7 Denis Diderot, Paris (France), 206 pp
Bacour C, et al. (2002a) Reliability of the estimation of vegetation characteristics by inversion of
three canopy reflectance models on airborne POLDER data. Agronomie, 22:555566
Bacour C, Jacquemoud S, Tourbier Y, Dechambre M, Frangi JP (2002b) Design and analysis of
numerical experiments to compare four canopy reflectance models. Remote Sens. Environ.
79:7283
Bacour C, Baret F, Beal D, Weiss M, Pavageau K (2006) Neural network estimation of LAI, fAPAR, fCover and LAIxCab, from top of canopy MERIS reflectance data: principles and validation. Remote Sens. Environ. 105:313325
Banari A, Huete AR, Morin D, Zagolski F (1996) Effets de la couleur et de la brillance du sol sur
les indices de vegetation. Int. J. Remote Sens. 17(10):18851906
Baret F, Guyot G (1991) Potentials and limits of vegetation indices for LAI and APAR assessment.
Baret F, Clevers JGPW, Steven MD (1995) The robustness of canopy gap fraction estimates
from red and near infrared reflectances: a comparison of approaches. Remote Sens. Environ.
54:141151
Baret F, Knyazikhin J, Weiss M, Myneni R, Pragnère A (1999) Overview of Canopy Biophysical
Variables Retrieval Techniques. ALPS99, Meribel, France, CNES, Paris
Baret F, Weiss M, Troufleau D, Prevot L, Combal B (2000) Maximum information exploitation for
canopy characterisation by remote sensing. Remote sensing in agriculture. Aspects of Applied
Biology. Association of Appied Biologists, pp. 7182
Baret F, et al. (2006a) Evaluation of the representativeness of networks of sites for the global validation and inter-comparison of land biophysical products. Proposition of the CEOS-BELMANIP.
IEEE Trans.Geosci. Remote Sens. 44(7:special issue on global land product validation):1794
1803
Baret F, et al. (2006b) Algorithm Theoretical Basis Document for MERIS Top of Atmosphere Land
Products (TOA VEG), INRA-CSE, Avignon
Baret F, et al. (2007) LAI, fAPAR and fCover CYCLOPES global products derived from VEGETATION. Part 1: principles of the algorithm. Remote Sens. Environ. 110:275286
Berk A, et al. (1998) MODTRAN cloud and multiple scattering upgrades with application to
AVIRIS. Remote Sens. Environ. 65:367375
Bicheron P, Leroy M (1999) A method of biophysical parameter retrieval at global scale by inversion of a vegetation reflectance model. Remote Sens. Environ. 67:251266
Bonnans, JF, Gilbert, JC, Lemarechal, C, and Sagastizabal, CA (2006) Numerical Optimization.
Theoretical and Practical Aspects (Second Edition). Springer, New-York. 494pp.
Braswell BH, et al. (1996) Extracting ecological and biophysical information from AVHRR optical
data: an integrated algorithm based on inverse modeling. J. Geophys. Res. 101(D18):23335
23348
198
F. Baret, S. Buis
Breon FM, Maignan F, Leroy M, Grant I (2002) Analysis of hot spot directional signatures measured from space. J. Geophys. Res. 107(D16):AAC 1 115
Chen JM, Leblanc SG (1997) A four-scale bidirectional reflectance model based on canopy architecture. IEEE Trans. Geosci. Remote Sens. 35:13161337
Chen JM, et al. (2002) Derivation and validation of Canada wide coarse resolution leaf area index
maps using high resolution satellite imagery and ground masurements. Remote Sens. Environ.
80:165184
Combal B, Ochshepkov SL, Sinyuk A, Isaka H (2000) Statistical framework of the inverse problem
in the retrieval of vegetation parameters. Agronomie 20:6577
Combal B, Baret F, Weiss M (2001) Improving canopy variables estimation from remote sensing data by exploiting ancillary information. Case study on sugar beet canopies. Agronomie
22:215
Combal B, et al. (2002) Retrieval of canopy biophysical variables from bi-directional reflectance
data. Using prior information to solve the ill-posed inverse problem. Remote Sens. Environ.
84:115
Danson FM, Rowland CS, Baret F (2003) Training a neural network with a canopy reflectance
model to estimate crop leaf area index. Int. J. Remote Sens. 24(23):48914905
Deng F, Chen JM, Chen M, Pisek J (2006) Algorithm for global leaf area index retrieval using
satellite imagery. IEEE Trans. Geosci. Remote Sens. 44(8):22192229
Faivre R, Fischer A (1997) Predicting crop reflectances using satellite data observing mixed pixels.
J. Agric., Biol. Environ. Stat. 2:87107
Fang H, Liang S (2005) A hybrid inversion method for mapping leaf area index from MODIS
data: experiments and application to broadleaf and needleleaf canopies. Remote Sens. Environ.
94:405424
Fang H, Liang S, Kuusk A (2003) Retrieving leaf area index using a genetic algorithm with a
canopy radiative transfer model. Remote Sens. Environ. 85:257270
Fernandes R, Leblanc SG (2005) Parametric (modified least squares) and non parametric (TheilSen) linear regressions for predicting biophysical parameters in the presence of measurements
errors. Remote Sens. Environ. 95:303316
Gao W, Lesht BM (1997) Model inversion of satellite-measured reflectance obtaining surface biophysical and bidirectional characteristics of grassland. Remote Sens. Environ. 59:461471
Garabedian P (1964) Partial differential equations. Wiley, New York, 672pp.
Garrigues S, Allard D, Baret F, Weiss M (2006) Quantifying spatial heterogeneity at the landscape
scale using variogram models. Remote Sens. Environ. 103:8196
Gobron N, Pinty B, Verstraete MM, Govaerts Y (1997) A semi-discrete model for the scattering of
light by vegetation. J. Geophys. Res.
Gobron N, Pinty B, Verstraete M, Govaerts Y (1999) Level 2 vegetation index algorithm theoretical
Basis. Joint Research Center, Ispra, Italy
Gobron N, Pinty B, Verstraete M, Widlowski JL (2000) Advanced vegetation indices optimized for
upcoming sensors: design, performances and applications. IEEE Trans. Geosci. Remote Sens.
38(6):24892505
Goel NS (1989) Inversion of canopy reflectance models for estimation of biophysical parameters
from reflectance data. In: G. Asrar (Editor), Theory and Applications of Optical Remote Sensing. Wiley, New York, pp. 205251
Goel NS, Deering DW (1985) Evaluation of a canopy reflectance model for LAI estimation through
its inversion. IEEE Trans. Geosci. Remote Sens. GE-23(5):674684
Goel NS, Thompson RL (1984) Inversion of vegetation canopy reflectance models forestimating
agronomic variables. III. Estimation using only canopy reflectance data as illustrated by the
SUIT model. Remote Sens. Environ. 15:223236
Goel NS, Strebel DE, Thompson RL (1984) Inversion of vegetation canopy reflectance models for
estimating agronomic variables. II. Use of angle transforms and error analysis as illustrated by
the SUIT model. Remote Sens. Environ. 14:77111
Gong P, Wang DX, Liang S (1999) Inverting a canopy reflectance model using a neural network.
Int. J. Remote Sens. 20(1):111122
199
Green RO, Boardman J (2001) Exploration of the relationship between information content and
signal to noise ratio and spatial resolution in AVIRIS spectral data. In: RO Green (ed), AVIRIS
Workshop, Pasadena, CA
Hansen MC, et al. (2002) Towards an operational MODIS continuous firld percent tree coveralgorithm: examples using AVHRR and MODIS data. Remote Sens. Environ. 83:309319
Hansen MC, DeFries RS (2004) Detecting long-term global forest change using continuous fields
of tree-cover maps from 8-km Advanced Very High Resolution Radiometer (AVHRR) data for
the years 198299. Ecosystems 7:695716
Hapke B (1981) Bidirectional reflectance spectroscopy. 1. Theory. J. Geophys. Res. 86:30393054
Huang D, et al. (2006) The importance of measurement error for deriving accurate reference leaf
area index maps and validation of the MODIS LAI products. IEEE Trans. Geosci. Remote Sens.
44(7):18661871
Huete AR (1988) A soil adjusted vegetation index (SAVI). Remote Sens. Environ. 25:295309
Huete AR, Liu HQ, Batchily K, van Leeuwen W (1997) A comparison of vegetation indices over
a global set of TM images for EOS-MODIS. Remote Sens. Environ. 59:440451
Jacquemoud S (1993) Inversion of the PROSPECT+SAIL canopy reflectance model from AVIRIS
equivalent spectra: theoretical study. Remote Sens. Environ. 44:125
Jacquemoud S, Baret F (1990) PROSPECT: a model of leaf optical properties spectra. Remote
Jacquemoud S, Baret F, Hanocq JF (1992) Modeling spectral and directional soil reflectance.
Jacquemoud S, Baret F, Andrieu B, Danson M, Jaggard K (1995) Extraction of vegetation biophysical parameters by inversion of the PROSPECT+SAIL model on sugar beet canopy reflectance
data. Application to TM and AVIRIS sensors. Remote Sens. Environ. 52:163172
Jacquemoud S, Bacour C, Poilve H, Frangi JP (2000) Comparison of four radiative transfer models to simulate plant canopies reflectance: direct and inverse mode. Remote Sens. Environ.
74:471481
Kimes DS, Nelson RF, Manry MT, Fung AK (1998) Attributes of neural networks for extracting
continuous vegetation variables from optical and radar measurements. Int. J. Remote Sens.
19(14):26392663
Kimes DS, Knyazikhin Y, Privette JL, Abuelgasim AA, Gao F (2000) Inversion methods for
physically-based models. Remote Sens. Rev. 18:381439
Knyazikhin Y, et al. (1998a) Estimation of vegetation canopy leaf area index and fraction of absorbed photosynthetically active radiation from atmosphere-corrected MISR data. J. Geophys.
Res. 103(D24):3223932256
active radiation from MODIS and MISR data. J. Geophys. Res.103(D24):3225732275
Knyazikhin Y, et al. (1999) MODIS leaf area index (LAi) and fraction of photosynthetically active
radiation absorbed by vegetation (FPAR) product (MOD15) algorithm theoretical basis document. http://eospso.gsfc.nasa.gov/atbd/modistables.html
Kotz B, Baret F, Poilve H, Hill J (2005) Use of coupled canopy structure dynamic and radiative transfer models to estimate biophysical canopy characteristics. Remote Sens. Environ.
95:115124
Kuusk A (1991a) Determination of vegetation canopy parameters from optical measurements. Remote Sens. Environ. 37:207218
Kuusk A (1991b) The inversion of the Nilson-Kuusk canopy reflectance model, a test case. Int.
Geosci. Remote Sens. Symp. (IGARSS91), Helsinki, Finland, pp 15471550
Kuusk A (1995) A fast invertible canopy reflectance model. Remote Sens. Environ. 51:342350
Lacaze R (2004) Land Surface Level3 Algorithm: The inversion of a vegetation model by a neural
network. short report. http://smsc.cnes.fr/POLDER/SCIEPROD/lsp2algol3.htm
Lacaze R (2005) POLDER-2 Land surface level 3 products. User manual and product validation,
Medias/CNES, Toulouse
200
F. Baret, S. Buis
Lauvernet C, Baret F, Hascoet L, Buis S, Le Dimet FX (2008) Multitemporal-patch ensemble inversion of coupled surface-atmosphere radiative transfer models for land surface characterization.
Remote Sens. Environ. (in press)
Leprieur C, Verstraete MM, Pinty B (1994) Evaluation of the performance of various vegetation
indices to retrieve vegetation cover from AVHRR data. Remote Sens. Rev. 10:265284
Leshno M, Ya Lin V, Pinkus A, Shocken S (1993) Multilayer feedforward networks with non
polynomial activation function can approximate any function. Neural Networks 6:861867
Lewis P, et al. (2004) Calibration of and L-system model of winter wheat for remote sensing
modelling and inversion. In: C. Godin (ed), 4th Int. Workshop on Functional-Structural Plant
Models, Montpellier, France, pp 257261
Liang S (2004) Quantitative Remote Sensing of Land Surfaces. Wiley, Hoboken, NJ
Liu W, et al. (2002) Relating soil surface moisture to reflectance. Remote Sens. Environ. 81:238246
Lotsch A, Tian Y, Friedl MA, Myneni RB (2003) Land cover mapping in support of LAI and fPAR
retrievals from EOS-MODIS and MISR: classification methods and sensitivities to errors. Int.J.
Remote Sens. 24(10):19972016
Lucht W (1998) Expected retrieval accuracies of bidirectional reflectance and albedo from EOSMODIS and MISR angular sampling. J. Geophys. Res. 103(D8):87638778
Makowski, D, Hillier, J, Wallach, D, Andrieu, B Jeuffroy, MH (2006) Parameter estimation for crop
models. In D. Wallach, D. Makowski and J.W Jones (eds.), Working with dynamic crop models:
evaluation, analysis, parameterization, and applications, pp. 101149. Elsevier, Amsterdam
Morisette J, et al. (2006) Validation of global moderate resolution LAI Products: a framework
proposed within the CEOS Land Product Validation subgroup. IEEE Trans. Geosci. Remote
Sens. 44(7):18041817
Nilson T, Kuusk A (1989) A reflectance model for homogeneous plant canopy and its inversion.
Pinty B, Verstraete MM (1991) Extracting information on surface properties from bidirectional
reflectance measurements. J. Geophys. Res. 96:28652874
Pinty B, Verstraete MM, Dickinson RE (1990) A physical model of the bidirectional reflectance of
vegetation canopies. 2. Inversion and validation. J. Geophys. Res. 95(D8):1176711775
Price J (1994) How unique are spectral signatures? Remote Sens. Environ. 49:181186
Price JC (1990) On the information content of soil reflectance spectra. Remote Sens. Environ.
33:113121
Privette JL, Emery WJ, Schimel DS (1996) Inversion of a vegetation reflectance model with NOAA
AVHRR data. Remote Sens. Environ. 58:187200
Qiu J, Gao W, Lesht BM (1998) Inverting optical reflectance to estimate surface properties of
vegetation canopies. Int. J. Remote Sens. 19(4):641656
Renders J-M, Flasse SP (1996) Hybrid methods using genetic algorithms for global optimization.
IEEE Trans. Systems, Man, and Cybernetics 26:243258
Richardson AJ, Wiegand CL, Wanjura DF, Dusek D, Steiner JL (1992) Multisite analyses of
spectral-biophysical data for sorghum. RSE, 41:71:82
Rondeaux G, Steven MD, Baret F (1996) Optimization of soil adjusted vegetation indices. Remote
Roujean JL, Lacaze R (2002) Global mapping of vegetation parameters from POLDER multiangular measurements for studies of surface-atmosphere interactions: a pragmatic method and
validtion. J. Geophys. Res. 107(D12):ACL 6 114
Saltelli, A, Tarantola, S, Campolongo, F, Ratto, M (2004) Sensitivity Analysis in practice. A guide
to assessing scientific models. Wiley & Sons, Chichester. 217pp.
Schwartz MD, Zimmermann G (2005) A new GLM-based method for mapping tree cover continuous fields using regional MODIS reflectance data. Remote Sens. Environ. 95:428443
Sellers PJ (1985) Canopy reflectance photosynthesis and transpiration. Int. J. Remote Sens.
3:13351372
Shoshany M (1991) The equifinality of bidirectional distribution functions of various microstructures. Int. J. Remote Sens. 12(11):22672281
201
Smith JA (1992) LAI inversion from optical reflectance using neural network trained with multiple
scattering model. In: WR (ed), IGARS92. IEEE, South Shore Harbour Resort and Conference
Center, Houston, TX, pp 757759
Smith JA (1993) LAI inversion using backpropagation neural network trained with multiple scattering model. IEEE Trans. Geosci. Remote Sens. 31(5):11021106
Soler CFS, Blaise F, de Reffye P (2001) A physiological plant growth simulation engine based on
accurate radiant energy transfer. 4116, INRIA, Montbonnot-Saint-Martin, France
Stenberg P (2006) A note on the G-function for needle leaf canopies. Agric. For. Meteorol. 136
(12):76
Tarantola A (2005) Inverse problem theory problem and methods for model parameter estimation.
Soc. Ind. Appl. Math., 358 pp
Tarentola A (1987) Inverse problem theory. Methods for data fitting and model parameter estimation. Elsevier, New York
Teillet PM, Gauthier RP, Staenz K, Fournier RA (1997) BRDF equifinality studies in the context
of forest canopies. In: G Guyot, T Phulpin (eds), Physical Measurements and Signatures in
Remote Sensing, Balkema, Rotterdam, Courchevel, France, pp 163170
Verhoef W (1984) Light scattering by leaf layers with application to canopy reflectance modeling:
the SAIL model. Remote Sens. Environ. 16:125141
Verhoef, W (2002) Improved modeling of multiple scattering in leaf canopies: The model SAIL++.
In J. A. Sobrino (ed.), 1st Int. Symp. on Recent Advances in Quantitative Remote Sensing,
pp. 1120, Universitat de Valencia, Spain
Vermote EF, Tanre D, Deuze JL, Herman M, Mocrette JJ (1997) Second simulation of the satellite
signal in the solar spectrum, 6S: an overview. IEEE Trans. Geosci. Remote Sens. 35(3):675
686
Verstraete M, Pinty B (1996) Designing optimal spectral indexes for remote sensing applications.
Weiss M, Baret F, Myneni R, Pragnère A, Knyazikhin Y (2000) Investigation of a model inversion
technique for the estimation of crop charcteristics from spectral and directional reflectance data.
Agronomie 20:322
Weiss M, et al. (2002a) Validation of neural net techniques to estimate canopy biophysical variables
from remote sensing data. Agronomie 22:547554
Weiss M, et al. (2002b) Evaluation of kernel-driven BRDF models for the normalization of
Alpilles/ReSeDA POLDER data. Agronomie 22:531536
Wiegand CL, Gerberman AH, Gallo KP, Blad BL, Dusek D (1990) Multisite analyses of spectralbiophysical data for corn. Remote Sens. Environ. 33:116
Wiegand CL, et al. (1992) Multisite analysis of spectral biophysical data for wheat. Remote Sens.
Environ. 42(1):122
Zhang Y, Shabanov N, Knyazikhin Y, Myneni RB (2002a) Assessing the information content of
multiangle satellite data for mapping biomes I. Theory. Remote Sens. Environ. 80:435446
Zhang Y, Tian Y, Myneni RB, Knyazikhin Y, Woodcock CE (2002b) Assessing the information
content of multiangle satellite data for mapping biomes I. Statistical analysis. Remote Sens.
Environ. 80:418434
Zhang Q, et al. (2005) Estimating light absorption by chlorophyll, leaf and canopy in a deciduous broadleaf forest using MODIS data and a radiative transfer model. Remote Sens. Environ.
99(3):357
Chapter 8
Knowledge Database and Inversion

Abstract Physical remote sensing models usually need dozens of parameters to

have reasonable degree of precision. With limited information provided by remotely
sensed observations, it is an ill-posed problem to estimate parameters through model
inversion. Many researches have developed inversion models and algorithms. We
developed a priori knowledge based inversion strategy and algorithm. The spectrum
knowledge database of typical land surface objects has been established to provide
the prior knowledge of model parameters. Some approaches are presented in this
chapter, which include the uncertainty and sensitivity matrix for analysis of observation data and parameters, the model inversion method supported by the knowledge
database, the scaling correction on estimated parameters. Some study directions in
model inversion, such as how to accumulate and use spatial and temporal change
knowledge, how to validate the parameter inversion results, are also discussed.
8.1 Questions on and Possible Answers to Physical Remote

Sensing Model Inversion
One of the primary problems of remote sensing science is the retrieval of information on land surface parameters from remotely sensed data. Since satellite remote sensing deals with a complex system coupling atmosphere and land surface,
physical remote sensing models usually need several to tens parameters to describe
the relation between land surface parameters and remote sensing signals for having reasonable degree of model precision. But, with limited information provided
by remotely sensed observations, it remains a challenge to estimate parameters
through remote sensing model inversion. In mathematics, it is an ill-posed problem
Research Center for Remote Sensing & GIS
Beijing Normal University, Beijing, China
wangjd@bnu.edu.cn
203
204
J. Wang, X. Li
to estimate more model parameters from less observations. In past years, many researchers have developed new inversion models and algorithms (Verstraete et al.,
1996; Myneni et al., 1995). Some inversion research has developed better cost
function for physical model inversion. In addition, different kinds of mathematical
optimization algorithms have been developed (Kimes et al., 2000; Kunsk, 1991).
Verstraete et al. (1996) pointed out the main limitation and the necessary observations for obtaining successful model inversion. Kimes et al. (1991) presented the
knowledge-based expert system for inferring vegetation characteristics. Fang et al.
(2005) used a genetic algorithm to estimate leaf area index (LAI) of vegetation
canopy with radiative transfer model inversion. Neural network methods were also
used for model inversion (Smith, 1993). Liang (2004) summarized the main research
achievements in estimating land surface biophysical variables and surface radiation
budgets.
As advances in the field of multi-angular remote sensing progress, bidirectional
reflectance distribution function (BRDF) models can be inverted to estimate the
structural parameters and spectral component signatures of land surface cover types,
such as the MODIS albedo and LAI products. Some operational algorithms have
been implemented to generate products of land surface parameters, such as albedo,
LAI, fraction of photosynthetically active radiation (FPAR), net primary production
(NPP) from MODIS and MISR observations at the spatial resolution of 250 m or
1 km. The surface BRDF and albedo product from POLDER has been developed
at a pixel resolution of approximately 6 km. Some validation work on these data
products has been carried out and it has been found that there are still some uncertainties on model, observation and reference data which influence the accuracy of
these products (Morisette et al., 2006; Tan et al., 2005).
However, when we want to improve the estimation accuracy of retrieved surface
parameters to meet requirements of applications at different spatial scales, the inversion of physical remote sensing models is a very difficult problem that still requires
further studies from the viewpoint of both information theory (Li et al., 1998) and
the comprehensive practice of model inversion (Privette et al., 1997; Wanner et al.,
1997; Li et al., 2000b; Liang, 2004). The real physical system that couples the atmosphere and land surface is extremely complex and it requires many parameters to
describe it faithfully. Any physical model can only be an approximation of this real
system, and a good model will have many important parameters to capture the major variations of the real system. However, remotely sensed observations are usually
more or less correlated. The remotely sensed signal, no matter how fine its spectral
and angular resolution, contains only limited information. Therefore, BRDF model
inversion problems, such as those in geoscience generally, are usually underdetermined, making the use of a priori knowledge necessary.
As the ancient philosopher Confucius pointed out, Our knowledge consists of
two parts what we know, and what we know we dont know. In the case that
remote sensing signals contain limited but valuable information, it is important to
extract information about what we dont know or what is uncertain, rather than to
invert all model parameters at the same time, pretending that we know nothing.
Using this principle in earlier work, we expressed a priori knowledge of model
8 Knowledge Database and Inversion
205
parameters as the best guesses for the associated uncertainties, and developed the
a priori knowledge based inversion strategy and algorithm (Li et al., 1997). The results were encouraging, and thus we tried to formalize the approach, and to establish
the a priori knowledge database of typical land surface parameters (Li et al., 2002).
8.1.1 A Priori Knowledge-based Inversion Strategy

We developed a priori knowledge based inversion strategy and algorithm (Li et al.,
1998, 2001) called, Multi-stage, Sample-direction dependent, Target-decision
(MSDT) (Li et al., 1997). There are three questions should be answered in order to run the strategy. The first question is how to express the a priori knowledge
and to make it available in model inversion algorithm. The second question is how
to divide the model parameters set into subsets, in order to invert the parameters by
using the most sensitive data. The third question is how to accumulate the knowledge during the inversion procedure, when we have one scene observing data or a
data set containing more scenes of continuous observations.
In our previous study (Li et al., 1998), the a priori knowledge of model parameters is expressed as a joint probability density, while a priori knowledge of the
model accuracy and measurement noise is expressed as a conditional joint probability density. In inversion model, the a priori probability density function (PDF) of
the observations can be defined, and can be used in the cost function based Bayesian
inference theory. An important feature of Bayesian inversion is that there is no prerequisite number of independent observations for a successful inversion. So long as
new observations are acquired, a priori probability density in parameter space can
be modified to obtain posterior density, allowing knowledge to be accumulated.
Taking the land surface spectral albedo inversion as an example, Li et al. (2001)
showed how the a priori knowledge significantly improves the retrieval of surface
bidirectional reflectance and spectral albedo from satellite observations. In the paper, the a priori knowledge are extracted from field measurements of bidirectional
reflectance factors for various surface cover types in red and near-infrared bands.
Bidirectional reflectance and albedo are retrieved by the kernel-driven BRDF model
inversion that uses surface reflectance observations derived from orbiting satellites.
A priori knowledge is applied when noise and poor angular sampling reduce the
accuracy of model inversion. In such cases, a priori knowledge can indicate when
retrieved kernel weights or albedos are outside expected bounds, leading to a closer
examination of the data. If data are noisy, a priori knowledge can be used to smooth
the data. If the data exhibit poor angular sampling, a priori knowledge can be used
according to Bayesian inference theory to yield a posteriori estimates of the unknown kernel weights, where Bayesian theory is applied in the data space rather
than in the parameter space. Extensive studies and simulations using 73 sets of field
observations and 395 space-borne observation sets from the POLDER instrument
demonstrates the importance of a priori information in improving inversions and
BRDF retrievals.
206
J. Wang, X. Li
The further studies introduce Tarantolas inversion cost function in to BRDF

model inversion algorithm. Tarantolas inversion theory has also been widely applied in geophysical inversion (Tarantola, 1987) and atmospheric remote sensing
(Rodgers, 1976), because the inverse problems in those fields are even more illposed than in land surface remote sensing. Based on Tarantolas theory, the cost
function used in land surface parameters inversion is defined as the parameters
posterior probability density, and can be expressed as

1
pM (X) = const exp ( f (X) Yobs )T CD1 ( f (X) Yobs )
2

1
(X Xprior )
(8.1)
+ (X Xprior )T CM
where Yobs is for observation data and Xprior is for a priori knowledge of parameters. The principle of the cost function is very similar to Bayesian inversion. One
of its advantages is its ability to clearly express the errors of the model, the observed data, and the parameters initial value, with co-variances matrixes CD and
CM respectively. This creates a new challenge on forming the error co-variances of
model, data and parameters, this requires a great deal of a priori knowledge.
Previous studies have demonstrated the effectiveness of the inversion model, using simulated data and field measurements (Yan et al., 2001; Wang et al., 2000). In
these studies, the parameters initial values can be estimated by our field measurements, while the co-variance was set as the parameters standard deviation under the
assumption that their estimated values have a normal distribution.
8.1.2 Uncertainty and Sensitivity Matrix of Parameters for Model

Inversion
It is important to examine the uncertainty and sensitivities of the parameters in
model inversion, since we expect to estimate the values of parameters that are sensitive to observations and model, while giving certain values for other parameters
that have relatively fewer uncertainties. To make the analysis of parameter uncertainty and sensitivity more effective, we defined the uncertainty and sensitivity matrix (USM), which is an objective expression of the prior knowledge.
In order to construct the USM, we assume that a BRDF model has N spectral
bands and K structural parameters and L spectral parameters of component materials, making K + N L in total. Since the multiangular observations have M samples,
the USM will have M N rows and K + N L columns. This matrix is too large to
effectively use for calculations. Because the structural parameters are independent
on the spectral band, we decomposed the matrix into a structural matrix that has
M N rows, K columns, where the N matrix of M L spectral parameters corresponds to N bands, making a total N + 1 matrices. An element of the USM can be
expressed as (Li et al., 1997),
USM[i][ j] =
207
BRDF( j|i)
BRDF(i)
(8.2)
where BRDF( j|i) is the maximum difference of BRDF as a function of the jth parameter within its uncertainty, given the ith geometry of illumination and viewing;
BRDF(i) is BRDF as predicted by the model at the ith geometry, when all parameters at their best guess values. These best guess values of parameters are also
from the accumulation of prior knowledge, and they can be updated when further
parameter estimates are obtained during the MSDT inversion procedure.
This USM definition has three advantages: (1) the uncertainty of the initial guess
for being inverted parameter is taken into account; (2) the USM is less dependent on
the initial guess; (3) all elements of the USM have the same units, and are therefore
quantitatively comparable. This USM has thus been used widely in our recent model
inversion studies. It can also be applied for parameters sensitivity analysis for models
and data in relative studies (Li et al., 1997; Gao and Zhu, 1997; Yan et al., 2001).
8.1.3 Getting the Prior Knowledge on Typical Land Surface

Parameters
As stated above, after developing the model inversion strategy, we recognized that
the most important issue in remote sensing model inversion is accumulation of prior
knowledge regarding the model parameters. Given that we are not only concerned
with the validation of the model, but also with how to apply the inversion algorithm
for land surface parameter estimation at the scale of remote sensing images, we have
to consider establishing a very powerful prior knowledge database that assembles
the initial estimates of the parameters and also their co-variance values during the
inversion procedure. These initial parameter values and their co-variance matrix are
also necessary for supporting data assimilation algorithms.
In practice, we firstly classify the a priori knowledge into different levels: the
general knowledge about the land surface, or global knowledge, knowledge related to land cover type, and target-specific knowledge. The means to accumulate
knowledge at these levels may be different, but should include the following:
1.
2.
3.
4.
5.
Applicable forward model(s)

Physical limits and probability density in model parameter space
Statistics of model accuracy and noise in remote sensing signals
Seasonal change associated with land cover types or targets
Confidence of the above knowledge
Note that even a single observation can change the a priori PDF of more than one
parameter significantly. Li et al. (1998) provided an example on how the accumulation of knowledge is achieved in parameter space. However, the required numerical
integration in parameter space is time-consuming whenever parameters set is large,
as is the case with the retrieval of surface BRDF and albedo from satellite data.
208
J. Wang, X. Li
Our recent approaches are to establish an a priori knowledge database of the

parameters of typical land surface targets, including the spatial and temporal distribution knowledge of parameters. This knowledge database also includes the forward
models, remote sensing data and a priori knowledge at both global and land cover
type related levels. Based on the knowledge base, it is expected that the accuracy of
land surface parameters estimation at the scale of remote sensing observations will
improve (Wang et al., 2003).
8.2 The Spectrum Knowledge Database of Typical Land Surface

Objects
Establishing the spectrum knowledge database of typical land surface objects is significant for the development of quantitative remote sensing. The research on spectral features of objects is the foundation of remote sensing applications. Land cover
and land use classification and image interpretation are usually based on the spectrum recognizing method. Many kinds of spectrum matching techniques have been
developed. However, users cannot always obtain the desired accuracy for classification and identification. One of the problems is that the spectral data of objects
measured at different scales are not comparable. It is rare to have the correct relation
between the spectrum of objects measured indoors or in the field, and the spectral
image data acquired through remote sensing observations. The available spectrum
data measured indoors and in the field do not have enough corresponding descriptors to clearly determine the related spectral environmental variables. This can result
in some confusion in spectrum applications using spectra matching and other image
processing algorithms.
The second problem is that we need to integrate the physical models and measurement data into a close linked system to allow models and data to be effectively applied for land surface parameters estimation. When the linked system is
developed, the prior knowledge of the model parameters can be extracted from the
database and the model prediction and parameters inversion algorithm can also be
included in the system. That is our approach to convert the individual measurement
data sets to be the knowledge database with the stated goal of having effective physical model explanations and predictions.
The establishment of the spectrum knowledge database of typical land surface
objects is supported by Chinas National High Technology Research and Development Program. The spectrum knowledge database consists of the measuring spectral
data set and related environmental variables of objects at different observing scales,
the physical models set and a priori knowledge set which provide the main geographical background data for models.
209
8.2.1 The Definition of Spectrum at Three Measured Scales

To avoid confusion in spectral matching for land cover type recognition, we clearly
defined and measured three kinds of spectrum for different observed objects, while
considering the application requirements. The three spectrums are the spectrums of
materials, endmembers and remote sensed pixels.
The spectrum of materials is usually measured under well controlled measurement conditions in the laboratory. The measured objects can be crop leaves, mineral
samples, and water samples. They are essential spectrum data of the spectrum database, and are usually taken as given variables in physical models.
The so-called spectrum of endmembers is the spectrum of components of remotely sensed pixel, it is the basic parameter in general remote sensing models. The
spectrum of endmembers is usually measured in the field, where the surface of the
measured object is relatively uniform, and where the measuring FOV of sensor is
less than pixel size. For example, in the scene integrated geometric optical BRDF
model on canopy reflectance, the sunlit crown, sunlit background, shaded crown and
shaded background are endmembers of the modeled forest pixel. Their spectrums
are usually the main parameters of scene integrated models, which may differ from
the spectrum of leaves and soils.
The spectrum of remotely sensed pixels comes from remotely sensed observations, and is usually the main concern of users. Note that when we consider different
modeling scales, the measured endmember spectrum and the spectrum of remotely
sensed pixels should be used according to the scale of the application (Li et al.,
2002; Wang et al., 2003).
8.2.2 Typical Land Surface Objects Spectrum Database

In the spectral data set, for every typical object the measured data includes not only
its surface spectra, but also the environmental variables of each observed object.
The variables are all physical model parameters, which can be used to predict the
spectra of a given surface object by a physical model. In the first step of the database
establishment, the main typical land surface objects include three land cover types:
the main crops growing in China, which are winter wheat, rice, maize, cotton and
cole; rocks and minerals; and bodies of water.
The collected data consist of two parts. One part is the individual spectrum data
sets measured during the last 20 years. In order to keep this data available, we compiled the data following our own data collection standards. For instance, some of
the data were measured in the laboratory and in standard well controlled measuring
conditions, such as for minerals.
The other part consists of new measurements. We first determined the environmental variables to be measured, which are dependent on the mature physical models, and widely required parameters for remote sensing applications. We established
the technical criteria for remote sensing experiment instruments, laboratories and
210
J. Wang, X. Li
field sites, and also the technical criteria and requirements for the measurement of
objects spectrum and environmental variables. All new data measurements follow
these criteria and regulations, ensuring that the measurements of surface spectra and
environmental variables are made together. For those objects which surface features
may change with time, such as crop vegetation canopy, we made measurements during each of their main growth stages across their entire growing season. These data
sets with temporal changes can be used to extract the temporal change information
of the canopy surface reflectance and canopy structural parameters, which is the
most important prior knowledge for canopy reflectance model inversion and data
assimilation.
8.2.3 From Spectrum Database to Knowledge Base

To make the spectrum database applicable for quantitative remote sensing, we further describe how to develop the spectrum database to the spectrum knowledge base.
We do this by integrating the spectrum database, the remote sensing image base, the
remote sensing model base, and the geographic background database to create the
spectrum knowledge database. The remote sensing model base is the key part of
the knowledge base. The models describe the relationship between the spectrum
reflectance measurements and related environmental variables. The spectra of materials, of endmembers, of remote sensing pixels and the measuring environment
variables can be linked by models when the scale of observation is known. The geographic background knowledge data sets provide the prior knowledge of the model
parameters for running the models.
A significant function of the knowledge base is to make surface spectra simulation or prediction. Because there are too many types of land surface objects, the
amount of objects we can measure is always limited. The potential change of the
objects are infinite, we can not include all the cases for even one given object. For
example, for winter wheat, we almost cannot obtain all the spectrum for its entire
growing seasons from the different regions where it is planted. Considering the application of spectrum database, users may occasionally require the spectra of winter
wheat for a given date from a specific place, which may not be stored in the database.
The simulation or prediction of spectrum is achieved by using physical-based
remote sensing models, and applicable empirical models on objects, such as crop
simulation models, as well as the knowledge base. Therefore, our knowledge base
developed the ability to simulate spectra. When the spectra of an object requested
by the user is not in the database, the remote sensing models can be used to simulate
the spectra, based on measured spectrum and environmental variables of similar
land surface objects saved in the knowledge base. For example, our database stored
some standard spectrum of winter wheat at its different growing stages, while the
user requires the spectrum at a specific stage in order to predict its growing state
from his remotely sensed image. The spectrum from existing spectrum and from
related structural parameters of several growing seasons will be calculated with a
211
spectral prediction modular. The crop simulation model will provide information
about the growing tend, the geographic knowledge base will provide the information
about phenological and regional characteristics. The spectral database will provide
the spectrum of material and components, and also the structural parameters of a
typical related winter wheat. In this way, the modular can interpolate and extend the
spectrum in temporal and spatial scales, and also predict the spectrum at a given field
of view of sensor, a given sunlight, and under the specific atmospheric conditions.
The simulated spectra will provide the user with a reference. Spectra data extension
is an important feature of our knowledge base. It can also be available as a research
platform for studies on object surface spectra prediction and model validation.
At present, in our knowledge base, the model base includes general models, physical models, and some special application models. The geographic background data
includes the following: the DEM of 1:100,0001:250,000 of the demonstration region, and land use map of the demonstration region, the base data of 1:4,000,000
covering the whole of China (DEM, physiognomy, vegetation, soil, geosciences,
rivers and lakes), phenological phase of typical crops, Chinese geological map
(1:5,000,000), such as Nonmetals Metals Mineral Resource Map, Mineral Resource
Map and China water resource map (1:4,000,000).
From the measured database, we can construct the prior knowledge of land surface parameters by the statistical data analysis. The a priori knowledge of parameters can be expressed as their mean and variance at spatial and temporal scales of
the accumulated data. The establishment of the spectrum knowledge base allows
knowledge based inversion strategy and algorithm to be used to process remotely
sensed image (Wang and Li, 2004).
8.2.4 The Internet-based Spectrum Knowledge Database Service

System
The internet-based spectrum knowledge database service system includes the software system and its corresponding hardware environment. This system consists of
several modules, such as spectrum querying, model calculation, knowledge querying, image management, application demonstrations, and system management. The
spectrum and models in our knowledge database were all imported according to the
data quality level regulations.
The system also has several demonstrations of typical field remote sensing applications, which include the following: production estimation and growth monitoring
of crops in North China; precise agriculture demonstration of winter wheat, demonstrations of cotton growth monitoring; rock and mineral mapping using airborne
hyperspectral remotely sensed data; water quality evaluation of the Huangpu River
in Shanghai with airborne hyperspectral remotely sensed data. By employing the explanations in the demonstration system, users can learn how to apply the spectrum
and other data to extract required land surface parameters from remotely sensed
images.
212
J. Wang, X. Li
8.3 Land Surface Parameters Inversion Supported

by the Knowledge Database
Based on the spectrum database, we propose our methods to estimate land surface
parameters, which make effective use of the spectrum knowledge. There are two
kinds of questions that need to be considered. One is how to abstract the spatial and
temporal distribution of the prior knowledge of parameters from database. When we
estimate vegetation parameters from remotely sensed data, the data we used can be
from a scene covering large region, so we need to have the spatial distribution prior
knowledge of parameters around the same region to support the prior knowledgebased inversion algorithm. When we want to understand the temporal changes of
crop growth parameters during its growing seasons, we need to have the knowledge
with temporal distribution of estimated parameters, such as when we perform data
assimilation using a crop growth model.
The second kind of questions we should consider is how to introduce the information on the spatial and temporal distribution of parameters into the inversion algorithm. The model we use for the inversion should be not only suitable for processing
one special scene remotely sensed data, but also for the data that changes with time.
A dynamic model should then be developed, and a data assimilation algorithm also
should be developed. The new method should be able to perform time sequential
matching for better estimation of parameters at given time and date.
8.3.1 Extract Spatial and Temporal Distribution of Land Surface

Parameters
From the spectrum database, we can extract the spatial and temporal distributed a
priori knowledge of land surface parameters by means of statistical data analysis, as
well as model predictions. To do this, we should first understand the kinds of spatial
and temporal data that are available, and then how to make use of them.
The prior knowledge data of parameters, which are applied in inversion and data
assimilation algorithms, can be at several spatial and temporal scales.
The prior knowledge data at different spatial scales can be categorized in two
types. One type of data relate to the observed objects, but at different resolutions at
which the data is acquired. The spatial resolution of acquired data could be at tens
of centimeters, at tens of meters, or even at several kilometers. Another type of data
relate to the spatial distribution variability from the same or different types of land
surface objects. The spectrum database provides knowledge data by mainly focusing on the former kinds of data. For the latter, some geographic background data
with several certain scales can be available. Looking at the example of inversion of
vegetation canopy reflectance model, the function of prior knowledge provided by
spectrum database can be the following: first, spectrum database can provide the
spectrum of materials or endmembers, such as those of leaves and soil from the
213
background of crop canopy, which can be used to fix some parameter values, if a
model has too many parameters and there are only limited observations to be available for estimate the all model parameters. Secondly, the spectrum database can
provide some statistical information of parameters, such as the parameters mean
value and standard deviation, which can be used as initial guesses and the uncertainty index of parameters to be retrieved.
For prior knowledge data at temporal scale, the spectrum database can provide
some referencable parameters for dynamic models, while retrieving the variable information of land surface parameters. Looking again at the example of crop canopy
reflectance model inversion, due to crop growth over time, the observed spectrum
will change. The dynamic change spectral observations can provide an opportunity to retrieve biophysical and biochemical parameters of the crop surface. Thus,
the dynamic spectral data of continuing growing vegetation can be used to extract
some specific knowledge from the temporal sequence. This type of knowledge may
be used to compensate for the limited information from remotely sensed data. In
addition, the knowledge is useful for data assimilation. The continuous temporal
knowledge of the parameters can be applied to compare with the dynamic model
prediction, and then to provide reference feedback to adjust the models parameter
estimation for the data assimilation procedure. This should result in a better estimation for parameters when their values change with time and with natural features,
such as the leaf area index in the crop growth model.
8.3.2 Inversion Model Based on Bayesian Network to Integrate

Physical Model and Prior Knowledge
Recent research has shown that combining the empirical formula method and the
physical model inversion into a new hybrid inversion scheme for estimating surface parameters will be a promising trend (Fang and Liang, 2005; Liang, 2004).
Following previous work on hybrid inversion, we build a new hybrid inversion
scheme which uses a Bayesian Network (BN) to determine the mapping relationship between simulated reflectances and their corresponding biophysical parameters
(Qu et al., 2005). As a hierarchical probability model, BN can not only be used
as a non-parameters regression model, but can also be used to deduce information
from multi-layered parameters (Marcot et al., 2001). This differs from other nonparameters regression methods, such as Neural Network. In our approach, we focus
on the incorporation of prior knowledge derived from spectrum database and physical model into a unified framework. A simple Bayesian Network is illustrated as
Fig. 8.1.
Using the Bayesian Network to retrieve parameters, the posterior probability density distribution of A can be calculated using the observed data and their ancillary
parameters, the following Eq. (8.3) can be derived using Bayesian theorems.
214
J. Wang, X. Li
Fig. 8.1 A simple Bayesian

Network
Ancillary
Parameter space
p(A|B = bi ,C = ck ) p(C = ck )p(A|C = ck )p(B = bi |A)
Data
(8.3)
Where p(A|C = ck ) presents the prior knowledge of interested parameters, p(C = ck )

represent the probability distribution of other variables, such as time, location, elevation, land use, and other assistant information, which can affect the ancillary
parameters, p(A|C = ck ) is the probability density distribution of the parameters to
be derived after obtaining the above information. The two probability distributions
can be obtained from the spectrum database, and the quantitative influence between
them, i.e., p(A|C) can be obtained statistically by using data simulated by physical
models. By extending the Bayesian theorem into the Bayesian Network, which
uses a multifactor deducing method, the prior knowledge about the land surface
parameters can be extracted from the spectrum database, and then can be combined
with the physical model to retrieve the information on the desired parameters.
By integrating the observed data and new information into a unified framework
to infer knowledge, this new hybrid inversion scheme has shown that it can incorporate more information besides model parameters into the process of remote sensing
model inversion to retrieve biophysical and biochemical parameters. The process
of extracting knowledge from the historical database, and using it in the retrieving
information from remote sensing models and remotely sensed data is the summingup of prior knowledge and new information. In this process, the information from
newly obtained data can be updated and added to the existing knowledge on parameters.
In our preliminary research, our proposed inversion approach was used and validated through retrieving the biophysical and biochemical parameters of winter
wheat. We abstracted the prior knowledge of canopy reflectance model parameters
from the spectrum database to obtain crop growth parameter distribution at different
growth stages. We developed the inversion model based on the Bayesian Network to
integrate the physical model and prior knowledge to estimated LAI and chlorophyll
(a + b) with simulated BRDF spectral data, and to estimated LAI with ETM image
(Qu et al., 2005).
215
8.4 Parameters Scaling and Validation

Scaling effects are the basic issues in remote sensing research. We studied the scaling effect in thermal infrared remote sensing models and the reciprocity principle
when these basic physical principles are applied in processing special problems in
remote sensing (Li et al., 2000a). When we estimate model parameters by remote
sensing model inversion, one important question is whether the parameters require a
uniform estimation value when observations are at different spatial scales. In our experiment of model inversion, one parameter may have different estimated values if
the available remote sensing data are at different scales of observation. In such case,
it is still a problem on how to improve the precision of model parameters estimation.
Taking the estimation of crop planting area as example, we usually obtain different
estimated area value from AVHRR data or from TM data. And the difference may
not follow a certain formula when land cover types from the observed regions are
different. This is mainly due to a greater amount of mixed pixels in AVHRR data,
and the heterogeneity of land surface. It is very similar to the fractal problems, such
as the coastline measurement using the data from different spatial resolution. In
order to describe the difference resulting from different scales of observation, we
proposed the concept of histo-variogram, defined as the total fractal dimension,
and applied it to derive a formula for scaling correction on land surface parameter
estimates (Zhang et al., 2003).
We use LULC data as an example to study land use area estimation method
using down-scaling. For the crop area estimates, we defined the initial area S0 which
means the observed area at initial measured scale, stated its relation to the total
fractal dimensions (d) and coefficients ( f (S0 )), and then applied it to up-scaling
and down-scaling in crop area estimation. The estimated crop area of standard pixel
(S ) at different scale () is expressed as
S = f (s0 ) d2
(8.4)
where ( f (S0 )) is the function of the initial area (S0 ) and with the same unit as S0 .
is for measured scale, can be expressed with fractal of standard pixel.
The relation between fractal (), coefficient (I) and standardized area (S 0 ) is:

(I i S 0 )/I i = 1.62D i + 0.03 0
(8.5)
i 0
where I is the f (S0 ) expressed in Eq. (8.4), can be calculated by Eq. (8.4) with the
initial measured scale i in the process of scaling-up, 0 is for finer measured scale
than i , S 0 is the crop area we want to estimate by down-scaling method at the
measured scale 0 .
We can obtain the estimated area when the scale of the data is changing based on
the down-scaling method. In this way, using LULC data, we obtain the result of crop
plant area estimation result by the down-scaling method, where the relative error is
less than 5% when the estimated pixel size is 16 times that of the original data.
216
J. Wang, X. Li
8.5 Summary and Discussion

The previous research we described are our main ideas on remote sensing model
inversion strategy and algorithm, base on the analysis of the special ill-posed inversion problem in land surface parameters inversion. The available land surface
spectrum knowledge database provides useful models and a priori knowledge of
model parameters. We are presently researching the following:
1. How to use spatial and temporal change knowledge in remote sensing model
inversion? Based on our previous research, we introduce a dynamic model to obtain a time consequence of model parameters. The idea on data assimilation will
then be applied to improve the precision and reliability of parameters estimation.
2. How to validate and evaluate the parameters inversion result when the raw data
are at different scales? This is still a major problem filled with many uncertainties. The problem regards the scaling effect of models, parameters and observations.
3. How to meet users requirements for land surface parameters estimated by remote sensing data, so as to provide input for common land model (CLM)? In this
case, the a priori knowledge of the parameters spatial distribution and temporal
signature will be obtained from the knowledge base, and then used to invert these
dynamic changing parameters by introducing the main idea of data assimilation.
References
94(3):405424
Gao F, Zhu Q (1997) Process system of measured bidirectional reflectance in Changchun laboratory. J. Remote Sens. 1(Suppl):123130
Kimes DS, Harrison PR, Ratcliffe PA (1991) A knowledge-based expert system for inferring vegetation characteristics. Int. J. Remote Sens. 12:19872020
Kimes DS, Knyazikhin Y, Privette J, Abuelgasim A, Gao F (2000) Inversion methods for
Kuusk A (1991) Determination of vegetation canopy parameters from optical measurements.
Li X, Yan G, Liu Y, Wang J, Zhu C (1997) Uncertainty and sensitivity matrix of parameters in
inversion of physical BRDF model. J. Remote Sens. 1(Suppl):113122
Li X, Wang J, Hu B, Strahler AH (1998) On utilization of prior knowledge in inversion of remote
sensing models. Sci. China (Series D) 41(6):580586
Li X, Wang J, Strahler AH (2000a) Scale effects and scaling-up by geometric-optical model. Sci.
China (Series E) 43(Suppl):1722
Li X, Gao F, Wang J, Strahler AH (2000b) Estimation of the parameter error propagation in inversion based BRDF observations at single sun position. Sci. China (Series E) 43(Suppl):916
Li X, Gao F, Wang J, Strahler AH (2001) A priori knowledge accumulation and its application to
linear BRDF model inversion. J. Geophys. Res. 106(D11):1192511935
217
Li X, Wang J, Gao F, Strahler A, Su L (2002) Accumulation of spectral BRDF knowledge for quantitative remote sensing of land surfaces. Proc. The First Int. Symposium on Recent Advances
in Quantitative Remote Sensing, Torrent, Valencia, Spain, 16 September 2002
Liang S (2004) Quantitative Remote Sensing of Land Surfaces. Wiley, New York
Marcot BG, Holthausen RS, Raphael MG, Rowland M, Wisdom M (2001) Using Bayesian belief
networks to evaluate fish and wildlife population viability under land management alternatives
from an environmental impact statement. Forest Ecol. Manage. 153(13):2942
Morisette JT, Baret F, Privette JL, Myneni RB, et al. (2006) Validation of global moderateresolution LAI products: a framework proposed within the CEOS land product validation subgroup. IEEE Trans. Geosci. Remote Sens. 44(2):18041817
Myneni RB, Maggion S, Iaquinta J, et al. (1995) Optical remote sensing of vegetation: modeling,
caveats and algorithms. Remote Sens. Environ. 51:169188
Privette JL, Eck TF, Deering DW (1997) Estimating spectral albedo and nadir reflectance
through inversion of simple BRDF models with AVHRR/MODIS-like data. J. Geophys. Res.
102(D24):2952929542
Qu Y, Wang J, Song J, Lin H (2005) A hybrid inversion scheme to estimate biophysical parameters
of winter wheat: simulation and validation. Proc. ISPMSRS05, Beijing,, 1719 October 2005,
pp 743746
Rodgers CD (1976) Retrieval of atmospheric temperature and composition from remote measurement of thermal radiation. Rev. Geophys. Space Phys. 14(4):609624
Smith JA (1993) LAI inversion using a back propagation neural network trained with a multiple
scattering model. IEEE Trans. Geosci. Remote Sens. 31:11021106
Tan B, Hu J, Huang D, et al. (2005) Assessment of the broadleaf crop leaf area index product from
the Terra MODIS instrument. Agric. Forest Meteorol. 135:124134
Tarantola A (1987) Inverse Problem Theory Methods for Data Fitting and Model Parameter
Estimation. Elsevier, Amsterdam, The Netherlands, 613pp
Verstraete MM, Pinty B, Myneni RB (1996) Potential and limitation of information extraction on
the biosphere from satellite remote sensing. Remote Sens. Environ. 52:201214
Wanner W, Strahler AH, Hu B, Lewis P, Muller J-P, Li X, Barker Schaaf CL, Barnsley MJ (1997)
Global retrieval of bidirectional reflectance and albedo over land from EOS MODIS and MISR
data: theory and algorithm. J. Geophys. Res. 102(D24):1714317162
Wang J, Li X (2004) The spectrum knowledge base of typical objects and remote sensing inversion
of land surface parameters. J. Remote Sens. 8(Suppl):47 (in Chinese)
Wang J, Li X, Sun X, Liu Q (2000) Component temperatures inversion for remote sensing pixel
based on directional thermal radiation model. Science in China (Series E) 43(Suppl):4147
Wang J, Zhang L, Zhang H, Li X, Liu S (2003) Current progress of constructing the spectrum
knowledge base of typical objects of land surfaces, ESA SPECTRA workshop in France (Italy),
November 2003
Yan G, Wang J, Li X (2001) Making use of a priori knowledge of vegetation spectrum in inversion
for canopy structure parameters. Proc. IGARSS01, Sydney, Australia, 913 July 2001
Zhang H, Jiao Z, Yang H, Li X, et al. (2003) Research on scale effect of histogram. Sci. China
(Series D) 45(10):112
Chapter 9
Retrieval of Surface Albedo from Satellite

Sensors
Crystal Schaaf, John Martonchik, Bernard Pinty, Yves Govaerts, Feng Gao,
Alessio Lattanzio, Jicheng Liu, Alan Strahler, and Malcolm Taberner
Abstract Observations from a number of polar-orbiting and geostationary satellite

sensors are now being used to produce operational land surface albedo products
for range of modeling applications. The MODIS, MISR and Meteosat algorithms
are presented as examples of the current strategies being employed to best exploit
multi-day sequential, multi-angular instantaneous, and multi-temporal observations
and accurately specify the reflective qualities of the underlying surface. While these
retrievals represent a major advance in the remote sensing of the spatial and temporal
heterogeneity of the surface, issues such as atmospheric correction, directional-tohemispherical conversion, and spectral interpolation remain to confound the satellite signal and introduce uncertainties and variability within and between products.
Nevertheless, the potential of using multiple products and fusing recent observations
with remotely sensed historical data must be explored as a realistic way to meet the
needs of the modeling community.
Keywords: MODIS MISR Meteosat albedo reflectance anisotropy
Crystal Schaaf, Jicheng Liu and Alan Strahler

Department of Geography and Environment, Boston University, Boston, USA
schaaf@bu.edu
Feng Gao
Earth Resources Technology, Jessup, USA
John Martonchik
Jet Propulsion Laboratory, Pasadena, USA
Bernard Pinty and Malcolm Taberner
Global Environment Monitoring Unit, IES, EC Joint Research Centre, Ispra (VA), Italy
Yves Govaerts
EUMETSAT, Darmstadt, Germany
Alessio Lattanzio
Makalumedia GMBH, Darmstadt, Germany
219
220
C. Schaaf et al.
9.1 Introduction
The availability of a large number of directional observations sampling the viewing
hemisphere over a particular land surface can effectively capture its surface
anisotropy and thus be used to accurately compute the surface albedo of that surface.
While numerous samples may be possible in the field or laboratory, remotely sensed
retrieval methods based on data from individual satellites usually must suffice with a
limited number of directional reflectances of the surface, and the producers of such
data sets must acknowledge that these observations may not necessarily represent
a well-distributed sampling (Privette et al., 1997). Therefore a model is usually
adopted to characterize the surface anisotropy a model which can be inverted with
a finite set of angular samples and then be used to predict surface reflectance in
any sun-view geometry and derive surface albedo (Roujean et al., 1992; Walthall
et al., 1985; Rahman et al., 1993; Engelsen et al., 1996; Wanner et al., 1997;
Pinty et al., 2000a; Breon et al., 2002; Maignan et al., 2004).
The acquisition of directional measurements from an individual sensor is determined by its scanning configuration and the platforms orbital characteristics
(Barnsley et al., 1994). However, cloud obscuration always reduces the number of
clear-sky observations possible. Therefore, in the case of a single field of view sensor such as the MODerate Resolution Imaging Spectroradiaometer (MODIS), on
board the polar orbiting Terra and Aqua platforms, an adequate directional sampling of surface reflectances can only be obtained by the accumulation of sequential observations over a specified time period. Multi-angular instruments such as the
Multiangle Imaging SpectroRadiometer (MISR) instrument (also on board the Terra
platform) obtain sufficient simultaneous directional observations to specify the surface anisotropy whenever a cloud-free acquisition is possible. Geostationary sensors
(such as Meteosat) must trade numerous acquisitions under different illumination
conditions during a day for directional observations to obtain the angular information necessary to sample the surfaces directional characteristics. Since 2000, all of
these approaches have been implemented operationally to produce robust surface
albedo fields for use in climate, hydrological, biogeochemical, and weather prediction models.
9.2 Background
As a key land physical parameter controlling the surface radiation energy budget (Dickinson, 1983, 1995), global surface albedo with an absolute accuracy of
0.020.05 is required by climate models at a range of spatial and temporal scales
(Henderson-Sellers and Wilson, 1983). Land cover-based schemes have historically been adopted in most of the land surface models and climate models for the
parameterization and specification of surface albedo (Bonan et al., 2002; Sellers
et al., 1996). Natural landscapes, however, are a collection of nested objects in a
hierarchy and various processes control the biophysical characteristics at different spatial scales (Woodcock and Harward, 1992; Collins and Woodcock, 2000).
9 Retrieval of Surface Albedo from Satellite Sensors
221
Therefore, land surface models usually allow a sub-grid specification of land cover
proportions to account for the heterogeneity of surface properties within a grid
(Dickinson et al., 1995; Bonan et al., 2002), while the climate models are generally
implemented at coarser spatial resolutions. However, the increasing spatial resolution of modern climate models makes it necessary to examine the spatial features
of global surface albedo and the effect of spatial scales on the albedo specification. Therefore, a consistent and accurate global albedo data set is essential to the
investigation of the sensitivity of climate to various types of forcing and to the
identification of the effects of human activities. Satellite remote sensing represents
the only efficient way to compile such consistent global albedo characterizations.
Historically, global albedo data sets have been derived from the Advanced Very
High Resolution Radiometer (AVHRR) (Csiszar and Gutman, 1999) and the Earth
Radiation Budget Experiment (ERBE) radiometer (Li and Garand, 1994). With
the advent of routine albedo products derived from MODIS (Gao et al., 2005;
Schaaf et al., 2002; Lucht et al., 2000), MISR (Martonchik, 1997; Martonchik
et al., 1998b), CERES (Clouds and the Earths Radiant Energy System), POLDER
(Polarization and Directionality of the Earths Reflectances) which is currently on
board PARASOL (Polarization & Anisotropy of Reflectances for Atmospheric Sciences) coupled with Observations from a Lidar (PARASOL) (Leroy et al., 1997;
Hautecoeur and Leroy, 1998; Bicheron and Leroy, 2000; Maignan et al., 2004),
and Meteosat (Pinty et al., 2000a,b), albedo data sets with spatial resolutions of
500 m to 20 km and temporal frequencies of daily to monthly are now available. Although the retrieval of albedo from these instruments represents a major advance in
sensing the spatial and temporal surface heterogeneity, issues such as atmospheric
correction, directional-to-hemispherical conversion, and spectral interpolation can
still confound the satellite signal and introduce uncertainties. Most of these satellite products rely on sophisticated radiative transfer methods (Vermote et al., 1997;
Berk et al., 1998; Liang et al., 1999; Liang, 2000) and bidirectional modeling
(Roujean et al., 1992; Walthall et al., 1985; Rahman et al., 1993; Engelsen et al.,
1996; Wanner et al., 1995; Wanner et al., 1997; Martonchik et al., 1998b; Pinty
et al., 2000a, b) to obtain accurate surface quantities.
The modeling community has enthusiastically begun to utilize these global and
regional satellite albedo products as they have become available (Oleson et al., 2003;
Zhou et al., 2003; Tian et al., 2004; Roesch et al., 2004; Knorr et al., 2001; Myhre
et al., 2005a, b). With 5 or more years of data now available, interannual variations
can be explored and short-term climatologies prepared which compensate for transient cloudiness or snowcover (Moody et al., 2005; Gao et al., 2005; Barlage et al.,
2005). However, there remains the need to generate analogous surface albedo products prior to year 2000 and in particular over the last 25 years or so where Earth
observing systems from space (e.g., the series of weather satellites) have been acquiring relevant data. Unfortunately, the design of the large majority of these global
observation systems for environmental applications has been driven solely by demands in the domain of meteorology and weather forecasting and, as a consequence,
these sensors do not fulfill some basic requirements for quantitative remote sensing
applications over land, such as those related to the accurate sensor characterization,
geolocation, and calibration.
222
C. Schaaf et al.
However, despite a number of technological limitations, these historical weather

sensors constitute the only possible solution left to remote sensing scientists to assess such quantities as surface albedos at a global scale over the past decades. The
required sequential accumulation of data over multiple days, i.e., for different view
conditions, as adopted for MODIS sensors for inferring flux quantities from a number of instantaneous radiance measurements, can be extended to exploit the AVHRR
series data archive (see for instance dEntremont et al., 1999). An analogous strategy of sequential accumulation (but over every single day, i.e., for different solar
illumination conditions), can be envisaged in the case of the archived measurements
collected by sensors placed on a geostationary orbit. The exploitation of the reciprocity principle then allows the production for every sample area, of daily accumulated datasets of radiances measured at different viewing angles (see for instance,
Lattanzio et al., 2006). Assuming thus that the geophysical system under investigation does not suffer from drastic changes during the period of data accumulation,
e.g., multiple hours (days) for geostationary (polar) orbiting sensors, the temporal
sampling of the radiance field for a given location can be interpreted as an angular
sampling.
9.3 MODIS Albedo and Anisotropy Algorithm

The operational MODIS albedo and anisotropy algorithm makes use of a kerneldriven, linear model of the Bidirectional Reflectance Factors (BRFs), which relies
on the weighted sum of an isotropic parameter and two functions (or kernels) of
viewing and illumination geometry (Roujean et al., 1992) to estimate the Bidirectional Reflectance Distribution Function (BRDF). One kernel is derived from radiative transfer models (Ross, 1981) and the other is based on surface scattering
and geometric shadow casting theory (Li and Strahler, 1992). The kernel weights
selected are those that best fit the cloud-cleared, atmospherically corrected surface
reflectances available for each location over a 16-day period (Lucht et al., 2000;
Schaaf et al., 2002). This model combination (Ross-Thick/Li-Sparse-Reciprocal or
RTLSR) has been shown to be well suited to describing the surface anisotropy of
the variety of land covers that are distributed world-wide (Privette et al., 1997;
Lucht et al., 2000) and is similar to the kernel-driven schemes used to obtain
anisotropy and albedo information by the POLDER (Leroy and Hautecoeur, 1998;
Bicheron and Leroy, 2000; Maignan et al., 2004) satellite sensor. Once an appropriate anisotropy model has been retrieved, integration over all view angles results
in a Directional Hemispherical Reflectance (DHR) or a black-sky albedo at any
desired solar angle and a further integration over all illumination angles results in
a BiMemispherical Reflectance (BHR) under isotropic illumination or a white-sky
albedo. These albedo quantities are intrinsic to a specific location and are governed
by the character and structure of its land cover. They can be combined with appropriate optical depth information to produce an actual (blue-sky) albedo for a specific time such as would be measured at the surface by field sensors under ambient
223
illumination. The anisotropy models can also be used to compute surface reflectances at any other view or solar zenith angle desired. The spectral acquisitions can
also be combined via narrow to broadband conversion coefficients (Liang et al.,
1999; Liang, 2000) to provide broadband anisotropy information and thus broadband albedos similar to those routinely collected in the field with pyranometers and
commonly used in large-scale models.
The MODIS instruments on both Aqua and Terra have a 16-day repeat cycle
and provide measurements on a global basis every 12 days. The 16-day period has
also been chosen as an appropriate tradeoff between the availability of sufficient
angular samples and the temporal stability of surface (Wanner et al., 1997; Gao
et al., 2001). This assumption becomes tenuous during periods of strong phenological change such as vegetation greenup, senescence, or harvesting. By overlapping
processing of the data such that retrievals are attempted every 8 days (based on
all clear observations over the past 16 days), some of the phenological variability can be more accurately captured. Other periods of rapid change at the surface
such as ephemeral snowfall also provide challenges in retrieving appropriate surface albedos. The MODIS algorithm addresses this by determining whether the
majority of the clear observations available over a 16-day period represent snowcovered or snow-free situations and then retrieving the albedo of the majority condition accordingly.
For those locations where the full anisotropic model described above can not be
confidently retrieved due to poor or insufficient input observations, a backup algorithm is employed. This method (Strugnell and Lucht, 2001; Strugnell et al., 2001)
relies on a global database of archetypal anisotropic model based on a land cover
classification and historical high quality full model retrievals. This a priori data base
is then used as a first guess of the underlying anisotropy and any available observations are used to constrain the model. While considered a lower quality result, Jin
et al. (2003a, b) and Salomon et al. (2006) have found that this backup method often
performs quite well under normal situations (e.g., Fig. 9.1).
In view of the often insufficient angular sampling available, a synergistic use
of multi-sensor observations has offered the best opportunity to improve both the
coverage and the quality of global anisotropy and albedo retrievals. Terra has a descending equatorial crossing time of 10:30 a.m., while Aqua is flying in an ascending orbit with a 1:30 p.m. equatorial crossing time. By combining MODIS observations from both Terra and Aqua, more high-quality, cloud-free observations
(under varying solar zenith angle) are available to generate better constrained model
retrievals (see Fig. 9.1). Since the MODIS-Terra and MODIS-Aqua have similar instrument characterizations and utilize the same atmospheric correction algorithm,
the combination of these data is fairly straightforward. However, the calibration and
geolocation of both instruments must be continually monitored for compatibility and
the quality of the aerosol retrieval from each sensor and its effect on the respective
atmospherically-corrected surface reflectances must also be accounted for. In general, the combined Terra and Aqua MODIS product processing stream begins with
a detailed quality check of each atmospherically corrected surface reflectance and
then assigns various penalty weights to the individual observations according to
224
C. Schaaf et al.
MODIS Surface Albedo vs. Boulder, Colorado, 2003
0.90
0.81
Ground Site (Daily at LSN)

Ground Site (16-Day Average)
0.72
TERRA (Full Inv.)

TERRA (Mag. Inv.)
0.63
Surface Albedo
AQUA + TERRA (Full Inv.)

AQUA + TERRA (Mag. Inv.)
0.54
0.45
0.36
0.27
0.18
0.09
0.00
17 33 49 65 81 97 113 129 145 161 177 193 209 225 241 257 273 283 305 321 337 353 369
Julian Day
Fig. 9.1 MODIS 16-day broadband albedo as compared to field data from the Boulder BSRN
tower site
the quality flag contained in each surface reflectance product (Schaaf et al., 2002).
Thus, the quantified uncertainty of the sensor-specific surface reflectances is directly integrated into the retrieval. Results from the combined Terra-MODIS and
Aqua-MODIS algorithm (Fig. 9.2) indicate that the increase in the number of observations does result in more higher quality retrievals and can decrease the use of
backup retrievals by as much as 50% (Salomon et al., 2006).
The MODIS BRDF/Albedo standard operational products (Lucht et al., 2000;
Schaaf et al., 2002; Gao et al., 2005) provide the best fit RTLSR model parameters
describing the surface anisotropy, black sky and white sky albedo quantities,
the nadir (view-angle-corrected) surface reflectance of each location, and extensive quality information. The best fit RTLSR model parameters are retrieved
for the first seven spectral bands of MODIS and three additional broadbands
(0.30.7 m, 0.75 m, 0.35 m). These anisotropy models are then used to compute white sky albedo and black sky albedo at local solar noon for the same seven
spectral bands and three broadbands. The anisotropy models are also used to correct
surface reflectances for view angle effects and provide BRFs at a common nadir
view angle (Fig. 9.3). These Nadir BRDF-Adjusted Reflectances (NBAR) are computed for the seven spectral bands and are used as the primary input for the MODIS
Land Cover and Land Cover Dynamics Products due to their stability and temporal
consistency (Friedl et al., 2002; Zhang et al., 2003). In addition to the standard 500 m
and 1 km tiled products in a sinusoidal projection, these same science data sets are
225
Fig. 9.2 Quality improvement possible by combining MODIS observations from both the Terra
and Aqua platforms. Top panel shows Terra alone (green high quality, red lower quality) while
bottom panel shows Terra and Aqua (March 2006)
MODIS Reflectance (MODO9GHK) 2004-126
Nadir BRDF-Adjusted Reflectance (NBAR) 2004-126
NIR(0.1-0.4) Red (0.0-0.2) Green (0.0-0.18)
NIR(0.1-0.4) Red (0.0-0.2) Green (0.0-0.18)
MODO9GHK. View Angular effects
Nadir BRDF-Adjusted Reflectance (NBAR).
between two swaths
Angular effect is removed
(North China Plain)
(North China Plain)
Fig. 9.3 Application of the MODIS anisotropy model parameter product to correct adjoining surface reflectance swaths for view angle effects by generating Nadir BRDF-Adjusted Reflectances
(NBAR) 5 May 2004
226
C. Schaaf et al.
Fig. 9.4 MODIS global white sky albedo (from Terra and Aqua) March 2006
also routinely produced at a 0.05 spatial resolution in a global geographic (latitude/longitude) projection specifically for use by global modelers (Gao et al., 2005).
In Fig. 9.4, the global false color field of spectral white sky albedo (March 2006)
captures the seasonal variation due to vegetation phenology and snow cover extent.
9.4 MISR Albedo and Anisotropy Algorithm

The Multi-angle Imaging SpectroRadiometer (MISR) on the EOS Terra platform
consists of nine pushbroom cameras, viewing symmetrically about nadir in forward to aftward directions along the spacecraft track. Image data are acquired
with nominal view zenith angles relative to the surface reference ellipsoid of
0.0 , 26.1 , 45.6 , 60.0 , and 70.5 in four spectral bands (446, 558, 672, and 866 nm)
and with a crosstrack ground spatial resolution of 275 m to 1.1 km and a swath width
of about 400 km (Diner et al., 1998). After these data are radiometrically calibrated,
georectified, and averaged to a uniform resolution of 1.1 km, the land data undergo
a series of processing steps, resulting in a myriad of surface parameters.
The basic land surface products currently being generated include the spectral
hemispherical-directional reflectance factor (HDRF) at the nine MISR view angles
and the associated BHR. The HDRF is a measure of the directional reflectance of the
surface under ambient atmospheric illumination (i.e., direct plus diffuse radiation).
It is the ratio of the directionally reflected radiance from the surface to the reflected
radiance from an ideal lambertian target under identical illumination conditions as
the surface. The BHR is the HDRF integrated over all reflection angles in the upward hemisphere, i.e., it is the surface albedo under ambient atmospheric illumination. Related MISR surface parameters are the spectral BRFs at the nine MISR view
227
angles and the DHR. The BRF and the DHR characterize the surface in the same
way as the HDRF and BHR, respectively, but are defined for the condition of direct
(i.e., collimated beam) illumination only. Thus, the top-of-atmosphere (TOA) MISR
radiances are first atmospherically corrected to produce the HDRF and the BHR,
surface reflectance properties as would be measured at ground level but at the MISR
spatial resolution. The HDRF and BHR then are further atmospherically corrected to
remove all diffuse illumination effects, resulting in the BRF and DHR. In addition
to these spectral surface reflectance products, the BHR and DHR, integrated over
the wavelength region of Photosynthetically Active Radiation (PAR) (400700 nm),
are also computed. The determination of these surface products requires that the atmosphere be sufficiently characterized in order for the correction process to occur.
This characterization is accomplished by means of aerosol retrieval, a process performed on a region 17.6 17.6 km in size, containing the 1.1 1.1 km size subregions (Martonchik et al., 1998a, 2002a). After a surface BRF is determined at the
subregion scale it is fitted to the three parameter modified RahmanPintyVerstraete
(MRPV) empirical model (Rahman et al., 1993; Engelsen et al., 1996), which provides a convenient representation of the surface scattering characteristics The details
of the retrieval methodologies used to derive these various surface products have
been described by Martonchik et al. (1998b).
The unique capabilities of MISRs multiple cameras allow for a simultaneous
sampling of the surface anisotropy. By coupling the angular information with the
spectral information, the MISR observations can be exploited to capture ephemeral
effects such as springtime snow cover. On 17 April 2001 MISR observed a rural
part of Manitoba and Saskatchewan about 110 km north of the US border (Path 34,
Orbit 7083). Most of MISRs imaging data have a resolution of 1.1 km, but all nine
cameras in the red band (672 nm) and all four bands in the nadir camera take global
data at the higher resolution of 275 m. Figure 9.5 shows two false color images of the
Canadian scene at 275 m resolution, one emphasizing spectral information and the
other, angular information. The image on the left is a multispectral color composite
in which the MISR green band (558 nm), red band, and near IR band (866 nm)
nadir view imagery are colored blue, green, and red. Here, vegetation appears red
due to its high reflectivity in the near IR band and low reflectivity in the green and
red bands. The image on the right is a multiangular color composite in which the
60 forward view, the nadir view, and 60 aftward view images are colored blue,
green and red, respectively, essentially color coding the angular signature of the
scene. Thus, for example, a region with a reflectance predominately in the nadir
direction will appear green. Prominent features in both images are the Assiniboine
and QuAppelle rivers, running southward and eastward, respectively.
The bidirectional reflectance factors for three sites marked by yellow arrows in
Fig. 9.5 are displayed in Fig. 9.6. The solar zenith angle is 42 and the azimuth
angles of the BRFs are about 32 from the principal plane. The northern-most site
is reddish in color in both composite images in Fig. 9.5, indicating vegetation with
substantial backscatter. This backscatter signature is more clearly shown in the BRF
plot in Fig. 9.6 The BHR (i.e., actual albedo), is 0.08 in the red band and the NDVI
is 0.49, implying moderately dense vegetation. The eastern-most site also has a
228
C. Schaaf et al.
Fig. 9.5 Two false color images (275 m resolution) of an area (240 175 km) in central Canada on
17 April 2001 centered on the SaskatchewanManitoba border. The left image is a multispectral
composite in which red (more like purple/blue in the pdf/doc versions of the images) indicates vegetation. The image on the right is a multiangular composite in which green indicates predominate
scattering in the nadir direction
Moderately dense vegetation (albedo = 0.08, NDVI = 0.49)
Snowy forest (albedo = 0.18, NDVI = 0.24)
Agricultural field with light snow (albedo = 0.18, NDVI = 0.13)
0.25
BRF (red band)
0.20
0.15
0.10
0.05
80
60
40
20
20
40
60
80
View Zenith angle (deg)
Fig. 9.6 Bidirectional reflectance factors for three sites marked by yellow arrows in Fig. 9.5
vegetative character, colored purple and red in the multiangular and multispectral
composites, respectively. However, the BRFs for this site are higher and with more
forward scattering than for the previous site, and with an increased red band BHR
of 0.18 and a lower NDVI of only 0.13. The brightening of the BRF, the increase
229
in forward scattering, and the decrease in the NDVI is indicative of the presence of
some snow in and among small scale vegetation, probably agricultural. The last site
in Fig. 9.5 is located in Duck Mountain Provincial Park and is colored green in the
multiangular composite and red in the multispectral composite. This color combination implies vegetation but with strong scattering in the nadir direction. The BRF
plot in Fig. 9.6 shows this signature in more detail, and it is quite different than those
from the other two sites. Here, the vegetation is a forest with snow on the ground
between the trees. When viewing in or near the nadir direction, the snow is highly
visible and the BRF is at its highest. As the view angle progressively increases, the
ratio of snow cover to tree structure decreases, lowering the reflectance until, at the
extreme off-nadir view angles, the BRF is virtually all canopy with its characteristic
pronounced backscatter.
During the 7 weeks from 14 August to 29 September 2000, numerous orbits of
MISR data were analyzed and compiled for southern Africa, as part of the dry season
campaign of the Southern Africa Regional Science Initiative (SAFARI-2000), an international effort to study linkages between land and atmospheric processes. During
this period a number of AERONET sunphotometer sites (Holben et al., 1998) were
operational over the region, providing independent determinations of aerosol optical depth which were compared to those retrieved using MISR data (Diner et al.,
2001). This validation study produced very favorable results, allowing considerable
confidence to be placed in the subsequent atmospheric correction procedures and
in the quality of the retrieved surface products. Figure 9.7 is a true color, 1.1 km
resolution mosaic of the surface DHR for southern Africa, derived from 27 orbital
swaths accumulated during this time period. The bright feature in the center is the
Makgadikgadi Pans, an extensive salt bed in Botswana.
In the interior part of southern Africa, much of the land can be classified as
savanna and grassland. Figure 9.8 shows the HDRFs in all four MISR bands for
grassland not far from Johannesburg. The grass is dried out, as can be discerned
from the monotonically increasing HDRF with wavelength. Using data from 15
August (Path 168, Orbit 3509), this particular site was positioned on the extreme
western edge of MISRs orbital swath, providing multispectral measurements within
1 of the retro-solar direction (direct backscatter). The resulting hotspot, due to an
almost complete lack of shadowing within the structured surface, can be seen very
clearly in Fig. 9.8 as an enhancement of the HDRF in all bands at 49 view zenith
angle, which is also the solar zenith angle. The hotspot, while not common in MISR
surface retrievals, does occur for a wide range of latitudes, appearing at different
camera view angles, depending on the season.
In addition to the standard MISR products available at 275 m or 1.1 km, many are
also available in a format of monthly global maps at a spatial resolution of 0.5 in
both latitude and longitude. An example of this type of map is displayed in Fig. 9.9,
showing surface DHR in natural color for the month of September 2005. The individual 0.5 pixels are created by averaging all 1.1 km DHR values accumulated
within that pixel for that month. The white specks evident in some areas are fill pixels where no 1.1 km DHR values were available for the entire month, due mainly to
cloud activity.
230
C. Schaaf et al.
Fig. 9.7 MISR true color, 1.1 km resolution mosaic of the surface directional-hemispherical reflectance (DHR) for southern Africa (14 August29 September 2000)
DRY GRASSLAND, SOUTH AFRICA
Blue
Green
Red
Near-IR
0.45
0.40
0.35
HDRF
0.30
0.25
0.20
0.15
0.10
0.05
0.00
80
60
40
20
20
40
60
80
View Zenith angle (deg)
Fig. 9.8 Spectral and angular variation of HDRF for dried grassland on 15 August 2000 for a site
near Johannesburg. Note the hotspot at the view zenith backscatter angle of 49 , which is also the
solar zenith angle
231
Fig. 9.9 Surface DHR in natural color for the month of September 2005
9.5 Meteosat Albedo and Anisotropy Algorithm

The cornerstone of the retrieval algorithm for geostationary satellites developed by
Pinty et al. (2000a) relies on the temporal sampling of geostationary satellites (data
acquired every 15 or 30 min from sunrise to sunset) as if it were an instantaneous
angular sampling of the radiance field emerging at the top of the atmosphere. The
frequency of measurements of the same Earth location is indeed a unique capability
offered by geostationary satellites that thus translates into an increasing number of
conditions or positive constraints to be satisfied by the retrieval algorithm.
The physics of the Meteosat retrieval aims at solving an inverse radiation transfer problem simultaneously with respect to the lower boundary condition, i.e., the
surface bidirectional reflectance factor (BRF), and the aerosol optical thickness
(Martonchik et al., 1998b, 2002b; Pinty et al., 2000a). All other effects due, for instance, to water vapour and ozone in the atmosphere are accounted for via a prescription of gas concentrations taken from either climatology and/or weather forecast
models (re) analysis. In order to simplify the problem further, the gaseous absorption processes are treated separately from the aerosol-scattering-absorbing effects
by the specification of two distinct atmospheric layers, one for the representation of
the molecular absorption only and the other for the modeling of the coupled surfaceaerosol radiation transfer processes. The inverse algorithm is basically focusing on
the estimates of key variables, namely the aerosol load and surface scattering properties, for which the a priori knowledge is quite limited or somewhat uncertain and
the level of variability is quite high.
The mathematics of the retrievals is established in such a way that the amplitude of the surface BRF is propagated to the top of the atmospheric scattering layer
while its shape is modulated by the atmospheric scattering and absorbing properties
232
C. Schaaf et al.
(Pinty et al., 2000a). This is made possible thanks to (1) the mathematical formulation of the surface BRF model, namely the RPV model (Engelsen et al., 1996;
Rahman et al., 1993) which separates the amplitude from the shape of the surface
BRF and (2) the decomposition of some atmospheric functions like the upward and
downward diffuse transmission with a Fourier expansion limited to the first two
components. This approach proved to be computer efficient and accurate for modeling the radiance field at the top of a scattering-only atmosphere (Martonchik et al.,
2002b). In this way, the angular field of the BRF at the top of the scattering atmosphere can be simply expressed as a sum of contributions invoking the coupling
between the surface BRF shape and atmospheric scattering functions that can all be
pre-computed and called during the retrieval procedure. For a given set of measured
BRF values at the top of the atmosphere the latter procedure itself solves (1) a linear equation to calculate first the amplitude values of the surface BRF for the given
pre-computed scattering functions and (2) a second order cost function estimating
the closeness between the measured and the modeled BRF values at the top of the
atmosphere.
The retrieval procedure then ends up with an identification of probability distribution functions of the acceptable solutions, i.e., those satisfying one or multiple
criteria depending on the number of degrees of freedom and the distributions of
uncertainties in both the observations and in the forward model. The selection of
the most probable solution for any given set of measured BRF amongst the set
of acceptable solutions can be performed using various criteria including the identification of the solution corresponding to the arithmetic mean of the distribution
of the amplitude of the surface BRF values. The solution retained is thus a set of
model variables and parameters describing the surface scattering problem, such as
the parameters characterizing the shape of the surface BRF and the effective aerosol
loads, associated with the selected value for the amplitude of the surface BRF. The
aerosol loads together with the surface scattering properties are given via effective
optical thickness values for a prescribed aerosol type corresponding to average standard aerosol conditions regarding their detailed properties and vertical distribution
as well. Since the retrieval strategy delivers the optimized set of the RPV model
parameters characterizing the surface BRDF, one can generate DHR or black-sky
albedo for any solar angle and/or BHR or white-sky albedo products (Pinty et al.,
2000a).
The abundance of cloudy conditions occurring during a day over the Earth disk
sections sensed by geostationary satellites motivates the implementation of a procedure screening conditions that do not correspond to clear-sky cases. Pinty et al.
(2000b) suggested adoption of an angular consistency check by which the daily top
of the atmosphere radiance series for each individual pixel is used in an attempt to
fit the MRPV model. This is based on a recursive filtering technique which identifies sequentially during the day, the outliers deviating significantly from the fitted
MRPV model solutions. In the vast majority of the cases, larger (lower) radiance
values than the MRPV fitted solutions are associated with cloudy (shadowed) conditions and the spatio-temporal fields of these outliers were shown to, indeed, display very consistent cloud fields and associated shadows along the course of the day
233
(Pinty et al., 2000a). Measurements acquired at solar and view zenith angles larger
than 70 or corresponding to cloud conditions are rejected. A minimum of six valid
clear sky observations are necessary for the activation of the retrieval procedure.
These cloudy as well as additional undesired conditions translate into incomplete
geographic surface albedo map products. This caveat can be overcome by applying
a time composite algorithm selecting, over a given composite time period, the particular day delivering the albedo value which is the closest to the average of the
ensemble of values retrieved during that same time period. In this way, the geophysical values for each pixel can be delivered with all the relevant information
used to generate them in the retrieval algorithm such as, for example, the number
of observations used to perform the retrieval and estimation of the retrieval uncertainties among others. Sensors on board geostationary satellites sample the scattered
solar radiance fields in a usually single, large (according to todays standards) spectral band (see Fig. 9.10), e.g., 0.41.1 m for Meteosat and GMS, 0.050.8 m
Fig. 9.10 Examples of sensor spectral responses on board geostationary satellites in the solar domain (red line). The green solid line illustrates typical reflectance of green vegetation
234
C. Schaaf et al.
for GOES, which prevents us benefiting from advanced atmospheric correction algorithms based on multi-spectral information, as is possible for the MODIS and
MISR sensors. Due to the fast spatio-temporal variability of the cloud, water vapor
and aerosol fields, the crucial problem of accurately assessing/removing the undesired effects induced by these atmospheric components on the measured radiances
remains to be solved on the sole basis of the available information gathered by the
sensor or other sources of information. For those components, such as ozone, that
have a non-negligible but still limited impact on the surface retrievals and whose
variability remains somewhat small, the use of climatological values is generally acceptable. In most if not all cases to be addressed, the availability of a spectrally large
single band only renders the partitioning between the surface and atmospheric contributions quite difficult since the scattering and absorption effects in the atmosphere
are wavelength-dependent and coupled with spectrally variant surface properties.
To date, the constraints imposed by operational exploitation infrastructures have
not favored the processing of multi-sensor measurements assembled via data fusion
procedures. This state of affairs encourages the development of albedo retrieval algorithms relying on the analysis of data (and data strings) collected by each geostationary sensor in stand alone mode. In turn, this places stringent requirements on
the reliability and overall performance of the retrieval algorithm which thus, on the
basis of a single spectrally large band, must be able to identify cloud occurrence
and then solve, as well as possible, the coupled surface-atmosphere radiation transfer problem. In that context, multiple sensitivity test have to be conducted in order
to optimize the crucial choices to be made such as, for instance, between the length
of the period to perform sequential data accumulation, e.g., between a few hours
and a few days, and the impacts of the assumption hindering this multi-angular data
emulation procedure, e.g., no drastic changes in the geophysical system under investigation. Meteosat data processed with this algorithm have already been used in
a variety of applications (see for instance, Pinty et al., 2000c; Knorr et al., 2001;
Myhre et al., 2005b).
9.6 Fusion of Modern and Historical Surface Albedo Products

Documenting the Earth climate and its variability requires access to reliable and
accurate long time series of environmental products. Hence, archived meteorological satellite observations could contribute to the generation of climatic data records
providing that (1) significant efforts are devoted to the improvement of the spectral
characterization and calibration of these radiometers and (2) developing state of the
art retrieval algorithms. The first point aims at reducing and controlling as much as
possible the impact of measurements uncertainties on the accuracy of the retrievals
while the second point similarly at model uncertainties. Previous studies have, indeed, already demonstrated the possibility to perform post-launch improvements of
the radiometer characteristics (e.g., Govaerts, 1999). In the specific case of surface
albedo retrieval, the science context set up by the requirements in the exploitation
235
of data acquired by modern sensors such as those on board Terra and Meteosat
Second Generation for instance has motivated a large number of studies addressing
these points. The associated improved knowledge translates, in turn, into the development and use of better approaches for the optimal exploitation of measurements
taken by old generation sensors.
While active efforts are underway within the NASA community to couple
MODIS observations to the historical AVHRR archives (Pedelty et al., 2007) considerable progress has already been made in exploiting historical geostationary
satellite data to establish a historical data set of global surface albedo values. This
requires the back processing and analysis of measurements assembled by the fleet
of geostationary satellites over the past 25 years or so. The quality of these retrievals
can be assessed by various means including first, the comparison of these retrieved
albedo values against those operationally generated since year 2000 from modern
and technologically advanced instruments such as MODIS and MISR (Pinty et al.,
2004) and second, the intercomparison of surface albedo products generated over
geographical regions of overlap that are, therefore, sampled simultaneously by
two adjacent sensors together placed on geostationary orbit but located at different
longitudes (Govaerts et al., 2004).
In order to conduct comparison exercises of relevance for climate model applications, surface albedo products have to be made available over large spectral regions covering the energetically relevant solar domain [0.33.0 m] split, whenever
possible, into its broad visible [0.30.7 m] and near-infrared [0.73.0 m] parts.
Achieving this step, usually called spectral conversion, requires developing appropriate tools to transform albedo product values, estimated over and weighted by the
spectral response of the geostationary sensor, into values representative of the desired broad spectral range of interest (see for instance Liang (2000) and Govaerts
et al. (2006)). One possible solution consists in approximating a parametric expression relating the measurements from the sensor to those that would be provided
by an ideal rectangular shape sensor covering the solar domain of interest. Such
an expression can be established on the basis of (1) a large number of simulations
of top of atmosphere radiance fields of various geophysical situations that can be
expected for the region of interest (e.g., vegetation with varying density, bare soils
with different brightness, snow surfaces, coupled with a diversity of atmospheric
conditions) and (2) a multi-regression analysis fitting at best the sensor-like values
against those representative of the desired solar domain. This spectral conversion
constitutes quite a delicate step and, as a matter of fact, its reliability relies strongly
on the degree of coincidence between the distributions of the simulated and actual
conditions. Its impact on the uncertainty of the final albedo products also depends
crucially on the sensor spectral response function since the spectral conversion basically assumes that strong correlations exist between radiances taken across various
wavelengths of the solar spectrum.
Preliminary attempts to compare surface albedo products from modern sensors, such as MODIS and MISR, against those generated by the retrieval algorithm outlined here for geostationary satellites result into quite positive conclusions.
Figure 9.11 illustrates an example of results to be expected when comparing
236
C. Schaaf et al.
Fig. 9.11 Scatterplot (density plot) between the MODIS white sky albedo and Meteosat BHR
values retrieved between 2065 W longitude and 40 S 40 N latitude during the first 2 weeks of
year 2001. This plot includes the high quality flag MODIS products only and the outliers from the
two distributions have been removed (less than 5% of the total number of valid retrievals). The full,
dashed, and dotted lines feature the fit obtained using the slope of the means, the linear regression,
and the primary eigenvector, respectively
Meteosat and MODIS (considering high quality flags only) surface albedo products,
in units of white sky albedo (BHR), over a large geographical region extending from
Southern Europe and covering the entire African continent (between 20 W65 N
longitude and 40 S40 N latitude. This figure is built from the analysis of products available during the first 2-week period of year 2001 after removal of outliers
detected in large majority along the coastlines. The 10-day composite Meteosat
products have been re-mapped into the MODIS Climate Modeling Grid at a spatial resolution of 0.05 . The MODIS and Meteosat spectral albedo products were
both converted into an ideal rectangular shape (0.41.1 m) in order to (1) make
the best possible use of the available spectral information for both sensors, i.e.,
one large band in the Meteosat case and four narrow bands well distributed over this
spectral interval in the MODIS case, and (2) minimize the uncertainties associated
with the required spectral conversion. All three indicators used to characterize the
statistical differences between the MODIS and Meteosat products, i.e., the slope of
237
the means (full line), the regression line (dashed line) and the primary eigenvector
(dotted line), show limited variability around the one-to-one line. The observed statistical differences are largely within the range of the systematic error/uncertainty
due to the calibration knowledge (about 6% for Meteosat-7) (Govaerts et al., 2004)
and/or model approximations (for instance, the decoupling of the gaseous absorption from the aerosol scattering effects).
Figure 9.12 illustrates results from a comparison between the broadband black
sky (DHR) albedo products generated, at 30 solar angle, by the analysis of the daily
radiances collected during the first 10-day period of May 2001 by both GOES West
(GOES-10) and GOES East (GOES-8) over the common land regions of the Earth
disk that they jointly sample (see top panel in Fig. 9.12). These results illustrate the
robustness of the albedo retrieval algorithm since its application yields differences
between data sets from adjacent sensors that are well within the range of their estimated uncertainty level (about 1015% for the GOES sensors) of their broadband
albedo products. Figure 9.12 confirms earlier comparison results obtained by analyzing albedo products generated by two adjacent Meteosat sensors (Govaerts et al.,
2004).
The conclusions drawn from Figs. 9.11 and 9.12 suggest the generation of historical series of surface albedo products based on the exploitation of the fleet of
geostationary satellites. Figure 9.13 is a demonstration example of an output from
such an initiative which assembles broadband products (DHR at 30 Sun zenith angle) retrieved for the first 10-day period of May 2001 from five different satellites,
namely GOES West and East, Meteosat-7 and Meteosat-5, and GMS-5. Table 9.1
provides statistics about these retrievals and the estimated uncertainties associated
with each satellite retrievals. The measurement error includes both the radiometric
uncertainties and approximations in the forward model. The estimated error on the
DHR values is then derived from the uncertainty on the retrieved surface parameters. It thus looks feasible to build global albedo products for the last 25 years or
so, for those places covered by archived data. These preliminary results open new
avenues for the exploitation of geostationary archive data and prototype the fusion
of such the generated products.
9.7 Summary
The MODIS, MISR and Meteosat algorithms represent three complementary strategies for characterizing land surface reflectance anisotropy and obtaining measures of
land surface albedo. Each algorithm makes use of the unique capabilities of its sensor to capture the spectral, spatial, temporal, and angular information necessary to
accurately specify the reflective qualities of the underlying surface cover. With more
than 6 years of MODIS and MISR observations now available, as well as the opportunity to utilize the historical geosynchronous satellite record, the modeling and
data analysis communities enjoy unprecedented access to consistent, high-quality
albedo and anisotropy data of the Earths land surface.
238
C. Schaaf et al.
Fig. 9.12 Comparison between broadband DHR (30 ) values retrieved over the common geographical region covered by both GOES West (GOES-10) and GOES East (GOES-8). The top
panel corresponds to the density plot of the two DHR distributions for the period 110 May 2001.
The bottom panel documents the histogram of the relative differences between the two distributions. The vertical lines colored in blue feature the mean value of these differences (dash-dotted)
and one standard deviation from the mean (dashed)
239
Fig. 9.13 Top panel: Location of operational geostationary satellites which archive data and are
used to derived products shown in the bottom panel. The circles show the 60 viewing angle limit.
Bottom panel: Illustration of the broadband surface albedo map derived from the application of the
geostationary satellite retrieval algorithm on measurements taken simultaneously by GOES-8/10,
Meteosat-5/-7 and GMS-5 over the period 110 May 2001
Table 9.1 Number of days processed during the 110 May 2001 period for each satellite.
< Img/day > is the mean number of measurements available per day (note that some of the GOES
images did not provide the nominal geographical coverage). <Meas. R. Err.> is the average measurement relative error, i.e., including both the radiometric error and forward model uncertainty.
<DHR R. Err.> is the mean estimated DHR relative error
Satellite
GOES-10
GOES-8
MET-7
MET-5
GMS-5
Nbr days
<Img/day>
<Meas. R. Err.>
<DHR R. Err>
10
10
10
10
10
22.9
13.7
17.3
16.3
9.9
5.2%
6.8%
7.4%
10.0%
8.4%
12.5%
14.4%
8.7%
10.1%
10.5%
240
C. Schaaf et al.
Acknowledgements Authors Schaaf, Strahler, and Liu are supported by NASA (under grant
NNG04HZ14) and by their colleagues on the MODIS Science Team while Martonchik is supported by the MISR Science Team. Authors Pinty, Govaerts, Lattanzio, and Taberner are grateful to
the Japan Meteorological Agency (JMA) and the Satellite Services Group of the National Oceanic
and Atmospheric Administration (NOAA) for providing the GMS-5 and GOES-8/-10 data, respectively. Their contributions would not have been possible without the support of the Global Environment Monitoring unit of the Institute for Environment and Sustainability at the Joint Research
Centre, and EUMETSAT.
References
Barlage M, Zeng X, Mitchell K, Wei H (2005) A global 0.05-deg maximum albedo
dataset of snow-covered land based on MODIS observations. Geophys. Res. Lett. 32,
doi:10.1029/2005GL022881
Barnsley MJ, Strahler AH, Morris KP, Muller J-P (1994) Sampling the surface bidirectional reflectance distribution function (BRDF), 1, evaluation of current and future satellite sensors.
Remote Sens. Rev. 8:271311
Berk A, Bernstein LS, Anderson GP, Acharya PK, Robertson DC, Chetwynd JH, Adler-Golden
SM (1998) MODTRAN cloud and multiple scattering upgrades with application to AVIRIS.
Bicheron P, Leroy M (2000) Bidirectional reflectance distribution function signatures of major
biomes observed from space. J. Geophys. Res. 105:2666926681
Bonan GB, et al. (2002) The land surface climatology of the NCAR community land model coupled to the NCAR community ClimateModel. J. Clim. 15:31233149
Breon FM, Maignan F, Leroy M, Grant I (2002) Analysis of hot spot directional signatures measured from space. J. Geophys. Res. 107(16):42824296
Collins JB, Woodcock CE (2000) Geostatistical methods for analysis of multiscale variation in
spatial data. Geographical Analysis 32(1):5063
Csiszar, I, Gutman, G (1999) Mapping global land surface albedo from NOAA/AVHRR. J. Geophys. Res. 104:62156228
dEntremont R, Schaaf CB, Lucht W, Strahler A (1999) Retrieval of red spectral albedo and bidirectional reflectance using AVHRR HRPT and GOES satellite observations of New England.
J. Geophys. Res. 104:62296239
Dickinson RE (1983) Land surface processes and climate-surface albedos and energy balance.
Adv. Geophys. 25:305353
Dickinson RE (1995) Land processes in climate models. Remote Sens. Environ. 51:2738
Diner DJ, Beckert JC, Reilly TH, Bruegge CJ, Conel JE, Kahn RA, Martonchik JV, Ackerman TP,
Davies R, Gerstl SAW, Gordon HR, Muller J-P, Myneni RB, Sellers PJ, Pinty B, Verstraete MM
(1998) Multi-angle Imaging SpectroRadiometer (MISR) instrument description and experiment
overview. IEEE Trans. Geosci. Remote Sen. 36:10721087
Diner DJ, Abdou WA, Bruegge CJ, Conel JE, Crean KA, Gaitley BJ, Helmlinger MC, Kahn RA,
Martonchik JV, Pilorz SH, Holben BN (2001) MISR aerosol optical depth retrievals over southern Africa during the SAFARI-2000 dry season campaign. Geophys. Res. Lett. 28:31273130
Engelsen O, Pinty B, Verstraete M M, Martonchik JV (1996) Parameteric bidirectional reflectance
fractor models: evaluation, improvements, and applications. Technical Report EUR 16426 EN.
EC Joint Research Centre
Friedl MA, McIver DK, Hodges JCF, Zhang X, Muchoney D, Strahler AH, Woodcock CE, Gopal
S, Schnieder A, Cooper A, Baccini A, Gao F, Schaaf C (2002) Global land cover from MODIS:
algorithms and early results. Remote Sens. Environ. 83:287302
Gao F, Schaaf C, Strahler AH, Lucht W (2001) Using a multi-kernel least variance approach to
retrieve and evaluate albedo from limited BRDF observations. Remote Sens. Environ. 76:5766
241
Gao F, Schaaf C, Strahler A, Roesch A, Lucht W, Dickinson R (2005) The MODIS BRDF/Albedo
Climate Modeling Grid Products and the variability of albedo for major global vegetation types.
J. Geophys. Res. 110: D01104, doi:10.1029/2004JD00519
Govaerts Y (1999) Correction of the Meteosat-5 and -6 VIS vand relative spectral response with
Meteosat-7 characteristics. Int. J. Remote Sens. 20:36773682
Govaerts Y, Clerici M, Clerbaux N (2004) Operational calibration of the Meteosat radiometer VIS
band. IEEE Trans. Geosci. Remote Sens. 43:3190031914
Govaerts Y, Lattanzio A, Pinty B, Schmertz J (2004) Consistent surface albedo retrieved from two adjacent geostationary satellite. Geophys. Res. Lett. 31:L15201,
doi:10.1029/2004GL020418 03
Govaerts Y, Pinty B, Taberner M, Lattanzio A (2006) Spectral conversion of surface albedo derived from Meteosat first generation observations. IEEE Geosci. Remote Sens. Lett. 3:2327,
doi:10.1109/LGRS.2005.854202
Henderson-Sellers A, Wilson MF (1983) Surface albedo data for climatic modeling. Rev. Geophys.
Space Phys. 21:17431778
Holben BN, Eck TF, Slutsker I, Tanre D, Buis JB, Setzer A, Vermote E, Reagan JA, Kaufman YJ,
Nakajima T, Lavenu F, Jankowiak I, Smirnov A (1998) AERONET A federated instrument
network and data archive for aerosol characterization. Remote Sens. Environ. 66:116
Hautecoeur O, Leroy MM (1998) Surface bi-directional reflectance distribution function observed
at global scale by POLDER/ADEOS. Geophys. Res. Lett. 22:41974200
Jin Y, Schaaf CB, Woodcock CE, Gao F, Li X, Strahler AH, Lucht W, Liang S (2003a) Consistency of MODIS surface BRDF/Albedo retrievals: 1. Algorithm performance. J. Geophys. Res.
108(D5):4158, doi:10.1029/2002JD002803
Jin Y, Schaaf CB, Woodcock CE, Gao F, Li X, Strahler AH, Lucht W, Liang S (2003b) Consistency
of MODIS surface BRDF/Albedo retrievals: 2. Validation. J. Geophys. Res. 108(D5):4159,
doi:10.1029/2002JD002804
Knorr W, Schnitzler KG, Govaerts Y (2001) The role of bright desert regions in shaping North
African climate. Geophys. Res. Let. 28:34893492
Lattanzio A, Govaerts Y, Pinty B (2006) Consistency of surface anisotropy characterization with
Meteosate observations. Adv. Space Res. doi:10.1016/j.asr.2006.02.049
Leroy M, Deuze JL, Breon FM, Hautecoeur O, Herman M, Buriez JC, Tanre D, Bouffies S,
Chazette P, Roujean J-L (1997) Retrieval of atmospheric properties and surface bidirectional
reflectances over land from POLDER/ADEOS. J. Geophys. Res. 102:1702317037
Li Z, Garand L (1994) Estimation of surface albedo from space: a parameterization for global
application. J. Geophys. Res. 99:83358350
Liang S (2000) Narrowband to broadband conversions of land surface albedo I: algorithms. Remote
Liang S, Strahler AH, Walthall CW (1999) Retrieval of land surface albedo from satellite observations: a simulation study. J. Appl. Meteorol. 38:712725
Lucht W, Schaaf CB, Strahler AH (2000) An algorithm for the retrieval of albedo from space using
semiempirical BRDF models. IEEE Trans. Geosci. Remote Sens. 38:977998
Maignan F, Breon FM, Lacaze R (2004) Bidirectional reflectance of Earth targets: evaluation of
analytical models using a large set of spaceborne measurements with emphasis with the hot
spot. Remote Sens. Environ. 90:210220
Martonchik JV (1997) Determination of aerosol optical depth and land surface directional reflectances using multi-angle imagery. J. Geophys. Res., 102:1701517022.
Martonchik JV, Diner DJ, Kahn RA, Ackerman TP, Verstraete MM, Pinty B, Gordon HR (1998a)
Techniques for the retrieval of aerosol properties over land and ocean using multiangleimaging.
Martonchik JV, Diner DJ, Pinty B, Verstraete MM, Myneni RB, Knyazikhin Y, Gordon HR (1998b)
Determination of land and ocean reflective, radiative, and biophysical properties using multiangle imaging. IEEE Trans. Geosci. Remote Sens. 36:12661281
242
C. Schaaf et al.
Martonchik JV, Diner DJ, Crean KA, Bull M (2002a) Regional aerosol retrieval results from MISR.
Martonchik JV, Pinty B, Verstraete MM (2002b) Note on an improved model of surface BRDFatmospheric coupled radiation. IEEE. Trans. Geosci. Remote Sens. 40:16371639
Moody EG, King MD, Platnick S, Schaaf CB, Gao F (2005) Spatially complete global spectral
surface albedos: value-added datasets derived from Terra MODIS land products. IEEE Trans.
Myhre G, Kvalevag MM, Schaaf CB (2005a) Radiative forcing due to anthropogenic vegetation change based on MODIS surface albedo data set. Geophys. Res. Lett. 32:L21410,
doi:10.1029/2005GL024004
Myhre G, Govaerts Y, Haywood JM, Berntsen TK, Lattanzio A (2005b) Radiative effect of surface albedo change from biomass burning. Geophys. Res. Lett. 32:L20812,
doi:10.1029/2005GL022897
Oleson KW, Bonan GB, Schaaf C, Gao F, Jin Y, Strahler A (2003) Assessment of global
climate model land surface albedo using MODIS data. Geophys. Res. Lett. 30(8):1443,
doi:10.1029/2002GL016749
Pedelty J, Devadiga S, Masuoka E, Brown M, Pinzon J, Tucker C, Vermote E, Prince S, Nagol J,
Justice C, Roy D, Ju J, Schaaf C, Liu J, Privette J, Pinheiro, A (2007) Generating a Long-term
Land Data Record from the AVHRR and MODIS Instruments, Proceedings, IEEE International
Geosciences and Remote Sensing Symposium (IGARSS07), Barcelona, Spain, 2327 July,
2007
Pinty B, Roveda F, Verstraete MM, Gobron N, Govaerts Y, Martonchik J, Diner D, Kahn R (2000a)
Surface albedo retrieval from METEOSAT Part 1: theory. J. Geophys. Res. 105:1809918112
Pinty B, Roveda F, Verstraete MM, Gobron N, Govaerts Y, Martonchik J, Diner D, Kahn R (2000b)
Surface albedo retrieval from METEOSAT Part 2. Appl. J. Geophys. Res. 105:1811318134
Pinty B, Verstraete MM, Gobron N, Roveda F, Govaerts Y (2000c) Do manmade fires affect
the Earths surface reflectance at continental scales? Eos Trans. American Geophys. Union
81:388389
Pinty B, Taberner M, Liang S, Govaerts Y, Martonchik JV, Lattanzio A, Schaaf CB, Verstraete MM,
Dickinson RE, Gobron N, Widlowski J-L (2004) Intercomparison of surface albedo products
from various spaceborne sensors. In: Proc. of the Workshop on Inter-Comparison of Large
Scale Optical and Infrared Sensors, ESA ESTEC, Noordwijk, The Netherlands, 1214 October
2004. ESA ESTEC
Privette JL, Eck TF, Deering DW (1997) Estimating spectral albedo and nadir reflectance
through inversion of simple BRDF models with AVHRR/MODIS-like data. J. Geophys. Res.
102:2952929542
Rahman H, Pinty B, Verstraete MM (1993) Coupled surface-atmosphere reflectance (CSAR)
model 2: semiempirical surface model usable with NOAA advanced very high resolution
radiometer data. J. Geophys. Res. 98:2079120801
Roesch A, Schaaf C, Gao F (2004) Use of Moderate-Resolution Imaging Spectroradiometer bidirectional reflectance distribution function products to enhance simulated surface albedos. J.
Geophys. Res. 109(D12), doi:10.1029/2004JD004552
Ross JK(1981) The Radiation Regime and Architecture of Plant Stands,Junk W (ed), Norwell,
MA: Artech House, p. 392
Roujean J-L, Leroy M, Deschamps PY (1992) A bidirectional reflectance model of the Earths
Salomon J, Schaaf CB, Strahler AH, Gao F, Jin Y (2006) Validation of the MODIS Bidirectional
Reflectance Distribution Function and albedo retrievals using combined observations from the
aqua and terra platforms. IEEE Trans. Geosci. Remote Sens. 44:No. 6
Sellers PJ, Los SO, Tucker CJ, Justice CO, Dazlich DA, Collatz CJ, Randall DA (1996) A revised
land surface parameterization (SiB2) for atmospheric GCMs, part II, the generation of global
fields of terrestrial biospheric parameters from satellite data. J. Clim. 9:706737
243
Schaaf CB, Gao F, Strahler AH, Lucht W, Li X, Tsang T, Strugnell NC, Zhang X, Jin Y, Muller J-P,
Lewis P, Barnsley M, Hobson P, Disney M, Roberts G, Dunderdale M, Doll C, dEntremont
R, Hu B, Liang S, Privette JL (2002) First operational BRDF, Albedo and Nadir Reflectance
Products from MODIS. Remote Sens. Environ. 83:135148
Strugnell N, Lucht W (2001) An algorithm to infer continental-scale albedo from AVHRR data,
land cover class and field observations of typical BRDFs. J. Climate 14:13601376
Strugnell N, Lucht W, Schaaf C (2001) A global albedo data set derived from AVHRR data for use
in climate simulations. Geophys. Res. Lett. 28:191194
Tian Y, Dickinson RE, Zhou L, Myneni RB, Friedl M, Schaaf CB, Carroll M, Gao F (2004) Land
boundary conditions from MODIS data and consequences for the albedo of a climate model.
Geophys. Res. Lett. 31, doi:10.1029/2003GL019104
Vermote EF, Tanre D, Deuze JL, Herman M, Morcrette JJ (1997) Second simulation of the satellite
signal in the solar spectrum: an overview. IEEE Trans. Geosci. Remote Sens. 35:675686
Walthall CL, Norman JM, Welles JM, Campbell G, Blad BL (1985) Simple equation to approximate the bidirectional reflectance from vegetation canopies and bare soil surfaces. Appl. Opt.
24:383387
Wanner W, Li X, Strahler AH (1995) On the derivation of kernels for kernel-driven models of
Wanner W, Strahler AH, Hu B, Lewis P, Muller J-P, Li X, Barker Schaaf CL, Barnsley MJ (1997)
Global retrieval of bidirectional reflectance and albedo over land from EOS MODIS and MISR
data: theory and algorithm. J. Geophys. Res. 102:1714317162
Woodcock CE, Harward VJ (1992) Nested-Hierarchical Scene Models and image segmentation.
Int. J. Remote Sens. 13(16):31673187
Zhang X, Friedl MA, Schaaf CB, Strahler AH, Hodges JCF, Gao F, Reed BC, Huete A (2003)
Monitoring vegetation phenology using MODIS. Remote Sens. Environ. 84:471475
Zhou L, Dickinson RE, Tian Y, Zeng X, Dai Y, Yang Z-L, Schaaf CB, Gao F, Jin Y, Strahler A,
Myneni RB, Yu H, Wu W, Shaikh M (2003) Comparison of seasonal and spatial variations of
albedos from Moderate-Resolution Imaging Spectroradiometer (MODIS) and Common Land
Model. J. Geophys. Res. 108(D15):4488, doi:10.1029/2002JD003326
Chapter 10
Modeling and Inversion in Thermal Infrared

Remote Sensing over Vegetated Land Surfaces
Frederic Jacob, Thomas Schmugge, Albert Olioso, Andrew French,
Dominique Courault, Kenta Ogawa, Francois Petitcolin, Ghani Chehbouni,
Ana Pinheiro, and Jeffrey Privette
Frederic Jacob
Formerly at Remote Sensing and Land Management Laboratory
Purpan Graduate School of Agriculture, Toulouse, France
Now at Institute of Research for the Development
Laboratory for studies on Interactions between Soils Agrosystems Hydrosystems
UMR LISAH SupAgro/INRA/IRD, Montpellier, France
frederic.jacob@supagro.inra.fr
Thomas Schmugge
Gerald Thomas Professor of Water Resources
College of Agriculture
New Mexico State University, Las Cruces, NM, USA
Albert Olioso and Dominique Courault
National Institute for Agronomical Research
Climate Soil Environment Unit
UMR CSE INRA/UAPV, Avignon, France
Andrew French
United States Department of Agriculture/Agricultural Research Service
US Arid Land Agricultural Research Center, Maricopa, AZ, USA
Kenta Ogawa
Department of Geo-system Engineering, University of Tokyo
Japan
Francois Petitcolin
ACRI-ST, Sophia Antipolis, France
Ghani Chehbouni
Institute of Research for the Development
Center for Spatial Studies of the Biosphere
UMR CESBio CNES/CNRS/UPS/IRD, Toulouse, France
Ana Pinheiro
Biospheric Sciences Branch, NASAs GSFC, Greenbelt, MD, USA
Jeffrey Privette
NOAAs National Climatic Data Center, Asheville, NC, USA
245
246
F. Jacob et al.
Abstract Thermal Infra Red (TIR) Remote sensing allows spatializing various land
surface temperatures: ensemble brightness, radiometric and aerodynamic temperatures, soil and vegetation temperatures optionally sunlit and shaded, and canopy
temperature profile. These are of interest for monitoring vegetated land surface
processes: heat and mass exchanges, soil respiration and vegetation physiological activity. TIR remote sensors collect information according to spectral, directional, temporal and spatial dimensions. Inferring temperatures from measurements
relies on developing and inverting modeling tools. Simple radiative transfer equations directly link measurements and variables of interest, and can be analytically
inverted. Simulation models allow linking radiative regime to measurements. They
require indirect inversions by minimizing differences between simulations and observations, or by calibrating simple equations and inductive learning methods. In
both cases, inversion consists of solving an ill-posed problem, with several parameters to be constrained from few information.
Brightness and radiometric temperatures have been inferred by inverting simulation models and simple radiative transfer equations, designed for atmosphere and
land surfaces. Obtained accuracies suggest refining the use of spectral and temporal
information, rather than innovative approaches. Forthcoming challenge is recovering more elaborated temperatures. Soil and vegetation components can replace
aerodynamic temperature, which retrieval seems almost impossible. They can be
inferred using multiangular measurements, via simple radiative transfer equations
previously parameterized from simulation models. Retrieving sunlit and shaded
components or canopy temperature profile requires inverting simulation models.
Then, additional difficulties are the influence of thermal regime, and the limitations
of spaceborne observations which have to be along track due to the temperature fluctuations. Finally, forefront investigations focus on adequately using TIR information
with various spatial resolutions and temporal samplings, to monitor the considered
processes with adequate spatial and temporal scales.
10.1 Introduction
Using TIR remote sensing for environmental issues have been investigated the
last three decades. This is motivated by the potential of the spatialized information for documenting the considered processes within and between the Earth
system components: cryosphere [12], atmosphere [36], oceans [79], and land
surfaces [10]. For the latter, TIR remote sensing is used to monitor forested areas [1114], urban areas [1517], and vegetated areas. We focus here on vegetated
areas, natural and cultivated. The monitored processes are related to climatology,
meteorology, hydrology and agronomy: (1) radiation, heat and water transfers at the
soilvegetationatmosphere interface [1824]; (2) interactions between land surface
and atmospheric boundary layer [25]; (3) vegetation physiological processes such as
transpiration and water consumption, photosynthetic activity and CO2 uptake, vegetation growth and biomass production [2639]; (4) soil processes such as respiration
10 Modeling and Inversion in Thermal Infrared Remote Sensing
247
and CO2 uptake, evapotranspiration and water depletion, spatio-temporal variability

of soil moisture [3943]; (5) long-term dynamics of land cover [44], land surface
radiative budget [4548], water shortage and drought [49].
TIR remote sensing allow retrieving emissivity and temperature, with various
complexity degrees presented in Section 10.2. The remotely sensed information
is collected from operational and prospective sensors, listed in Section 10.3. This
information is characterized by temporal and spatial dimensions (Section 10.3.1),
as well as by spectral and directional dimensions (Section 10.3.2). Then, inferring
emissivity and temperature consists of developing and inverting modeling tools,
by exploiting the dimensions of the collected information (Section 10.4). Based
on TIR fundamentals (Section 10.4.1), simple radiative transfer equations directly
link measurements to emissivities and temperatures of interest (Section 10.4.2),
and simulation models describe the influence of radiative regime on measurements
(Section 10.4.3). However, simple radiative transfer equations must be parameterized, and simulation models require significant information. Further, inversion is
not trivial: most of simulation models are not directly invertible, and the numerous parameters to be constrained from remote sensing often make inversion an illposed problem (Section 10.4.4). The several solutions proposed to overcome these
difficulties are assessed using validations, intercomparisons, and sensitivity studies
(Section 10.5).
Current limitations and proposed solutions are presented with an increasing
complexity for the temperatures of interest (Section 10.6). Atmospheric perturbations are corrected by inverting modeling tools for atmosphere, and surface
brightness temperature measurements are simulated using modeling tools for land
surfaces (Section 10.6.1). Surface emissivity effects are removed using simple radiative transfer equations (Section 10.6.2). Reported performances suggest accuracies
rather close to requirements, though refinements are necessary. Recovering temperature for the one source modeling of heat transfers is still not trivial, since the
required parameterization significantly varies in time and space (Section 10.6.3).
Recent studies suggested focusing on more elaborated temperatures: soil and vegetation components, optionally sunlit and shaded, and canopy temperature profile.
Their retrieval is a forthcoming challenge, with efforts on measuring, modeling and
inversion (Section 10.6.4). The paper ends with forefront investigations about space
and time issues in TIR remote sensing: monitoring land processes with adequate
spatial scales and temporal samplings, by using available remote sensing observations (Section 10.7).
10.2 Land Surface Emissivity/Temperature from TIR Remote

Sensing
This section defines the various terms considered in TIR remote sensing, which
are related to land surface emissivity and temperature. We focus on their physical definitions and various interests. The corresponding equations are detailed in
Section 10.4.
248
F. Jacob et al.
Surface brightness temperature is equivalent to the radiance outgoing from the

target, by assuming a unity emissivity [50], and corresponds to the basic TIR
remote sensing measurement. It is recovered from at sensor measurements performing atmospheric corrections. It can be assimilated, using modeling tools for
land surface, into process models such as SVAT and crop models [18, 38, 39, 43].
Ensemble waveband emissivity is needed to derive radiometric temperature from
brightness temperature [50, 51]. It is also useful for retrieving ensemble broadband emissivity, a key parameter for land surface radiative budget [5254].
Ensemble radiometric temperature is emissivity normalized [50, 51]; and corresponds to kinetic temperature for an homogeneous and isothermal surface [55]. It
is used to estimate surface energy fluxes and water status from spatial variability
indicators: the vegetation index / temperature triangle [41, 5658]; or the albedo /
temperature diagram [23, 37, 59, 60]. It is also used for retrieving soil and vegetation temperatures from two source energy balance modeling [19, 24].
Aerodynamic temperature is air temperature at the thermal roughness length [50].
It is the physical temperature to be used with one source models of surface energy
fluxes based on excess resistance [6163]. These can be SVAT models [39, 64];
or energy balance models [22, 23, 37, 59, 60, 65, 66].
Soil and vegetation temperatures correspond to kinetic [67] or radiometric [68]
temperatures. They are often used for two-source modeling. The latter can be
SVAT models [43, 67, 69]; or energy balance models [20, 70, 71]. Retrieving
these temperatures requires an adequate estimation of directional ensemble emissivity.
Sunlit and shaded components are refinements of soil and vegetation temperatures. They can significantly differ, according to various factors which drive the
thermal regime: the water status, the solar exposure resulting from the canopy
geometry and the illumination direction. These components are of interest for
understanding canopy directional brightness and radiometric temperatures [58,
7274].
Canopy temperature profile, from the soil surface to the top of canopy, is the
finest temperature one can consider. Similarly to sunlit and shaded components
for soil and vegetation temperatures, this thermal regime is considered for understanding canopy directional brightness and radiometric temperatures, in relation
with local energy balance within the canopy [7578].
The seek accuracies vary from one application to another,according to the sensitivities of process models. For temperature, the goal is accuracy better than 1 K [79].
For emissivity, the goal is absolute accuracy better than 0.01 [80]. Recovering both
relies on exploiting the dimensions of the TIR remotely sensed information.
10.3 Available Information from TIR Remote Sensing

The four dimensions of the remotely sensed information are temporal and spatial (Section 10.3.1), and spectral and directional (Section 10.3.2). Due to orbital
249
rules and technological limitations, current spaceborne sensors cannot provide full
information over these dimensions. Further, the latter can be linked, according to
the mission objectives: a daily monitoring with sun-synchronous sensor requires a
kilometric resolution with an across track angular sampling. Exploratory missions
with airborne and ground-based sensors are under progress, for assessing the potential of original remotely sensed information. Table 10.1 provides an overview of
the main operational and prospective sensors. We deal here with recent, current and
forthcoming US and EU missions.
10.3.1 Temporal and Spatial Capabilities

The temporal dimension corresponds to the time interval between consecutive observations. It is of importance for monitoring land surface temperature and related
processes: radiative and convective transfers, soil respiration and vegetation physiological activity. The spatial dimension corresponds to the ground resolution of the
measurements. It is of importance for the meaning of surface temperature collected
over kilometric size pixels which include different land units. Both dimensions are
strongly correlated for current TIR spaceborne sensors: high temporal samplings for
finer monitoring correspond to coarse spatial resolutions with larger heterogeneity
effects, and reversely.
The highest temporal samplings are provided by geostationary sensors: 1530
min with GOES Imager [81] and MSG/SEVIRI [82], corresponding to ground resolutions between 2 and 4 km. Intermediate scales correspond to kilometric resolution
sensors onboard sun-synchronous platforms, providing daily nighttime and daytime
observations: NOAA/AVHRR [83], ADEOS/GLI [84], and Terra-Aqua/MODIS
[85]. A 3 day temporal sampling with a 1 km resolution has been provided by
ERS/ATSR-1 and -2, and ENVISAT/AATSR [86]. The highest spatial resolutions
are 60 and 120 m from Landsat/TM & ETM [87], and 90 m from Terra/ASTER
[88]; with 16-day temporal samplings. ASTER and Landsat/ETM missions have
limited lifetimes, with currently no follow on TIR high spatial resolution missions
from space.
Regarding current possibilities, new spaceborne sensors are demanded, to monitor land processes with adequate temporal and spatial scales. Past missions IRSUTE
and SEXTET proposed 4060 m spatial resolutions with a 1-day revisit [89, 90] and
SPECTRA proposed 50 m with 3 days [91]. MTI mission offers a 20 m resolution
with a 7-day revisit [92], but the military context restricts the data access. Airborne
prospective observations have allowed studying temporal and spatial issues, with
metric resolutions and adjustable revisits. Let us cite the airborne missions TIMS
[93], DAIS [94], MAS [95] and MASTER [96]; and the airborne-based ReSeDA
program [9799].
250
F. Jacob et al.
Table 10.1 Nominal characteristics for operational and prospective sensors; in relation with recent,
current and forthcoming US and EU missions. VZA means View Zenith Angle, VAA means View
Azimuth Angle. Across (respectively along) track means viewing directions in a plan perpendicular
(respectively parallel) to the satellite path.
Sensor
Daytime Spatial Spectral

sampling resolution features
Directional
features
Spaceborne
MSG
SEVIRI
15 mn
3 km
GOES 10 and 12
Imager
30 mn
24 km
NOAA 1517
AVHRR / 3
1 day
Terra-Aqua
MODIS
1 MIR: 3.9 m
1 latitude5 TIR: 8.7, 9.7, 10.8, 12, 13.4 m dependent VZA
1 MIR: 3.7 m
2 TIR: 10.8, 12 m
1 latitudedependent VZA
1 km
1 MIR: 3.8 m
2 TIR: 11, 12 m
Across track
VZA: 55
1 day
1 km
3 MIR: 3.8, 3.95, 4.1 m

3 TIR: 8.6, 11, 12 m
Across track
VZA: 55
ADEOS
GLI
1 day
1 km
1 MIR: 3.7 m
3 TIR: 8.6, 10.8, 12 m
Across track
VZA: 40
ERS-ATSR 1 and 2
ENVISAT-AATSR
3 days
12 km
1 MIR: 3.7 m
2 TIR: 10.8, 12 m
Along track
VZA: 0, 55
Landsat 57
TM and ETM
16 days
120 m
1 TIR: 11.5 m
Close nadir
VZA
Terra
ASTER
16 days
90 m
5 TIR: 8.3, 8.6, 9.1, 10.7, 11.3 m
Close nadir
VZA
TIMS
(multispectral)
15 m
6 TIR: 8.4, 8.8, 9.2, . . .

. . . 9.9, 10.7, 11.7 m
Across track
VZA: 38
DAIS
(multispectral)
15 m
6 TIR: 8.7, 9.7, 10.5, . . .

. . . 11.4, 12.0, 12.7 m
Across track
VZA: 26
MAS / MASTER
(multispectral)
15 m
10 TIR: 7.8, 8.2, 8.6, 9.1, 9.7, . . . Across track

. . . 10.1, 10.6, 11.3, 12.1, 12.9 m VZA: 40
SEBASS
(hyperspectral)
15 m
MIR: [2.55.3] m
TIR: [7.613.5] m
Spectral resolution > 0.1 m
Close nadir
VZA
Two temperature
Box method
50 cm
1 broadband over [813] m
Nadir VZA
Hyperspectral FTIR
BOMEM suite
Few cm
Optical spectral range: [220] m Nadir VZA

Spectral resolution: 1 cm1
Goniometric
systems
Few cm
1 broadband over [813] m
Airborne
Ground-based
VZA [090 ]
VAA [0360 ]
251
10.3.2 Spectral and Directional Capabilities

The spectral dimension corresponds to the number and location of sensor wavebands within the TIR and optionally the MIR domains. The directional dimension
corresponds to the number and angular distribution of viewing directions. Both dimensions are used for recovering emissivities and temperatures via modeling tools.
The basic spectral configuration corresponds to TM and ETM, with 1 channel.
Richer information is provided via two channels with GOES Imager, AVHRR and
the ATSR suite; three channels with MODIS and GLI; and five channels with
SEVIRI and ASTER. Additional MIR information can be combined with TIR
information, to be used with continuous observations from geostationary sensors
(SEVIRI, GOES Imager), or day night observations from sun-synchronous sensors
(AVHRR, MODIS).
The basic directional configuration corresponds to SEVIRI, GOES Imager, TM,
ETM, and ASTER; with a single viewing direction. Richer information is collected
from across track viewing with AVHRR, MODIS, GLI; and along track viewing
with the ATSR suite. Across track viewing allows a daily monitoring, while sampling the angular dynamic within a given temporal window (16 days for MODIS).
This is of interest for stable surface properties such as emissivity. For surface temperature which fluctuates, capturing the angular dynamic requires almost simultaneous observations. This is possible with ATSR along track bi-angular observations
only, which is limited.
Future spaceborne missions will pursue current ones for long-term records: the
GOES suite [100], NPOESS/VIIRS following AVHRR and MODIS [101]. MTI
provides original information: 2 MIR/3 TIR bands, 0 and 50 along track. At
the airborne level, the spectral dimension has been investigated with multispectral
(TIMS, DAIS, MAS & MASTER) and hyperspectral (SEBASS [102]) sensors, and
the directional dimension has been assessed with video cameras (see [103] with the
ReSeDA program). At the ground level, the spectral dimension has been explored
with hyperspectral sensors (FTIR BOMEM [104]), or with broadband radiometers
[105107], and the directional dimension has been examined with goniometric systems [58, 108, 109].
In the context of monitoring land processes, the various types of information
presented here are valuable for recovering land surface emissivity and temperature.
Using this information requires designing modeling tools and inversion methods,
either under development for prospective studies or with operational capabilities.
10.4 Developing Modeling Tools and Inversion Methods

Modeling tools aim at forwardly simulating, with different complexities, measured
brightness temperature from emissivities and temperatures of interest. Table 10.2
provides an overview of the modeling tools currently used. Based on TIR fundamentals (Section 10.4.1), simple radiative transfer equations directly link measurements
252
F. Jacob et al.
Table 10.2 Listing of the modeling tools currently used, with an increasing complexity. The second rightmost column gives the related medium, and the rightmost column gives the types of land
surface emissivity and temperature currently investigated with each tool. The modeling of atmospheric radiative transfer is considered here in the context of performing atmospheric corrections.
Modeling
tools
Literature
examples
Related
medium
Investigated land surface

temperatures and emissivities
Simple radiative transfer equations

Atmospheric
radiative transfer
Eq. 10.6
[81, 178]
Atmosphere
Brightness temperature
(Atmospheric corrections)
Composite surface
radiative transfer
Eq. 10.7
[12, 158]
Land surface
Ensemble emissivity and

radiometric temperature
Split Window and

Dual Angle
Eq. 10.9
[125, 126]
Soil and vegetation

radiative transfer
Eq. 10.10
[68, 127]
Land surface
Soil and vegetation temperatures
Kernel-driven
radiative transfer
Eq. 10.11
[128, 129]
Land surface
Ensemble emissivity
Radiative transfer
MODTRAN [134]
Atmosphere
(Atmospheric corrections)
Radiative transfer
Prevots [139]
SAIL [74, 137]
Land surface
Ensemble emissivity
with sunlit and shaded components
Geometric-optics
Kimess [141]
Caselless [143]
Yus [73]
Land surface
with sunlit and shaded components
Geometric-optics
radiative transfer
CUPID [147]
Thermo [148]
Jias [149]
DART [76, 77]
Land surface
Canopy temperature profile
Monte Carlo
ray tracing
[127, 150, 151]
Land surface
Ensemble emissivity
Soil and vegetation emissivities
Atmosphere and Ensemble radiometric temperature

land surface
(atmospheric corrections)
Simulation models
to emissivities and temperatures of interest (Section 10.4.2), and simulation models

describe the influence of radiative regime on measurements (Section 10.4.3). Next,
inversion methods aim at backwardly retrieving emissivities and temperatures of
interest from measurements (Section 10.4.4).
253
10.4.1 Fundamentals in TIR Remote Sensing

The use of TIR remote sensing to infer the temperatures of interest involves an
aerodynamic issue for the related temperature, and a radiative issue for the other
temperatures.
10.4.1.1 Aerodynamic Issue

Aerodynamic temperature Taero is not radiative-based and cannot be remotely
sensed. It is required for one source modeling of surface energy fluxes, since it
corresponds to the value of the logarithmic-based air temperature profile Tair (z) at
thermal roughness length zoh [110]. For a negligible displacement height, sensible
heat flux H is expressed from the air temperature gradient between zoh and reference
level zre f :
H=
Tair (zoh ) Tair (zre f )

rah (zoh , zre f )
with
Taero = Tair (zoh )
(10.1)
where rah (zoh , zre f ) is aerodynamic resistance for heat between zoh and zre f [111].
Due to larger resistance for heat transfers, zoh is lower than mechanical roughness
length zom [112]. The link between both is the aerodynamic kB1 parameter [113]:

zom
kB1 = ln
(10.2)
zoh
The physical meanings of Taero and zoh are equivocal. Taero is an effective temperature for heat sources that are soil and vegetation [114]. zoh is an effective level for
which Tair = Taero . Their retrieval from remote sensing is not trivial (Section 10.6.3).
Nevertheless, Taero can be unequivocally derived from soil and vegetation temperatures Tsoil and Tveg , by merging one source and two source modeling [20, 115]:
Taero =
Tsoil
ra,soil
Tair (zre f )
rah
1
1
ra,veg + rah
veg
+ ra,veg
+
1
ra,soil
(10.3)
where ra,soil (respectively ra,veg ) is aerodynamic resistance from the soil (respectively vegetation) to zom , and rah is aerodynamic resistance from zom to zre f [111].
10.4.1.2 Radiative Issue

Apart from aerodynamic temperature, the land surface temperatures inferred from
TIR remote sensing are radiative-based. Then, fundamentals deal with the TIR
radiative regime within atmosphere and over land surfaces. This includes three
mechanisms which drive the wave matter interactions: emission, absorption, and
254
F. Jacob et al.
scattering. Emitted radiance L( , , T ) from a natural object at a kinetic temperature T is written:

L( , , T ) = ( , ) B( , T ) = ( , )
C1 5

exp TC2 1
(10.4)
is the monochromatic wavelength. is the emission direction. B( , T ) is the

blackbody emitted radiance, expressed from Plancks Law. C1 and C2 are first and
second radiative constants. Emissivity ( , ) is the conversion factor from thermodynamic to radiative energy, lower than 1 for natural objects. This so-called
e-emissivity definition is linked to emission mechanisms, since it is the ratio of the
actual to the blackbody emitted radiances for the same kinetic temperature. Under
local thermodynamic equilibrium, Kirchhoffs Law assumes emissivity and absorptivity are equal. For opaque elements, emissivity is then linked to hemisphericaldirectional reflectance ( j , ):
( j , ) = 1 ( j , )
(10.5)
( j , ) is the average of bidirectional reflectance over illumination angles [116].

This so-called r-emissivity definition is derived from Kirchhoffs Law, and therefore
linked to reflection mechanisms. Finally, emitted radiance from a given element can
be reflected by other elements, inducing changes in radiation path, called scattering
effects.
Within the atmosphere, scattering is negligible: the radiative regime is driven by
the temperature and density of absorbers and emitters (water vapor, CO2 , O3 , . . .).
A clear atmosphere behaves as an horizontally homogeneous medium: the radiative
regime primarily depends on vertical profiles for temperature and density of absorbers and emitters (Fig. 10.1). Over heterogeneous land surfaces with structured
patterns, the radiative regime is more complex than within atmosphere: soil and
vegetation act as emitters, absorbers and scatterers for canopy and atmospheric irradiances. Additional effects are surface and volume scatterings (Fig. 10.2). Surface
scattering corresponds to shadowing effects for a geometric medium, with sunlit
and shaded areas. Volume scattering corresponds to reflections between soil and
vegetation: radiation is trapped within the canopy.
TIR remotely sensed measurements result from the processes discussed above.
Sensor brightness temperature is driven by vertical profiles for temperatures and
densities of atmospheric constituents. Surface brightness temperature results from
the radiative regime over a heterogeneous and non isothermal area. Then, emissivity
and kinetic temperature are equivocal: the canopy acts as an effective medium with
ensemble emissivity and radiometric temperature [50]. Besides, e- and r-emissivities
differ according to vegetation amount, since spatial averaging for e-emissivity includes emitted radiance as an additional weighting factor [50, 117]. Due to its
simpler formulation, r-emissivity is preferred [68, 71, 74, 118120]. Further, the
measured brightness temperature results from emission, but also from absorption
and scattering of canopy and atmospheric irradiances. This induces spectral and
directional variations, driven by (1) radiative properties of soil and vegetation
255
Fig. 10.1 Atmospheric TIR radiative regime for an off nadir propagation. The key processes to be
considered for atmospheric corrections are emission and absorption by atmospheric constituents.
Within a horizontally homogeneous atmosphere, the radiative regime depends on the vertical fields
of temperature and density for emitters and absorbers. Regardless of considered layer (zi or zk ),
radiative regime is driven by atmospheric absorption (1), atmospheric emission (2), and surface
emission through atmosphere transmission (3). (Adapted from [264].)
(reflectance and emissivity), (2) surface scattering with sunlit and shaded areas, and
(3) volume scattering with the cavity effect. These three factors induce ensemble
emissivity is anisotropic, with values greater than that of vegetation as the latter
quantitatively increases [118, 121, 122].
Various modeling tools have been developed to simulate sensor and surface
brightness temperature measurements. The first way is using simple radiative transfer equations for directly linking measurements to emissivities and temperatures of
interest. The second way is using simulation models for understanding the influence
of the TIR radiative regime on the measured brightness temperature.
10.4.2 Simple Radiative Transfer Equations

Simple radiative transfer equations directly link TIR measurements to emissivities
and temperatures of interest. Their advantages are linearity and simplicity, but most
of them are limited to homogeneous media by assuming turbidity and azimuthal
isotropy.
256
F. Jacob et al.
Fig. 10.2 Surface (or geometric) and volume (or volumetric) scattering. Surface scattering induces shadowing effects with hotter and cooler elements. Volume scattering induces an increase of
brightness temperature by adding a component to emission. (Adapted from [129].)
Measured brightness temperature at the sensor level Tbrs is linked to surface

brightness temperature Tbs via the atmospheric radiative transfer equation:

B ( j , Tbrs ( , j )) = B ( j , Tbs ( , j )) a ( , j ) + B j , Tba ( , j )
(10.6)
is the view zenith angle. j is the equivalent waveband over the sensor channel
j [123]. B( , T ) is the blackbody emitted radiance, expressed from Plancks Law
(Eq. 10.4). a is the atmospheric transmittance, vertically integrated between the
surface and the sensor. B( j , Tba ) is the atmospheric upward radiance towards the
sensor.
Surface brightness temperature is expressed as the sum of canopy emission and
scattering of atmospheric irradiance, via the composite surface radiative transfer
equation:
257
B( j , Tbs ( , j )) = ( j , ) B( j , Trad ( )) + (1 ( j , )) B( j , Tba ( j )) (10.7)

B( j , Tba ) is the hemispherical average of atmospheric downward radiance. ( j , )
and Trad ( ) are ensemble emissivity and radiometric temperature. Ensemble emissivity can be expressed from emissivities of soil soil ( j ) and vegetation veg ( j ),
with the optional inclusion of a correction term d for the cavity effect [124]:
( j , ) = Fsoil ( ) soil ( j ) + Fveg ( ) veg ( j ) + 4 d Fveg ( ) Fsoil ( ) (10.8)

Fsoil ( ), Fveg ( ) are directional gap and cover fractions, with Fsoil ( ) = 1 Fveg ( ).
Brightness temperature measured from space can be linked to emissivity and
radiometric temperature by merging Eqs. 10.6 and 10.7. Another possibility is simultaneously considering atmospheric and surface effects: Split Window (SW) and
Dual Angle (DA) methods directly express radiometric temperature Trad as a spectral or angular difference between two brightness temperatures Tbrs at the sensor
level [125, 126]:
rs
rs
rs
rs
rs 2
Trad = Tb1
+ A(Tb1
Tb2
) + B(Tb1
Tb2
) +C
1 + 2
+ D(1 2 ) + E (10.9)
2
is surface emissivity. A, B, C, D, E are empirical coefficients. Indices 1 and 2 are

two spectral channels for SW method, or two view zenith angles for DA method.
The angular differencing uses variations in atmospheric transmittance between different paths for two view zenith angles. The spectral differencing uses variations
in atmospheric transmittance due to different water vapor absorptions for two spectrally close channels.
The emission term of Eq. 10.7 can be split into soil and vegetation components,
which yields the soil and vegetation radiative transfer equation [68, 119, 127]:
( j , ) B ( j , Trad ( )) = can ( ) soil ( j ) B( j , Tsoil )
+ ( , veg ( j )) B( j , Tveg )
(10.10)
Tsoil and Tveg are soil and vegetation radiometric temperatures [68]. can ( ) and
( , veg ( j )) are vegetation directional transmittance and fraction of emitted radiation. The angular effects can also be described with linear kernel driven approaches, by expressing the directional emission as a linear combination of generic
shapes [128]:
( j , ) B ( j , Trad ( )) =
N
i ( j ) Ki (Tveg , Tsoil , veg ( j ), soil ( j ), , s , s )
(10.11)
i=1
s is the solar zenith angle. s is the relative azimuth between illumination

and viewing directions. i, j are weighting coefficients. Kernels Ki describe gray
body isotropy, volume scattering, and surface scattering. Various kernel formulations may be proposed, by linearizing different sets of complex equations. Kernel
driven approaches are also used to derive ensemble r-emissivity from accurate
258
F. Jacob et al.
hemispherical - directional reflectance (Eq. 10.5): [129] expressed TIR BRDF as

a linear combination of generic shapes, following previous works over the solar
domain [130132].
10.4.3 Simulation Models

Simulation models mimic the TIR radiative regime within atmosphere and canopies,
to understand spatial, spectral and directional behaviors of brightness temperature
measurements. These models are classified here via an increasing complexity: radiative transfer, geometric-optics, geometric-optics/radiative transfer, and ray tracing.
Radiative transfer models are designed for turbid media (atmosphere, homogeneous canopies). Assuming turbidity and azimuthal isotropy, they split the medium
into a finite layer number, and account for volume scattering between layers. For the
atmosphere, volume scattering is negligible, and each layer is described with temperature and densities of absorbers and emitters. For canopies, soil and vegetation
layers are described with temperature; and with densities of absorbers, emitters and
scatterers, derived from LAI and LIDF. Brightness temperature is simulated using
the stream concept: transmittance, upward and downward radiances are computed
for each layer, and vertically integrated (see MODTRAN for atmosphere [133, 134]
and SAIL for canopy [68, 74, 135137]. Simulations can also be probabilistic calculations for photon interception, deduced from the directional gap fraction of each
layer [118, 138, 139].
Geometric-optics models are designed for structured patterns over land surfaces,
such as row crops of cotton or maize. Considering vegetation as an opaque medium,
they account for surface scattering with shadowing effects. Sunlit and shaded areas
are described via their cross sections, derived from canopy geometry (vegetation
height, row size, etc.), illumination and viewing directions, and directional gap fraction within and between rows. Canopy brightness temperature is computed from the
resulting spatial distribution of temperature [73, 121,140143].
The finest radiosity models are geometric-optics/radiative transfer models,
designed for complex land surfaces. By accounting for both volume and surface
scattering, they are appropriate to vegetation patchworks. They can conjugate a
radiative transfer and a geometric-optic module [144, 145]. They can be more complex, such as 3-Dimensional mock-ups based models. This allows a finer description
of the radiative regime within canopies, but requires significant information about
the micro-scale conditions. Examples are CUPID [146, 147]; Thermo [72, 148];
Jias model [149], and DART [76, 77]. Further, accounting for convective and energetic transfers allow understanding their influence on the radiative regime, such as
with DART-EB [78].
The finest modeling degree is Monte Carlo ray tracing, which stochastically calculates photon trajectories within turbid or geometric atmosphere and canopies.
A photon is tracked from birth (emission or penetration within medium) to death
(absorption or escape from medium), with scattering based on probabilistic wave
259
matter interactions. Millions of simulations describe spectral, directional and spatial behaviors. Ray tracing is used to assess the influence of multiple scattering on
spatial aggregation and angular dynamics, over heterogeneous and non isothermal
land surfaces [127, 150, 151].
10.4.4 Inversion Methods

Retrieving variables from measurements is an inverse problem. Given a set of m
measurements M for a physical system, with k known parameters K and p unknown
parameters P to be retrieved, direct F and inverse F 1 problems are written [152]:

M1
M1
P1
P1
..
.
.
1 .
. = F .. , [K1 Kk ] .. = F .. , [K1 Kk ](10.12)
Mm
Pp
Pp
Mm
Inversion is possible if there are more independent equations than unknowns

(m p). Direct inversion analytically writes the inverse problem. This is possible
for simple radiative transfer equations (Section 10.4.2), but not for most simulation
models (Section 10.4.3). For the latter, indirect inversion numerically sets parameters such as simulations agree with observations [153]. It has been improved for
accuracy and rapidity, by calibrating neural networks, lookup tables, genetic algorithms or regression trees [152, 154]. Inversion can be a well-posed problem, when
solving an overdetermined equation system using optimization techniques. However, it is usually an ill-posed problem, with several parameters to be constrained
from few observations. Proposed solutions use a priori information about soil and
vegetation properties, or parameter ranges [152, 155, 156].
Inversion over the TIR domain is not as developed as over the solar domain. This
results from (1) additional micrometeorological complex influences, and (2) the lack
of high resolution data. Atmospheric simulation models have been inverted calibrating neural networks [51], SW and DA methods (Eq. 10.9) [125, 126], or the
atmospheric radiative transfer equation (Eq. 10.6) [157, 158]. Over land surfaces,
simulation models have been assessed in the forward mode [7274, 77, 144]. No investigation was found about their indirect inversion, but they can serve as references
for parameterizing simple radiative transfer equations which are directly invertible.
Thus, various formulations have been assessed for the soil and vegetation radiative
transfer equation (Eq. 10.10), optionally accounting for multiple scattering and non
linearities [68, 71, 118, 119, 127]. Further, inverting simple radiative transfer equations is often an ill-posed problem. For instance, inverting the composite surface
radiative transfer equation (Eq. 10.7) from N multispectral observations includes
N emissivities and radiometric temperature. Similarly, inverting the soil and vegetation radiative transfer equation (Eq. 10.10) or linear kernel driven approaches
(Eq. 10.11) from multiangular observations requires angular parameters: ensemble,
soil and vegetation emissivities; vegetation transmittance.
260
F. Jacob et al.
10.5 Assessing Modeling Tools and Inversion Methods

Modeling tools and inversion methods have been assessed experimentally through
validation exercises, and theoretically via sensitivity studies.
Validation exercises have been conducted over databases collected in the framework of various international programs such as FIFE [159], EFEDA [160], HAPEX
[161], ReSeDA [97], JORNEX [162], FLUXNET [163], DAISEX [164], SALSA
[165], SMACEX [166]. Assessments over these various datasets allow accounting
for different biomes and climates. Some exercises were ground-based [73, 104,
145]. Most of them were airborne-based [94, 96, 103, 157, 167169,170174], for
assessments in actual conditions by reducing spatial heterogeneity effects. Few validations were conducted using spaceborne observations with hectometric resolutions
[175179]; and with kilometric ones over areas almost homogeneous [180182].
Original exercises based on classifications were designed for kilometric scale heterogeneities [126, 183], while new improvements for the solar domain should be
implemented over the thermal one [184]. Complementary to validations, intercomparisons are now feasible thanks to multisensor missions such as Terra. This allows
accounting for larger panels of environmental situations [158].
Validations and intercomparisons have also been performed using simulated
datasets. This allow considering more conditions than measured datasets, and focusing on physics modeling without measurement intrinsic errors [81, 82, 126, 185].
Simulated datasets are necessary when dealing with elaborated temperatures: aerodynamic, soil and vegetation, sunlit and shaded components, and canopy temperature profile [68, 71, 73, 76, 118]. Indeed, validating the latter using measured
datasets is not trivial, since the corresponding ground-based measurements are difficult to implement.
Additionally to validations and intercomparisons, sensitivity studies allow assessing information requirements such as accuracies on remotely sensed information, medium structural and radiative properties. Examples are (1) accuracy on
atmospheric status for retrieving brightness temperature [171, 178], (2) accuracy
on observations, atmospheric status and land use for recovering ensemble emissivity and radiometric temperature [12, 157, 158, 169, 182, 186188], (3) accuracy on
canopy structural parameters and radiative properties for deriving soil and vegetation temperatures [68, 118, 189]. Finally, sensitivity studies of simulation models
provide valuable information about the pertinent parameters for inversion [73, 76],
with innovative approaches over the solar domain based on adjoint models (Baret
et al., this issue).
10.6 Current Capabilities and Future Directions

From the basic materials presented before, we focus now on current investigations,
via an increasing temperature complexity. Success and failures suggest future directions.
261
10.6.1 Surface Brightness Temperature

Surface brightness temperature is derived from that at the sensor level by inverting
modeling tools for atmosphere. It is simulated using modeling tools for land surfaces. In both cases, these tools are simple radiative transfer equations or simulation
models.
10.6.1.1 Atmospheric Radiative Regime and Related Corrections

Atmospheric corrections for the retrieval of surface brightness temperature can be
performed inverting simulation models, via the calibration of the atmospheric radiative transfer equation (Eq. 10.6) for a given atmosphere [12, 81, 157, 158, 171,
172, 177, 178]. An operational context faces two challenges: reducing computation
time to process millions observations, and accurately characterizing the atmospheric
status.
To reduce by a third-order computation time for simulation models without accuracy degradation, [190] implemented correlated-K methods, by quickly integrating
waveband atmospheric absorption and emission. Predictor-based models accurately
compute the latter for a range of reference profiles, to next differencing current ones
and nearest predictors [191]. Multilayer computation based on water vapor continuum absorption can replace simulation models, with an accuracy degradation lower
than 1 K [81]. Computation time can also be reduced via inversion by including
a range of atmospheres into the simulation set. Expressing transmittance and upwelling radiance of Eq. 10.6 from atmosphere water vapor content and mean temperature yields an accuracy degradation lower than 2 K [180]. Neural networks can
replace Eq. 10.6 considering atmospheric profiles and view zenith angle, with an
accuracy degradation lower than 0.5 K [186, 192].
The atmospheric status can be well documented using ancillary information:
measured profiles allow reaching a 1 K accuracy [171, 172, 177, 178], but meteorological networks are not dense enough for regional inversion. One alternative
is profile simulation from meteorological models [193, 194]. Such information is
soon available with a 3 h sampling, and a 0.25 latitude/longitude griding to be
re-sampled to sensor resolutions via interpolation procedures [12, 195]. The relief
influence is handled using digital elevation models, now available with decametric
resolutions and metric accuracies [196]. Also, the TIR observations to be corrected
can inform about the atmospheric status. Atmosphere absorption and emission can
be retrieved from multispectral and hyperspectral observations, using variabilities
of atmospheric properties [80, 197]. Thus, water vapor content was adjusted from
ASTER multispectral observations, such as emissivity spectrum is flat over vegetation or water [185]. It was also inferred from the ATSR-2 SW channels with a
0.2 g. cm2 accuracy, using the SWVCR which relies on the spatial variability of
SW surface brightness temperatures [198].
Solar or TIR observations collected onboard the same platform also provide
coincident information about the atmospheric status. [199] expressed water vapor
262
F. Jacob et al.
content as a polynomial of MODIS near infrared radiance ratios, with a 0.4 g. cm2
accuracy. Atmospheric sounders allow inferring profiles of temperature and water
vapor density, using Eq. 10.6 or neural networks [3, 4]. Previous sounders such as
TOVS permitted to reach a 0.4 g. cm2 accuracy on water vapor content [200]. New
sounders such as IASI [201], with finer spectral samplings and spatial resolutions,
should provide accuracies better than 1 K and 10% for atmospheric profiles of temperature and humidity.
10.6.1.2 Land Surface Radiative Regime and Related Measurements

Surface brightness temperature is simulated using simple radiative transfer equations or simulation models. The former provide easy and efficient solutions for assimilating TIR remote sensing data into land process models. The latter are fine and
accurate solutions for understanding TIR remotely sensed measurements.
To constrain land process model parameters, surface brightness temperature can
be simulated using simple radiative transfer equations coupled with SVAT models.
[39] coupled the composite surface radiative transfer equation (Eq. 10.7) with a crop
and a one source SVAT model. The latter calculated ensemble radiometric temperature by closing the surface energy budget. R-emissivity was estimated using the
SAIL TIR version of [136], documented by the crop model for vegetation structural parameters. Similarly, [67] coupled the soil and vegetation radiative transfer
equation (Eq. 10.10) with a two source SVAT model. The latter calculated soil and
vegetation temperatures by closing the energy budget for each, while setting soil
and vegetation emissivities to nominal values.
Calculating surface brightness temperature from simulation models requires information about vegetation structure (row crop, LAI, LIDF, cover fraction), soil
and vegetation radiative properties (emissivity, reflectance), and thermal regime
(canopy temperature distribution). The latter can be derived from a SVAT model,
which solves local energy budget according to meteorological conditions (solar
position, wind speed, air temperature), vegetation status (leaf stomatal resistance),
and soil moisture. Then, simulation models mimic the radiative regime using more
or less complex descriptions of the thermal regime: a unique vegetation temperature [73], soil and vegetation temperatures with optional sunlit and shaded components [74, 137], additional vegetation layer temperatures for specific crops [145], or
canopy temperature profiles [78].
Simulation models are currently under development, verification and analysis
[73, 74, 76, 78, 144, 145]. Current investigations focus on spectral behaviors [120],
but especially on directional effects which allow normalizing multiangular observations (Fig. 10.3). For instance, [58, 72] angularly normalized water stress indices
over row structured crops. Similarly, [202] normalized across track observations
from sun-view geometry effects, for a daily monitoring at the continental scale.
263
0
80
30
330
60
300
34
40
33
60
20
32
270
90
31
30
29
240
120
28
150
210
180
Fig. 10.3 Simulating measured brightness temperature over a maize canopy in row structure, with
a resulting angular dynamic about 8 K [73]. Black star indicates the solar direction. The brightness
temperature maximum value is located in the solar direction. However, this hot spot effect is not
systematical (Section 10.6.4.)
10.6.1.3 Partial Conclusions

The various methods developed to perform atmospheric corrections are of interest, since they were designed for optimizing the collected information according
to sensor configurations. Measured or simulated profiles are tributary to their representativeness, and coincident information relies on strong assumptions. Despite
these limitations, significant progresses were made the last decades, with accuracies now close to 1 K. Current investigations focus on refinements rather than new
developments.
Simulating brightness temperature is ongoing for describing brightness temperature measurements, according to the various land surface behaviors: geometric like,
radiative transfer like, or both. Validation results emphasized good performances
with accuracies close to 1 K, though significant documentations are required about
thermal regime, medium structure and radiative properties. Such simulation models
will be of interest for future designs of inversion methods, conjointly to the solar
domain (Section 10.4.4).
264
F. Jacob et al.
10.6.2 Ensemble Emissivity and Radiometric Temperature

Ensemble radiometric temperature is derived by directly inverting composite surface
radiative transfer equation (Eq. 10.7), or indirectly inverting simulation models via
differencing equations (Eq. 10.9). The first way is two-step based and requires previous atmospheric corrections. The second way is one-step based by simultaneously
correcting atmosphere and surface effects. In both cases, performances depend on
characterizing these effects. Inverting Eq. 10.7 is an ill-posed problem, with N equations from channel measurements and N + 1 unknowns being channel emissivities
and radiometric temperature. Proposed solutions consist of adding an N + 1 equation. They are reported here via an increasing amount of information, according to
the spectral, directional and temporal dimensions.
10.6.2.1 Single-Channel TIR Instantaneous Observations
Radiometric temperature is derived from single channel observations using two step
approaches. After atmospheric corrections, inverting the composite surface radiative
transfer equation (Eq. 10.7) requires estimating waveband emissivity. The latter is
inferred using in-situ observations, nominal values proposed by literature, or solar
remotely sensed observations. This have been investigated for ground-based and
airborne sensors during field experiments, and for spaceborne sensors such as the
Landsat TM series.
Considering ensemble emissivity increases with vegetation amount, it can be
linked to NDVI [203], or to cover fraction (Eq. 10.8) neglecting spatial variabilities
for soil and vegetation emissivities [177, 199]. However, low correlations were observed between AVHRR emissivities and cover fraction [188]; and between ASTER
broadband emissivity and MODIS solar albedo [46]. Indeed, the link between emissivity and vegetation amount depends on canopy structure, cavity effect, and optical
properties of soil and vegetation [136]. Besides, emissivity may decrease with the
vegetation amount, according to the type of soil and the vegetation water status
[120].
Good results were reported with TM and DAIS (1 K over semi-arid agricultural
areas [174, 177]), but the use of in situ information at the local scale raises the question of method applicability. A promising way is using additional MIR data, which
contain information on water content. For optimizing the temporal monitoring, another possibility is deriving single channel emissivity from multispectral ones, by
conjointly using different sensors such as Landsat and ASTER. However, this is
contingent upon the temporal stability of surface conditions between consecutive
satellite overpasses.
10.6.2.2 Dual-Channel and Dual-Angle TIR Instantaneous Observations
Radiometric temperature is recovered from dual channel and dual angle observations using SW and DA one step approaches, which require accounting well for
265
atmospheric and surface effects [125, 126]. Most investigations deal with the SW
method, since multispectral observations are more usual than multiangular ones.
Various versions have been proposed for Eq. 10.9: linear or quadratic forms
(B = 0 or B = 0), optional inclusion of emissivity (C = 0 or C = 0, D = 0 or D = 0),
expressing coefficients from atmospheric water vapor content. Larger freedom degrees perform better [125, 126, 174, 183, 204]. Wavebands around 11 and 12 m
are the most appropriate for SW and DA assumptions, since they correspond to
low variations of emissivity, spectrally and spatially [172, 174]. Calibration relies
on simulations from emissivity spectral libraries and atmospheric radiative transfer
[82, 125, 126, 204]. Operational use requires documentation. Atmospheric water
vapor content is inferred from climatological database [183], the SWVCR [198]
or near infrared radiance ratios [199]. Emissivities are derived from classifications
[181, 182, 205], or from Eq. 10.8 with nominal values for soil and vegetation emissivities [172, 199].
Several validation exercises reported accuracies better than 1 K. Excellent results
were obtained from TIMS without a priori information [172]. Using classificationbased knowledge of emissivity can perform well [181, 182], though significant
subclass variabilities were observed [206, 207]. However, a 1 K accuracy usually
requires local information on surface conditions for emissivity effects. Further, the
lack of such information can induce errors up to 3 K [125, 126, 174, 183, 199].
10.6.2.3 Multispectral and Hyperspectral TIR Instantaneous Observations

Radiometric temperature is derived from multispectral and hyperspectral observations using two step approaches. After atmospheric corrections, the ill-posed
problem can be solved using either a priori information, or the spectral variability captured over the whole TIR range. This last possibility is very different from
the two channel SW differencing which aims at avoiding emissivity variations.
For multispectral observations, the NEM approach sets maximum emissivity to
a nominal value [208], where the latter can be derived from Eq. 10.8 using a priori information about soil and vegetation emissivities [173]. The adjusted ANEM
relies on land use [168, 170], and the MIR NEM is extended to MIR observations
[209]. Rather than using a priori information, other approaches aim at benefiting
the variability captured from multispectral and hyperspectral data, the latter providing finer spectral samplings. The TES algorithm derives minimum emissivity
from the spectral variations, via an empirical relationship verified over most land
surfaces for the TIR and the MIR domains [104, 157, 210212]. Capturing larger
variabilities with hyperspectral data increase TES accuracy up to 0.5 K [104]. To
derive absolute emissivity from relative spectral variations, the alpha residuals logarithmically linearize Plancks equation, with an optional improvement based on
Taylor expansion for hyperspectral observations [213]. Taylor expansion also provides derivative approaches, such as the Grey Body method [123]. Finally, multispectral and hyperspectral observations are useful for deriving broadband emissivity
via NTB conversions [52, 54, 207, 214].
266
F. Jacob et al.
Fig. 10.4 Validation against sample-based laboratory measurements (squares), of ASTER/TES

emissivity retrievals (circles), for eleven days over a gypsum site at the White Sands National
Monument in New Mexico. TES derived emissivity spectra were averaged over the 11 acquisitions,
where the corresponding standard deviation ranges from 103 to 102 . (From [265].)
Several validation exercises reported 1 K accuracies for NEM, with or without

a priori information [168, 170, 173]. Good results were obtained for the TES algorithm (Fig. 10.4), with accuracies better than 0.01 on emissivity [176] and 1 K on
temperature [173]. Similar performances were reported by [158] when intercomparing ASTER/TES and MODIS/TISIE retrievals (Fig. 10.5, the TISIE concept is
presented below).
10.6.2.4 Multispectral MIR and TIR Consecutive Observations

Solving the ill-posed problem to invert Eq. 10.7 is also possible using temporal
series from geostationary or sun-synchronous daytime/nighttime observations. Assuming emissivity is stable between consecutive observations yields more equations
than unknowns. Then, investigations rely on using TIR observations only [215], or
MIR/TIR observations [12, 186, 188, 216].
TTM is a two step approach for inverting Eq. 10.7 over TIR consecutive observations [217]. Assuming surface emissivity is constant yields 2N equations for
N + 2 unknowns: N channel emissivities and two radiometric temperatures. Then,
two channels are enough for solving the ill-posed problem. TTM performs better
267
Fig. 10.5 Intercomparison, over a Savannah landscape (Africa) and a semiarid rangeland (Jornada), of surface radiometric temperature retrievals from the MODIS/TISIE and ASTER/TES algorithms. Differences were lower than 0.9 K. (From [158].)
268
F. Jacob et al.
with the SEVIRI finest temporal sampling, and with observations near daily temperature extrema [187, 215]. TISIE is a two step approach for inverting Eq. 10.7 over
TIR and MIR daytime/nighttime observations. Raising emissivity ratios to specific
powers yields relative variations independent of radiometric temperature. Assuming TISIE are stable between consecutive observations, MIR r-emissivities can be
retrieved, and next TIR ones. Various TISIE versions were designed for AVHRR,
MODIS, SEVIRI [12, 186, 188, 218]. Day Night Pair is a one step approach for inverting both Eqs. 10.6 and 10.7 over TIR and MIR daytime/nighttime observations.
The system of 2N equations with N + 2 unknowns can be solved with k additional
unknowns, as long as k N 2. Thus, the 7 MODIS channels allow recovering five
unknowns about atmospheric and surface effects [216].
The accuracies reported for these methods range from 0.5 to 2.5 K, and are
slightly worse for TTM. They correspond to sensitivity studies for TTM and TISIE
[186188], to validation exercises over various study sites for Day Night Pair
(Fig. 10.6) [181, 182], and to intercomparisons against ASTER/TES retrievals for
TISIE [158].

Radiometric temperature is derived by indirectly inverting simulation models
through differencing equations (Eq. 10.9), or directly inverting simple radiative
transfer equations (Eqs. 10.6 and 10.7). Solving the ill-posed problem depends on
1
0.95
0.9
sand1
Emissivity
0.85
sand2
sand3
0.8
sand4
sahara_em20
0.75
sahara_em22
sahara_em23
0.7
sahara_em29
sahara_em31
0.65
sahara_em32
0.6
3
8
9
10
Wavelength (um)
11
12
13
14
15
(a)
Fig. 10.6 Validation against sample-based laboratory spectra (lines), of Day Night Pair-based
MODIS retrievals over the Sahara Desert (points). (From [182].)
269
the available remote sensing information. A priori information is needed to obtain

good results with instantaneous single channel or dual channel / dual angle data. It
can be avoided with more remote sensing information: instantaneous TIR and MIR
data, instantaneous multispectral TIR data, or temporal series with dual channel
data. Similarly, more surface and atmospheric parameters can be recovered with a
larger amount of remotely sensed information.
Reported accuracies have increased these last years, and are now closer to that
required for further applications, i.e., 1 K. For instance, differences between retrievals from ASTER/TES and MODIS/TISIE were found lower than 0.9 K, though
both methods differ in terms of using spectral, directional and temporal information
[158]. However, ASTER/TES, MODIS/Split Window and MODIS/Day Night Pair
were found very different over Northern America [206]. Current efforts are refinements rather than new concepts: disaggregation methods should allow benefiting
the synergy between IASI hyperspectral and AVHRR kilometric sensors onboard
METOP. Then, the next challenge is retrieving more elaborated temperatures, discussed below.
10.6.3 Aerodynamic Temperature

Aerodynamic temperature Taero and thermal roughness length zoh are equivocal variables which cannot be directly recovered from remote sensing (Section 10.4.1).
Therefore, investigations have aimed at substituting aerodynamic temperature by
radiometric temperature Trad , by parameterizing a correcting factor in the sensible
heat flux expression (Eq. 10.1).
The physical meaning of sensible heat flux (Eq. 10.1) can be preserved using the
Taero Tair (zre f )
, empirically expressed from LAI [62]. Howmultiplicative factor
Trad Tair (zre f )
ever, studying this factor from simulations and measurements for growing sparse
vegetation showed significant variations according to meteorological and surface
conditions [63].
The correction factor can also be included in the kB1 parameter (Eq. 10.2),
which is then called thermal kB1 [113]. It includes corrections for (1) the difference between thermal and mechanical roughness lengths, and (2) the difference
between radiometric and aerodynamic temperatures. According to environmental
conditions, thermal kB1 varies from a vegetation type to another, and up to 100%
in relative terms [113, 219]. Parameterizations based on near surface wind speed
and temperature gradients depend on sensible heat flux [220]. Overall, formulating the thermal kB1 seems almost impossible, since it is driven by several factors
which vary in time and space: vegetation structure and water stress, meteorological
conditions, canopy temperature profile and solar position [221229].
A potential way for characterizing the kB1 parameter would be multiangular TIR remote sensing [230]. However, using this information for deriving soil
and vegetation temperatures seems more pertinent. First, these temperatures are
270
F. Jacob et al.
functionally equivalent to aerodynamic temperature (Eq. 10.3). Second, they provide information for monitoring vegetation photosynthesis and soil respiration.
10.6.4 Directional Emissivity and Soil/Vegetation Temperatures

Single directional radiometric temperature can be split into soil and vegetation components by closing energy balance for each. However, this relies on strong assumptions about vegetation water status [231233]. The promising approach is then using
multiangular TIR observations [20, 68, 71, 118, 119, 189]. No literature was found
about retrieving soil and vegetation temperatures by inverting simulation models
over multiangular data. Nevertheless, such models have been used for parameterizing the soil and vegetation radiative transfer equation which is directly invertible (Eq. 10.10). These parameterizations were designed considering the [814] m
spectral range.
10.6.4.1 Parameterizing the Soil/Vegetation Radiative Transfer Equation
Inverting Eq. 10.10 requires estimating the involved parameters (ensemble emissivity ( j , ), vegetation transmittance can ( ), and vegetation fraction of emitted
radiance ( , veg ( j ))), with optional simplifications for easier use. Various complexity degrees have been proposed, listed in Table 10.3 (P1P8). This introduces
two new parameters. Hemispherical gap fraction f is directional gap fraction integrated over illumination angles, to account for atmospheric thermal irradiance down
to the soil via the vegetation. The cavity effect coefficient is the ratio of canopy to
vegetation hemispherical-directional reflectance, to account for radiation trapping
within the canopy.
The finest parameterization (P1), proposed by [118], accounts for multiple scattering and cavity effect. The latter, which does not depend on LAI, was previously
calculated as a function of LIDF and view zenith angle, by using the probabilistic
simulation model from [139]. Similarly, [127] introduced effective directional gap
ef f
ef f
ef f
( ) and cover Fveg
( ) fractions (P2). Fsoil
( ) included single scattering of soil
Fsoil
ef f
emission by vegetation. Fveg ( ) included single scattering of vegetation emission
by soil and vegetation. Half complex parameterizations (P3P5) account for multiple scattering, with optional linearizations [68, 71]. The simplest parameterizations
(P6P8) do not account for cavity effect nor multiple scattering. They differ by their
linearization degrees, and their assumptions about soil and vegetation emissivities
[71, 119, 189].
Some of these parameterizations were assessed in direct mode for simulating
directional ensemble emissivity and radiometric temperature [68]. Apart from simplest versions, most provided close results for directional canopy emissivity, with
discrepancies lower than 0.01. For radiometric temperature, all provided similar results, with differences lower than 1 K. Next, differences decreased with atmospheric
irradiance which compensates emission (Eq. 10.7). However, this is minor under
271
Table 10.3 Listing of existing parameterizations (P1P8) for the soil and vegetation radiative
transfer equation (Eq. 10.10), with a decreasing complexity. The spectral dependence was removed
since these parameterizations were designed considering the [814] m spectral range. Labels fi
refer to specific functions proposed by the corresponding references. The dependence on LAI and
LIDF is implicitly included into directional gap and cover fractions Fsoil ( ) and Fveg ( ), the cavity
effect coefficient ( ), and hemispherical gap fraction f [68].
Standard formulation
( ) B (Trad ( )) = can ( ) soil B(Tsoil ) + ( , veg ) B(Tveg )

Label
From
Formulations

( ) = f3 Fsoil ( ), soil , veg , f , ( )
can ( ) = Fsoil ( )

( , veg ) = f4 Fsoil ( ), Fveg ( ), soil , veg , f , ( )
Remarks
P1
[118]
P2
[127]
ef f
ef f
( ) = Fsoil
( ) soil + Fveg
( ) veg
e
f
f
can ( ) = Fsoil ( )
ef f
( , veg ) = Fveg
( ) veg

ef f
Fsoil ( ) = f1 Fveg ( ), f

ef f
Fveg
( ) = f2 Fsoil ( ), f
Accounts for
multiple scattering
and cavity effect
Effective
parameterization
P3
[68]
( ) = f5 (Fsoil ( ), Fveg ( ), soil , veg , f )

can ( ) = f6 Fsoil ( ), soil , veg , f

( , veg ) = f7 Fsoil ( ), Fveg ( ), soil , veg , f
Accounts for
multiple scattering
P4
P5
[71]
Linearizing P3 considering B(T ) T 4

Linearizing P3 considering B(T ) T
Accounts for
multiple scattering
P6
[119]
( ) = Fsoil ( ) soil + Fveg ( ) veg

can ( ) = Fsoil ( )
( , veg ) = Fveg ( ) veg
P7
P8
[71]
Linearizing P6 considering B(T ) T

Simplifying P7 considering veg = soil = 1
Accounts for
multiple scattering
and cavity effect
clear sky conditions, with irradiance lower than 30 W. m2 between 8 and 14 m

[234]. Finally, [127] observed analytical formulation P2 significantly diverged from
a ray tracing reference when soil and vegetation emissivities were very different.
10.6.4.2 Inverting the Soil/Vegetation Radiative Transfer Equation

The various parameterizations reported above have been assessed in inverse mode
considering dual angle observations, nadir and 45 or 55 off nadir. Given soil
and vegetation emissivities, dual angle measurements allow retrieving component
temperatures. Off nadir angles above 45 are required to capture large angular dynamics and reduce observation errors [108, 109, 119, 144, 235237]. Dual angle
observations at 0 and 55 correspond to the ATSR suite viewing configuration.
272
F. Jacob et al.
Error in leaves temperature retrieval (C)
Error in soil temperature retrieval (C)
Converging conclusions were reported about the documentation requirements for

canopy structural parameters. Poorly estimating LIDF can result in errors on temperatures up to 1 K [118]. LAI must be known within 5% (respectively 10%) for
a 0.5 K accuracy on vegetation (respectively soil) temperature [118, 119]. Similarly,
a 78% relative error on directional cover fraction can induce errors on soil and
vegetation temperatures from 1 to 3 K [189]. Such recommendations have to be
compared with current accuracies on LAI retrievals from solar remote sensing, i.e.,
around 20% [238].
Diverging conclusions were reported about the documentation requirements for
canopy radiative properties, the parameterization degree to be considered, and the
performances. First, [118] concluded a 0.01 accuracy is necessary for soil and vegetation emissivities, whereas [71] claimed using unity values has no consequence.
Second, [71, 119] concluded simple parameterizations similarly performed than
complex ones, while multiple scattering and cavity effect can be neglected. Conversely, [68] reported it is necessary to account for multiple scattering (Fig. 10.7).
Third, [68] concluded a 1 K accuracy can be reached on both temperatures, and
better for vegetation than soil; whereas [71] reported lower errors for soil (<2 K)
than vegetation (<4 K).
SAIL IRT - Mod1

Mod4 - Mod1
Mod3 - Mod1
Mod2 - Mod1
SAIL IRT - Mod1

Mod4 - Mod1
Mod3 - Mod1
Mod2 - Mod1
2
0
LAI
LAI
Fig. 10.7 Performance intercomparison for the different parameterizations listed in Table 10.3.
Mean errors on soil (left) and vegetation (right) temperature retrievals are plotted as functions
of LAI, along with standard deviations (bars). Reference Mod 1 is the probabilistic simulation
model of [139]. SAIL IRT is the TIR version of the SAIL radiative transfer model from [68].
Mod 2 is the parameterization P6, Mod 3 the P3 and Mod 4 the P1. (From [68].)
273

Parameterizing the soil and vegetation radiative transfer equation still is in debate, as
is the required documentation. Additionally to parameters to be well estimated, both
structural (LAI, LIDF, directional gap and cover fractions, hemispherical gap fraction, cavity effect), and radiative (ensemble emissivity and radiometric temperature),
inversion includes emissivity and radiometric temperature for soil and vegetation.
When focusing on finer temperatures such as sunlit and shaded components
for soil and vegetation, or canopy temperature profile, complex simulation models
with additional unknowns have to be considered (Section 10.4.3). Then, attractive
modeling tools are linear kernel driven approaches for accuracy and feasibility
(Section 10.4.2). Such tools have been inverted and validated against multidirectional observations, with promising results about the potential for adjusting the captured angular dynamic [103, 128].
Regarding the aforementioned challenges, ongoing multiangular observations are
of prime interest. Maximum temperatures were captured in nadir and solar directions [58, 73, 108, 109]. Solar peak corresponds to the hot spot effect, with more
sunlit surfaces in the solar direction. Nadir peak corresponds to a larger fraction
of hot soil with a cooler vegetation. These measurements at the ground level have
to be used along with modeling tools, for designing observation configurations and
inversion methods.
10.7 Forefront Investigations: Pertinently Using TIR Remote

Sensing
This section extends the discussion to its scientific context, with the use of land surface temperature for monitoring exchanges of heat, water and mass between soil,
vegetation and atmosphere. For instance, the difficulties faced with aerodynamic
temperature can be overcome using space or time differencing energy balance models. New difficulties are related to time and space issues. First, monitoring land
surface processes from TIR remote sensing requires accounting for the fluctuating
nature of surface temperature. Second, current spaceborne spatial resolutions do not
systematically provide the appropriate scales for process modeling. This yields new
approaches such as remote sensing data assimilation and aggregation/disaggregation
procedures.
Regarding the insufficient accuracy on surface temperature and the difficult
use of aerodynamic temperature (Sections 10.6.2 and 10.6.3), several studies
have suggested using soil and vegetation temperatures with two source models
(Section 10.6.4). However, one source modeling with ensemble radiometric temperature still is very pertinent: time and space differencing models have been
developed this last decade to minimize errors on surface temperature. Spatially
274
F. Jacob et al.
is it possible adjusting air temperatures from surface temperatures to make the

gradient consistent [239]. Similarly, evaporative fraction can be derived from the
radiometric temperature dynamic given an albedo range [59]. Temporally, surface
fluxes can be estimated regarding the potential extreme situations for the considered
surfaces [22]. Similarly, time differencing between morning geostationary observations minimizes error on surface temperature [19]. These approaches have been
widely validated for various environmental conditions, showing the pertinence of
such differencing concepts [19, 22, 23, 25, 60, 66, 240]. However, they are limited
by the agreement between the assumed and captured variabilities.
Monitoring land surfaces is performed using dynamic models which simulate the
considered processes, such as the functioning of cultivated and natural vegetation
[241, 242]. Assimilating remotely sensed information allows constraining model
trajectory, to obtain consistent series for the considered processes and related variables. Model parameters and initial variables (respectively state variables) are adjusted (respectively readjusted) to make simulations and observations agreeing over
a temporal window (respectively at a given time) [39, 243]. Minimizing differences
between simulations and observations relies on stochastic methods and adjoint models [38, 244]. Solar and radar information can be directly assimilated, since the considered variables are almost temporally stable [242, 245248]. Assimilating TIR
observations requires adding a SVAT model, due to surface temperature fluctuations [29, 38, 39, 67, 137]. Then, the SVAT model documents a simple radiative
transfer equation for simulating surface brightness temperature (Section 10.6.1). To
avoid using complex SVAT models, secondary variables such as energy fluxes or
soil moisture can be assimilated in place of primary variables such as surface temperature [34, 249, 250].
Monitoring land surfaces from remote sensing faces the inadequacy of sensor
spatial resolutions to match the process scales (Section 10.3.1). TIR remote sensors
perform at hectometric resolutions with poor temporal samplings, or at kilometric resolutions with daily revisiting. Then, high spatial resolution observations provide valuable information for understanding heterogeneity effects and aggregation
processes [21, 63, 119, 251260]. This is of interest for hydrology and meteorology, when the considered processes can be monitored at a kilometric resolution.
Agricultural issues require higher spatial resolutions, in accordance with the field
scale. Since TIR hectometric resolution remote sensors do not provide sufficient
temporal sampling, a solution is disaggregating daily kilometric resolution variables [261]. By using the well known vegetation index/temperature triangle, it is
possible to disaggregate TIR observations from solar ones, since the latter always
have higher resolutions [56]. However, this is limited by soil moisture conditions
[41, 258]. Other possibilities are statistical or deterministic disaggregation. Statistical procedures have been applied at the landscape unit scale for solar observations
[262]. Deterministic procedures have been proposed for TIR geostationary observations [115, 263].
275

Significant progresses were achieved for the retrieval of ensemble emissivity and radiometric temperature, with accuracies close to requirements for many applications.
Current efforts are refinements rather than new concepts. Regarding the difficult
use of aerodynamic temperature, it is recommended focusing on soil and vegetation
temperatures which are functionally equivalent. Then, the next challenge is deriving
soil and vegetation temperatures with optional sunlit and shaded components, and
canopy temperature profile from the soil to the top of canopy via vegetation layers.
Soil and vegetation temperatures can be recovered parameterizing and inverting simple radiative transfer equations, where simplicity is preferred for operational
applications. More elaborated temperatures require finer approaches, with efforts
to be made on measuring and modeling. Ground-based goniometric systems allow
capturing complex angular dynamics of brightness temperature. Simulation models
currently under improvement are increasingly accurate. The next step is developing inversion methods such as inductive learning and lookup tables, already implemented over the solar domain. Expected difficulties result from the temperature
profile which is driven by micro meteorological conditions. Given surface temperature fluctuations require along track observations from space, recovering temperature profile seems limited to few layers, optionally sunlit and shaded. This should
allow designing optimal viewing configurations for spaceborne sensors, in terms of
observation number and angular distribution.
Finally, current focus on spatial and temporal issues are important. TIR spaceborne sensors do not provide optimum temporal monitoring and spatial scales. Adequate spatial resolutions and revisit rates still do not exist, despite the numerous
missions proposed the last decade. While waiting for such information, it is necessary to develop aggregation and disaggregation methods, to benefit from the combination of spatial scales from high resolution sensors and daily monitoring from
kilometric resolution sensors.
Acknowledgements This review article was possible thanks to numerous interactions, discussions and scientific exchanges, in the framework of various collaborations supported by several
programs: the US ASTER Project of NASAs EOS-Terra Program (P.I. Thomas Schmugge), the
US NASA EOS Grant 03-EOS-02 (P.I. Andrew French), the French PNTS program (project P.Is.
Frederic Jacob and Albert Olioso), the French PNBC program (project P.I. Jean Claude Menaut),
the French Inter Region MIP / PACA program (project P.I. Dominique Courault), and finally the
program from the Department of Research, Development, and International Relations of Purpan
Graduate School of Agriculture. Many thanks to the ISPMSRS 2005 sponsors for permitting this
symposium.
Glossary
AATSR
ANEM
ASTER
Advanced ATSR
Adjusted NEM
Advanced Spaceborne Thermal Emission and Reflection Radiometer
276
F. Jacob et al.
ATSR
AVHRR
Along-Track Scanning Radiometer

Advanced Very High Resolution Radiometer
DA
DAIS
DAISEX
DART
DART-EB
Dual Angle differencing method

Digital Airborne Imaging Spectrometer
DAIS EXperiment
Discrete Anisotropic Radiative Transfer
DART-Energy Balance
EFEDA
ENVISAT
ERS
ETM
European Field Experiment in Desertification threatened Areas

ENVIronment SATellite
European Remote Sensing
Enhanced Thematic Mapper
FIFE
FLUXNET
FTIR
First ISLSCP Field Experiment

Flux Network
Fourier Transform Infra Red spectroradiometer
GLI
GLobal Imager
HAPEX
HIRS
Hydrology Atmosphere Pilot EXperiment

High-resolution Infrared Radiation Sounder
IASI
IRSUTE
ISLSCP
Infrared Atmospheric Sounding Interferometer

Infra Red Satellite Unit for the Thermal Environment
International Satellite Land Surface Climatology Project
JORNEX
JORNada EXperiment
LAI
LIDF
Leaf Area Index

Leaf Inclination Distribution Function
MAS
MASTER
METOP
MIR
MODIS
MODTRAN
MSG
MTI
MODIS Airborne Simulator

MODIS / ASTER airborne simulator
METeorological OPerational
Middle Infra Red: from 3 to 5 m
MODerate resolution Imaging Spectroradiometer
MODerate resolution atmospheric TRANsmission
Meteosat Second Generation
Multispectral Thermal Imager
NDVI
NEM
NOAA
NPOESS
NTB
Normalized Difference Vegetation Index

Normalized Emissivity Method
National Oceanic and Atmospheric Administration
National Polar Orbiting Environmental Sensor Suite
Narrowband To Broadband conversion
ReSeDA
REmote SEnsing Data Assimilation (European research program)
SAIL
SALSA
SVAT
SW
SWVCR
Scattering by Arbitrarily Inclined Leaves

Semi Arid Land Surface Atmosphere
(MexicoUnited StatesFrance joint research program)
Spatially Enhanced Broadband Array Spectrograph System
Spinning Enhanced Visible and Infrared Imager
Soil Moisture Atmosphere Coupling Experiment
Surface Processes and Ecosystem Changes Through Response
Analysis (proposal for a European Space Agency mission)
SoilVegetationAtmosphere Transfer
Split Window differencing method
Split Window Variance Covariance Ratio
TES
TIMS
TIR
TIROS
TISIE
TM
TOVS
TTM
Temperature Emissivity Separation

Thermal Infrared Multispectral Scanner
Thermal Infra Red: from 7 to 14 m
Television Infra Red Observation Satellite
Temperature Independent Spectral Indices of Emissivity
Thematic Mapper
TIROS Operational Vertical Sounders
Two Temperature Method
VIIRS
Visible Infrared Imaging Radiometer Suite
SEBASS
SEVIRI
SMACEX
SPECTRA
277
References
1. Romanov P, Tarpley D (1993) Automated monitoring of snow cover over South America
using GOES imager data. Int. J. Remote Sens. 24(5):11191125
2. Kay J, Gillespie A, Hansen G, Pettit E (2003) Spatial relationships between snow contaminant content, grain size, and surface temperature from multispectral images of Mt. Rainier,
Washington, DC (USA). Remote Sens. Environ. 86(2):216231
3. Aires F, Chedin A, Scott N, Rossow W. (2002) A regularized neural net approach for retrieval of atmospheric and surface temperatures with the IASI instrument. J. Appl. Meteorol.
41:144159
4. Aires F, Rossow W, Chedin A, Scott N (2002) Remote sensing from the infrared atmospheric
sounding interferometer instrument 2. Simultaneous retrieval of temperature, water vapor,
and ozone atmospheric profiles. J. Geophys. Res. 107(D22):4620
5. Prigent C, Aires F, Rossow W (2003) Land surface skin temperatures from a combined
analysis of microwave and infrared satellite observations for an all weather evaluation of the
differences between air and skin temperatures. J. Geophys. Res. 108(D10):4310
6. Hong G, Heygster G, Kunzi K (2005) Intercomparison of deep convective cloud fractions
from passive infrared and microwave radiance measurements. IEEE Geosci. Remote Sens.
Lett. 2(1):1824
7. Sunlner M, Michael K, Bradshaw C Hindell M (2003) Remote sensing of Southern Ocean
sea surface temperature: implications for marine biophysical models. Remote Sens. Environ.
84(2):161173
278
F. Jacob et al.
8. Gould R, Arnone R (2004) Temporal and spatial variability of satellite sea surface temperature and ocean colour in the Japan/East Sea. Int. J. Remote Sens. 25(78):13771382
9. Simpson J, Yueh L, Schmidt A, Harris A (2005) Analysis of along track scanning
radiometer-2 (ATSR-2) data for clouds, glint and sea surface temperature using neural networks. Remote Sens. Environ. 98(23):152181
10. Quattrochi D, Luvall J (2003) Thermal remote sensing in land surface processes. Taylor &
Francis, London
11. Justice CO, Giglio L, Korontzi S, Owens J, Morisette JT, Roy D, Descloitres J, Alleaume S,
Petitcolin F, Kaufman Y (2002) The MODIS fire products. Remote Sens. Environ. 83:244
262
12. Petitcolin F, Vermote E (2002) Land surface reflectance, emissivity and temperature from
MODIS middle and thermal infrared data. Remote Sens. Environ. 83:112134
13. Kasischke ES, Hewson JH, Stocks B, van der Werf G, Randerson J (2003) The use of ATSR
active fire counts for estimating relative patterns of biomass burning a study from the boreal
forest region. Geophys. Res. Lett. 30(18):1969, doi:10.1029/2003GL017859
14. Manso-Delgado L, Aguirre-Gomez R, Alvarez R (2003) Multitemporal analysis of land surface temperature using NOAA-AVHRR: preliminary relationships between climatic anomalies and forest fires. Int. J. Remote Sens. 20:44174423
15. Lo C, Quattrochi D, Luvall J (1997) Application of high-resolution thermal infrared remote
sensing and GIS to assess the urban heat island effect. Int. J. Remote Sens. 18(2):287304
16. Lagouarde JP, Moreau P, Irvine M, Bonnefond JM, Voogt JA, Solliec F (2004) Airborne experimental measurements of the angular variations in surface temperature over urban areas:
case study of Marseille (France). Remote Sens. Environ. 93:443462
17. Mestayer PG, Durand P, Augustin P, Basting S, Bonnefond JM, Benech B, Camp-istron B,
Coppalle A, Delbarre H, Dousset B, Drobinski P, Druilhet A, Frejafon E, Grimmond, CSB,
Groleau D, Irvine M, Kergomard C, Kermadi S, Lagouarde JP, Lemonsu A, Lohou F,
Long N, Masson V, Moppert C, Noilhan J, Offerle B, Oke TR, Pigeon G, Puygrenier V,
Roberts S, Rosant JM, Sad F, Salmond J, Tal-baut M, Voogt J (2005) The urban boundarylayer field campaign in Marseille (UBL/CLU-escompte): set-up and first results. BoundaryLayer Meteorol. 114(2):315365
18. Olioso A, Taconet O, Ben Mehrez M (1996) Estimation of heat and mass fluxes from IR
brightness temperature. IEEE Trans. Geosci. Remote Sens. 34:11841190
19. Norman J, Kustas W, Prueger J, Diak GR (2000) Surface flux estimation using radiometric
temperature: a dual-temperature-difference method to minimize measurement errors. Water
Res. Res. 36:22632274
20. Chehbouni A, Nouvellon Y, Lhomme J, Watts C, Boulet G, Kerr Y, Moran M, Goodrich D
(2001) Estimation of surface sensible heat flux using dual angle observations of radiative
surface temperature. Agric. Forest Meteorol. 108:5565
21. Wassenaar T, Olioso A, Hasager C, Jacob F, Chehbouni A (2002) Estimation of evapotranspiration over heterogeneous pixels. In: Sobrino J (ed.) Proceedings of the First International Symposium on Recent Advances in Quantitative Remote Sensing, September 2002,
Valencia, Spain, pp 458465
22. Su Z (2002) The Surface Energy Balance System (SEBS) for estimation of turbulent heat
fluxes. Hydrol. Earth Sci. Syst. 6:8599
23. Jacob F, Olioso A, Gu X, Su Z, Seguin B (2002) Mapping surface fluxes using visible, near
infrared, thermal infrared remote sensing data with a spatialized surface energy balance
model. Agronomie: Agric. Environ. 22:669680
24. French A, Jacob F, Anderson M, Kustas W, Timmermans W, Gieske A, Su B, Su H, McCabe
M, Li F, Prueger J, Brunsell N (2005) Surface energy fluxes with the Advanced Spaceborne
Thermal Emission and Reflection radiometer (ASTER) at the Iowa 2002 SMACEX site
(USA). Remote Sens. Environ. 99:5565
25. Courault D, Lacarrère P, Clastre P, Lecharpentier P, Jacob F, Marloie O, Prevot L, Olioso A
(2003) Estimation of surface fluxes in a small agricultural area using the three-dimensional
atmospheric model Meso-NH and remote sensing data. Can. J. Remote Sens. 29:741754
279
26. Olioso A, Carlson T, Brisson N (1996) Simulation of diurnal transpiration and photosynthesis of a water stresses soybean crop. Agric. Forest Meteorol. 81:4159
27. Moran M, Vidal A, Troufleau D, Qi J, Clarke T Jr., Pinter PJ, Mitchell T, Inoue Y, Nealeg
MU (1997) Combining Multifrequency Microwave and Optical Data for Crop Management.
28. Moran M, Inoue Y, Barnes E (1997) Opportunities and Limitations for Image-Based Remote
Sensing in Precision Crop Management. Remote Sens. Environ. 61:319346
29. Cayrol P, Kergoat L, Moulin S, Dedieu G, Chehbouni A. (2000) Calibrating a coupled
SVAT vegetation growth model with remotely sensed reflectance and surface temperature.
J. Appl. Meteorol. 39:24522472
30. Bastiaanssen W, Molden D, Makin I (2000) Remote sensing for irrigated agriculture: examples from research and possible applications. Agric. Water Manag. 46:137155
31. Droogers P, Bastiaanssen W (2002) Irrigation performance using hydrological and remote
sensing modeling. J. Irrig. Drain. Eng. 128:1118
32. Hale R, Duchon C (2003) Use of AVHRR-derived surface temperatures in evaluating a landatmosphere model. Int. J. Remote Sens. 24:45274541
33. Jacobs A, Ronda R, Holtslag A (2003) Water vapour and carbon dioxide fluxes over bog
vegetation. Agric. Forest Meteorol. 116:103112
34. Schuurmans J, Troch P, Veldhuizen A, Bastiaanssen W, Bierkens M (2003) Assimilation
of remotely sensed latent heat flux in a distributed hydrological model. Adv. Water Res.
26:151159
35. Nijbroek R, HoogenBoom G, Jones J (2003) Optimizing irrigation management for a spatially variable soybean field. Agric. Syst. 76:359377
36. Panda R, Behera S, Kashyap P (2003) Effective management of irrigation water for wheat
under stress conditions. Agric. Water Manag. 63:3756
37. Bastiaanssen W, Chandrapala L (2003) Water balance variability across Sri Lanka for assessing agricultural and environmental water use. Agric. Water Manag. 58:171192
38. Demarty J, Ottle C, Braud I, Olioso A, Frangi JP, Gupta H, Bastidas L (2005) Constraining
a physically based Soil-Vegetation-Atmosphere Transfer model with surface water content
and thermal infrared brightness temperature measurements using a multiobjective approach.
Water Resour. Res. 41 (2005) W01011, doi:10.1029/2004WR003695
39. Olioso A, Inoue Y, Ortega-Farias S, Demarty J, Wigneron JP, Braud I, Jacob F,
Lecharpentier P, Ottle C, Calvet JC, Brisson N (2005) Future directions for advanced evapotranspiration modeling: Assimilation of remote sensing data into crop simulation models
and SVAT models. Irrig. Drain. Syst. 19(34):355376
40. Capehart W, Carlson T (1997) Decoupling of surface and near-surface soil water content:
A remote sensing perspective. Water Resour. Res. 33:13831395
41. Gillies R, Carlson T, Cui J, Kustas W, Humes K (1997) Verification of the triangle method
for obtaining surface soil water content and energy fluxes from remote measurements of the
Normalized Vegetation Index NDVI and surface radiant temperature. Int. J. Remote Sens.
18:31453166
42. Inoue Y, Olioso A, Choi W (2004) Dynamic change of CO2 flux over bare soil field and its
relationship with remotely sensed surface temperature. Int. J. Remote Sens. 25(10):1881
1892
43. Coudert B, Ottle C, Boudevillain B, Demarty J, Guillevic P (2005) Contribution of thermal
infrared remote sensing data in multiobjective calibration of a dual source SVAT model.
J. Hydrometeorol. 7:404420
44. Sobrino J, Raissouni N (2000) Toward remote sensing methods for land cover dynamic
monitoring. Application to Morroco. Int. J. Remote Sens. 21:353366
45. Wilber A, Kratz D, Gupta S (1999) Surface emissivity maps for use in satellite retrievals of
longwave radiation. NASA/TP-1999-209362, NASA
46. Zhou L, Dickinson R, Ogawa K, Tian Y, Jin M, Schmugge T (2003) Relations between albedos and emissivities from MODIS and ASTER data over North African Desert. Geophys.
Res. Lett. 30(20):2026, doi:10.1029/2003GL018069
280
F. Jacob et al.
47. Zhou L, Dickinson R, Tian Y, Jin M, Ogawa K, Yu H, Schmugge T (2003) A sensitivity

study of climate and energy balance simulations with use of satellite derived emissivity
data over Northern Africa and the Arabian Peninsula. J. Geophys. Res. 108(D24):4795,
doi:10.1029/2003JD004083
48. Jin M, Liang S (2006) Improving Land surface emissivity parameter of land surface model
in GCM. J. Climate 19:28672881
49. Wan Z, Wang P, Li X (2004) Using MODIS Land Surface Temperature and Normalized
Difference Vegetation Index products for monitoring drought in the southern Great Plains,
USA. Int. J. Remote Sens. 25(1):6172
50. Norman J, Becker F (1995) Terminology in thermal infrared remote sensing of natural surfaces. Remote Sens. Rev. 12:159173
51. Dash P, Gottsche FM, Olesen FS, Fischer H (2002) Land surface temperature and emissivity
estimation from passive sensor data: theory and practice-current trends. Int. J. Remote Sens.
23(13):25632594
52. Ogawa K, Schmugge T, Jacob F, French A (2003) Estimation of land surface window (8
12 m) emissivity from multi-spectral thermal infrared remote sensing a case study in a
part of Sahara Desert. Geophys. Res. Lett. 30:10671071
53. Ogawa K, Schmugge T (2004) Mapping Surface Broadband Emissivity of the Sahara Desert
Using ASTER and MODIS Data. Earth Interactions 8 Paper No. 7
54. Wang K, Wan Z, Wang P, Sparrow M, Liu J, Zhou X, Haginoya S (2005) Estimation of Surface Long Wave Radiation and Broadband Emissivity Using Moderate Resolution Imaging
Spectroradiometer (MODIS) Land Surface Temperature/Emissivity Products. J. Geophys.
Res. Atmos. 110(D11, D11109) Paper No. 10.1029/2004JD005566
55. Becker F, Li Z (1995) Surface temperature and emissivity at various scale: definition, measurement and related problem. Remote Sens. Reviews 12:225253
56. Kustas W, Norman J, Anderson M, French A (2003) Estimating subpixel surface temperatures and energy fluxes from the vegetation index-radiometric temperature relationship.
57. Nishida K, Nemani RR, Glassy JM, Running SW (2003) Development of an Evapotranspiration Index From Aqua/MODIS for Monitoring Surface Moisture Status. IEEE Trans.
58. Luquet D, Begue A, Vidal A, Clouvel P, Dauzat J, Olioso A, Gu X, Yu T (2003) Using
multidirectional thermography to characterize water status of cotton. Remote Sens. Environ.
84:411421
59. Roerink GJ, Su Z, Menenti M (2000) S-SEBI: a simple remote sensing algorithm to estimate
the surface energy balance. Phys. Chem. Earth (B) 25(2):147157
60. Gomez M, Sobrino J, Olioso A, Jacob F (2005) Retrieval of evapotranspiration over the
Alpilles test site using PolDER and thermal camera data. Remote Sens. Environ. 96:399
408
61. Kustas W, Humes K, Norman J, Moran M (1996) Single and dual source modeling of surface
energy fluxes with radiometric surface temperature. J. Appl. Meteorol. 35:110121
62. Chehbouni A, Lo Seen D, Njoku E, Monteny B (1996) Examination of the difference between radiative and aerodynamic surface temperatures over sparsely vegetated surfaces. Remote Sens. Environ. 58:177186
63. Chehbouni A, Lo Seen D, Njoku E, Lhomme JP, Monteny B, Kerr Y (1997) Estimation of
sensible heat flux over sparsely vegetated surfaces. J. Hydrol. 188189:855868
64. Habets F, Noilhan J, Golaz C, Goutorbe JP, Lacarrère P, Leblois E, Ledoux E, Martin E,
Ottle C, Vidal-Madjar D (1999) The ISBA surface scheme in a macroscale hydro-logical
model applied to the Hapex-Mobilhy area. Part I: model and database. J. hydrol. 217:7596
65. Su Z, Menenti M, Pelgrum H, van den Hurk B, Bastiaanssen W (1999) Remote sensing of
land surface fluxes for updating numerical weather predictions. In: Nieuwenhuis, Vaughan,
Molenaar (eds), Operational Remote Sensing for Sustainable Development, Balkema,
Rotterdam
281
66. Li J, Su Z, van den Hurk B, Menenti M, Moene A, de Bruin H, Baselga Yrisarry J, Ibanez M,
Cuesta A. (2003) Estimation of sensible heat flux using the Surface Energy Balance System
(SEBS) and ATSR measurements. Phys. Chem. Earth 28:7588
67. Demarty J, Ottle C, Braud I, Olioso A, Frangi JP, Bastidas L, Gupta H (2004) Using a
multiobjective approach to retrieve information on surface properties used in a SVAT model.
J. Hydrol. 287:214236
68. Francois C (2002) The potential of directional radiometric temperatures for monitoring soil
and leaf temperature and soil moisture status. Remote Sens. Environ. 80:122133
69. Braud I, Dantas Antonino A, Vauclin M, Thony J, Ruelle P (1995) A Simple Soil Plant Atmosphere Transfer model (SiSPAT) development and field verification. J. Hydrol. 166:213
250
70. Boulet G, Chehbouni A, Braud I, Vauclin M (1999) Mosaic versus dual source approaches
for modeling the surface energy balance of a semi-arid land. Hydrol. Earth Sci. Syst.
3(2):247258
71. Merlin O, Chehbouni A (2004) Different approaches in estimating heat flux using dual angle
observations of radiative surface temperature. Int. J. Remote Sens. 25:275289
72. Luquet D, Vidal A, Dauzat J, Begue A, Olioso A, Clouvel P (2003) Using directional TIR
measurements and 3D simulations to assess the limitations and opportunities of water stress
indices. Remote Sens. Environ. 90:5362
73. Yu T, Gu X, Tian T, Legrand M, Baret F, Hanocq JF, Bosseno R, Zhang Y (2004) Modeling directional brightness temperature over a maize canopy in row structure. IEEE Trans.
74. Verhoef W, Xiao Q, Jia L, Su Z (2006) Earth observation modeling based on layer scattering
matrices. IEEE Trans. Geosci. Remote Sens. (submitted).
75. Smith J Jr, Ballard, J.R (1999) Physics Based Modeling and Rendering of Vegetation in the
Thermal Infrared. In: 1999 IEEE Workshop on Photometric Modeling for Computer Vision
and Graphics 34
76. Guillevic P, Gastellu-Etchegorry JP, Demarty J, Prevot L (2003) Thermal infrared radiative
transfer within three-dimensional vegetation cover. J. Geophys. Res. -Atm. 108(D8):4248,
doi:10.1029/2002JD02247
77. Gastellu-Etchegorry JP, Martin E, Gascon F (2004) DART: a 3D model for simulating satellite images and studying surface radiation budget. Int. J. Remote Sens. 25(1):7396
78. Belot A, Gastellu-Etchegorry JP, Perrier A (2005) DART-EB, a 3-D model of mass and
energy transfers in vegetated canopy. Remote Sens. Environ. (submitted)
79. Seguin B, Becker F, Phulpin T, Gu X, Guyot G, Kerr Y, King C, Lagouarde J, Ottle C,
Stoll M, Tabbagh T, Vidal A (1999) IRSUTE: a minisatellite project for land surface heat
flux estimation from field to regional scale. Remote Sens. Environ. 68:357369
80. Hernandez-Baquero ED (2000) Characterization of the Earths Surface and Atmosphere
from Multispectral and Hyperspectral Thermal Imagery. Ph.D. thesis, Air Force Institute of
Technology, Wright-Patterson, AFB, OH, 270pp.
81. French A, Norman J, Anderson M (2003) A simple and fast atmospheric correction for
space-borne remote sensing of surface temperature. Remote Sens. Environ. 87:326333
82. Sobrino J, Romaguera M (2004) Land surface temperature retrieval from MSG1-SEVIRI
data. Remote Sens. Environ. 92:247254
83. Petitcolin F, Nerry F, Stoll MP (2002) Mapping temperature independent spectral indice
of emissivity and directional emissivity in AVHRR channels 4 and 5. Int. J. Remote Sens.
23:34733491
84. Nakajima T, Nakajima T, Nakajima M, Fukushima H, Kuji M, Uchiyama A, Kishino M
(1998) Optimization of the Advanced Earth Observing Satellite (ADEOS) II Global Imager
(GLI) channels by use of radiative transfer calculations. Appl. Opt. 37(15):31493163
85. Justice C, Vermote E, Townshend J, Defries R, Roy D, Hall D, Salomonson V, Privette J,
Riggs G, Strahler A, Lucht W, Myneni R, Knyazikhin Y, Running S, Ne-mani R, Wan Z,
Huete A, van Leeuwen W, Wolfe R, Giglio L, Muller JP, Lewis P, Barnsley M (1998) The
MODerate Imaging Spectroradiometer (MODIS): land remote sensing for global change
research. IEEE Trans. Geosci. Remote Sens. 36:12281249
282
F. Jacob et al.
86. Llewellyn-Jones D, Edwards M, Mutlow C, Birks A, Barton I, Tait H (2001) AATSR:

global-change and surface-temperature measurements from Envisat. Eur Space Agency Bull
105:1121
87. Goward S, Masek J, Williams D, Irons J, Thompson R (2001) The Landsat 7 mission
Terrestrial research and applications for the 21st century. Remote Sens. Environ. 78:312
88. Yamaguchi Y, Kahle A, Tsu H, Kawakami T, Pniel M (1998) Overview of Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER). IEEE Trans. Geosci. Remote
Sens. 36:12821289
89. Becker F, Seguin B, Phulpin T, Durpaire J (1996) IRSUTE, a small satellite for water budget
estimate with high resolution infrared imagery. Acta Astronautica 39:883897
90. Lagouarde J (1997) SEXTET: a high resolution and high repetitivity infra red instrument on
a mini satellite platform. Proposal to CNES call for ideas
91. Menenti M, Rast M, Baret F, van den Hurk B, Knorr W, Mauser W, Miller J, Moreno J,
Schaepman M, Verstraete M.M (2004) Understanding vegetation response to climate variability from space: recent advances towards the SPECTRA Mission. In: Meynart R (ed)
Proceedings of SPIE: Sensors, Systems, and Next-Generation Satellites VII. Vol 5234, pp
7685
92. Weber PG, Brock BC, Garrett AJ, Smith BW, Borel CC, Clodius WB Bender SC, Kay RR,
Decker ML (1999) MTI Mission Overview. In: Proceedings of SPIE Conference on Imaging
Spectrometry V. Vol 3753, pp 340346
93. Palluconi F, Meeks G (1985) Thermal Infrared Multi-spectral Scanner (TIMS): an investigators guide to TIMS data. JPL Publications N 8532
94. Muller A, Gege P, Cocks T (2001) The airborne imaging spectrometers used in daisex. In:
DAISEX Final Results Workshop, ESTEC, Holland, 1516 March 2001, ESA SP-499, pp
36
95. King MD, Menzel WP, Grant PS, Myers JS, Arnold GT, Platnick SE, Gumley LE, Tsay
SC, Moeller CC, Fitzgerald M, Brown KS, Osterwisch FG (1996) Airborne scanning spectrometer for remote sensing of cloud, aerosol, water vapor and surface properties. J. Atm.
Oceanic Technol. 13:777794
96. Hook S, Myers J, Thome K, Fitzgerald M, Kahle A (2001) The MODIS/ ASTER airborne
simulator (MASTER) a new instrument for Earth science studies. Remote Sens. Environ.
76:93102
97. Prevot L, Baret F, Chanzy A, Olioso A, Wigneron J, Autret H, Baudin F, Besse-moulin P,
Bethenod O, Blamont D, Blavoux B, Bonnefond J, Boubkraoui S, Bouman B, Braud I,
Bruguier N, Calvet J, Caselles V, Chauki H, Clevers J, Coll C, Company A, Courault D,
Dedieu G, Degenne P, Delecolle R, Denis H, Desprats J, Ducros Y, Dyer D, Fies J,
Fischer A, Francois C, Gaudu J, Gonzalez E, Gouget R, Gu X, Guerif M, Hanocq J,
Hautecoeur J, Haverkamp R, Hobbs S, Jacob F, Jeansoulin R, Jongschaap R, Kerr Y,
King C, Laborie P, Lagouarde J, Laques A, Larcena D, Laurent G, Laurent J, Leroy M,
McAneney J, Macelloni G, Moulin S, Noilhan J, Ottle C, Paloscia S, Pampaloni P, Podvin T,
Quaracino F, Roujean J, Rozier C, Ruisi R, Susini C, Taconet O, Tallet N, Thony J,
Travi Y, Van Leewen H, Vauclin M, Vidal-Madjar D, Vonder O, Weiss M (1998) Assimilation of multi-sensor and multi-temporal remote sensing data to monitor vegetation and
soil: the Alpilles ReSeDA project. In: Tsang L (ed) IGARSS 98 International Geoscience
and Remote Sensing Symposium, IEEE, Institute of Electrical and Electronics Engineers,
Piscataway (USA), Sensing and managing the environment. Vol 5, pp 23992401
98. Olioso A, Prevot L, Baret F, Chanzy A, Braud I, Autret H, Baudin F, Bessemoulin P,
Bethenod O, Blamont D, Blavoux B, Bonnefond J, Boubkraoui S, Bouman B, Bruguier N,
Calvet J, Caselles V, Chauki H, Clevers J, Coll C, Company A, Courault D, Dedieu G,
Degenne P, Delecolle R, Denis H, Desprats J, Ducros Y, Dyer D, Fies J, Fischer A,
Francois C, Gaudu J, Gonzalez E, Gouget R, Gu X, Guerif M, Hanocq J, Hautecoeur
J, Haverkamp R, Hobbs S, Jacob F, Jeansoulin R, Jongschaap R, Kerr Y, King C,
Laborie P, Lagouarde J, Laques A, Larcena D, Laurent G, Laurent J, Leroy M, McAneney
J, Macelloni G, Moulin S, Noilhan J, Ottle C, Paloscia S, Pampaloni P, Podvin T,
99.
100.
101.
102.
103.
104.
105.
106.
107.
108.
109.
110.
111.
112.
113.
114.
115.
116.
117.
118.
283
Quaracino F, Roujean J, Rozier C, Ruisi R, Susini C, Taconet O, Tallet N, Thony J, Travi Y,

Van Leewen H, Vauclin M, Vidal-Madjar D, Vonder O, Weiss M, Wigneron J (1998) Spatial aspects in the Alpilles-ReSeDA project. In: Marceau D (ed.) Scaling and Modeling in
Forestry: Application in Remote Sensing and GIS, Ed. D Marceau, Universite de Montreal,
Quebec pp 92102
Baret F (2000) ReSeDA, assimilation of multisensor and multitemporal remote sensing
data to monitor soil and vegetation functioning. Final Report EC project ENV4CT960326,
INRA, France, 59p.
National Research Council (2004) Committee on NASA-NOAA Transition from Research
to Operations: Satellite Observations of the Earths Environment: Accelerating the Transition of Research to Operations. National Academies Press, Washington, DC
Yu Y, Privette J, Pinheiro A (2005) Analysis of the NPOESS VIIRS Land Surface Temperature Algorithm Using MODIS Data. IEEE Trans. Geosci. Remote Sens. 43:23402350
Vaughan R, Calvin W, Taranik J (2003) SEBASS hyperspectral thermal infrared data: surface emissivity measurement and mineral mapping. Remote Sens. Environ. 85:4863
Liu Q, Gu X, Li X, Jacob F, Hanocq JF (2000) Study on airborne experiment for directional
patterns of thermal infrared radiation. Science in China(E) 30:8997
Payan V, Royer A (2004) Analysis of the Temperature Emissivity Separation (TES) algorithm applicability and sensitivity. Int. J. Remote Sens. 25:1537
Rubio E, Caselles V, Badenas C (1997) Emissivity measurements of several soils and vegetation types in the 814 m wave band: analysis of two fields method. Remote Sens. Environ.
59:490521
Kant Y, Badarinath K (2002) Ground-based method for measuring thermal infrared effective
emissivity and perspectives on the measurement of land surface temperature from satellite
data. Int. J. Remote Sens. 23(11):21792191
Rubio E, Caselles V, Coll C, Valor E, Sospedra F (2003) Thermal infrared emissivities of
natural surfaces: improvement on the experimental set-up and new measurements. Int. J.
Remote Sens. 24(24):53795390
Li ZL, Zhang R, Sun X, Su H, Tang X, Zhu Z, Sobrino JA (2004) Experimental system
for the study of the directional thermal emission of natural surfaces. Int. J. Remote Sens.
25(1):195204
Coret L, Briottet X, Kerr Y, Chehbouni G (2005) Experimental study of directional aggregation over heterogeneous surfaces in the thermal infrared. Remote Sens. Environ. (submitted).
Garrat J, Hicks B (1973) Momentum, heat and water vapor transfer to and from natural and
artificial surfaces. Quart. J. Roy. Meteorol. Soc. 99:680687
Ham J, Heilman J (1991) Aerodynamic land surface resistances affecting energy transport
in a sparse crop. Agric. Forest Meteorol. 53:267284
Garrat J (1978) Transfer characteristics for a heterogeneous surface of large aerodynamic
roughness. Quart. J. Roy. Meteorol. Soc. 104:491502
Troufleau D, Lhomme J, Monteny B, Vidal A (1997) Sensible heat flux and radiometric
surface temperature over sparse Sahelian vegetation. I. An Experimental analysis of the kB
parameter. J. Hydrol. 188189:815838
Choudhury B (1989) Estimating evaporation and carbon assimilation using infrared temperature data: vistas in modeling. In: Theory and Applications of Optical Remote Sensing.
Wiley Interscience, Paris, France, pp 629690
Norman J, Divakarla M, Goel N (2000) Algorithms for Extracting Information from Remote
Thermal-IR Observations of the Earths Surface. Remote Sens. Environ. 51:157168
Martonchik JV, Bruegge CJ, Strahler AH (2000) A Review of Reflectance Nomenclature for
Remote Sensing. Remote Sens. Rev. 19:920
Li X, Strahler A, Friedl M (1999) A conceptual model for effective directional emissivity
from non-isothermal surfaces. IEEE Trans. Geosci. Remote Sens. 37:25082517
Francois C, Ottle C, Prevot L (1997) Analytical parameterization of canopy directional
emis-sivity and directional radiance in the thermal infrared. Application on the retrieval
of soil and foliage temperatures using two directional measurements. Int. J. Remote Sens.
18:25872621
284
F. Jacob et al.
119. Chehbouni A, Nouvellon Y, Kerr Y, Moran M, Watts C, Prevot L, Goodrich D, Rambal S

(2001) Directional effect on radiative surface temperature measurements over a semi-arid
grassland site. Remote Sens. Environ. 76:360372
120. Olioso A, Soria G, Sobrino J, Duchemin B (2006) Evidences of low land surface thermal infrared emissivity in presence of dry vegetation. IEEE Geosci. Remote Sens. Lett. 4(1):112
116
121. Sutherland R, Bartholic J (1977) Significance of vegetation in interpreting thermal radiation
from a terrestrial surface. J. Appl. Meteorol. 16:759763
122. Colton A (1996) Effective thermal parameters for a heterogeneous lans surface. Remote
123. Barducci A, Pippi Y (1996) Temperature and emissivity retrieval from remotely sensed images using the Grey Body Emissivity method. IEEE Trans. Geosci. Remote Sens. 34:681
695
124. Valor E, Caselles V (1996) Mapping land surface emissivity from NDVI: application to
European, African and South American areas. Remote Sens. Environ. 57:167184
125. Sobrino J, Sòria G, Prata AJ (2004) Surface temperature retrieval from Along Track
Scanning Radiometer 2 data: Algorithms and validation. J. Geophys. Res. 109D11101,
doi:10.1029/2003JD004212
126. Sòria G, Sobrino J (2006) Envisat/AATSR derived land surface temperature over a heterogeneous region. Remote Sens. Environ. 111(4):409422
127. Chen L, Li ZL, Liu Q, Chen S, Tang Y, Zhong B (2004) Definition of component effective
emissivity for heterogeneous and non-isothermal surfaces and its approximate calculation.
Int. J Remote Sens. 25(1):231244
128. Su L, Li X, Friedl M, Strahler A, Gu X (2002) A kernel-driven model of effective directional
emissivity for non-isothermal surfaces. Prog. Natural Sci. 12(8):603607
129. Snyder W, Wan Z (1998) BRDF models to predict spectral reflectance and emissivity in the
infrared. IEEE Trans. Geosci. Remote Sens. 36:214225
130. Roujean J, Leroy M, Deschamps P (1992) A bidirectional reflectance model of the Earths
131. Wanner W, Li X, Strahler A (1995) On the derivation of kernels for kernel-driven models of
132. Engelsen O, Pinty B, Verstraete M, Martonchik J (1996) Parametric bidirectional reflectance
factor models : evaluation, improvements and applications. Report EUR16426EN, European
Commission, Joint Researches Center, Space Application Institute, ISPRA, Italy
133. Berk A, Bernstein L, Anderson G, Acharya P, Robertson D, Chetwynd J, Adler-Golden S
(1998) MODTRAN cloud and multiple scattering upgrades with application to AVIRIS.
134. Anderson G, Berk A, Acharya P, Matthew M, Bernstein L, Chetwynd J, Dothe H, AdlerGolden S, Ratkowski A, Felde G, Gardner J, Hoke M, Richtsmeier S, Pukall B, Mello J,
Jeong L (2000) MODTRAN 4.0: radiative transfer modeling for remote sensing. In: Algorithms for multispectral, hyperspectral and ultraspectral imagery VI: Proceedings of SPIE ,
pp 176183
135. Verhoef, W. (1984) Light scattering by leaf layers with application to canopy reflectance
modeling: the SAIL model. Remote Sens. Environ. 16:125141
136. Olioso A (1995) Simulating the relationship between thermal infrared emissivity and the
Normalized Difference Vegetation Index. Int. J. Remote Sens. 16:32113216
137. Olioso A, Chauki H, Courault D, Wigneron J (1999) Estimation of evapotranspiration and
photo-synthesis by assimilation of remote sensing data into SVAT models. Remote Sens.
Environ. 68:341356
138. Kimes D (1980) Effects of vegetation canopy structure on remotely sensed canopy temperatures. Remote Sens. Environ. 10:165174
139. Prevot L (1985) Modelisation des e changes radiatifs au sein des couverts vegetaux Application a` la teledetection Validation sur un couvert de Mas. Ph.D. thesis, p 185, Universite
Paris VI, Paris, France
285
140. Jackson R, Reginato R, Pinter P (1979) Plant canopy information extraction from composite
site reflectance of row crop. Appl. Opt. 18:37753782
141. Kimes D, Kirchner J (1983) Directional radiometric measurements of row crop temperatures. Int. J. Remote Sens. 4:299311
142. Sobrino JA, Caselles V (1990) Thermal infrared radiance model for interpreting the directional radiometric temperature of a vegetative surface. Remote Sens. Environ. 33:193199
143. Caselles V, Sobrino JA, Coll C (1992) A physical model for interpreting the land surface
temperature obtained by remote sensors over incomplete canopies. Remote Sens. Environ.
39:203211
144. Coret L, Briottet X, Kerr Y, Chehbouni G (2004) Simulation study of view angle effects
on thermal infrared measurements over heterogeneous surfaces. IEEE Trans. Geosci. and
145. Du Y, Liu Q, Chen L, Liu Q, Yu, T (2007) Modeling directional brightness temperature of
the winter wheat canopy at the ear stage. IEEE Trans. Geosci. Remote Sens. 45(11):3721
3739
146. Norman J, Campbell G (1983) Application of a plant-environment model to problems in
irrigation. In: Hillel D (ed.) Advances in irrigation, Academic Press, New York, pp 155188
147. Norman J, Arkebauer T (1991) Predicting canopy light-use efficiency from leaf characteristics. In: Ritchie J, Hanks R (eds), Modeling plant and soil systems, ASA, Madison, WI, pp
125143
148. Dauzat J, Rapidel B, Berger A (2001) Simulation of leaf transpiration and sap flow in virtual
plants: description of the model and application to a coffee plantation in Costa Rica. Agric.
Forest Meteorol. 109:143160
149. Jia L (2004) Modeling heat exchanges at the land atmosphere interface using multi-angular
thermal infrared measurements. Ph.D. thesis, Wageningen University, Wageningen, The
Netherlands
150. Su H, Zhang R, Tang XZ, Sun X (2000) Determination of the effective emissivity for the
regular and irregular cavities using Monte-carlo method. Int. J. Remote Sens. 21(11):2313
2319
151. Su L, Li X, Liang S, Strahler A (2003) Simulation of scaling effects of thermal emission
from non-isothermal pixels with the typical three dimension structure. Int. J. Remote Sens.
24(19):37433753
152. Kimes D, Knyazikhin Y, Privette J, Abuelgasim A, Gao F (2000) Inversion methods for
153. Chen J, Li X, Nilson T, Strahler A (2000) Recent advances in geometrical optical modeling
and its applications. Remote Sens. Rev. 18:227262
154. Qin W, Liang S (2000) Plane-parallel canopy radiation transfer modeling: recent advances
and future directions. Remote Sens. Rev. 18:281305
155. Nolin A, Liang S (2000) Progress in bidirectional reflectance modeling and applications for
surface particulate media: snow and soils. Remote Sens. Rev. 18:307342
156. Lucht W, Roujean J (2000) Considerations in the parametric modeling of BRDF and albedo
from multiangular satellite sensor observations. Remote Sens. Rev. 18:343379
157. Schmugge T, French A, Ritchie J, Rango A, Pelgrum H (2002) Temperature and emissivity
separation from multispectral thermal infrared observations. Remote Sens. Environ. 79:189
198
158. Jacob F, Petitcolin F, Schmugge T, Vermote E, Ogawa K, French A (2004) Comparison
of land surface emissivity and radiometric temperature from MODIS and ASTER sensors.
159. Hall F, Huemmrich K, Goetz S, Sellers P, Nickeson P (1992) Satellite remote sensing of
surface energy balance : success, failures and unresolved issues in FIFE. J. Geophys. Res.
97:1906119089
160. Bolle H, et al. (1993) EFEDA: European Field in Desertification threatened Area. Annales
Geophysicae 11:173189
286
F. Jacob et al.
161. Goutorbe J, et al. (1994) HAPEX-SAHEL: a large scale study of land-atmosphere interactions in the semiarid tropics. Annales Geophysicae 12: 5364
162. Havstad K, Kustas W, Rango A, Ritchie J, Schmugge T (2000) Jornada Experimental Range:
a unique arid land location for experiments to validate satellite system. Remote Sens. Environ. 74:1325
163. Baldocchi D, Falge E, Gu LH, Olson R, Hollinger D, Running S, Anthoni P, Bern-hofer C,
Davis K, Evans R, Fuentes J, Goldstein A, Katul G, Law B, Lee XH, Malhi Y, Meyers T,
Munger W, Oechel W, Paw UKT, Pilegaard K Schmid HP, Valentini R, Verma S (2001)
FLUXNET: a new tool to study the temporal and spatial variability of ecosystem-scale carbon dioxide, water vapor, and energy flux densities. Bull. Am. Meteorol. Soc. 82-11:2415
2434
164. Berger M, Rast M, Wursteisen P, Attema E, Moreno J, Mueller A, Beisl U, Richter R,
Schaepman M, Strub G, Stoll MP, Nerry F, Leroy M (2001) The DAISEX campaigns in
support of a future land-surface-processes mission ESA Bulletin. Remote Sens. Environ.
105:101111
165. Goodrich DC, Chehbouni A, Goff BF, Macnish R, Maddock T, Moran MS,
Shut-tleworth WJ, Williams DG, Watts CJ, Hipps LJ, Cooper DI, Schieldge J, Kerr YH,
Arias H, Kirkland M, Carlos R, Cayrol P, Kepner W, Jones B, Avissar R, Begue A,
Bonnefond JM, Boulet G, Branan B, Brunel JP, Chen LC, Clarke T, Davis MR, Debruin H,
Dedieu G, Elguero E, Eichinger WE, Everitt J, Garatuza-Payan J, Gempko VL, Gupta H,
Harlow C, Hartogensis O, Helfert M, Holifield C, Hymer D, Kahle A, Keefer T, Krishnamoorthy S, Lhomme JP, Lagouarde JP, Lo Seen D, Luquet D, Marsett R, Monteny B,
Ni W, Nouvellon Y, Pinker R, Peters C, Pool D, Qi J, Rambal S, Rodriquez J, Santiago F,
Sano E, Schaeffer SM, Schulte M, Scott R, Shao X, Snyder KA, Sorooshian S, Unkrich CL,
Whitaker M, Yucel I (2000) Preface paper to the Semi-Arid Land-Surface- Atmosphere
(SALSA) Program special issue. Agric. Forest Meteorol. 105:320
166. Njoku E, Lakshmi V, ONeill P (2004) Soil moisture field experiment special issue. Remote
167. Schmugge T, Hook S, Coll C (1998) Recovering surface temperature and emissivity from
thermal infrared multispectral data. Remote Sens. Environ. 65:121131
168. Coll C, Caselles V, Rubio E, Sospedra F, Valor E (2000) Temperature and emissivity separation from calibrated data of the Digital Airborne Imaging Spectrometer. Remote Sens.
Environ. 76:250259
169. Buongiorno M, Realmuto V, Doumaz F (2002) Recovery of spectral emissivity from thermal
infrared multispectral scanner imagery acquired over a mountainous terrain: a case study
from mount Etna, Sicily. Remote Sens. Environ. 79:123133
170. Coll C, Caselles V, Rubio E, Valor E, Sospedra F, Baret F, Prevot L, Jacob F (2002) Temperature and emissivity extracted from airborne multi-channel data in the ReSeDA experiment.
Agronomie: Agric. Environ. 22:567574
171. Jacob F, Gu X, Hanocq JF, Baret F (2003) Atmospheric corrections of single broadband
channel and multidirectional airborne thermal infrared data. Application to the ReSeDA
Experiment. Int. J. Remote Sens. 24:32693290
172. Coll C, Caselles V, Valor E, Rubio E (2003) Validation of temperature-emissivity separation and split-window methods from TIMS data and ground measurements. Remote Sens.
Environ. 85:232242
173. Coll C, Valor E, Caselles V, Niclòs R (2003) Adjusted Normalized Emissivity Method for
surface temperature and emissivity retrieval from optical and thermal infrared remote sensing data. J. Geophys. Res. 108(D23):4739, doi:10.1029/2003JD003688
174. Sobrino J, Jimenez-Munoz J, El-Kharraz J, Gomez M, Romaguera M, Sòria G (2004)
Single-channel and two-channel methods for land surface temperature retrieval from DAIS
data and its application to the Barrax site. Int. J. Remote Sens. 25(1):215230
175. Schmugge T, Jacob F, French A, Ogawa K (2002) Quantitative estimates of emissivity using ASTER data. In: Sobrino J (ed.) Proceedings of the First International Symposium on
Recent Advances in Quantitative Remote Sensing, September 2002, Valencia, Spain, pp
585589
287
176. Schmugge T, Ogawa K, Jacob F, French A, Hsu Y, Ritchie J, Rango A (2003) Validation of
emissivity estimates from ASTER data. In: Proceedings of 2003 International Geoscience
and Remote Sensing Symposium, Toulouse, France, 2125 July 2003. Vol III , pp 1873
1875
177. Sobrino J, Jimenez-Munoz J, Paolini L (2004) Land surface temperature retrieval from
Land-sat TM 5. Remote Sens. Environ. 90:434440
178. Li F, Jackson T, Kustas W, Schmugge T, French A, Cosh M, Bindlish R (2004) Deriving
land surface temperature from Landsat 5 and 7 during SMEX02/SMACEX. Remote Sens.
Environ. 92:521534
179. Jimenez-Munoz JC, Sobrino JA, Gillespie A, Sabol D, Gustafson TG (2006) Improved land
surface emissivities over agricultural areas using ASTER NDVI. Remote Sens. Environ.
103:474487
180. Jimenez-Munoz JC, Sobrino JA (2003) A generalized single-channel method for retrieving land surface temperature from remote sensing data. J. Geophys. Res. 108(D22):4688,
doi:10.1029/2003JD003480
181. Wan Z, Zhang Y, Zhang Q, Li ZL (2002) Validation of the land surface temperature products
retrieved from Terra Moderate Resolution Imaging Spectroradiometer data. Remote Sens.
Environ. 83:163180
182. Wan Z, Zhang Y, Zhang Q, Li ZL (2004) Quality assessment and validation of the MODIS
global land surface temperature. Int. J. Remote Sens. 25:261274
183. Coll C, Caselles V, Galve JM, Valor E, Niclòs R, Sanchez J, Rivas R (2005) Ground measurements for the validation of land surface temperatures derived from AATSR and MODIS
184. Weiss M, Baret F, Garrigues S, Lacaze R (2007) LAI and fAPAR CYCLOPES global products derived from VEGETATION. Part2: validation and comparison with MODIS collection
4 products. Remote Sens. Environ. 110(3):317331
185. Tonooka H (2001) An atmospheric correction algorithm for thermal infrared multispectral
data over land a water vapor scaling method. IEEE Trans. Geosci. Remote Sens. 39:682
692
186. Dash P, Gottsche FM, Olesen FS (2002) Potential of MSG for surface temperature
and emissivity estimation: considerations for real-time applications. Int. J. Remote Sens.
23(20):45114518
187. Peres L, DaCamara C (2004) Land surface temperature and emissivity estimation based
on the two-temperature method: sensitivity analysis using simulated MSG/SEVIRI data.
188. Dash P, Gottsche FM, Olesen FS, Fischer H (2005) Separating surface emissivity and temperature using two-channel spectral indices and emissivity composites and comparison with
a vegetation fraction method. Remote Sens. Environ. 96:117
189. Jia L, Li Z, Menenti M, Su Z, Verhoef W, Wan Z (2003) A practical algorithm to infer
soil and foliage component temperatures from bi-angular ATSR-2 data. Int. J. Remote Sens.
24:47394760
190. Kratz DP, Chou MD, Yan MMH, Ho CH (1998) Minor Trace Gas Radiative Forcing Calculations Using the k Distribution Method with One-Parameter Scaling. J. Geophys. Res.
103(D24):3164731656
191. Saunders R, Brunel P, English S, Bauer P, OKeeffe U, Francis P, Rayer P (2005) Rttov-8,
science and validation report. Technical report, EUMESAT (2005) report NWPSAF-MOTV-007, v1.2, 02/03/2005
192. Gottsche FM, Olesen F (2002) Evolution of neural networks for radiative transfer calculations in the terrestrial infrared. Remote Sens. Environ. 80:157164
193. Derber J, Parrish D, Lord S (1991) The new global operational analysis system at the National Meteorological Center (NMC). Weather Forecast. 6:538547
194. Derber J, Wu W (1998) The use of TOVS cloud-cleared radiances in the NCEP SSI analysis
system. Monthly Weather Rev. 126:22872299
288
F. Jacob et al.
195. Schroedter M, Olesen FS, Fischer H (2003) Determination of land surface temperature distributions from single channel IR measurements: an effective spatial interpolation method
for the use of TOVS, ECMWF, and radiosonde profiles in the atmospheric correction
scheme. Int. J. Remote Sens. 24:11891196
196. Hirano A, Welch R, Lang H (2002) Mapping from ASTER stereo image data: DEM validation and accuracy assessment. ISPRS J. Photogramm. Remote Sens. 55:115
197. Gu D., Gillespie AR, Kahle AB, Palluconi FD (2000) Autonomous atmospheric compensation (AAC) of high resolution hyperspectral thermal infrared remote sensing imagery. IEEE
198. Li ZL, Li J, Su Z, Wan Z, Zhang R (2003) A new approach for retrieving precipitable water
from ATSR2 split-window channel data over land area. Int. J. Remote Sens. 24(24):5095
5117
199. Sobrino J, El Kharraz J, Li ZL (2003) Surface temperature and water vapor retrieval from
MODIS data. Int. J. Remote Sens. 24(24):51615182
200. De Felice T, LLoyd D, Meyer D, Baltzer T, Piraino P (2003) Water vapour correction of the
daily 1 km AVHRR global land dataset: part I validation and use of the water vapour input
field. Int. J. Remote Sens. 24(11):23652375
201. Rabier F, Fourrie N, Chafa D, Prunet P (2001) Channel selection methods for infrared
atmospheric sounding interferometer radiances. Quart. J. Roy. Meteorol. Soc. 128:117
202. Pinheiro A, Privette J, Mahoney R, Tucker C (2004) Directional effects in a daily
AVHRR land surface temperature dataset over Africa. IEEE Trans. Geosci. Remote Sens.
42(9):19411954
203. van de Griend A, Owe M (1993) On the relationship between thermal infrared emissivity
and the Normalized Difference Vegetation Index for natural surfaces. Int. J. Remote Sens.
14 :11191131
204. Kerr Y, Lagouarde J, Nerry F, Ottle C (2004) Land surface temperature retrieval techniques
and applications: case of the AVHRR. In: Thermal remote sensing in land surface processes.
Boca Raton, FL: CRC, pp 33109 + 14 colour plates
205. Snyder W, Wan Z, Zhang Y, Feng YZ (1998) Classification-based emissivity for land surface
temperature measurement from space. Int. J. Remote Sens. 19:27532774
206. Payan V, Royer A (2004) Spectral emissivity of northern land cover types derived with
MODIS and ASTER sensors in MWIR and LWIR. Research Note, Can. J. Remote Sens.
30(2):150156
207. Peres L, DaCamara C (2005) Emissivity maps to retrieve land-surface temperature from
MSG/SEVIRI. IEEE Trans. Geosci. Remote Sens. 43(8):18341844
208. Gillespie A (1985) Lithologic mapping of silicate rocks using TIMS. In: The TIMS data
user workshop, June 1819, 1985. JPL Publication N 86-38, Jet Propulsion Laboratory,
Pasadena, CA (1985), pp 2944
209. Mushkin A, Balick LK, Gillespie AR (2005) Extending surface temperature and emissivity retrieval to the mid-infrared (35 m) using the Multispectral Thermal Imager (MTI).
Remote Sens. Environ. 98 141151
210. Gillespie A, Rokugawa S, Matsunaga T, Cothern S, Hook S, Kahle A (1998) A temperature and emissivity separation algorithm for Advanced Spaceborne Thermal Emission and
Reflection radiometer (ASTER) images. IEEE Trans. Geosci. Remote Sens. 36:11131126
211. Liang S (2001) An optimization algorithm for separating land surface temperature and
emissivity from multispectral thermal infrared imagery. IEEE Trans. Geosci. Remote Sens.
39:264274
212. Mushkin A, Balick LK, Gillespie AR (2002) Temperature/emissivity separation of MTI data
using the Terra/ASTER TES algorithm. In: Shen SS, Lewis PE (eds) (2002) Proceedings of
SPIE Algorithms and Technologies for Multispectral, Hyperspectral, and Ultraspectral
Imagery VIII. Vol 4725, pp 328337
213. Gu D, Gillespie AR (2000) A new approach for temperature and emissivity separation. Int.
J. Remote Sens. 21:21272132
289
214. Ogawa K, Schmugge T, Jacob F, French A (2002) Estimation of broadband land surface
emissivity from multispectral thermal infrared remote sensing. Agronomie: Agric. Environ.
22:695696
215. Peres L, DaCamara C (2004) Inverse problems theory and application: analysis of the twotemperature method for land-surface temperature and emissivity estimation. Geosci. Remote
Sens. Lett. 1(3):206210
216. Wan Z, Li ZL (1997) A physics-based algorithm for retrieving land-surface emissivity and
temperature from EOS/MODIS data. IEEE Trans. Geosci. Remote Sens. 35:980996
217. Watson K (1992) Two-temperature method for measuring emissivity. Remote Sens. Environ.
42:117121
218. Boyd D, Petitcolin F (2004) Remote sensing of the terrestrial environment using middle
infrared radiation (3.05.0 m). Int. J. Remote Sens. 25(17):33433368
219. Huntingford C, Verhoef A, Stewart J (2000) Dual versus single source models for estimating
surface temperature of African Savannah. Hydrol. Earth Sci. Syst. 4(1):185191
220. Kustas W, Choudhury B, Moran M, Reginato R, Jackson R, Gay L, Weaver H (1989) Determination of sensible heat flux over sparse canopy using thermal infrared data. Agric. Forest
Meteorol. 44:197216
221. Moran M, Kustas W, Vidal A, Stannard D, Blanford J, Nichols W (1994) Use of groundbased remotely sensed data for surface energy balance evaluation of a semiarid rangeland.
Water Res. Res. 30:13391349
222. Stewart J, Kustas W, Humes K, Nichols W, Moran M, de Bruin H (1994) Sensible heat flux
radiometric surface temperature relationship for eight semiarid areas. J. Appl. Meteorol.
33:11101117
223. Sun J, Mahrt L (1995) Determination of surface fluxes from the surface radiative temperature. J. Atmos. Sci. 52(8):10961106
224. Cahill A, Parlange M (1997) On the Brutsaert temperature roughness length model for sensible heat flux estimation. Water Res. Res. 10:23152324
225. Lhomme J, Troufleau D, Monteny B, Chehbouni A, Bauduin S (1997) Sensible heat flux and
radiometric surface temperature over sparse Sahelian vegetation II. A model for the kB1
parameter. J. Hydrol. 188189:839854
226. Verhoef A, De Bruin H, Van Den Hurk B (1997) Some practical notes on the kB1 . J. Appl.
Meteorol. 36:560572
227. Qualls R, Yates D (2001) Directional radiometric temperature profiles within a grass canopy.
Adv. Water Res. 24:145155
228. Molder M, Lindroth A (2001) Dependence of kB1 factor on roughness Reynolds number
for barley and pasture. Agric. Forest Meteorol. 106:147152
229. Molder M, Kellner E (2002) Excess resistance of bog surfaces in central Sweden. Agric.
230. Prevot L, Brunet Y, Paw UK, Seguin B (1993) Canopy modeling for estimating sensible heat
flux from thermal infrared measurements. In: Workshop on Thermal Remote Sensing of the
Energy Balance Over Vegetation in Conjunction with Other Sensor, La Londe les Maures,
2023 September. Vol 93, pp 1730
231. Norman J, Kustas W, Humes K (1995) Source approach for estimating soil and vegetation
energy fluxes in observations of directional radiometric surface temperature. Agric. Forest
Meteorol. 77:263293
232. Kustas W, Norman J (1999) Evaluation of soil and vegetation heat flux predictions using
a simple two-source model with radiometric temperatures for partial canopy cover. Agric.
233. Kustas W, Norman J (2000) A two source energy balance approach using directional radiometric temperature observations for sparse canopy covered surfaces. Agron. J. 92:847854
234. Olioso A (1995) Estimating the difference between brightness and surface temperatures for
a vegetal canopy. Agric. Forest Meteorol. 72:237242
235. Lagouarde JP, Kerr Y, Brunet Y (1995) An experiment study of angular effects on surface
temperature for various plant canopies and bare soils. Agric. Forest Meteorol. 77:167190
290
F. Jacob et al.
236. Verbrugghe M, Cierniewski J (1998) Influence and modeling of view angles and micro-relief
on surface temperature measurements of bare agricultural soils. ISPRS J. Photogramm. Remote Sens. 53:166173
237. Sobrino J, Cuenca J (1999) Angular variation of thermal infrared emissivity for some natural
surfaces from experimental measurements. Appl. Opt. 38(18):39313936
238. Weiss M, Baret F, Leroy M, Hautecoeur O, Bacour C, Prevot L, Bruguier N (2002) Evaluation of Neural Network techniques to estimate canopy biophysical variables from remote
sensing data. Agronomie: Agric. Environ. 22:547553
239. Bastiaanssen W, Menenti M, Feddes R, Holtslag A (1998) A remote sensing surface energy
balance algorithm for land (SEBAL). I: formulation. J. Hydrol. 212213:198212
240. Su Z, Yacob A, Wen J, Roerink G, He Y, Gao B, Boogaard H, van Diepen C (2003) Assessing soil moisture with remote sensing data: theory, experimental validation, and application
to drought monitoring over the North China Plain. Phys. Chem. Earth 28:89101
241. Brisson N, Gary C, Justes E, Roche R, Mary B, Ripoche D, Zimmer D, Sierra J, Bertuzzi P,
Burger P, Bussière F, Cabidoche Y, Cellier P, Debaeke P, Gaudillère J, Henault C, Maraux F,
Seguin B, Sinoquet H (2003) An overview of the crop model STICS. Eur. J. Agron. 18:309
332
242. Mougin E, Lo Seen D, Rambal S, Gaston A, Hiernaux P (2004) A regional sahelian grassland model to be coupled with multispectral satellite data. I: model description and validation. Remote Sens. Environ. 52:181193
243. Pellenq J, Boulet G (2004) A methodology to test the pertinence of remote-sensing data
assimilation into vegetation models for water and energy exchange at the land surface.
Agronomie 24:197204
244. Lauvernet C, Baret F, Le Dimet FX (2003) Assimilating high temporal frequency SPOT data
to describe canopy functioning: the ADAM project. In: Geoscience and Remote Sensing
Symposium, 2003. IGARSS 03. Proceedings. 2003 IEEE International. Vol 5, pp 3184
3186
245. Delecolle R, Maas S, Guerif M, Baret F (1992) Remote sensing and crop production model:
present trends. ISPRS J. of Photogramm. Remote Sens. 47:145161
246. Weiss M, Troufleau D, Baret F, Chauki H, Prevot L, Olioso A, Bruguier N, Brisson N (2001)
Coupling canopy functioning and radiative transfer models for remote sensing data assimilation. Agric. Forest Meteorol. 108:113128
247. Moulin S, Kergoat L, Cayrol P, Dedieu G, Prevot L (2002) Calibration of a coupled canopy
functioning and SVAT model in the ReSeDA experiment. Towards the assimilation of
SPOT/HRV observations into the model. Agronomie: Agric. Environ. 22:681686
248. Prevot L, Chauki H, Troufleau D, Weiss M, Baret F, Brisson N (2003) Assimilating optical
and radar data into the STICS crop model for wheat. Agronomie: Agric. Environ. 23:297
303
249. Ottle C, Vidal-Madjar D (1994) Assimilation of soil moisture inferred from infrared remote
sensing in a hydrological model over the HAPEX-MOBILHY region. J. Hydrol. 158:241
264
250. van den Hurk B, Bastiaanssen W, Pelgrum H. van Meijgaard E (1997) A new methodology for assimilation of initial soil moisture fields in weather prediction models using
METEOSAT and NOAA data. J. Appl. Meteorol. 36:12711283
251. Chehbouni A, Njoku E, Lhomme JP, Kerr Y (1995) Approaches for averaging surface parameters and fluxes over heterogeneous terrain. J. Climate 8:13861393
252. Njoku E, Hook S, Chehbouni A (1996) Effects of surface heterogeneity on thermal remote
sensing of land parameters, Chapter 2. In: Scaling up in hydrology using remote sensing.
John, West Sussex, pp 1931
253. Su Z, Pelgrum H, Menenti M (1999) Aggregation effects of surface heterogeneity in land
surface processes. In: Su Z, Menenti M (eds), Hydrology and Earth Science System. Vol 3,
pp 549563
254. Bouguerzaz FA, Olioso A, Raffy M (1999) Modeling radiative and energy balance on heterogeneous areas from measured radiances. Can. J. Remote Sens. 25(4):412424
291
255. Kustas W, Norman J (2000) Evaluating the effects of subpixel heterogeneity on pixel average
fluxes. Remote Sens. Environ. 74:327342
256. Chehbouni A, Watts C, Kerr Y, Dedieu G, Rodriguez JC, Santiago F, Cayrol P, Boulet G,
Goodrich D (2000) Methods to aggregate turbulent fluxes over heterogeneous surfaces: application to SALSA data set in Mexico. Agric. Forest Meteorol. 105:133144
257. French A, Schmugge T, Kustas W, Brubaker K, Prueger J (2003) Surface energy fluxes over
El Reno, Oklahoma using high resolution remotely sensed data. Water Res. Res. 39:1164
258. Brunsell N, Gillies R (2003) Scale issues in land atmosphere interactions: implications for
remote sensing of the surface energy balance. Agric. Forest Meteorol. 117:203221
259. Kustas W, Li F, Jackson T, Prueger J, MacPherson J, Wolde M (2004) Effects of remote
sensing pixel resolution on modeled energy flux variability of croplands in Iowa. Remote
260. Lyons T, Halldin S (2004) Surface heterogeneity and the spatial variation of fluxes. Agric.
261. Norman J, Anderson M, Kustas W, French A, Mecikalski J, Torn R, Diak G, Schmugge T,
Tanner B (2003) Remote sensing of surface energy fluxes at 101 m pixel resolutions. Water
Res. Res. 39:1221
262. Cardot H, Faivre R, Goulard M (2003) Functional approaches for predicting land use with
the temporal evolution of coarse resolution remote sensing data. J. Appl. Stat. 30(10):1185
1199
263. Anderson M, Norman J, Diak G, Kustas W, Mecikalski J (1997) A two-source time integrated model for estimating surface fluxes using thermal infrared remote sensing. Remote
264. Guillevic P (1999) Modelisation des bilans radiatif et e nergetique des couverts vegetaux.
Ph.D. thesis, Universite Paul Sabatier Toulouse III, 181pp.
265. Schmugge T, Ogawa K (2006) Validation of Emissivity Estimates from ASTER and MODIS
Data. In: Geoscience and Remote Sensing Symposium, 2006. IGARSS 06. Proceedings.
2006 IEEE International. Vol I, pp 260262
Chapter 11
Spectrally Consistent Pansharpening

Ari Vesteinsson, Henrik Aanaes, Johannes R. Sveinsson,
and Jon Atli Benediktsson
Abstract Several widely used methods have been proposed for fusing high
resolution panchromatic data and lower resolution multi-channel data. However,
many of these methods fail to maintain spectral consistency of the fused high resolution image, which is of high importance to many of the applications based on
satellite data. Additionally, most conventional methods are loosely connected to the
image forming physics of the satellite image, giving these methods an ad hoc feel.
In this chapter, a method for image fusion of satellite images is given. The method
is based on the properties of imaging physics in a statistically meaningful way.
Based on our analysis, it is seen that spectral consistency is a direct consequence
of imaging physics and hence guaranteed by our method. This is achieved while
exploiting the high resolution single-channel data in what can be seen as a statistical
optimal way, yielding a framework to which additional constraints can be added
in a straight forward manner. In this chapter we exploit this framework and add
some simple optimization terms for smoothing the fused image. Specifically, the
method is based on the observation that any given channel of the satellites imaging
device can be seen as an inner-product between the radiated light arriving at the
sensor and the spectral response function of that channel. This gives a simple inner
product space encompassing the relationship between the different channels as well
as imposing spectral consistency. Normal distributed statistics inducing the same
norm as the above mentioned inner product are used for regularization. This yields
a framework to which additional constraints are added in a straight forward manner.
Ari Vesteinsson, Johannes R. Sveinsson, and Jon Atli Benediktsson

Department of Electrical and Computer Engineering,
University of Iceland, Hjardarhaga 2-6, 107 Reykjavik, Iceland
{sveinsso, benedikt}@hi.is
Henrik Aanaes
Informatics and Mathematical Modelling,
Technical University of Denmark, Denmark
{haa}@imm.dtu.dk
293
294
A. Vesteinsson et al.
Here we add a simple term to the discussed framework for smoothing the image,
although more elaborate terms might be preferable. Computationally, we achieve
a solution for the derived objective function via stochastic optimization, i.e., using
the Metropolis algorithm in conjunction with simulated annealing.
Apart from contributing with a novel analysis giving more insight into the
image fusion problem, the method proposed in this chapter has been applied to images from the IKONOS satellite. These experimental results validate the proposed
method.
11.1 Introduction
In this chapter Remote Sensing will be understood to mean taking images from
a great distance. In the modern context remote sensing started with the aerial photographs, originally taken from balloons and kites, but later from airplanes and later
still from satellites. Remote sensing satellites are usually on sun-synchronous orbits
(as opposed to communication satellites which are usually geostationary) but differ
in the layout of their orbits, resolution of their sensors and their spectral sensitivity
(i.e., the width and number of their frequency-bands).
The first photographs were black and white (gray-scale), in the 1940s the color
photograph was invented and later it became possible to record electromagnetic
radiation outside the visible spectrum. In order to take color images1 the satellites photosensors restrict the incoming radiation with band selective filters, filtering through the different frequency components of the spectrum: Red, green and
blue. These images are called multispectral images, as opposed to panchromatic images which cover the whole visible spectrum, and hyperspectral images which cover
many more frequency bands.
Image fusion is the subset of data fusion dealing with merging images, data fusion has been defined by Wald [1] as
a formal framework in which are expressed means and tools for the alliance of data originating from different sources. It aims at obtaining information of greater quality; the exact
definition of greater quality will depend upon the application.
In the context of image fusion different sources simply means multiple images
and any other a priori information that is available. The most common meaning
of quality in image fusion has been visual improvement, i.e., taking into account
how the human perceives the fusion product. Another frequently used meaning of
greater quality is improved classification accuracy, in particular automated classification. Many other objectives have also been identified, see [2, 3].
Image fusion can be done on several levels: Pixel level, feature level, object level
and decision level, depending on the intended use of the fused image. Pixel fusion is
1
Note that in this chapter the terms color image and multispectral image will be used interchangeably.
11 Spectrally Consistent Pansharpening
295
the combination of images pixel-by-pixel, feature fusion is the merging of features

that have been extracted from the image (e.g., long, bright, square), object fusion is
used for finding objects in the images (e.g., roads, lights, houses) and decision fusion
is the use of multiple images in order to make a decision (e.g., go to house). This
chapter is only concerned with pixel level fusion and when the terms image fusion
or fusion are used below, pixel level fusion is intended. In the current context,
fusion is the next step after preprocessing and the step before further processing,
such as segmentation and classification.
Historically, a great number of methods have been utilized in image fusion. Pohl
et al. [2] provides a good overview of many such methods. There are many ways
to categorize image fusion methods. In this chapter four categories will be looked
at: Frequency methods, color transformation methods, statistical methods, and hierarchial methods. Frequency methods split the images into frequency parts and then
join frequency parts from different images together. Color transformation methods
transform the images to a new coordinate system which is more easily comprehended in terms of the human visual experience, and then combine those parts that
enhance this experience. Statistical methods use the correlation between features
of the image and the fact that some features are more dominant than others. Finally, hierarchial or scale space methods, create a pyramid of scaled down images
and traverse this pyramid in reverse to scale up images, thus, in effect, fusing the
images.
Where applicable, the focus of this chapter will be on pansharpening rather than
on image fusion in general.
Frequency Methods: The idea behind the frequency methods is that the high
frequency part of panchromatic image contains the spatial details of the image
while the (low) frequency part of the multispectral image contains its spectral
details. By combining the two frequency parts an image containing details and
color can be created. This idea is heuristically supported by the fact that a low
resolution image can be considered to have been created by the averaging (Low
Pass Filtering (LPF)) of a high resolution image. As a consequence of the averaging process details of the high resolution image become less pronounced or
smoothed out.
Belonging to this category are methods like High Pass Filtering (HPF), High Frequency Modulation (HFM) and the Modified Brovey Transform (MBT). HPF can
be subdivided into High Pass Filtering Substitution (HPFS) and High Pass Filtering Addition (HPFA). In both cases the high-frequency part of the panchromatic
and the low-frequency part of the multispectral image are isolated. Then, for HPFS,
the fused image is constructed by adding the two isolated parts together. On the
other hand, for HPFA, the high-frequency component of the panchromatic image is
added to the high-frequency component of the multispectral image before combining the two parts. The high and low pass operations can either be achieved by kernel
methods, e.g., Gaussian or Laplacian kernels, or by using the Fast Fourier Transform (FFT). HPFS and HPFA are commonly used to evaluate new fusion methods,
see, e.g., [4].
296
With HFM, each low resolution channel is boosted with the ratio of the highfrequency to low-frequency filtered versions of the high resolution panchromatic
image.
The MBT is similar to the HFM, but instead of fusing each channel separately, the
intensity component of the color image is fused with the high resolution panchromatic image in MBT. Actually, this is a product of combining the frequency and
color transformation methods, which will be discussed shortly.
For the frequency category of methods, it is important to find the best cut-off
frequency for the high and low pass filters. One way of accomplishing this is by
using an entropy measure. However, due to the imperfections of the filters, such as
the windowing effect, this filtering will not be perfect. Also, there is an inherent
problem due to the difference in resolutions, i.e., the frequency spectrum of the
lower resolution image will contain aliasing components that are not present in the
higher resolution image. Additionally, these methods are very sensitive to any high
frequency noise that could be present in the panchromatic image.
Color Transformation Methods: The Intensity-Hue-Saturation (IHS) method and
its close relatives belong to the color transformation methods. Since it was first
applied to image fusion, by Carper et al. [5], the IHS method has been the most
popular fusion technique. This is mainly due to how simply it can be implemented,
how fast it can be executed, and how visually pleasing the fusion result is.
A color image is natively represented by three color channels: The red, green and
blue channels (i.e., RGB). The IHS methods describes how to transform these three
channels into a new coordinate system, where a coordinate is described by three
values: The intensity, hue and saturation. The intensity measures the strength of
the color, the hue is a measure of the dominant wavelength in the spectrum, and the
saturation is a measure of the purity of the color. The idea of the IHS methods is to
substitute the intensity component with the panchromatic image and transform the
result back into the RGB representation, producing a high resolution color (RGB)
image.
There exist several algorithms for transforming a color image from the RGB
color space to the IHS color space, not all giving the same results. In fact, IHS
fusion sometimes goes under different names that reflect how the transformation
is done, e.g., Hue-Saturation-Value2 (HSV) and Hue-Saturation-Lightness3 (HSL).
Two IHS algorithms are given by
1
1
1
I
R
3
3
3
2 2 2 2
v1 = 6
6
6 G
1
1
v2
B
0
2
2

v2
H = arctan
and S = v12 + v22 ,
v1
2
3
Value is defined as max(R, G, B).

Lightness is defined as (max(R, G, B) min(R, G, B))/2.
(11.1)
297
and
R+G+B
% 3
arccos(a)
if G R
H=
2 arccos(a) if G < R
I=
a=
(2B G R)/2
(11.2)
(B G)2 + (B R)(G R)
3 min(R, G, B)
.
S = 1
R+G+B
The major shortcoming of the IHS method is that it produces color distortion in the
fused image. In [6] it is demonstrated how this color distortion results in an incorrect
saturation value. There it is in fact shown that the distortion at a given pixel is
proportional to the difference of the intensity values of the images. Zhang [7] points
out that for IKONOS images, the panchromatic spectrum extends into the infra-red.
Therefore, the difference between the intensity values is even greater than in the
fusion of earlier Landsat and SPOT images. Furthermore, the color degradation is
even more pronounced.
To counteract the differences in the intensity values, histogram equalization
methods have been used. These methods bring the intensities closer together on
the average. Potentially, on the other hand, they leave out some distortion at the
individual pixels.
Statistical Methods: One example of a statistical method is the Principle Components Substitution (PCS) method. It applies principle component analysis to the
multivariate image data, transforming them into a set of uncorrelated components.
The PCS method is similar to the IHS method, except that in the PCS, the first principle component is replaced by the intensity component of the panchromatic image.
Then, by reverting to a RGB representation, a fused high resolution color image is
created. Shettigara was amongst the first to use the PCS method for image fusion
[8]. Zhang [7] explains that the problem with the PCS method is that the first principal component usually has the greatest variance. Therefore, the first component
dominates in the fused image, possibly resulting in color distortion.
Hierarchial Methods: The common idea for the hierarchial methods is to reduce
stepwise the resolution of the panchromatic image, reaching the low resolution of
the multispectral image. Then, this path is retracted with the multispectral image,
producing a high resolution multispectral image. There are several ways to perform
the reduction, e.g., using Laplacian and Gaussian kernels, morphological operators,
the wavelet transform, or the curvelet transform. The current focus in data fusion
algorithms seems to be concentrated on the hierarchical family of methods [9], and
the wavelet transform in particular.
The main drawback of the hierarchical methods is the assumption that the spectral content of the panchromatic and multispectral images coincide exactly, i.e., the
same path would have been generated with an original high resolution multispectral image. This is not necessarily always the case.
298
Other Fusion Methods: Many other fusion methods exist, e.g., the Color Normalized transform (CN), the Spectral Balance Preserving fusion (SPB), and more.
All the methods that have been discussed above are derived from a certain perspective of what an image is, rather than serving a given objective. In this chapter,
these two approaches will be combined, i.e., an algorithm will be derived from how
an image is formed and then improved to suite a given objective.
In this chapter we present a fusion algorithm, called SCP (Spectrally Consistent
Pansharpening), which utilizes the known spectral responses of the photosensors to
fuse multispectral and panchromatic satellite images at the pixel level into a highresolution multispectral image. The intention is to create an image using as much
information as feasible from both all the available images and all the available a
priori knowledge. The created image maintains spectral consistency with the original multispectral image. Then, a more elaborate model for the imaging physics is
introduced and a fusion framework is developed in which it is possible to, simultaneously, keep spectral consistency and smooth the final image. It is noteworthy that the
smoothing will be used as a constraint, but any other constraint or constraints could
have been selected. Scene smoothness, however, is a common assumption in remote
sensing. Finally, the fusion problem is modeled as a minimization problem where
the objective function is defined as some energy functions (defined in Section 11.3).
The SCP method is tested in experiments by fusing a low resolution hyperspectral image with a high resolution panchromatic image. The chapter is organized as
follows. In Section 11.2, the spectrally consistent pansharpening method is introduced. A fusion framework in which it is possible to, simultaneously, keep spectral
consistency and smooth the final image is introduced in Section 11.3. Experimental
results are given in Section 11.4, and conclusions drawn in Section 11.5.
11.2 Spectrally Consistent Pansharpening (SCP)

Spectral consistency of a high resolution multispectral image with respect to a low
resolution multispectral image is the requirement that the integration of the spectral
density over the same geographical area gives the same spectral value. This amounts
to saying that the sum of pixel values over the same geographical area is the same
for both images. Put yet another way: down-sampling the high resolution image, by
averaging, down to the resolution of the low resolution image should give an image
identical to the low resolution image. The maintenance of spectral consistency between the fused image and the original multispectral image is of great importance
for many applications. For example, it is very important in thematic classification,
where objects are being classified on the bases of their spectral values, both relatively within the image, as well as absolutely with respect to some know spectral
signatures. Maintaining spectral consistency is a non-trivial task and most conventional fusion algorithms fail in doing so.
It is important to base the fusion of images on solid mathematical and physical
foundations. In this chapter, a model of the image formation process will be derived.
299
q*F
*
A
w
Fig. 11.1 Construction of pixel value
This model will help to understand how the observation channels can be combined
to produce a fused image of full spectral integrity with respect to the original color
channels. In deriving the image formation model, it will be assumed to be noise free
and deterministic.
11.2.1 Image Formation Model

A sensors observation is captured to a digital media in the form of a set of pixels, each having a given value (often referred to as a Digital Number (DN)). The
number of stored pixels depends on the resolution of the sensors. Furthermore, the
value of the pixel depends upon the amount of light that comes through the sensor.
The output of the sensor due to the accumulated incident spectral energy can be expressed in terms of the spectral energy density, q, and the sensors spectral response,
F, integrated over the pixel area, A, and all wavelengths, w, as
&& &
A w
F q = pixel value,
(11.3)
see Figure 11.1 [3]. It is reasonable to assume that the sensors spectral response is
constant for all the pixels. Therefore, (11.3) can equally be expressed as
&
&
&
q=
F
A
F S = pixel value,
(11.4)
where S is the total spectral energy density for A.
11.2.2 Spectral Inner Product Space (W)

Considering functional analysis, it is seen that the integral in (11.4) defines an inner
product between a spectral response and the spectral density, i.e.4
4
For readers not familiar with functional analysis, consider the discrete version/approximation of
the spectral response and the spectral density. In this case the integral is substituted by a summation,
and (11.4) reduces to the standard inner product between two real vectors as know from linear
algebra.
300
F, S =
&
w
F S.
(11.5)
In this context, the spectral responses for the red (R), green (G), blue (B), nearinfrared (NIR) and panchromatic (P) channels along side the spectral density (S) can
be thought of as vectors in an infinite dimensional inner product space, W. Please
note that, in this setting there is also no difference between the spectral density and
the spectral responses, hence, e.g., the inner product
R, G =
&
w
R G,
(11.6)
between the spectral responses for the red and the green channel makes perfect
sense.
As seen from the above the pixel value of a given pixel and a given channel,
can be expressed as an inner product, i.e., F, S. This, however, implies that even
though S is infinite dimensional, we can only observe the parts of S spanned by the
spectral response functions. That is,
Wobs = span{R, G, B, NIR, P},
(11.7)
with Wobs denoting the observable part of W. Hence for the purpose at hand we
can just as well operate in the four-dimensional space Wobs , which is a subset of
W. The interesting here is that the inner product defined in (11.4) on elements in W
induces an inner product on Wobs . Hence from the above we can calculate the inner
product between the different spectral response functions, which intuitively tells us
how much information one channel has about another.
1. Projections in W: The problem at hand is to find out how the high-resolution
panchromatic measurements (P) can be transformed into high-resolution RGBmeasurements. Figure 11.2 shows how the spectral density Si of a pixel i projects
FP
P1
PL
SP1
P2
SR
SP2
R2
R1
FR
Fig. 11.2 Projections resulting from the assumption that SP1 and SP2 lie as close as possible to SR
301
onto the spectral response function of the panchromatic channel FP giving the value
Pi of pixel i in the panchromatic image. In the same way the spectral density projects
into R, G, and B measurements of the corresponding spectral channels. Note that
when Pi is known all that can be said about Si is that it lies in a hyperplane perpendicular to Fi , intersecting Fi at the point Pi .
In order to derive the relationship between panchromatic measurements and R,
G, and B measurements, it is simplest to start with a high to low resolution ratio of
2 and focus on one of the spectral channels, for instance, the red. Figure 11.2 shows
three spectral densities, SR , SP1 and SP2 , and their corresponding measurements R,
P1, and P2, respectively. The relationship between SR and {SP1 , SP2 } is simply
SR = SP1 + SP2 ,
(11.8)
where linearity of the sensors has been assumed. As a side note: The extension of
this idea to a ratio of M high resolution pixels for each low resolution pixel is simply
M
SR = SPi .
i=1
Now, it is important to remember that the measurements have been normalized

and it can be assumed that the normalizing constant for the high-resolution image
is twice that of the low-resolution image (given the 2-to-1 resolution ratio). So the
normalized version of (11.8) is
SR =
SP1 + SP2
.
2
(11.9)
From the above, the possible spectral densities corresponding to a measurement

have been restricted to a hyperplane in W. This, however, does not fully specify
where the true spectral density lies in W, for that additional knowledge is needed.
2. Problem Regularization: In order to extract a unique solution out of the set of
possible solutions for three given measurements (P1, P2, and R), a priori assumptions are added to the problem specification. One possible such a priori assumption
is that the distance from SP1 and SP2 to SR should be as small as possible. The
justification for this assumption is that if the spectrum of the neighbors of the low
resolution pixel can be assumed related by the normal distribution, then it is most
likely that the high resolution pixels lie as close as possible to the low resolution
pixel. This can also be called a smoothness assumption. This assumption results
in the two intersections shown with closed-circles in Figure 11.2. The projection
of these two intersections onto FR gives the estimated high resolution R values, R1
and R2 (see Figure 11.2). The assumption amounts to the same as saying that the
difference in the spectral values lies only in their intensity component (not hue or
saturation).
3. Fusion Formula and Results: Using the assumption displayed in Figure 11.2,
the high resolution R values can be calculated by
302
R1 = R + (P1 PL ) cos( )
R2 = R + (P2 PL ) cos( ),
(11.10)
where the cos( ) is calculated from the inner-product, using

FP , FR = |FP | |FR | cos

FP , FR
= arccos
|FP | |FR |
'
FP FR dw
= arccos
.
|FP | |FR |
(11.11)
11.3 SCP Fusion Framework

This section introduces a more elaborate model for the imaging physics as well as
developing a fusion framework in which it is possible to, simultaneously, keep spectral consistency and smooth the final image. Note that here smoothing will be used
as an constraint, but any other constraint or constraints could have been selected;
scene smoothness, however, is common assumption in remote sensing. The fusion
problem is modeled as a minimization problem where the objective function is the
energy of a Gibbs distribution [11]. In this implementation the energy consists of
three energy terms: One for the spectral consistency constraint, one for the smoothness constraint, and one for the image formation, or imaging physics (IP) constraint.
The minimization of the objective function is sought using stochastic optimization.
11.3.1 Objective Function

The Gibbs distribution, P( f ), is defined as
1
P( f ) = Z 1 exp( U( f )),
T
(11.12)
where T is called the temperature, Z is a normalizing constant, called the partition

function:
1
Z = exp( U( f )),
(11.13)
T
f F
and U( f ) is called the energy function, defined by
U( f ) =
Vc ( f ).
cC
(11.14)
303
aspectral
Pspectral
asmooth
1/T
Psmooth
aIP
PIP
Fig. 11.3 Calculating the energy function
over the set C. As mentioned above, the energy function is specified in terms of three
energies, namely, Pspectral , Psmooth , and PIP , i.e.,
U( f ) = spectral Pspectral ( f ) + smooth Psmooth ( f ) + IP PIP ( f ),
(11.15)
where spectral , smooth , and IP are the weights for the respective energy functions, chosen such that energy function is minimized. Figure 11.3 displays the above
schematically. The derivation and application of these three energies is the subject
of the following subsections.
1. Spectral Constraint: The reason for applying the spectral constraint is to
enforce the spectral consistency between the original low resolution multispectral
image and the new, constructed, high resolution image. Pspectral ( f ) measures the difference between the average of the pixels in the fused image and the corresponding
pixel in the original low resolution multispectral image, i.e.,
Pspectral ({CH CL }) = (CL
1
CH )2 ,
n2 {C
C }
H
(11.16)
where CL is the low resolution pixel, the set {CH CL } contains the corresponding
high resolution pixels, and n2 is the number of elements in that set, i.e., the resolution
ratio between the high and low resolution images is n.
2. Smoothness Constraint (SP): The smoothness constraint is applied to enforce
that the new constructed image is smooth. This is to counteract the blockyness remaining in the image after the SCP fusion. The SP is defined by the smoothness
energy, Psmooth , which is given as
Psmooth ( fi ) =
1
( fi fi )2
k i
N
i S,
(11.17)
where fi is the value of the pixel i in the fused image, Ni is the neighborhood for
pixel i, k is the number of pixels in the neighborhood, and S contains all the images
304
pixels. Here a simple 4-neighborhood will be used, i.e., k = 4. On the other hand,
the Euclidian three-dimensional distance is used as a measure of the distance measure between fi and fi . The larger objects should become smoothed without loss of
detail.
3. Imaging Physics (IP) Constraint: By using (1) it can been shown that the IP
constraint is the extension of the theory presented earlier. Consider the subspace of
the frequency space, W, spanned by the red (R), green (G), blue (B), and panchromatic (P) spectral values, and denote the mean of the P-channel for a corresponding
low resolution pixel as P . Now, by assuming that the pixels are normally distributed
in the frequency space, with a mean and variance

RL

GL
CP
C
C =
BL , C = T 2 ,
P
CP
P
respectively, where
R2
R G R, G R B R, B

G2
G B G, B ,
C = G R G, R
B R B, R B G B, G
B2
R P R, P
CP = G P G, P ,
B P B, P
x2 is the variance for channel x {R, G, B, P}, and ., . denotes the inner product.
Then the conditional distribution of a high resolution RGB pixel given P, is a normal
distribution with the mean given as

RL
CP
C|P = GL + 2 (P P )
(11.18)
P
B
L
and variance
T
CP CP
.
P2
This distribution gives that the panchromatic energy can be expressed as

C|P = C
PIP ( fi ) =
1
( fi C|P )
log(det(C|P )) + ( fi C|P )T C|P
(11.19)
(11.20)
for all i S. By applying the simplifying assumption that the standard deviations of
all channels are the same, i.e.,
R = G = B = P = ,
then C|P becomes the new pixel value in the SCP fusion algorithm, namely
305
R, P
RL
C|P = GL + G, P (P P ).
B, P
BL
(11.21)
However, contrary to earlier definition, the distributions variance will be utilized

here as well.
4. Selection of a Search Algorithm: The selection of a quadratic potential function
and a Gaussian observation model, guarantees that the a posteriori energy function
is convex and, consequently, has only one extremum. Therefore, in the search for
a global minimum there is no risk of getting trapped in a local minimum. Consequently, there is no need to apply global minimization algorithms. In this case, the
choice of a local minimization algorithm is between the ICM and Metropolis algorithms. For the images that are going to be used, it is known that the value of each
pixel is stored as 11 bits, i.e., a pixel can take 211 = 2,048 possible values. Since the
ICM algorithm tests every possible configuration on each pass, it will test LS labels,
or 1275 million labels in the current case. This immediately rules out the use of
ICM, since it would be impossibly slow to execute. The Metropolis algorithm, on
the other hand, only tests S number of labels on each pass. Therefore, it is chosen in
this case.
11.4 Experimental Results

The data set used in this paper consists of two images of Reykjavik, Iceland acquired
by the IKONOS earth imaging satellite on 9 August 2001. One image is a multispectral image of 4 m resolution (see Figure 11.4a). The other is a panchromatic image
of 1m resolution (see Figure 11.4b).
The panchromatic image is 976-by-640 pixels. It consists of a single channel (i.e.,
an intensity image) and is free from speckle noise. From the histogram of the image,
it is evident that there is some compression of a curtail around 0 and a slight saturation in the intensity values, indicating that the normalization procedure included
compressing the data inside the available range. Individual cars are distinguishable
on the road and in parking lots. The multispectral image is 244-by-160 pixels. It
consists of the R, G and B channels, i.e., red, green and blue, and it is also free
from speckle noise. There is a slight peak in the intensity values at 0 suggesting a
compression of a curtail. Most individual cars on the road are distinguishable. On
the other hand, that is often not the case for cars in parking lots. Visually, some adjacent houses are more easily distinguished on the multispectral images than on the
panchromatic one, which is due to the distinct colors of their rooftops.
IKONOS can produce imagery of the same geographical area every 3 days. Its
spectral range covers the visible band into the infrared band, see Table 11.1. Notice
that the spectral range of the panchromatic channel is significantly larger than the
combined spectral range for the red (R), green (G), blue (B) and near-infrared (NIR)
channels, see Table 11.1 and Figures 11.5 and 11.6.
The SCP algorithm was tested and compared to the IHS method [5]. Results are
shown in Figure 11.7 and Table 11.2.
306
Fig. 11.4 (a) The multispectral image. (b) The panchromatic image
Table 11.1 IKONOSs spectral bands

1 m black-and-white
Panchromatic
0.450.90 m.
4 m multispectral
Blue
Green
Red
Near-infrared
0.450.52 m
0.510.60 m
0.630.70 m
0.760.85 m
The SCP fusion framework algorithm was used on the data. First, the SCP algorithm was performed to get the initial image for the optimization problem, shown in
Figure 11.7. The results after 500 iterations using the spectral, IP, and smoothness
constraints and a fixed non-zero temperature for a 256-by-256 copped out area is
shown in Figure 11.8 and numerically results are given in Table 11.3. The energies
for each constraint for this experiment are given in Figure 11.9.
Comparing the results from all the experiments with the SCP fusion framework
shows that none of them actually produce better result than the SCP algorithm, at
307
Fig. 11.5 IKONOS multispectral sensors spectral responses
Fig. 11.6 IKONOS panchromatic sensors spectral response
least not with respect to the metrics used. Two conclusions can be drawn from this
fact. First, that a new minima of the objective function was not found, and perhaps
did not exist. Second, this could simply imply that the smoothness constraint (the
additional ingredient) was not a good a priori assumption about the ground truth,
eventhough it did produce a more visually pleasing image. However, all of the experiments did compare favorably to the results of the IHS fusion algorithm.
308
Fig. 11.7 The SCP fusion image, which is also used as the initial image in the SCP framework
algorithm
Table 11.2 Quality metrics calculated for individual channels, fused image down-sampled
RMSE
Correlation
SCP IHS SCP

Red
0.006 0.029 1.000
Green 0.006 0.037 1.000
Blue 0.004 0.034 1.000
IHS
0.955
0.930
0.886
309
Fig. 11.8 The fused image

after 500 iterations using the
spectral, IP, and smoothness
constraints and a fixed nonzero temperature
Table 11.3 Quality metrics for the image using smoothness, spectral and IP constraints using a
fixed non-zero temperature
Metric
Result
Correlation with original red

Correlation with original green
Correlation with original blue
Root-mean-squared-difference with original red
Root-mean-squared-difference with original green
Root-mean-squared-difference with original blue
Correlation with original panchromatic
0.997
0.996
0.995
0.008
0.009
0.008
0.879
150
IP
Spectral
Smooth
energy
100
50
50
100
150
200
250
300
350
400
450
500
iteration
Fig. 11.9 After 500 iterations using the spectral, IP, and smoothness constraints and a fixed nonzero temperature
310
11.5 Conclusion
The underlying objective of this chapter has been to approximate an image produced
by multispectral sensors with the characteristics (spectral response) of the original
sensors and the resolution of a panchromatic sensor. What has been achieved is an
image with the spectral integrity of the original color image intact and with some details of the panchromatic image included. Unfortunately, the resulting image from
the SCP method is blocky. However, that is as good as it gets using the available
raw information. In order to further improve the fused image, additional knowledge (a priori) has to be introduced. By introducing this information into the fusion
process, it might no longer be possible to guarantee spectral consistency with respect
to the original color image. However, given that the new information is correct, the
spectral correctness with respect to the ground truth should in fact have increased.
This introduction of new information leads to a construction of SCP fusion framework where multiple constraints can be combined together to make up the objective
function for the fusion problem. The underlying algorithm uses known spectral response functions to get a good initial image, as well as utilizing them to calculate the
IP constraint. Furthermore, the algorithm employs stochastic optimization to locate
a minimum of the objective function. The alpha version of the SCP fusion framework was shown to be promising for combining both data and a priori assumptions.
The Focus of our future work in SCP and SCP fusion framework could be to make
the framework more generally useful by for example automatically determining the
model parameters and including the weights for the objective functions. Possible
solutions could use the resulting classification accuracy as an input to the parameter
estimation procedure.
Acknowledgements The work was partially supported by the Icelandic Research Council and the
Research Fund of the University of Iceland.
References
1. Wald L (1999) Some terms of reference in data fusion. IEEE Trans. Geosci. Remote Sens.
37(3):1901193
2. Phol C, Genderen JLV (1998) Multisensor image fusion in remote sensing: concepts, methods
and applications. Int. J. Remote Sens. 19(5):823854
3. Munechika CK, Warnick JS, Salvaggio C, Schott JR (1993) Resolution enhancement of multispectral image data to improve classification accuracy. Photogramm. Eng. Remote Sens.
59(1):6772.
4. Tsai VJD (2003) Frequency-based fusion of multiresolution images. Geoscience and Remote
Sensing Symposium, July 2003, pp. 3665 3667
5. Carpter W, Lillesand T, Kiefer R. The use of intensity-hue-saturation transformations for
merging SPOT panchromatic and multispectral image data. Photogramm. Eng. Remote Sens.
56(4):459467
6. Tu T-M, Su S-C, Shyu H-C, Huang PS (2001) A new look at IHS-like image fusion methods.
Inform. Fusion 3(2):177186
311
7. Zhang Y (2002) Problems in the fusion of commercial high-resolution satellite as well as

Landsat 7 images and initial solutions. Proceedings of the ISPRS Technical Commission IV
Symposium 2002
8. Shettigara VK (1992) A generalized component substitution technique for spatial enhancement of multispectral images using a higher resolution data set. Photogramm. Eng. Remote
Sens. 58(5):561567
9. Schowengerdt RA (1997) Remote sensing models and methods for image processing, 2nd ed.
Academic Press, San Diego, CA
10. Space Imaging. IKONOS [online]. Available at: http://www.spaceimaging. com/products/
ikonos/
11. Geman S, Geman D (1984) Stochastic relaxation, Gibbs distributions, and Bayesian restoration of images. IEEE Trans. Pattern Anal. Machine Intell. PAMI-6(6):721741
12. Vesteinsson A, Sveinsson JR, Aanes H, Benediktsson JA (2005) Spectral consistent satellite
image fusion: using a high resolution panchromatic and low resolution multi-spectral images.
IEEE International Geoscience and Remote Sensing Symposium (IGARSS05), vol 4, Seoul,
Korea, 2529 July 2005, pp 28342837
Chapter 12
Data Assimilation Methods for Land Surface

Variable Estimation
Shunlin Liang and Jun Qin
Abstract Estimating land surface variables from remote sensing data is an ill-posed
problem. Integration of observations from multiple satellite sensors with different
spectral, spatial, temporal and angular signatures is now an important research frontier. Data assimilation (DA), integrating not only remotely sensed data products,
but also other measurements and land dynamic models, is an advanced set of techniques for innovative parameter estimation. After a brief introduction, we describe
the basic principles of DA, and then provide in-depth discussions of some relevant
issues while using DA. The latest applications of DA for estimation of soil moisture,
energy balance, carbon cycle and agricultural productivity are summarized.
12.1 Introduction
Despite the abundance and variety of remote sensing measurements, land surface
characterization from satellite observations is still very challenging. There are multiple sources of surface information, such as remote sensing data and derived products, in situ measurements, and land surface model outputs. Innovative techniques
are needed to merge these information sources and optimize the use of satellite
measurements for robust surface products and greater predictability. Data assimilation (DA) is a mathematical approach that enables use of all available information
within a given time window to estimate various unknowns. The information that
can be incorporated includes observational data, existing a priori information, and,
very importantly, a dynamic model that describes our system and encapsulates current theoretical understanding. The model brings consistency to the observational
Shunlin Liang
Department of Geography, University of Maryland, College Park, USA
sliang@umd.edu
Jun Qin
Institute for Geographical Science and Natural Resource Research, Beijing, China
313
314
S. Liang, J. Qin
data, and interpolates or extrapolates data into data-devoid regions in space and
time. The observational data, representing the actual state of the system, corrects
the trajectory of the imperfect model through adjusting model parameters. DA is
also called modeldata synthesis or datamodel fusion in different disciplines.
The benefits of DA for maximizing the scientific and economic value of remote
sensing data is summarized as follows (ONeill et al., 2004; Raupach et al., 2005):
1. Forecasting and error tracking. By regularly comparing forecasts with observations, extremely valuable error statistics can be built up, which in turn can be used
to improve the quality of the observations (e.g., by revealing biases in instrument
calibration) as well as the quality of the models. It serves as a model testing and
data quality control procedure.
2. Combining multiple data sources. Different observing systems (both in situ and
remote sensing data) have varying virtues and deficiencies. Such variety can
be preferentially exploited or contrasted to optimize the value of the resulting
data set.
3. Interpolating spatially and temporally sparse observations. The model provides
a way to propagate information in a consistent manner in space and time from
data-rich regions to data-poor regions. This capability is vital to successfully utilize satellite observations, which due to limited and sequential sampling provide
only an incomplete picture of the Earth. DA fills in missing pieces to achieve
a full global picture.
4. Inferring, from available observations, quantities not directly observable.
Through relationships expressed in the models governing equations, measured
parameters convey knowledge of those that are inadequately measured or completely lacking. For example, soil moisture vertical profile can be inferred from
the surface skin temperature or surface top-layer soil moisture content.
5. Forecasting. Predicting forward in time on the basis of past and current observations.
6. Designing observing systems. Decisions to deploy new satellite-borne instruments require critical assessment of the incremental value or benefit of the data
to be acquired by the new sensors. With careful design, DA experiments provide an objective, quantitative way to contribute to such assessment. In addition,
DA can optimize the sampling pattern from an observing system, and can target
observations to capture features of concern, such as a rapidly developing storm.
DA methods exploit data streams not only to validate model outputs, or directly to
infer fluxes, but principally to constrain internal model parameters to optimize values (i.e., parameter estimation). Different data sets constrain different components
of a model which is able to assimilate data across a range of space and time scales.
Another distinctive characteristic of DA is that uncertainties associated with the
observation, techniques, processing, representation, and accuracy are as important
in determining the final outcome as the measured values themselves. Thus, uncertainty estimates, for both measurements and model parameters, take on even greater
importance (Canadell et al., 2004).
The meteorological and oceanographic communities have been at the forefront in developing and using DA methods. In recent years, meteorologists and
12 Data Assimilation Methods for Land Surface Variable Estimation
315
oceanographers have tended to view DA as a model state estimation problem. The

land community has aggressively endeavored to catch up and apply DA methods in
recent years, and some example applications are shown in Section 12.4.
The remainder of this chapter presents a brief introduction to the basic principles
of land DA as this is a relatively new field. Following the introduction is a discussion
of some critical issues pertinent to land DA. Section 12.4 introduces recent DA
applications in several major disciplines.
12.2 Principles of Land DA

In this chapter, land DA is considered as essentially an estimation problem: that is
to acquire optimal estimates of model state variables of the land surface and its parameters given a set of remote sensing products, a land surface process model, and
any available a priori information. Various land DA schemes have different characteristics, but they may have the following common features: (1) a forward land
dynamic model that describes the time evolution of state variables such as surface
temperature, soil moisture and carbon stocks; (2) an observation model that relates
the model estimates of state variables to satellite observations and vice versa; (3) an
objective function that combines model estimates and observations along with any
associated prior information and error structure; (4) an optimization scheme that adjusts forward model parameters or state variables to minimize the discrepancy between model estimates and satellite observations; and (5) error matrices that specify
the uncertainty of the observations, model and any background information (these
are usually included in the objective function).
12.2.1 Dynamic Model

Land surface process models are often structured as a discrete-time nonlinear statespace model with additive noise as:
xt+1 = f (xt , ut , ) + wt
(12.1)
where x denotes the model state vector (e.g., soil moisture), u the external forcing
data (e.g., meteorological data), the model parameter vector (e.g., soil texture), wt
the model noise, and f () the model operator mapping the previous state xt to the
next state xt+1 . The differences among various land surface models are reflected in
different specifications of x, u and f ().
12.2.2 Measurement Process

DA methods assimilate the measured values into a dynamic model, so the measured quantities have to be linked with model variables. The measurement process
316
S. Liang, J. Qin
is a mathematical model that relates the model state vector (xt ) to the observation
vector (yt ):
(12.2)
yt = h(xt , ) + et
where et the observation noise, h() the observation operator, and is the parameter
set of observation operator. If remote sensing data products are used as the observations, the remote sensing reflectance, or emittance model of land surfaces, and
the coupled land and atmosphere system are the observation operator and is its
parameters.
12.2.3 Objective Function

The underlying principle of the DA method is to estimate the parameters and variables of the dynamic models by minimizing the differences between the predictions
of the model and the assimilated data products. The difference is often characterized
as the objective (or cost) function, and the common measure is the least squares.
DA methods usually attempt to minimize the objective function J
1
1
J(x) = (x xb )T B1 (x xb ) + (H(x) y)T R1 (H(x) y) = Jb + Jo
2
2
(12.3)
where y is the observation vector, x is the extended model state variables, xb is the
background field (or first guess), H is the model operator, and R is the observationerror covariance matrix and B is background-error covariance matrix.
The DA problem now becomes: vary x to minimize J(x), subject to the constraint
that the state variables must satisfy the dynamic model. The value of x at the minimum is the a posteriori estimate of x, including information from the observations
as well as the background. In Eq. (12.3), the first term Jb is to force the optimal
parameters as closely as possible to background fields, and the second term Jo is to
adjust parameters so that model outputs will be as close to the observations as possible. Specifying R and B depends on the relative accuracy of background information
and remote sensing data products. In extreme cases, if the errors of the first-guess
values are extremely large, the final estimates will be decided from the fitting of the
observations and will be close to the first-guess values.
12.2.4 Assimilation Algorithm

Since land DA is considered to be an estimation problem, an assimilation or estimation algorithm is needed to estimate model parameters or states by assimilating
observations into the dynamic model. This process provides three kinds of output:
optimal estimates for the model properties to be adjusted, uncertainty statements
about these estimates, and an assessment of how well the model fits the data, given
317
the data uncertainties. There are two types of assimilation algorithms popular in
current research and applications, namely cost function-based methods and sequential methods. These two methods have many different forms and will be discussed
in Section 12.3.5.
12.2.5 Error Model

Both dynamic models and measurements are not perfect and contain uncertainties
to some degree. These uncertainties or errors have to be characterized in the DA
system. Consequently statistical properties of wt in Eq. (12.1) and et in Eq. (12.2)
or R and B in Eq. (12.3) need to be estimated. The error terms in the dynamic and
observation models can, in principle, be quite general in form, including biases,
drifts, temporal correlations, extreme outliers and so on. Many extant methods
assume these error terms (e.g., wt and et ) have Gaussian distributions with zero
mean and no temporal correlation. Detailed discussions of observational errors are
given in Section 12.3.6.
12.3 Critical Issues in Land DA

12.3.1 Dynamic Models
Land DA uses land surface process models that describe the exchanges of momentum, energy and mass between soil, vegetation and atmosphere. These models, often
called soilvegetationatmosphere transfer (SVAT) models, have similar structures,
whether land surface, hydrological, ecological or crop growth models.
How to select an appropriate dynamic model is largely related to the objectives
of the DA study. For the same objective, multiple models might be available. For example, if we aim to estimate land surface energy balance components such as latent
and sensible heat fluxes by assimilating remotely sensed land surface temperature
(LST) products, almost any land surface process models may be appropriate. However, some models are rather complicated since they may describe other processes
such as those in the carbon cycle in addition to water and energy budgets and thus
contain more model parameters. Often a good strategy is to use a relatively simple
land surface model if possible.
A more general and important issue is the uncertainty of the model. All mathematical models are abstractions and approximations of a complex physical environment. When dealing with ecosystem models, for example, keep in mind that their
designs are not completely derived from rigorous natural laws and hence the mathematical descriptions of the biological processes are not universal. Thus, we are
uncertain not only about values of the numerous model parameters, but also about
the model parameterizations and errors.
318
S. Liang, J. Qin
DA aims to incorporate measured observations into a dynamic system model to

produce accurate estimates of the systems current (and future) state variables. The
neglected model uncertainties are interpreted, in the course of strongly constraining
parameter estimation, as variations in model parameters and this may result either
in unrealistic parameter estimates or in solutions deviating far from the data. In
reality, the model does not exactly reproduce system behavior. Significant errors can
arise due to a lack of resolution, and inaccuracies in physical parameters, boundary
conditions and forcing terms.
Many studies have been reported to address these issues in terms of weak constraints (Liaqat et al., 2003a; Losa et al., 2004; Natvik et al., 2001; Qin et al., 2007b).
As a rule, weakly constrained DA has been used for state estimation. However, if
the model parameters are poorly known, state estimation with fixed model parameters can produce unacceptable results. Weakly constrained parameter estimation
results from a combination of the adjoint method and the generalized inversion; consequently it can take into account errors inherent in the model and data, as well as
find optimal values of the poorly known model parameters. In addition, weakly constrained DA makes it possible to derive valuable supplementary information about
the model itself. This information is completely lost when the strong-constraint
scheme is used.
Chepurin et al. (2005) recently summarized four solutions to correct biases due to
model error in the meteorological community. The most straightforward approaches
to handling time-mean bias involve computing climatologies, and then introducing correction terms into the equations of motion. The second type of approach is
to correct for rapidly changing bias in a data-rich environment. It involves examination of the previous few updating cycles for a systematic trend, which is then
corrected. A third class of approaches, useful in linear one-dimensional problems,
involves pre-whitening the errors so that their frequency spectrum resembles white
noise. A fourth class of approaches is referred to as two-stage estimation. The
two-stage estimation algorithm begins with the assumption that a reasonable estimate of the bias may be made prior to estimating the state of the system itself, thus
allowing the estimation procedures for bias and state to be carried out successively.
12.3.2 Observation Vector

Which remotely sensed data products and other measurements are the most valuable for DA? There are two major schemes for assimilating remotely sensed data
products into dynamic models (Liang, 2004). The first approach assimilates the derived high-level land products (e.g., leaf area index (LAI), fractional photosynthetically active radiation absorbed by green vegetation (FPAR), LST, gross primary
production (GPP), biomass) from remote sensing observations (see Fig. 12.1). It is
equivalent to simplification of measurement process through static remote sensing
inversion. So the observation vector is directly the components of the model state
and h() is simplified as a matrix H with non-diagonal elements equal to 0:
319
Fig. 12.1 Illustration of assimilating remotely sensed high-level products into a land dynamic
model
Fig. 12.2 Illustration of assimilating the direct remote sensing observations to the coupled radiative
transfer and land surface dynamic model
yt = Hxt + et
(12.4)
The second approach assimilates the direct observations (radiance, reflectance,

brightness temperature, or vegetation indices) (see Fig. 12.2). In this approach, h()
is a radiative transfer (RT) model relating the state vector x to the remote sensing
signal (e.g., radiance) or its simple transformation (e.g., reflectance).
320
S. Liang, J. Qin
Both approaches have strengths and weaknesses. If the high-level product

retrieval is very accurate and its uncertainties well-characterized, then, use of such
data is preferable. However, most inversion processes are ill-posed (see Chapters 7
and 8), and the resulting products may be poor. This is particularly so if their error
characteristics cannot be specified accurately for DA (see Section 12.3.6). Another
issue is that most current land products are not continuous both spatially and temporally due to cloud contamination or the inversion failure due to an inadequate
number of good observations. Since many biogeophysical variables are continuous spatially and temporally, gaps need to be filled using various spatiotemporal
modeling techniques (Fang et al., 2007a; Fang et al., 2007c; Julien et al., 2006;
Moody et al., 2005).
On the other hand, there is no need for an inversion if direct observations are
used, but the forward RT models must be coupled with land surface process models.
With physically based RT models, an additional advantage of the second approach
is that consistency can be maintained between the direct physical meaning of model
parameters in both the RT and land surface process models. This is not necessarily
so for high-level products that may be derived under a different set of assumptions
(e.g., spatial distribution of foliage) to those assumed in the process model.
For using high-level products, not only those from optical and thermal observations (e.g., LAI/FPAR, albedo, LST) are considered but also those obtained from
LIDAR (e.g., canopy structure parameters and above-ground biomass) and RADAR
(above-ground biomass and moisture conditions) observations. Other products, such
as forest biomass inventories, soil carbon survey, and Eddy covariance flux measurements, atmospheric CO2 and tracers concentrations, nutrient fluxes and stocks
(particularly N) can also be used for assimilation.
A highly related issue is the compatibility of remotely sensed products and variables in land surface process models. Although substantial efforts have been made
from both remote sensing and modeling communities, the gaps still exist. For example, most land surface models separate broadband albedos into direct and diffuse
components (Dai et al., 2003), but remote sensing albedo products are usually total
broadband albedo (see Chapter 9). Most land surface models normally partition surface temperature into sunlit/shadow leaf and soil temperatures, but remote sensing
skin temperature is an effective temperature for each pixel (see Chapters 4 and 10).
A particular advantage of DA schemes is that they can be used to evaluate the
advantage gained (assessed through reduction in uncertainty) by the addition of
new observations. This will be of great use in assessing the value of different types
and sample availability of observations, as well as performing synthetic experiments
to evaluate new types of observations (e.g., from proposed monitoring networks or
space missions).
12.3.3 Target Variables

The target variables are the properties of the model to be adjusted or estimated in the
optimization process. They include any model property considered to be sufficiently
321
uncertain as to benefit from constraint by the data. Model properties which can be
target variables include: (1) model parameters (); (2) forcing variables, if there is
substantial uncertainty about them; (3) initial conditions on the state variables; and
(4) time-dependent components of the state vector. Land surface models are becoming more and more sophisticated with numerous parameters. Some parameters may
be easily estimated, while others may not be very sensitive to the cost (objective)
function. These target variables must be identified based on extensive sensitivity
experiments.
Selecting the target variables depends on the land surface model used in the DA
system, along with how many and which data products are assimilated. For example, Kaminski et al. (2002) optimized 24 parameters (light use efficiency and Q10
for heterotrophic respiration for each of 12 biomes) in a terrestrial biosphere model
coupled with an atmospheric transport model using CO2 data. Barrett (2002) optimized a set of parameters (turnover times, C allocation ratios, humification ratios,
and light use efficiency) in a terrestrial C cycle model. Rayner et al. (2005) estimated
56 process parameters plus an initial condition through terrestrial carbon cycle DA
system with a coupled ecosystem and atmospheric transport model. Williams et al.
(2005) estimated nine unknown parameter constants in the C box model and the
initial values of the five C pools using the ensemble Kalman filter.
For a given land surface model, a rigorous sensitivity study is absolutely required,
which enables determination of the parameters/state variables that are sensitive to
the assimilated data. The automatic differentiation (AD) tool called the Tangent
linear and Adjoint Model Compiler (Giering and Kaminski, 1998) as well as its
successor, Transformation of Algorithms in Fortran (Giering and Kaminski, 2002)
used to generate the adjoint code of the process model may provide an effective way
for sensitivity studies.
12.3.4 Multi-Objectives Optimization

Practical experience suggests that any single objective (cost) function, no matter
how carefully chosen, is often inadequate to properly measure all of the characteristics of the assimilated data sets deemed to be important (Demarty et al., 2005; Vrugt
et al., 2003). One strategy to circumvent this problem is to define several objective
functions that measure different (complementary) aspects of the system behavior and to use multi-objective optimization to identify the set of non-dominated,
efficient, or Pareto optimal solutions. Such an approach makes it more efficient to
assimilate multiple data products simultaneously.
The choice of fitting criterion or estimator is crucial. The very popular
weighted least-squares estimator is the simplest, and is amenable to very efficient
mathematical algorithms. However it may be seriously inappropriate in certain
circumstances. For example, the estimator of maximum likelihood may be a much
more reasonable choice when errors in the data or dynamical information are
non-Gaussian.
322
S. Liang, J. Qin
Generally speaking, the multi-objective problem can be stated as the following

minimizing formulation:
Min{J1 ( ), J2 ( ), . . . , Jm ( )}
(12.5)
where
Ji symbolizes
a single objective function with i = 1, . . . , m and =
)
(
1 , 2 , . . . , p a particular set of p parameters included in the feasible parameter space.
This issue is highly related to those discussed in Section 12.3.2. A single objective function corresponds to each high-level product or directly observed variable,
thus, we can combine both high-level products and direct observations. Even for the
same product, a single objective function can be a measure of errors (Gupta et al.,
1998), such as root mean square error, mean absolute error, maximum absolute error,
and so on.
The solution of the multi-objective formulation given in Eq. (12.5) does not
lead to a unique solution, but to a set of solutions, generally named Pareto set or
behavioral set (Gupta et al., 1998). There are many different algorithms to address this problem (Marler and Arora, 2004), such as the Multiobjective Shuffled
Complex Evolution Metropolis (MOSCEM) algorithm that is capable of solving the
multi-objective optimization problem for hydrologic models (Vrugt et al., 2003).
MOSCEM is available to the public. Other assimilation algorithms will be discussed
in the following Section 12.3.5.
12.3.5 Assimilation Algorithms

The target variables characterizing the land surface properties are estimated through
the assimilation algorithms. On one hand, there are many different assimilation
algorithms available in the literature from meteorological and oceanographic
DA communities, such as three- or four-dimensional variational algorithms (see
Section 12.3.5.1), and Kalman filters (Section 12.3.5.2). On the other hand, land
DA has different characteristics, such as handling only dozens of unknowns, while
the meteorological DA system typically manages 106 109 unknowns. Neural network and particle filtering techniques are briefly introduced in Sections 12.3.5.3
and 12.3.5.4.
Before moving on, the measurement sequence is defined as y1:t {yi }ti=1 and the
state sequence as x0:t {xi }ti=0 . With these definitions, the whole DA process as an
estimation problem can be stated as:
x0:t = g(y1:t )
(12.6)
where x0:t represents estimated state sequence, and g() the estimator, which can be
viewed as the assimilation algorithm. When a measurement sequence y1:t is inserted
into the estimator, we obtain a realization of the estimator (x0:t ).
323
12.3.5.1 Cost Function-Based Methods

This type of method is also called non-sequential or batch one. The most general
form of cost function-based methods can be expressed as the following constrained
optimization problem:
min J = x0 x0 2B1 + i=0 wi 2Q1 + i=0 ei 2R1
t1
x0:t
s.t. xi+1 = f (xi , ui , ) + wi , i = 0, . . . ,t 1

yi = h(xi , ui , ) + ei
(12.7)
i = 1, . . . ,t
where the weighting matrices Qi , Ri , and B can be regarded as covariances of the

distribution functions p(wi ), p(ei ), and p(x0 ). If the parameter vector needs to
*2
*
be estimated simultaneously, a term * *P1 can be added into Eq. (12.7) and

x0:t
the control vector becomes
.
This problem is the so-called weak constraint DA approach (see Section 12.3.2)
when the dynamic model is considered to be imperfect. The traditional strong constraint DA problem is just its special case in which the model is assumed to be
2
is removed in Eq. (12.7). The formulation
perfect and thus the term t1
i=0 wi Q1
i
statements can then be expressed as:
min J = x0 x0 2B1 + i=0 ei 2R1
t
x0:t
s.t. xi+1 = f (xi , ui , ) i = 0, . . . ,t 1

yi = h(xi , ui , ) + ei i = 1, . . . ,t
(12.8)
In the cost function-based problems, all data are treated simultaneously and the
minimization problem is solved only once, as presented in Eq. (12.7) or (12.8).
The problem becomes the usual optimization problem. It is computationally expensive when the number of unknowns is large. There are many non-sequential algorithms depending on whether the derivative information is used, but the key issue
is to incorporate an effective global searching algorithm. The typical algorithms include the shuffled complex evolutionary (SCE) method (Duan et al., 1993; Duan
et al., 1992; Duan et al., 1994), a very fast simulated annealing (SA) algorithm
(Ingber, 1989; Li et al., 2004), the differential evolutionary (DE) method (Storn
and Price, 1997; Storn and Price, 1996), and the genetic algorithm (GA) (Goldberg,
1989). Codes for these methods are available to the public. The common weakness
of these methods is their slow computational speed.
In order to solve both weak and strong constraint problems stated above, some
descent algorithms, such as the conjugate gradient method, can be used. These
approaches require the first-order derivative or even Hessian matrix of the cost function. To this end, the adjoint of the dynamic model has to be developed. However,
this process is tedious. It is encouraging that automatic differentiation (AD) techniques have been developed and applied to automatically generate the adjoint model
324
Fig. 12.3 Illustrations of evaluating the derivatives of the
cost function with AD techniques. F denotes the whole
codes to evaluate the cost
function. F represents the
adjoint codes of the original
codes, which can be used to
evaluate derivates easily.
S. Liang, J. Qin
(a)
J
x
F '*
dJ = 1
(b)
at the level of computer codes (Bischof et al., 1996; Bischof et al., 2002; Carmichael
et al., 1997; Dobmanin et al., 1995; Giering and Kaminski, 2002; Verma, 2000). AD
is very effective and easy to use. This dramatically conserves time and energy of DA
practitioners. The principle of AD is simple and based on two facts. First, any computer code statement can be regarded as a composition of elementary functions.
Second, chain rule can be used to differentiate this composition of elementary functions. Many software packages have been developed in accordance with the principles described above for FORTRAN and C computer languages. They are given at
the web site www.autodiff.org. The illustration of AD running process is presented
in Fig. 12.3. We have recently applied this method in estimating LAI from satellite
data (Qin et al., 2007a).
The advantages of cost function-based method are twofold. First, all data in a
batch window are used to estimate the state. Second, inequality, equality, and bound
constraints can be included explicitly. Its disadvantages are also apparent. First, the
adjoint model is generated to evaluate the derivative of the cost function if the highly
efficient descent optimization method is used. However, the development of the adjoint model requires that the dynamic model should be differentiable. This condition
can not usually be met in the land surface process modeling because of many discontinuous parameterizations. Instead, SCE, SA, DE, or GA could be used, but they
are computationally very slow. Second, the cost function-based method just uses
the inverses of covariance matrices as the weights, as seen in Eqs. (12.3) or (12.7)
and (12.8). Since the covariance is the second moment of one distribution, more
information included in the distributions p(wi ), p(ei ), and p(x0 ) is not used and
therefore wasted. If p(wi ), p(ei ), and p(x0 ) are normal distributions, no information is discarded since Gaussian distributions are completely characterized with the
first and second moments.
12.3.5.2 Sequential-Based Methods

To derive the optimal sequential assimilation scheme, assume that the background
states are represented by prior estimates. The sequential approaches are based on the
Bayesian inference that combines the prior knowledge of the state vector and the
measurement to obtain the posterior distribution of the state vector. Once obtaining
the posterior distribution, everything is known about the state vector. This process
can be expressed as follows.
p(x0:t |y1:t ) p(y1:t |x0:t )p(x0:t )
325
(12.9)
where p(x0:t ) represents the prior distribution of the state vector, p(y1:t |x0:t ) the measurement distribution, and p(x0:t |y1:t ) the posterior distribution. Typically, a Markov
assumption is applied to the prior. So the state vector at time t only depends on the
state vector at time t 1:
p(x0:t ) = p(x0 ) i=1 p(xt |xt1 )
t
(12.10)
where p(xt |xt1 ) is the evolution distribution, and p(x0 ) is the distribution of the
initial state vector (background or first guesses). Another important assumption is
that measurements are independent given the true state:
p(y1:t |x0:t ) = i=1 p(yt |xt )
t
(12.11)
Substituting Eqs. (12.9) and (12.10) into (12.8), we obtain:

p(x0:t |y1:t ) p(x0 ) i=1 p(yt |xt )p(xt |xt1 )
t
(12.12)
This equation implies that once new data is available, the previous estimate of the
state process could be sequentially updated without having to calculate from scratch.
However, this also means we have to store all state vectors up to time t, and the size
of x0:t will expand as time goes by, becoming very large. In fact, there often is
an interest in the filtering distribution p(xt |y1:t ), that is to estimate the probability
density of the current state vector conditioned on the measurements up to now. The
whole filtering process is straightforward.
The filtering density, p(xt |y1:t ), and the one step prediction, p(xt+1 |y1:t ), density
are recursively given by a measurement update according to
p(xt |y1:t ) =
p(yt |y1:t1 ) =
p(yt |xt )p(xt |y1:t1 )

p(yt |y1:t1 )
&
(12.13)
p(yt |xt )p(xt |y1:t1 )dxt
(12.14)
p(xt+1 |xt )p(xt |y1:t )dxt
(12.15)
and a time update according to

p(xt+1 |y1:t ) =
&
and the recursion is initiated by

p(x0 |y0:1 ) = p(x0 )
(12.16)
In the general case, one is normally unable to obtain an analytical expression of the
filtering density except under the assumption of linear model and observation, and
Gaussian error distributions. This leads to the prominent Kalman filter (KF). If the
dynamical model and measurement process are characterized as follows
326
S. Liang, J. Qin
xt+1 = Fxt + wt , wt N(0, Qt )

yt = Hxt + et , et N(0, Rt )
(12.17)
the KF can be expressed as

xt+1|t = F xt|t
Pt+1|t = FPt|t F T + Qt
xt+1|t+1 = xt+1|t + G(yt+1 H xt+1|t )
(12.18)
Pt+1|t+1 = (I GH)Pt+1|t
G = Pt|t1 H T (H T Pt|t1 H + Rt )1
where x denotes the estimate of the state vector.
There are two popular variants of Kalman filter. One is the Extended Kalman
filter (EKF) and the other is the Ensemble Kalman filter (EnKF). EKF handles the
cases where either H or F is nonlinear. Note that most land surface models (H) are
nonlinear. In this case, they can be defined as its tangent linear. The EnKF is a sophisticated sequential DA method. The EnKF applies an ensemble of model states
to represent the error statistics of the model estimate, uses ensemble integrations to
predict the error statistics forward in time, and employs an analysis scheme which
operates directly on the ensemble of model states when observations are assimilated. The EnKF efficiently manages strongly nonlinear dynamics and large state
spaces and is now used in realistic applications with primitive equation models for
the ocean and atmosphere. Originally proposed by Evensen (1994), the EnKF is
more recently reviewed by Evensen (2003), providing detailed information on the
formulation, interpretation and implementation of the EnKF.
Other sequential assimilation algorithms widely used in meteorological or
oceanographic DA communities include successive correction, optimal or statistical
interpolation, analysis correction, 3DVAR and 3DPSAS.
All these methods have been widely applied in a variety of fields (Evensen, 2003;
Houtekamer and Mitchell, 1998; Jones et al., 2004; Kumar and Kaleita, 2003; Qin
et al., 2006; Reichle et al., 2002; Wade and Eric, 2003).
12.3.5.3 Neural Network-Based Methods

In recent years, people have attempted to combine artificial neural network (ANN)
and DA method (Liaqat et al., 2003a, b; Tang and Hsieh, 2001; Wu et al., 2005;
Yu et al., 1997). The primary objective of these investigations is to complete or
approximate the dynamic models from measurements by estimating weights and
biases of ANN using DA method. The advent of the feed-forward neural network
model opens the possibility of hybrid neural-dynamical models via variational DA.
Such a hybrid model may be used in situations where some variables, difficult to
model dynamically, have sufficient data for empirical modeling with ANN. Liaqat
et al. (2001) used a neural network for constructing an arbitrary mapping function.
327
A neural network is trained by optimizing an object function composed of squared

residuals of differential equations at collocation points and squared deviations of the
observation data from the computed values. An assimilation problem is solved even
if the model differential equations do not express the observed phenomena exactly.
Since the dynamic model can be constructed directly from measurements and can
be approximated, the computational speed greatly improves.
12.3.5.4 Particle Filtering

Particle filtering (PF) is also called sequential Monte Carlo filtering based on point
mass (or particle) representations of probability densities, which can be applied
to any state-space model and which generalize the traditional KF methods (Ristic
et al., 2004; Smith et al., 2005). It is an important technique to manage the DA with
elements of nonlinearity and non-Gaussianity such that the underlying dynamics of
a physical system are modeled accurately. It has been applied in many engineering
fields and attracted attention from some DA practitioners since the posterior distribution of state vector can be represented with Monte Carlo samples. However, KF
and its variants just evaluate the mean and covariance of the posterior distribution.
PF better grasps the filtering density evolution of the nonlinear system in time than
KF and its variants do. PF itself also has many variants, among which the bootstrap
filter, also called sampling importance resampling filter (SIR), is the most easily
used. Its main steps include:
(i)
Step 1: for t = 0, sample {x0 }Ni=1 p(x0 );

(i)
(i)
(i)
(i)
(i)
(i)
Step 2: draw {xt+1 }Ni=1 p(xk+1 |xk ). That is xt+1 = f (xt , ut , ) + wt , wt

p(wt );
(i)
Step 3: compute weights ut+1 =
(i)
p(yt+1 |xt+1 )
N
(i)
p(yt+1 |xt+1 )
i=1
(i)
(i)
where p(yt+1 |xt+1 ) denotes the value of p(yt+1 h(xt+1 , ));

(i)
(i)
Step 4: resample {xt+1 }Ni=1 with replacement according to weights {ut+1 }Ni=1 in
+
(i)
order to get {xt+1 }Ni=1 with weights {1 N }Ni=1 ;
Step 5: set t = t + 1 and go to step 2.
Observe that SIR is easy to be implemented in DA practice.
12.3.6 Observation Error Matrix

Characterizing the errors in DA is extremely important since it significantly
affects the estimates (Daley, 1991). The observation error covariance matrix R
in Eqs. (12.3), (12.6) and (12.16) can in principle be quite general in form, including biases, temporal correlations, extreme outliers and so on. Many extant methods
328
S. Liang, J. Qin
assume the error to be Gaussian with zero mean and no temporal correlation.
However, more general error structures are very common, and the development of
methods for dealing with such errors is an active area of current research (Evensen,
2003; Raupach et al., 2005).
All NASA Earth Observing System (EOS) land surface high-level products have
been claimed to be validated, but most were based on limited ground truths
(Morisette et al., 2006). The error magnitudes, and their spatial and temporal distributions, have never been well specified. Assigning uncertainties to these high level
products is complex. There are a few studies reporting the accuracies of individual
products, but comprehensive modeling of spatial and temporal error structures, and
correlation among errors of different products has not yet been done. It raises the
challenge of evaluating the uncertainty properties of major products, and it is evident that this is an enormous goal. A range of issues identified by Raupach et al.
(2005) as needing to be addressed include:
The error magnitude rii for each high-level product, inclusive of all error sources
(in other words, the diagonal elements of the covariance matrix R)
The correlations among errors in different products, quantified by the offdiagonal elements of the covariance matrix R
The temporal structure of the errors: whether they are random in time or temporally correlated, and the possible presence of unknown long-term drifts or biases
The spatial structure of errors (random, slowly varying or bias as for temporal
structure)
The error distribution: normal (Gaussian), lognormal, skewed or the sum of multiple error sources with different distributions, such as a small Gaussian noise
together with occasional large outliers because of measurement corruption events
Possible mismatches between the spatial and temporal averaging implicit in the
model and the measurements (the scaling problem).
This same set of challenging issues (to define and specify the error models) pertains
to other data sets, such as the error properties of direct flux measurements and direct
measurements of carbon stores in addition to remote sensing of land surface properties. It is more straightforward to assign uncertainties to lower-level products than
to higher-level ones. This is a major factor for considering the assimilation of direct
observations (radiance/reflectance or vegetation indices), as noted above.
12.4 Recent DA Applications

DA is a powerful way to consistently combine measurements and dynamical models
for the accurate estimation of model parameters and the state vector. Since DA has
been explored and implemented in many applications, progress in the following
selected fields is summarized.
329
12.4.1 Soil Moisture Estimation

While microwave remote sensing provides the opportunity to map global soil
content, the revisiting frequency is limited. Moreover, the L band brightness temperature is related to only the surface soil moisture (top 5 cm) and yields little information about the root zone. A land surface process model forced by atmospheric data
can produce soil moisture and temperature profiles at the model time resolution.
It is obvious that just running the model without any constraints can lead to large
errors due to uncertainties in the model structure, model parameters, and external
forcing data. DA offers a means to consistently take advantage of both modeling
and observations. Two DA algorithms have been applied to estimate not only soil
moisture, but also latent heat flux (tightly associated with the root-zone moisture)
using L band microwave remote sensing data.
Entekhabi et al. (1994, 1998) first applied EKF to assimilate microwave remote
sensing data into a land surface model to subsequently retrieve soil water moisture and temperature profiles. Although synthetic data was used and the land surface
model was relatively simple, their research opened up prospects for land surface DA.
Many improvements have been made since then. Walker and Houser (2001) and
Walker et al. (2001a, b) compared direct insertion and EKF assimilation methods,
conducted retrieval experiments with field data, and then applied their algorithm on
the regional scale for initialization of climate and hydrological models. But surface
soil moisture data rather than satellite observations was assimilated in their studies.
Reichle et al. (2002) used EnKF to assimilate L-band (1.4 GHz) microwave
radiobrightness observations into a land surface model, compared it with the variational method, and investigated the influence of the ensemble size on the retrieved
results. Their research indicates that the EnKF is a flexible and robust DA option that
gives satisfactory estimates even for moderate ensemble sizes although the updating process is suboptimal. Crow (2003) and Crow and Wood (2003) applied EnKF
to assimilate L-band microwave data to correct for the impact of poorly sampled
rainfall on land surface model predictions of root-zone soil moisture and surface
energy fluxes. The results suggest the EnKF-based assimilation system is capable
of correcting a substantial fraction of model error in root zone (40 cm) soil moisture
and latent heat flux predictions associated with the use of temporally sparse rainfall
measurements as forcing data. The recent studies also applied EnKF to estimation
of soil moisture profiles (e.g., Merlin et al., 2006; Zhou et al., 2006). Margulis et al.
(2002) explained why the EnKF technique is so appealing for soil moisture estimation: (i) its sequential nature is well suited to real-time data streams and forecasting;
(ii) it not only provides an estimate of surface and profile soil moisture, but information about the statistical confidence of estimation; (iii) it is sufficiently straightforward to use with off-the-shelf models; and (iv) it is relatively efficient, making
its application to large-scale problems feasible.
The variational cost function-based approach has also been successfully applied
to hydrological studies in recent years. The key point in using the variational
assimilation method is to develop the adjoint of a dynamic model. This requires
the model should be differentiable, but land surface schemes usually are not.
330
S. Liang, J. Qin
Consequently some approximations have to be made. Reichle (2000) and Reichle

et al. (2001) first used the variational method to assimilate L-band microwave data
into a complicated land surface model to retrieve the soil moisture profile. The
results showed that the state estimates obtained from the assimilation algorithm improve significantly over prior model predictions derived without assimilating radiobrightness data.
Soil moisture estimation is also one of the major goals in the regional and global
land surface DA systems, such as the global land DA system (Rodell et al., 2004),
the North American Land DA system (Mitchell et al., 2004), and the South America
Land DA system (de Goncalves et al., 2006).
12.4.2 Energy Balance Fluxes Estimation

Accurate estimation of energy and momentum fluxes, especially sensible and latent heat fluxes, between the land surface and the atmospheric boundary layer, is
required in a wide variety of agricultural, hydrological, and meteorological applications (Courault et al., 2005; Su, 2002). Many methods, such as eddy correlation
and Bowen ratio, can be used to measure these fluxes at the field level. Their applicability, however, is limited on the fine spatial scale. Currently, the only way to
achieve this goal of mapping fluxes regionally is to use remote sensing techniques
that provide various spatial and temporal imageries covering large areas.
Estimation of surface energy balance components using remote sensing data can
be roughly divided into three categories: empirical, residual and DA. Empirical
methods directly build on the relationship between remote sensing products, such
as various vegetation indices and retrieved LST for the estimation. Residual methods of the energy budget couple some empirical formulas and physical mechanisms
to realize the estimation of evapotranspiration (ET or LE) and sensible heat flux
(H) by using remote sensing to directly estimate input parameters, such as the surface energy balance system (SEBS) (Su, 2002). In recent years, DA methods have
integrated remote sensing data and soilvegetationatmosphere transfer models for
estimating surface fluxes, and have achieved encouraging results. This new method
has been receiving considerable attention from researchers in recent years (Boni
et al., 2001a, b; Caparrini and Castelli, 2004; Caparrini et al., 2003; Castelli et al.,
1999) since it combines dynamic models and temporal remote sensing data based
on the control theory in order to accurately retrieve critical parameters for flux estimation. The first two methods more or less use only the instantaneous data and
empirical relationships.
Castelli et al. (1999) developed a simple land surface scheme and its adjoint,
defined a moisture index as control variables, used radiometric surface temperature
as observations, and performed DA experiments. Results indicated that this algorithm can retrieve land surface energy balance components effectively. Boni et al.
(2001a, b) investigated the impact of land surface temperature sampling frequency
on assimilation results and suggested that satellite remote sensing of land surface
331
temperature may be used to provide estimates of components of the surface energy

balance and land surface control on evaporation. Caparrini et al. (2004, 2003) replaced a moisture index with evaporative fraction as control variables, made many
assimilation experiments, and applied this algorithm on the regional scale using
AVHRR land surface temperature as observations.
12.4.3 Carbon Fluxes Estimation

It is increasingly recognized that DA (modeldata synthesis, datamodel fusion or
many other names) is one of the methods that can meet such a challenge. The US
Climate Change Science Program strategic plan and the North American Carbon
Program (NACP) plan repeatedly call for the development of DA methods for carbon cycle studies. A near-term priority identified by the US Carbon DA Program
Workshop Report (Fung et al., 2002) was urgent support for interdisciplinary teams
to develop component and coupled prototype carbon assimilation models. Canadell
et al. (2004) identified three fundamental research areas that require major development in order to provide policy relevant knowledge for managing the carbon-climate
system over the next few decades, with the first area being carbon observations and
multiple constraint data assimilation.
For carbon cycle studies, some efforts have been made to develop DA systems,
but most of these are based on top-down strategies that begin with measured
changes in atmospheric carbon concentrations and attempts to infer the spatial
distribution and magnitude of the net exchange. For example, the Carbon DataModel Assimilation (C-DAS) project at NCAR (National Center for Atmospheric
Research) is developing a carbon DA system based on their atmospheric transport
model (http://dataportal.ucar.edu/CDAS).
An equally important approach in carbon cycle science is the bottom-up approach (Cihlar et al., 2000) that starts with a specific land parcel to account for the
various pathways of carbon exchange between the ecosystem and the atmosphere,
and then scales up to much larger regions. It relies on both ecosystem process models and spatial data sets. The model can be developed at the local level and validated
using conventional measurements. Satellite observations provide spatial distribution
and frequent up-to-date information on the rate of change of the variables driving
the model. These models can in turn be used to estimate spatial and temporal variations in CO2 uptake and release over large areas, if appropriate input data sets
(e.g., vegetation and soils maps, weather data) are available.
Several obstacles stand in the way of extensive use of ecosystem process models
for extrapolation. The major issue is the requirement of much more information on
vegetation characteristics than is readily available or even known for many areas of
the globe. Many parameters are not measurable at earth system scales, because of
high small-scale spatial variability (e.g., any soil property), high temporal variability (e.g., stomatal conductance), and physical inaccessibility (e.g., most roots, deep
soil). Moreover, from the modeling perspective, some processes are well understood
332
S. Liang, J. Qin
(e.g., photosynthesis, decomposition of litter and soil organic matter), while other
processes are not (C-allocation among plant tissues, T-sensitivity of humus decomposition). As a result, large uncertainties arise in calculating terrestrial carbon cycle
using the bottom-up approach (Barrett, 2002).
Wang and Barrett (2003) developed a modeling framework that synthesizes various types of field measurements at different spatial and temporal scales to estimate
monthly means (and their standard deviations) of gross photosynthesis, total ecosystem production, net primary production (NPP), and net ecosystem production (NEP)
for eight regions of the Australian continent between 1990 and 1998. Williams et al.
(2005) developed a DA approach combining stock and flux observations with a dynamic model to improve estimates of ecosystem carbon exchanges. Rayner et al.
(2005) developed a terrestrial carbon cycle DA system (CCDAS) for determining
the space-time distribution of terrestrial carbon fluxes for the period 19791999.
Hazarika et al. (2005) integrated the MODIS LAI product with an ecosystem model
for accurate estimation of NPP. Validations of results in Australia and the USA show
that NPP estimated using the DA method to be more accurate than that generated
by the data forcing method. Their research demonstrates the utility of combining
satellite observations with an ecosystem process model to achieve improved accuracy in estimates and monitoring global net primary productivity.
Barrett et al. (2005) demonstrated that a multiple-constraints modeldata assimilation scheme using a diverse range of data types offers improved predictions of
carbon and water budgets at regional scales. Xu et al. (2006) applied the Bayesian
probability inversion and a Markov chain Monte Carlo (MCMC) technique to a terrestrial ecosystem model to analyze uncertainties of estimated carbon (C) transfer
coefficients and simulated C pool sizes. Their study shows that the combination of a
Bayesian approach and MCMC inversion technique effectively synthesizes information from various sources for assessment of ecosystem responses to elevated CO2 .
Sacks et al. (2006) used a modeldata synthesis approach with a simplified carbon
flux model to extract process-level information from 5 years of eddy covariance
data at an evergreen forest in the Colorado Rocky Mountains. Including water vapor
fluxes, in addition to carbon fluxes, in the parameter optimization did not yield significantly more information about the partitioning of the net ecosystem exchange of
CO2 into gross photosynthesis and ecosystem respiration. Sacks et al. (2007) used
the modeldata synthesis method to address fundamental questions about climate
effects on terrestrial ecosystem net CO2 exchange.
12.4.4 Crop Yield Estimation

Advance information on crop yield during the crop growing season is vital for effective crop management and for national food security policy. Agricultural harvested
grain yield is a reliable way to estimate crop yields by sampling field measurements
of standing crops. However, this method is both time consuming and costly, with
results that are not available until after harvest. In the last three decades, satellite
333
remote sensing data have been used to estimate crop yields over large areas because
these methods are more cost effective and more timely than traditional survey procedures (MacDonald and Hall, 1980).
Earlier studies were mostly based on empirical regression methods that relate
crop yield to remotely sensed surface reflectance and their combinations (i.e., vegetation indices). These relationships could be described with linear, cubic polynomial, or exponential regression (Jiang et al., 2003). Essentially a statistical model,
this method cannot predict the time-dependent processes of crop growth. Additionally, the relationship between yield and NDVI may not be accurate under extreme
weather conditions.
Mathematical crop growth models simulate fundamental processes such as photosynthesis, respiration, biomass partitioning, and water and nitrogen transfers
(Baret et al., 2000). This allows researchers to evaluate a wide array of alternatives, and to assemble processes in an integrated package. Along with supporting
better crop management decisions, mechanical crop growth models can simulate
the dynamics of LAI and other structural properties of the crop fields (e.g., height
and biomass). The combination of remote sensing and crop growth simulation models is increasingly recognized as a promising approach for monitoring growth and
estimating yield (Bauman, 1992).
The use of crop models is often limited by uncertainties in their input parameters
such as soil conditions, sowing date, planting density and initial field conditions.
Except in some controlled experimental fields, many of these parameters are poorly
known. Remote sensing can play a critical role in helping identify the field and crop
status from estimated biophysical parameters (Clevers and Leeuwen, 1996). Remote
sensing data, therefore, can be assimilated with crop growth models to improve their
overall performance.
Several assimilation schemes, of various degrees of complexity and integration,
have been developed in the last 10 years (Moulin et al., 1998). Various methods for
integrating a crop growth model with remote sensing data were described by Mass
(1988a, b) and were also reviewed by Fischer et al. (1997) and Moulin et al. (1998).
Mass (1993) compared the results of calibrating a crop simulation model using LAI
observations obtained either from field sampling or remote sensing. Winter wheat
yields were modeled more accurately using remotely sensed LAI observations than
field-sampled LAI observations (Mass, 1993). This difference appeared to result
from the apparent ability of the remotely sensed LAI observations to better represent
the photosynthetically active plant area in the crop canopy.
Bach et al. (2003) experimented with coupling a raster-based PROMET-V model
with the radiative transfer model GeoSAIL to predict biomass and yield. In their
study, LAI, fraction of brown leaves, and surface soil moisture were used as free
variables; surface reflectance was used as the control variable. Their assimilation
procedure produced improved biomass and yield results. Guerif and Duke (2000)
combined the SUCROS crop model with the SAIL canopy reflectance model for
accurate estimation of sugar beet yield. Ground measured reflectance was used to
match the predicted reflectance. One limitation of their study is that many crop
and soil parameters need to be obtained from field measurements. The SAIL model
334
S. Liang, J. Qin
has also been integrated with the EPIC crop model to estimate the yield of spring
wheat in North Dakota (Doraiswamy et al., 2003). Planting date is the only adjustable variable in this model. The estimated yields are mostly within 10% of the
NASS (National Agricultural Statistical Service) reports. In this work, climate data
are based on interpolation of weather station measurements. The crop area mask is
based on the 1 km AVHRR classification. NDVI is also calculated from AVHRR.
Although the AVHRR data set is easily accessible, using it will compromise the
precision of analyses owing to outdated calibrations and the application of partial atmospheric corrections. Doraiswamy et al. (2004) used a look-up table (LUT)
method to estimate LAI from 250 m MODIS reflectance data. The crop modeled
LAI was adjusted to fit the MODIS simulated LAI by changing planting time, time
when maximum LAI is attained, and the beginning of leaf senescence.
Jongschaap and Schouten (2005) estimated the regional wheat yield by assimilating SPOT data into a crop model. Microwave remote sensing data (ERS SAR
C-band) were used to estimate regional wheat flowering dates to calibrate a wheat
growth simulation model. Pauwels et al. (2007) assessed to what extent the results of
a fully coupled hydrologycrop growth model can be optimized through the assimilation of observed LAI and soil moisture values using the EnKF. A practical procedure using the variational optimization method to predict crop yield at the regional
scale from MODIS data was recently developed by Fang et al. (2007b). This method
outputs agronomic variables (yield, planting, emergence and maturation dates) and
biophysical parameters (e.g., LAI).
12.5 Summary
Though DA has reached maturity in meteorological and oceanographic applications,
the land community has just begun to employ it for estimation of land surface variables. Herein, basic DA principles have been described, critical issues in land DA
have been identified, and many of the latest applications in hydrology, carbon cycle,
and agriculture have been introduced. Because of the continuous improvement in
DA methods and computational technology, along with an available wealth of remote sensing observations and extensive ground observation networks, DA is likely
to become the best technique to monitor and map land surface environments by
integrating a priori knowledge with an enormous variety and sheer volume of data.
Revisiting the key issues addressed in this chapter, the following questions are
put forth:
1. Which remotely sensed data products and other measurements are the most valuable for land DA?
2. How are the differences between the constrained data sets and predictions of a
dynamic model characterized?
3. How can the model parameters/variables be estimated effectively?
4. How will the errors of the assimilated data be specified?
5. Which land surface properties can be estimated?
6. What types of dynamic models are suitable for DA?
335
To answer these questions, explorers from the community are needed to traverse
this new frontier, and we hope they will do so.
The computational issues of DA have not been broached since DA algorithms
are usually computationally expensive. However, the fast pace of computer science
advances promises to minimize such obstacles. Community efforts are needed to
build the practical tools so that few researchers have to start from scratch. A good example of such endeavor is the land information system developed by NASA (Kumar
et al., 2006). Ideally, more educational tools for this enterprise will follow.
Acknowledgements We are very grateful for Dr. John Townshend for providing valuable comments. S. Liang is partially funded by NASA under grant NNG04GL85G.
References
Bach H, Mauser W, Schneider K (2003) The use of radiative transfer models for remote sensing
data assimilation in crop growth models. In: J Stafford, A Werner (eds), Precision agriculture:
Papers from the 4th European Conference on Precision Agriculture, Berlin, Germany, 1519
June 2003
Baret F, Weiss M, Troufleau D, Prevot L, Combal B (2000) Maximum information exploitation for
canopy characterisation by remote sensing. Asp. Appl. Biol. 60(7182)
Barrett DJ (2002) Steady state turnover time of carbon in the Australian terrestrial biosphere. Glob.
Biogeochem. Cycles 16(4):Art. No. 1108
Barrett DJ, Hill MJ, Hutley LB, Beringer J, Xu JH, Cook GD, Carter JO, Williams RJ (2005)
Prospects for improving savanna biophysical models by using multiple-constraints model-data
assimilation methods. Aust. J. Bot. 53(7):689714
Bauman BAM (1992) Linking physical remote sensing models with crop growth simulation models
applied for sugarbeet. Int. J. Remote Sens. 14:25652581
Bischof C, Khademi P, Mauer A, Carle A (1996) Adifor 2.0: Automatic Differentiation of Fortran
77 Programs. IEEE Comp. Sci. Eng. 3(3):1832
Bischof CH, Hovland PD, Norris B (2002) Implementation of automatic differentiation tools. Proceedings of the 2002 ACM SIGPLAN Workshop on Partial Evaluation and Semantics-Based
Program Manipulation (PEPM-02), vol. 37, pp. 98107
Boni G, Castelli F, Entekhabi D (2001a) Sampling Strategies and Assimilation of Ground Temperature for the Estimation of Surface Energy Balance Components. IEEE Trans. Geosci. Remote
Sens. 39(1):165172
Boni G, Entekhabi D, Castelli F (2001b) Land data assimilation with satellite measurements for
the estimation of surface energy balance components and surface control on evaporation. Water
Resour. Res. 37(6):17131722
Canadell JG, Ciais P, Cox P, Heimann M (2004) Quantifying, understanding and managing the
carbon cycle in the next decades. Climatic Change 67(23):147160
Caparrini F, Castelli F (2004) Estimation of surface turbulent fluxes through assimilation of radiometric surface temperature sequences. J. Hydrometeorol. 5:145159
Caparrini F, Castelli F, Entekhabi D (2003) Mapping of land-atmosphere heat fluxes and surface
parameters with remote sensing data. Bound. Lay. Meteorol. 107:605633
Carmichael GR, Sandu A, Potra FA (1997) Sensitivity analysis for atmospheric chemistry models
via automatic differentiation. Atmos. Environ. 31(3):475489
Castelli F, Entekhabi D, Caporali E (1999) Estimation of surface heat flux and an index of soil
moisture using adjoint-state surface energy balance. Water Resour. Res. 35(10):31153125
336
S. Liang, J. Qin
Chepurin GA, Carton JA, Dee D (2005) Forecast model bias correction in ocean data assimilation.
Monthly Weather Rev. 133(5):13281342
Cihlar J, Denning AS, Gosz J (2000) Global Terrestrial Carbon Observation: Requirements, Present
Status, and Next Steps, Report of a Synthesis Workshop, 811 February 2000, Ottawa, Canada
Clevers JPGW, Leeuwen V (1996) Combined use of optical and microwave remote sensing data
for crop growth monitoring. Remote Sens. Environ. 56:4251
Courault D, Seguin B, Olioso A (2005) Review on estimation of evapotranspiration from remote
sensing data: from empirical to numerical modeling approaches. Irrig. Drain. Syst. 19:223249
Crow WT (2003) Correcting land surface model predictions for the impact of temporally sparse
rainfall rate measurements using an ensemble Kalman filter and surface brightness temperature
observations. J. Hydrometeorol. 4(5):960973
Crow WT, Wood EF (2003) The assimilation of remotely sensed soil brightness temperature imagery into a land surface model using Ensemble Kalman filtering: a case study based on ESTAR
measurements during SGP97. Adv. Water Resour. 26(2):137149
Dai Y, Zeng X, Dickinson RE, Baker I, Bonan GB, Bosilovich MG, Denning AS, Dirmeyer PA,
Houser PR, Niu G, Oleson KW, Schlosser CA, Yang Z.L (2003) The Common Land Model
(CLM). Bull. Am. Meteorol. Soc., 84(8):1013
Daley R (1991) Atmospheric data analysis. Cambridge University Press, New York.
de Goncalves LGG, Shuttleworth WJ, Burke EJ, Houser P, Toll DL, Rodell M, Arsenault K (2006)
Toward a South America Land Data Assimilation System: Aspects of land surface model spinup using the simplified simple biosphere. J. Geophys. Res. -Atmos. 111(D17)
Demarty J, Ottle C, Braud I, Olioso A, Frangi JP, Gupta HV, Bastidas LA (2005) Constraining a
physically based Soil-Vegetation-Atmosphere Transfer model with surface water content and
thermal infrared brightness temperature measurements using a multiobjective approach. Water
Resour. Res. 41(1):Art. No. W01011
Dobmanin M, Liepelt M, Schittkowski K (1995) Algorithm 746: PCOMP: a Fortran Code for
Automatic Differentiation. ACM Trans. Math. Software 21(3):233266
Doraiswamy PC, Hatfield JL, Jackson TJ, Akhmedov B, Prueger J, Stern A (2004) Crop condition
and yield simulation using Landsat and MODIS. Remote Sens. Environ. 92(4):548559
Doraiswamy PC, Moulin S, Cook PW, Stern A (2003) Crop yield assessment from remote sensing.
Photogramm. Eng. Remote Sens. 69(6):665674
Duan QY, Gupta VK, Sorooshian S (1993) Shuffled complex evolution approach for effective and
efficient global minimization. J. Optim. Theor. Appl. 76(3):501521
Duan QY, Sorooshian S, Gupta VK (1992) Effective and efficient global optimization for conceptual rainfall-runoff models. Water Resour. Res. 28(4):10151031
Duan QY, Sorooshian S, Gupta VK (1994) Optimal use of sce-ua global optimization method for
calibrating watershed models. J. Hydrol. 158:265284
Entekhabi D, Nakamura H, Njoku EG (1994) Solving the inverse problem for soil moisture and
temperature profiles by sequential assimilation of multifrequency remotely sensed observations. IEEE Trans. Geosci. Remote Sens. 32(2):438448
Evensen G (1994) Sequential data assimilation with a non-linear quasi-geostrophic model using
Monte Carlo methods to forecast statistics. J. Geophys. Res. 99(C5):1014310162
Evensen G (2003) The Ensemble Kalman Filter: theoretical formulation and practical implementation. Ocean Dyn. 53:343367
Fang H, Kim H, Liang S, Schaaf C, Strahler A, Townshend GRG, Dickinson R (2007a) Spatially
and temporally continuous land surface albedo fields and validation. J. Geophys. Res. 112, doi:
10.1029/2006JD008377
Fang H, Liang S, Hoogenboom G, Teasdale J, Cavigelli M (2007b) Crop yield estimation through
assimilation of remotely sensed data into DSSAT-CERES. Int. J. Remote Sens. doi: 10.1080/
01431160701408386
Fang H, Liang S, Townshend J, Dickinson R (2007c) Spatially and temporally continuous LAI data
sets based on an new filtering method: examples from North America. Remote Sen. Environ.
doi:10.1016/j.rse.2006.07.026
337
Fischer A, Kergoat L, Dedieu G (1997) Coupling satellite data with vegetation functional models:
review of different approaches and perspectives suggested by the assimilation strategy. Remote
Sens. Rev. 15:283303
Fung I, Kalnay E, Schimel D, Denning S, Doney S, Pawson S (2002) A US Carbon Data Assimilation Program Workshop Report, University of Maryland, College Park
Giering R, Kaminski T (1998) Recipes for Adjoint Code Construction. ACM Trans. Math.
Software 24(4):437474
Giering R, Kaminski T (2002) Generating recomputations in reverse mode AD. In: G Corliss,
A Griewank, FCL Hascoet, U Naumann (eds), Automatic differentiation of algorithms: from
simulation to optimization. Springer, Heidelberg, pp 283291
Goldberg DE (1989) Genetic algorithms in search, optimization and machine learning. AddisonWesley, Reading, MA
Guerif M, Duke CL (2000) Adjustment procedures of a crop model to the site specific characteristics of soil and crop using remote sensing data assimilation. Agric. Ecosyst. Environ. 81:5769
Gupta HV, Sorooshian S, Yapo PO (1998) Toward improved calibration of hydrologic models:
multiple and noncommensurable measures of information. Water Resour. Res. 34(4):751763
Hazarika MK, Yasuoka Y, Ito A, Dye D (2005) Estimation of net primary productivity by integrating remote sensing data with an ecosystem model. Remote Sen. Environ. 94(3):298310
Houtekamer PL, Mitchell HL (1998) Data assimilation using an Ensemble Kalman Filter technique. Monthly Weather Rev. 126(3):796811
Ingber L (1989) Very fast simulated re-annealing. J. Math. Comput. Model. 12(8):967973
Jiang D, Wang NB, Yang XH, Wang JH (2003) Study on the interaction between NDVI profile and
the growing status of crops. Chin. Geogr. Sci. 13(1):6265
Jones JW, Graham WD, Wallach D, Bostick WM, Koo J (2004) Estimating soil carbon levels using
an Ensemble Kalman filter. Trans. ASAE 47(1):331339
Jongschaap REE, Schouten LSM (2005) Predicting wheat production at regional scale by integration of remote sensing data with a simulation model. Agron. Sustain. Dev. 25(4):481489
Julien Y, Sobrino JA, Verhoef W (2006) Changes in land surface temperatures and NDVI values
over Europe between 1982 and 1999. Remote Sens. Environ. 103:4355
Kaminski T, Knorr W, Rayner PJ, Heimann M (2002) Assimilating atmospheric data into a terrestrial biosphere model: a case study of the seasonal cycle. Glob. Biogeochem. Cycles 16(4):Art
No. 1066
Kumar P., Kaleita AL (2003) Assimilation of near-surface temperature using Extended Kalman
Filter. Adv. Water Resour. 26:7993
Kumar SV, Peters-Lidard CD, Tian Y, Houser PR, Geiger J, Olden S, Lighty L, Eastman JL,
Doty B, Dirmeyer P, Adams J, Mitchell K, Wood EF, Sheffield J (2006) Land information system: an interoperable framework for high resolution land surface modeling. Environ. Model.
Software, 21(10):14021415
Li X, Koike T, Pathmathevan M (2004) A very fast simulated re-annealing (VFSA) approach for
land data assimilation. Comput. Geosci. 30(3):239248
Liang S (2004) Quantitative remote sensing of land surfaces. Wiley, New York
Liaqat A, Fukuhara M, Takeda T (2001) Application of neural network collocation method to data
assimilation. Comput. Phys. Commun. 141(3):350364
Liaqat A, Fukuhara M, Takeda T (2003a) Applying a neural network collocation method to an
incompletely known dynamical system via weak constraint data assimilation. J. Mar. Syst.
131(8):16961714
Liaqat A, Fukuhara M, Takeda T (2003b) Optimal estimation of parameters of dynamical systems
by neural network collocation method. Comput. Phys. Commun. 150(3):215234
Losa SN, Kivman GA, Ryabchenko VA (2004) Weak constraint parameter estimation for a simple
ocean ecosystem model: what can we learn about the model and data? J. Mar. Syst. 45(12):
120
MacDonald R, Hall F (1980) Global crop forecasting. Science 208:670679
Margulis S, McLaughlin DB, Entekhabi D, Dunne S (2002) Land data assimilation and soil moisture estimation using measurements from the Southern Great Plains 1997 Field Experiment.
Water Resour. Res. 38(12):1299, doi:10.1029/2001WR001114
338
S. Liang, J. Qin
Marler RT, Arora JS (2004) Survey of multiobjective optimization method for engineering. Struct.
Multidisc. Optim. 26(6):369395
Mass S (1988a) Using satellite data to improve model estimates of crop yield. Agron. J. 80:662
665
Mass SJ (1988b) Use of remotely-sensed information in agricultural crop growth models. Ecol.
Model. 41:247268
Mass SJ (1993) Within-season calibration of modeled wheat growth using remote sensing and field
sampling. Agron. J. 85(3):669672
Merlin O, Chehbouni A, Boulet G, Kerr Y (2006) Assimilation of disaggregated microwave soil
moisture into a hydrologic model using coarse-scale meteorological data. J. Hydrometeorol.
7(6):13081322
Mitchell KE, Lohmann D, Houser PR, Wood EF, Schaake JC, Robock A, Cosgrove BA, Sheffield J,
Duan QY, Luo LF, Higgins RW, Pinker RT, Tarpley JD, Lettenmaier DP, Marshall CH, Entin JK,
Pan M, Shi W, Koren V, Meng J, Ramsay BH, Bailey AA (2004) The multi-institution North
American Land Data Assimilation System (NLDAS): utilizing multiple GCIP products and
partners in a continental distributed hydrological modeling system. J. Geophys. Res. -Atmos.
109(D7):Art. No. D07S90
surface albedos: value-added datasets derived from terra MODIS land products. IEEE Trans.
Morisette JT, Baret F, Privette JL, Myneni RB, Nickeson JE, Garrigues S, Shabanov NV, Weiss M,
Fernandes RA, Leblanc SG, Kalacska M, Sanchez-Azofeifa GA, Chubey M, Rivard B,
Stenberg P, Rautiainen M, Voipio P, Manninen T, Pilant AN, Lewis TE, Iiames JS, Colombo R,
Meroni M, Busetto L, Cohen WB, Turner DP, Warner ED, Petersen GW, Seufert G, Cook R
(2006) Validation of global moderate-resolution LAI products: a framework proposed within
the CEOS Land Product Validation subgroup. IEEE Trans. Geosci. Remote Sens. 44(7):1804
1817
Moulin S, Bondeau A, Delecolle R (1998) Combining agricultural crop models and satellite observations; from field to regional scales. Int. J. Remote Sens. 19(6):10211036
Natvik LJ, Eknes M, Evensen G (2001) A weak constraint inverse for a zero-dimensional marine
ecosystem model. J. Mar. Syst. 28(12):1944
ONeill A, Mathieu PP, Zehner C (2004) Making the most of earth observation with data assimilation. ESA Bulletin 118:3338
Pauwels VRN, Verhoest NEC, De Lannoy GJM, Guissard V, Lucau C, Defourny P (2007) Optimization of a coupled hydrology-crop growth model through the assimilation of observed soil
moisture and leaf area index values using an ensemble Kalman filter. Water Resour. Res. 43(4)
Qin J, Liang S, Li X, Wang J (2007a) Development of the adjoint model of a canopy radiative transfer model for sensitivity study and inversion of leaf area index. IEEE Trans. Geosci. Remote
Sens. (revised)
Qin J, Liang S, Liu R, Zhang H, Hu B (2007b) A weak-constraint based data assimilation scheme
for estimating surface turbulent fluxes. IEEE Geosci. Remote Sens. Lett. 4(4):649653
Qin J, Yan G, Liu S, Liang S, Zhang H, Wang J, Li X (2006) Application of Ensemble Kalman
Filter to geophysical parameter retrieval in remote sensing: a case study of kernal-driven BRDF
model inversion. Sci. China, Series D, 49(6):632640
Raupach MR, Rayner PJ, Barrett DJ, DeFries RS, Heimann M, Ojima DS, Quegan S, Schmullius
CC (2005) Model-data synthesis in terrestrial carbon observation: methods, data requirements
and data uncertainty specifications. Glob. Change Biol., 11(3):378397
Rayner PJ, Scholze M, Knorr W, Kaminski T, Giering R, Widmann H (2005) Two decades of
terrestrial carbon fluxes from a carbon cycle data assimilation system (CCDAS). Glob. Biogeochem. Cycles 19(2):Art. No. GB2026
Reichle RH (2000) Variational Assimilation of Remote Sensing Data for Land Surface Hydrologic
Applications, MIT, Boston, MA
Reichle RH, McLaughlin DB, Entekhabi D (2001) Variational data assimilation of microwave
radiobrightness observations for land surface hydrology applications. IEEE Trans. Geosci.
Remote Sens. 39(8):17081718
339
Reichle RH, McLaughlin DB, Entekhabi D (2002) Hydrologic data assimilation with the Ensemble
Kalman Filter. Month. Weather Rev. 130:103114
Ristic B, Aruampalam S, Gordon N (2004) Beyond the Kalman Filter. Artech House,
Boston/London
Rodell M, Houser PR, Jambor U, Gottschalck J, Mitchell K, Meng CJ, Arsenault K, Cosgrove B,
Radakovich J, Bosilovich M, Entin JK, Walker JP, Lohmann D, Toll D (2004) The global land
data assimilation system. Bull. Am. Meteorol. Soc. 85(3):381
Sacks WJ, Schimel DS, Monson RK, 2007. Coupling between carbon cycling and climate in a
high-elevation, subalpine forest: a model-data fusion analysis. Oecologia 151(1):5468
Sacks WJ, Schimel DS, Monson RK, Braswell BH (2006) Model-data synthesis of diurnal and
seasonal CO2 fluxes at Niwot Ridge, Colorado. Glob. Change Biol. 12(2):240259
Smith A, Doucet A, Freitas ND, Gordon N (eds) (2005) Sequential Monte Carlo methods in practice. Springer, New York
Storn R, Price K (1997) Differential Evolution a simple and efficient heuristic for global optimization over continuous space. J. Global Optim. 11:341359
Storn R, Price KV (1996) Minimizing the real function of the ICEC96 contest by differential
evolution, IEEE conference on evolutionary computation, pp 842844
Su Z (2002) The Surface Energy Balance System (SEBS) for estimation of turbulent heat fluxes.
Hydrol. Earth Syst. Sci. 6(1):8599
Tang YM, Hsieh WW (2001) Coupling neural networks to incomplete dynamical systems via
variational data assimilation. Month. Weather Rev. 129(4):818834
Verma A (2000) An introduction to automatic differentiation. Curr. Sci. 78(7):804807
Vrugt JA, Gupta HV, Bastidas LA, Bouten W, Sorooshian S (2003) Effective and efficient algorithm for multiobjective optimization of hydrologic models. Water Resour. Res. 39(8):Art
No. 1214
Wade TC, Eric FW (2003) The assimilation of remotely sensed soil brightness temperature imagery
into a land surface model using Ensemble Kalman Filtering: a case study based on ESTAR
measurements during SGP97. Adv. Water Resour. 26(2):137149
Walker JP, Houser PR (2001) A methodology for initializing soil moisture in a global climate model: assimilation of near-surface soil moisture observations. J. Geophys. Res. -Atmos
106(D11):1176111774
Walker JP, Willgoose GR, Kalma JD (2001a) One-dimensional soil moisture profile retrieval by
assimilation of near-surface measurements: a simplified soil moisture model and field application. J. Hydrometeorol. 2(4):356373
Walker JP, Willgoose GR, Kalma JD (2001b) One-dimensional soil moisture profile retrieval by
assimilation of near-surface observations: a comparison of retrieval algorithms. Adv. Water
Resour. 24(6):631650
Wang YP, Barrett DJ (2003) Estimating regional terrestrial carbon fluxes for the Australian continent using a multiple-constraint approach I. Using remotely sensed data and ecological observations of net primary production. Tellus Ser B Chem. Phys. Meteorol. 55:270289
Williams M, Schwarz PA, Law BE, Irvine J, Kurpius MR (2005) An improved analysis of forest
carbon dynamics using data assimilation. Glob. Change Biol. 11(1):89
Wu J, Fukuhara M, Takeda T (2005) Parameter estimation of an ecological system by a neural
network with residual minimization training. Ecol. Model. 189(34):289304
Xu T, White L, Hui DF, Luo YQ (2006) Probabilistic inversion of a terrestrial ecosystem model:
analysis of uncertainty in parameter estimation and model prediction. Glob. Biogeochem. Cycles 20(2)
Yu TW, Iredell M, Keyser D (1997) Global data assimilation and forecast experiments using SSM/I
wind speed data derived from a neural network algorithm. Weather Forecast. 12(4):859865
Zhou YH, McLaughlin D, Entekhabi D (2006) Assessing the performance of the ensemble Kalman
filter for land surface data assimilation. Month. Weather Rev. 134(8):21282142
Chapter 13
Methodologies for Mapping Land Cover/Land

Use and its Change
Nina Siu-Ngan Lam
13.1 Introduction
Mapping and identifying land cover/land use and its change is the most important,
as well as the most widely researched, topic in remote sensing. Land cover/land use
has been used extensively to derive a number of biophysical variables, such as vegetation index, biomass, and carbon content (see other chapters). More importantly,
land cover/land use pattern and its change reflect the underlying natural and/or
social processes, thus providing essential information for modeling and understanding many different phenomena on the Earth. Knowledge of land cover/land use and
its change is also critical to effective planning and management of natural resources.
Mapping land cover/land use accurately and efficiently via remote sensing requires good image classification methods. Unfortunately, there are numerous factors
(e.g., image resolution and atmospheric condition) that could affect the effectiveness
and accuracy of the classification algorithms. Different land cover/land use classification methods may be needed for different problems under different environmental
conditions, making generalization and hence automation of the image classification
process across time and space extremely difficult. As a result, new and sophisticated
classification methods designed to improve the classification process continue to appear in the literature (e.g., Jensen, 2005; Gong, 2006). Newer approaches such as
fuzzy classification, artificial neural network, and object-based classification have
been developed and successfully applied (Definiens, 2004; Benz et al., 2004). However, these methods require extensive training and human supervision. We are still
far from being able to develop a common framework to successfully identify a variety of features in different landscapes and to generalize and automate the classification process.
Nina Siu-Ngan Lam

Department of Environmental Studies, Louisiana State University, Baton Rouge, USA
nlam@lsu.edu
341
342
N.S.-N. Lam
Extending the mapping and modeling of land cover/land use at one time period
to multiple time periods to analyze change will undoubtly add more complexity
and challenges. In addition to the above image classification issues, efficient methods are needed to ensure comparability and compatibility of images taken in different time periods. Many studies on change detection using remote sensing imagery
have already been reported in the literature. Lunetta and Elvidge (1998) provided
an excellent summary of the state of the science. Many others examined the performance of various techniques in various applications (e.g., Coppin and Bauer, 1996;
Jensen et al., 1993, 1997; Lu et al., 2005; Mas, 1999; Nackaerts et al., 2005; Yuan
et al., 1998).
Among the new techniques for land cover/land use classification and change
analysis, textural (spatial) analyses are gaining increasing attention from the remote sensing community (e.g., Briggs and Nellis, 1991; Dunn et al., 1991; Estreguil
and Lambin, 1996; Frank, 1984; Jupp et al., 1986; Lambin, 1996; Lambin and
Strahler, 1994; Pickup and Foran, 1987; Smits and Annoni, 2000; Crews-Meyer,
2002). We have seen new applications of old textural measures such as the spatial co-occurrence matrix, local variance, and others (Haralick et al., 1973; Clausi
and Jobanputra, 2006), as well as development of new textural analytical techniques such as wavelets (Daubechies, 1990; Muneeswaran et al., 2005). A number
of textural measures which had not been used for remote sensing applications before have recently been utilized for more accurate land cover/land use classification, such as fractals, variograms, lacunarity, and spatial autocorrelation statistics
(Lam, 1990; Lam and De Cola, 1993; Plotnick et al., 1993; Lam et al., 1998; Carr
and de Miranda, 1998; Carr, 1999; Dale, 2000; Dong, 2000; Franklin et al., 2000).
Although applications of these newer textural measures in change analysis have seldom been reported, we expect that the same textural and spatial methods that can be
used for identification of land cover/land use can also be used for change detection.
The purpose of this chapter is to introduce the use of textural/spatial measures
in land cover/land use classification and its potential for change analysis. Our main
notion is that the utilization of textural/spatial measures, in combination with original spectral information, will increase classification accuracy and have great potential for rapid change detection. The chapter is organized into four main sections. A
summary of the major textural/spatial methods is first provided in Section 13.2, with
a focus on those measures that can be applied directly to unclassified images. This
property is important for rapid image segmentation, classification, and change detection. Section 13.3 describes a number of examples that have utilized these measures
in land cover/land use classification. In Section 13.4, a framework for classifying
the various land cover/land use change detection methods is introduced. We argue
for the need to develop innovative methods for rapid and reliable change detection
especially during disastrous and unexpected events. We further argue that such need
could be best served by utilizing a metric of textural/spatial measures, in conjunction
with the original spectral information of the images. Section 13.5 describes a real
example of change analysis using textural measures. The prospect of utilizing textural measures in combination with other approaches for better and faster mapping
of land cover/land use and its change is summarized in the conclusion.
13 Methodologies for Mapping Land Cover/Land Use and its Change
343
13.2 Major Textural/Spatial Measures

13.2.1 Terminology
First, some clarification of the terms textural vs spatial measure is in order. The term
textural is more commonly used in pattern recognition and the general field of
image processing for raster/pixel data, while the term spatial is usually adopted in
geography, economics, statistics and other related disciplines and is derived mostly
for vector/polygonal data.
Texture is an important characteristic in many types of images. Despite its importance, a formal definition of texture does not exist. Haralick (1979) attempted
to characterize texture using two properties, the tonal primitive properties as well
as the spatial interrelationships among them. Under this two-layered tone-texture
concept, when an image has little variation of tonal primitives, the dominant property of that image is tone. On the contrary, when an image has wide variation of
tonal primitives, the dominant property of that image is texture. Haralick described
eight statistical approaches to measure image texture, which include autocorrelation functions, optical transforms, digital transforms, textural edgeness, structural
elements, spatial gray tone co-occurrence probabilities, gray tone run lengths, and
autoregressive models.
In remote sensing, texture is the spatial relationship exhibited by gray levels in a
digital image. Therefore, textural measures are measures that capture the spatial relationship among pixels. Spatial measures, which refer to measures mostly derived
from spatial statistics, have been used largely in geospatial applications for characterizing and quantifying spatial patterns and processes Although traditionally the
two fields of studies, textural analysis and spatial analysis, refer to quite different
sets of methods and analyses, they do intersect to a great extent in remote sensing.
In this chapter, we adopt the notion that the two terms, textural and spatial measures,
are interchangeable.
13.2.2 Criteria for Evaluating and Classifying Textural Measures

There are numerous textural measures in the literature, and it is beyond the scope of
this chapter to exhaust and evaluate each of them. Table 13.1 lists some of the commonly used measures, which illustrates clearly the diversity, as well as redundancy,
of these various measures. It is important to note that some measures are based
on vigorous statistical theory and mathematical derivation (e.g., fractals, wavelets,
spatial autocorrelation) (Lam et al., 1998), while others are based on simple geometric measurements with unknown statistical properties and/or theoretical minimum
or maximum (e.g., edge density). Some metrics measure one aspect of the landscape (e.g., landscape composition), while others measure another (e.g., landscape
configuration) (McGarigal and Mark, 1995). Baskent and Jordan (1995) classified
344
N.S.-N. Lam
Table 13.1 Some commonly used textural measures

Textural measures
References
First-order metrics (computed on the original data matrix)

using traditional statistical measures: mean, standard deviation, variance, correlation
First-order metrics using traditional texture measures: entropy, energy, contrast, homogeneity, angular second moment, Shannon diversity
First-order metrics using adapted spatial measures (from
ICAMS): fractal dimension, lacunarity, spatial autocorrelation, variogram
Second-order metrics: same set of metrics calculated on
matrix derived from the original matrix (e.g., gray-level
co-occurrence matrix, wavelet decomposed images)
Landscape indices designed for classified images (from
FRAGSTATS): area, density, edge, shape, proximity, connectivity, contagion/interspersion, diversity
Jensen (2005); Gong et al. (1992);

Woodcock and Strahler (1987)
Haralick (1979); Haralick et al.
(1973); Jensen (2005); Gong et al.
(1992)
Lam et al. (1998, 2002);
Quattrochi et al. (1997); Myint and
Lam (2005a, b)
Haralick (1979); Haralick et al.
(1973); Myint et al. (2002, 2004)
McGarigal and Mark (1995);
McGarigal (2002)
landscape indices into areal, linear, and topological. Yet another classification of
textural measures is based on whether the measures are applied directly to the original matrix (first-order textural measures), or to matrices derived from the original
matrix such as the gray-level co-occurrence matrix or wavelet decomposed images
(second-order textural measures) (Jensen, 2005). Hence, it is imperative to develop
useful criteria to evaluate and/or classify these textural measures. What is a good
textural measure? Although there may be no definite answers until each measure
is tested extensively for their discriminating and explanatory power, we suggest the
following criteria to evaluate and guide our understanding of these various measures.
Ideally, a good textural measure should have the following properties:
1. The textural measure should be conceptually simple and easy to calculate. For
example, statistical mean and standard deviation are concepts easily grasped
by most researchers and their statistical properties are well known. Extension
of these basic statistical measures in a spatial domain with some modifications
may provide a useful first approximation towards an understanding of the land
cover/land use pattern being studied. On the contrary, some measures may be
conceptually simple but require additional steps for calculation, such as edge
density, which needs an additional step to find the edges.
2. The textural measure should have theoretical maximum and minimum. For example, Morans I, a most commonly used spatial autocorrelation statistic, has
a range of 1. A Morans I value of 1 indicates a maximum positive spatial
autocorrelation; on the contrary, a 1 indicates a maximum negative spatial autocorrelation (Cliff and Ord, 1973).
3. The textural measure should reflect clearly and intuitively the characteristics
of the image pattern in a consistent manner. For example, a lower fractal
dimension value means a less spatially complex image, therefore, given an image
345
computed with a fractal dimension value of 2.3, we should be able to infer that
this image is far less complex than an image computed with a fractal dimension
value of 2.9, and a visual display of the two images should be able to reveal
the difference (Lam, 1990). Fractal dimension (D) also has the second property,
where D is expected to range from 2 to 3.0 for surfaces and 1.02.0 for lines
(Mandelbrot, 1982).
4. The statistical properties of the textural measure should be known to provide statistical confidence of the computed value. For example, theoretically a Morans
I value of 0 indicates a random pattern. If a pattern yields a computed value
of 0.2, can we determine if this value is statistically the same or different from
0 to conclude if the pattern is random or not? Fortunately, the statistical properties of Morans I are relatively well known and hypothesis testing of whether
a computed I value is significant can be conducted. Under the assumption of
randomization, the first and second moments of the Morans I value can be computed and the statistical significance of the value determined (Goodchild, 1986).
On the contrary, the statistical properties of fractal dimension are still not clear,
though it has well-defined theoretical minimum and maximum. Hence, it is difficult to judge, for example, if an image with a fractal dimension of 2.3 is significantly different from another image with a fractal dimension of 2.4. It is noted
that the statistical properties of most spatial measures are very difficult to derive and therefore remain unclear; many researchers have resorted to the Monte
Carlo approach to develop empirical probability functions for statistical hypothesis testing (e.g., Openshaw, 1989).
5. The textural measure should be computable globally for the entire study area
or locally for a local neighborhood. For example, some landscape metrics developed in FRAGSTATS are only computable at the landscape level, instead of at all
levels (patch, class, and landscape) (McGarigal, 2002), whereas mean, variance,
Morans I, and fractals can be applied both globally and locally to capture local
change (Lam, 2004; Emerson et al., 2005). This property refers only to whether
the measure can be computed at all levels; it does not necessarily imply that the
measure is useful in describing the landscape at all levels.
6. Finally, the textural measure should be applicable directly to both classified and
unclassified images. For example, the landscape metrics in FRAGSTATS were
developed exclusively for categorical maps (ONeill et al., 1998; McGarigal,
2002), or in other words, classified images, though some of the metrics can be
modified and applied to unclassified images. On the contrary, fractals, Morans I,
local variance (Woodcock and Strahler, 1987), variogram, lacunarity, and wavelet
measures can be applied to both unclassified and classified images.
This last property is considered very important to automated land cover/land use
classification and change detection for two reasons. First, if they can be applied
directly to unclassified images, land cove change could be detected first before
the tedious classification process. Only after the change is determined to be significant, then there is a need to identify or classify what the changes are. This is
considered a more efficient approach, especially for continuous environmental monitoring. Second, since these textural methods measure the spatial variations among
346
N.S.-N. Lam
pixels instead of comparing pixel by pixel, they are more likely to reflect dominant
changes rather than spurious changes that might have resulted from using images
taken in different time periods (Lambin, 1996; Smits and Annoni, 2000). If there
are only small and insignificant changes in land cover, it is expected that the spatial
relationship will not alter and the spatial index values will remain the same. On the
contrary, if there are significant land-cover changes, then it is expected that the spatial properties will be altered, and the spatial indices that are designed to measure
the spatial properties should be able to capture these changes.
13.2.3 Description of Selected Textural Measures

We describe below four textural measures, including spatial autocorrelation
(Morans I), fractal, lacunarity, and wavelet transform. All four methods have
already been implemented in a software module called ICAMS (Image Characterization And Modeling System), developed previously by the author and collaborators
(Quattrochi et al., 1997; Lam et al., 1998, 2002; Emerson et al., 1999). ICAMS was
developed mainly to provide spatial analytical tools, such as fractals, variograms,
and spatial autocorrelation, to visualize, measure, and characterize landscape patterns. Detailed descriptions of ICAMS can be found in a number of publications
(e.g., Quattrochi et al., 1997; Lam et al., 1997, 1998). The four methods were selected for illustration because they at least have properties 5 and 6 (can apply locally
and to unclassified images), which are important properties for rapid classification
and change detection. Furthermore, Morans I possesses all properties, fractals have
all but property 4. Lacunarity is an old measure but its use in land cover/land use
analysis is relatively new, its properties remain to be thoroughly studied (Myint and
Lam, 2005a, b). The mathematics of wavelet transformation is well defined. However, the properties of the textural measures computed on the wavelet transformed
images have seldom been studied, and will be a subject of our ongoing research.
13.2.3.1 Spatial Autocorrelation

Spatial autocorrelation statistic has been commonly used to measure the degree of
clustering, randomness, or fragmentation of a spatial pattern. The two most common
spatial autocorrelation measures for interval-ratio data are Morans I and Gearys C
(Cliff and Ord, 1973). Morans I is generally preferred over Gearys C, because
the values of the former are more intuitive (i.e., positive values for positive autocorrelation and vice versa). Morans I was also found to be generally more robust
(Goodchild, 1986). These two measures were originally developed for measuring
polygonal data (e.g., geographical regions) where the number of units or points
measured (n) is often smaller. It is only recently that these two measures were applied to raster data such as remotely sensed images, where the number of units being
347
measured (n n) is much larger (Emerson et al., 1999; Lam et al., 2002). Morans
I is calculated from the following formula:
I(d) =
ni nj wi j zi z j
w ni z2i
(13.1)
Z-value
0 90 181
where wi j is the weight at distance d so that wi j = 1 if point j is within distance

d from point i; otherwise, wi j = 0; zs are deviations from the mean for variable y, and w is the sum of all the weights where i = j. Morans I varies from
+1 for perfect positive correlation (a clumped pattern) to 1 for perfect negative
correlation (a checkerboard pattern). Figure 13.1 shows three simulated surfaces and
their corresponding Morans I and fractal dimension values. Morans I and fractal
2
2
Z-value
lu
Co
0 72 145 217
ns
lum
Co
ws
Ro
2
1
Z-value
0 72 145
mn
ws
Ro
1
0
ns
lum
Co
ws
Ro
Fig. 13.1 Three simulated surfaces mapped in three-dimensional and image forms. From top to
bottom, D = 2.1, 2.5, 2.9, and I = 1.0, 0.99, 0.82. (Modified from Lam et al., 2002.)
348
N.S.-N. Lam
dimension (D) have an inverse relationship, whereby a spatial pattern with a high
degree of fragmentation will have a low Morans I but a high fractal dimension.
13.2.3.2 Fractal Dimension

Fractals were derived mainly to overcome the difficulty in analyzing spatial forms
and processes by classical Euclidean geometry. The key parameter in fractals is
fractal dimension D, which is used to represent the complexity of spatial forms
and processes. The higher the D, the more complex is the curve or surface. The
D value of a curve can be any non-integer value between 1 and 2, and a surface
between 2 and 3. For example, coastlines have dimension values typically around
1.2, and topographic surface dimensions around 2.3. Dimension values for satellite
image surfaces have been reported to be much higher, and depending on the type of
landscapes examined, they can be as high as 2.72.9 (Lam, 1990; Jaggi et al., 1993).
Fractal dimension is derived from the concept of self-similarity, where a curve or
a surface is made up of copies of itself in a reduced scale (Mandelbrot, 1982). The
number of copies (m) and the scale reduction factor (r) can be used to determine
the dimensionality of the curve or surface, where D = log(m)/ log(r) (Falconer,
1988). Practically the D value of a curve is estimated by measuring the length of the
curve using various step sizes. The more irregular the curve, the greater increase in
length as step size decreases. Such an inverse relationship between total line length
and step size can be captured by a regression:
log(L) = C + B log(G)
(13.2)
where L is the line length, G is the step size, B is the slope of the regression, and C
is a constant. D can be calculated by: D = 1 B. For surfaces, D = 2 B.
Because of its attractive theoretical foundation, literally every major discipline
has found applications using the fractal concept in the past two decades, with numerous algorithms developed for computing the fractal dimension. Unfortunately, a
major problem in applying fractals is that different fractal measurement algorithms
yield different results. Often times, empirically computed fractal dimensions may
exceed the theoretical ranges. Moreover, fractal dimension is defined in various
ways in different algorithms. For example, some algorithms use only a single measurement to derive the dimension, instead of using multiple step sizes to derive the
dimension through regression analysis. FRAGSTATS defines fractal dimension as
the ratio between perimeter and area of a patch, which is very different from the
algorithms described below (Lam, 1990). The former definition of fractal dimension, though simple and easy to calculate, applies only to images that have already
been classified, whereas the algorithms described below (e.g., the triangular prism
algorithm) follow closely the original definition by Mandelbrot and can be applied
directly to unclassified images for textural comparison.
Lam (1990) demonstrated the use of three methods, including isarithm, triangular prism, and variogram methods, in measuring the spatial complexity of the
reflectance surfaces from remote sensing imagery (Goodchild, 1980; Clarke, 1986;
349
Mark and Aronson, 1984). In a subsequent benchmark study, Lam et al. (2002)
found that the modified triangular prism method was the most accurate and reliable method for estimating the fractal dimension of surfaces. Hence, the modified
triangular prism method is described as follows.
The modified triangular prism method (Clarke, 1986; Jaggi et al., 1993; Lam
et al., 2002) constructs triangles by connecting the heights or z-values at the four
corners of a grid cell to its center, with the center height being the average of the
pixels at the four corners. These triangular facets of the prism are then summed
to represent the surface area. In the second step, the algorithm increases the step
size from one pixel to two pixels, and the z-values at the four corners of the 2 2
composites are used to construct the prism. It is expected that as step size increases,
the prism surface area will increase, but at a decreasing rate, which can then be used
to determine the fractal dimension by a regression equation similar to Eq. (13.2):
Log A = K + (2 D) Log S, where A is the prism surface area, K is a constant, D is
the fractal dimension, and S is the pixel size.
13.2.3.3 Lacunarity
Despite the potential of fractals, Mandelbrot (1982) realized that fractal dimensions
are very far from providing a complete characterization of spatial forms. He introduced the term lacunarity (lacunar in Latin means gap) to further describe the
gappiness or texture of a spatial pattern. In other words, different fractal sets may
have the same fractal dimension values, but they may look different because they
have different lacunarities (Myint and Lam, 2005a).
Lacunarity represents the distribution of gap sizes; low lacunarity implies homogeneity as all gap sizes are the same, whereas high lacunarity implies heterogeneity
(Dong, 2000). Unfortunately, lacunarity is highly sensitive to scale, and depending
on the size of the gliding box used in computing the lacunarity value, the same
pattern can return with very different values, as objects that are homogeneous at a
small scale can be heterogeneous at a large scale (Plotnick et al., 1993). Myint and
Lam (2005a, b) compared several hypothetical patterns; three of them are shown
in Fig. 13.2. When a smaller gliding box of 3 3 is used, the small gap pattern
(Fig. 13.2) results in low lacunarity (1.05), the big gap pattern yields the highest
(1.40), and the random pattern yields a value in between (1.15). But when a bigger gliding box of 11 11 is used, the results are reverse, with the big gap pattern
yielding the lowest lacunarity (1.02) and the small gap pattern the highest (1.10).
Lacunarity value of the random pattern decreased slightly from 1.15 to 1.08. It is
also observed that the range of difference between the two patterns is much smaller
with bigger gliding box.
An algorithm for computing lacunarity using the gray-scale approach is described as follows (Voss, 1986; Myint and Lam, 2005a, b). Let P(m, L) be the probability that there are m intensity points within a cube size of L centered at an arbitrary
point in an image. Intensity points are points that fill the cube in each step. Hence,
we have
350
N.S.-N. Lam
Fig. 13.2 Three hypothetical binary patterns with different lacunarity values, see text for explanations. (Modified from Myint and Lam, 2005a.)
N
P(m, L) = 1
(13.3)
m=1
where N is the number of possible points in the cube of L. Suppose that the total
number of points in the image is M. If one overlays the image with cubes of side L,
then the number of cubes with m points inside the cube is (M/m)P(m, L). Hence
N
M(L) =
mP(m, L)
(13.4)
m=1
and
M 2 (L) =
m2 P(m, L)
(13.5)
m=1
Lacunarity (L) can be computed from the same probability distribution P(m, L),
and is defined as:
M 2 (L) (M(L))2
(13.6)
(L) =
(M(L))2
Unlike fractals, lacunarity has no theoretical maximum or minimum. The performance of the index, especially its high scale dependency, will need to be further
studied. However, a few studies have shown that adding a lacunarity layer in image
classification has dramatically improved accuracy (Myint and Lam, 2005a, b), indicating a promising approach towards more accurate, automated land cover/land use
mapping.
13.2.3.4 The Wavelet Transform Method
Pioneered by Mallat (1989) and Daubechies (1990), the wavelet method has found
numerous applications in a wide range of disciplines. The method has also recently
been demonstrated as a promising approach to increasing accuracy in image classification and image retrieval using remote sensing imagery (Manjunath and Ma,
1996; Zhu and Yang, 1998; Bian, 2003; Myint et al., 2004).
In brief, wavelets are translated and dilated versions of a common mathematical
function, called the mother wavelet. In the case of images, the translation refers to
351
Fig. 13.3 Multiresolution wavelet decomposition of a remote sensing image. (a) Original image,
(b) level-1 decomposition: upper left is approximate sub-image, clockwise from upper left are horizontal, diagonal, and vertical detailed sub-images, (c) level-2 decomposition image (From Myint
et al., 2004, reprint with permission from the American Society for Photogrammetry and Remote
Sensing)
the geographic location, and the dilation relates to different scales. By adjusting the
translation and dilation parameters, we can study the texture and scale locally. For
the 2D discrete wavelet transform, which is used for remote sensing image analysis,
the wavelet method will decompose an image into four sub-images: an approximate
image (low frequency) and three detailed images (high frequency horizontal, vertical, and diagonal). The approximate image can further be decomposed into another
level, resulting in a multi-resolution wavelet analysis. Figure 13.3 shows an example
of multiresolution wavelet decomposition. The coefficients of the four subimages
are computed by Eqs. (13.7)(13.10):
A(i, j) = h(k 2i)h(l 2 j) f (k, l)
(13.7)
H(i, j) = h(k 2i)g(l 2 j) f (k, l)
(13.8)
V (i, j) = g(k 2i)h(l 2 j) f (k, l)
(13.9)
D(i, j) = g(k 2i)g(l 2 j) f (k, l)
(13.10)
where f is the original image, A is the approximate image, H is the horizontal detailed image, V is the vertical detailed image, and D is the diagonal detailed image.
h(k), g(k) are the scaling filter and the wavelet filter, respectively, and k, l are the
number of rows and columns (Mallat, 1989; Daubechies, 1990).
After decomposition, indices can be computed for each sub-image at each level
to represent the texture of the image. In addition to mean and standard deviation,
Eqs. (13.11)(13.14) show other commonly used measures (Myint et al., 2002),
including log energy, Shannon index (SHAN), angular second moment (ASM), and
entropy:
352
N.S.-N. Lam
K
energy =
log(P(i, j)2 )
(13.11)
i=1 j=1
K
SHAN =
P(i, j) log(P(i, j))
(13.12)
i=1 j=1
ASM =
P(i, j)2
(13.13)
i=1 j=1
K
entropy =
|P(i, j)|2
Q(i, j) log |Q(i, j)|; Q(i, j) = ,
i=1 j=1
|P(i, j)|
(13.14)
i, j
where P(i, j) is the (i, j)th pixel wavelet coefficient value of a decomposed image
at a particular level. These computed textural indices are then used to discriminate
different types of land cover/land use.
13.2.4 Scale and Uncertainty in Textural Analysis

Textural measures must be computed from a group of pixels or objects. Hence, texture is very scale-dependent. The size of the moving window combined with the
resolution of the imagery plays a big part in determining what features are highlighted by these techniques. This scale and uncertainty issue has long been a central
concern across a number of disciplines that involve geospatial/environmental data,
textural analysis is not an exception. Scale variations are well known to constrain
the detail with which information can be observed, represented, analyzed, and communicated (Lam et al., 2004).
It should be noted that the term scale has different meanings, and depending
on the field of study, its meaning could be opposite. Lam and Quattrochi (1992)
outlined four different meanings of scale (Cao and Lam, 1997). Cartographic scale
refers to the degree of reduction in spatial dimensions that occurs when real-world
measurements are represented in hard-copy maps or computer screens. Operational
scale is an expression of the spatial or temporal dimensions over which a process operates, while observational scale refers to the dimensions within which a particular
phenomenon or process is observed. Measurement scale, commonly called resolution, refers to the smallest observable unit, such as pixels in remote sensing imagery.
In landscape ecology, observational scale is the extent, whereas measurement scale
is referred to as grain. Radiometric scales also exist within digital imagery. Older
Landsat Multispectral Scanner (MSS) images have a 6-bit grayscale depth, while
IKONOS images have a 16-bit depth. One of the basic goals of scale-related research is to be able to move up and down spatial scales, within disciplines and across
disciplines, so that the results concluded at one scale can be inferred to another
353
scale. Extrapolation of results across broad spatial scales remains the most difficult
problem in environmental research (ONeill et al., 1989; Turner et al., 1989; Lam
and Quattrochi, 1992; Quattrochi and Goodchild, 1997; Tate and Atkinsons, 2001).
Scale affects change detection. The myriad spatial, spectral, radiometric, and
temporal scales of remotely sensed imagery pose a real challenge to change detection, as techniques developed for imagery with a pixel resolution of 1 m (IKONOS
imagery) may not be applicable to imagery with a pixel resolution of 1 km (AVHRR
imagery). Since changes may occur at different scales, globally, regionally, or locally, and they may also occur rapidly or slowly, it is important to examine how
change detection methods and indices perform at different spatial scales.
Scale-related uncertainty in modeling results has significant impacts on decision
making, and basic research on decision making under uncertainty is necessary. Effective scale-related research requires interdisciplinary efforts of social, physical,
and computer scientists. Scale and scale-related uncertainty is a difficult problem
to tackle. Increasingly, it has been recognized that scale effects exist and can never
be eliminated, therefore strategies must be developed to understand and mitigate
the scale effects rather than to eliminate them. Two interrelated approaches were
suggested to mitigate the scale effects (Lam et al., 2004). The first approach is to
develop techniques to detect the scale ranges within which levels of observation are
phenomena scale-dependent. Techniques such as geographic variance, variograms,
correlograms, fractal analysis, and a number of textural methods have been proposed to detect the range of scales that yield the most information (Emerson et al.,
1999). The second approach is to develop a multi-scale assessment module so that
the same analysis can be conducted in multiple scales to compare the results and
estimate the uncertainty. A thorough benchmark study is very much needed to examine how textural methods perform at different spatial scales and resolutions in
land cover classification and change detection.
13.3 Land Cover Classification Using the Textural Approach:

Examples
13.3.1 Characterizing Land Cover in the Tropics
Read and Lam (2002) compared the performances of selected textural measures to
characterize different land covers using two sets of unclassified Landsat-TM data
(1986 and 1996/97) for a site in north-eastern Costa Rica. The purpose was to
determine whether landscape complexity can be captured by these methods, and
whether these methods can be used to reflect the degree of human disturbance. The
hypothesis was that complexity in a natural landscape decreases with increasing intensity of human activities. The methods evaluated were: (1) fractal dimension using the isarithm method, (2) fractal dimension using the modified triangular prism
method, (3) spatial autocorrelation using Morans I, (4) Shannons diversity index,
(5) contagion, and (6) fractal dimension from perimeter/area. The first three methods
354
N.S.-N. Lam
were available from ICAMS, and the last three were landscape metrics available
from FRAGSTATS.
The results revealed that fractal dimension using the triangular prism method and
Morans I could serve as indices for characterizing spatial complexity of LandsatTM data, whereas the landscape indices were not consistent. The fractal dimension
decreased along a gradient of increasing human disturbance: forestscrubpasture
agriculture. This study is among the first to examine how spatial indices can be used
to examine hypotheses related to land cover/land use and human disturbance in the
tropics.
13.3.2 Improving Urban Land Cover Classification

Emerson et al. (2005) examined the utility of local variance, fractal dimension, and
Morans I in improving urban land cover classification. Landsat ETM+ imagery of
Atlanta, Georgia obtained in 1999 was examined. Using the routines in ICAMS,
texture images were computed from the 15m panchromatic band using a 21 21
pixel moving window for every other row and column. This two-pixel increment
between rows and columns produced a 30m resolution texture image. The real number local variance, Morans I, and fractal dimension values (computed using the
modified triangular prism method) were converted to 8-bit image, and they were
added to the stack of multispectral bands for classification using a supervised maximum likelihood classification technique. Five land cover classes based on the USGS
Anderson Level 1 classification scheme were used, including low intensity urban,
high intensity urban, pasture/grassland, forest, and water.
Results show that classification accuracy improved with additional texture layer,
with the fractal dimension band being the most effective. By adding the fractal dimension band to the multispectral bands, the overall percent correctly classified
increased from 67.1% to 77.3%. Although not as effective as the fractal band, addition of local variance and Morans I still yielded an improved accuracy of 72.4% and
69.4%, respectively. The results show great promise, but further research is needed
to better utilize these indices.
13.3.3 Urban Feature Discrimination Using Wavelets

Based on a high-resolution ATLAS (Advanced Thermal Land Application Sensor)
image of Baton Rouge, Louisiana, USA, Myint et al. (2002, 2004) introduced the
wavelet approach for urban land cover classification. The ATLAS image had a
2.5 m pixel resolution and was acquired with 15 channels (0.4512.2 m) from a
NASA LearJet on May 7, 1999. Six urban land cover/land use classes with different
texture appearances were selected, including agriculture, commercial, woodland,
water body, single-family homes with less than 50% tree canopy (residential-1), and
single-family homes with more than 50% tree canopy (residential-2). These land
355
cover classes were based on classification by Lo et al. (1997), which was designed
for the purpose of urban planning, as information on surface vegetation and water
availability are crucial for city officials and environmental agencies in developing
better urban infrastructure. Based on previous studies, band 2 (0.520.60 m), band
6 (0.760.90 m), and band 12 (9.6010.20 m) were selected. In addition to these
three bands, principal component analysis band 1 (PCA1) was also examined to see
if a composite band could produce better accuracy.
Two segmented regions of each class were identified, and five training pixels
were then randomly selected from each region, leading to a total of 10 samples for
each class. Windows of 65 65, 33 33, and 17 17 pixels were selected using
these 10 pixels as centers. Textural measures of these samples were computed and
a linear discriminate analysis was applied to evaluate which measure is the most effective in discriminating the different land covers. Four different textural approaches
were evaluated, including the wavelet transform, spatial autocorrelation, spatial cooccurrence matrix, and fractals. It was found from both studies that the wavelet approach was the most accurate among all approaches considered (Myint et al. 2002,
2004). When 65 65 samples were used, the wavelet approach yielded 100% accuracy. The overall accuracy, however, decreased with smaller window sizes, with
an accuracy of 93% and 78%, respectively, for 33 33 and 17 17 samples. These
studies demonstrated the great potential of using the texture approach. They also
highlighted the importance of different scale parameters such as window size in affecting its performance. Future studies that systematically examine the effects of
scale on the certainty, or rather uncertainty, of the results are needed.
13.3.4 Lacunarity-Based Urban Classification

Myint and Lam (2005a, b) introduced the use of lacunarity in urban land cover/land
use classification. As mentioned in Section 13.2.3, lacunarity measures the gappiness of a pattern; low lacunarity implies homogeneity, whereas high lacunarity suggests heterogeneity. An IKONOS image at 4 m spatial resolution with four channels
acquired over Norman, Oklahoma, on March 20, 2000 was used. The selected landuse and land-cover classes in this study included single-family houses with less
than 50% tree canopy (residential-1), single-family houses with more than 50% tree
canopy (residential-2), commercial, woodland, agriculture, golf course, and water
body. Lacunarity measures for band 4 (0.760.90 m), band 3 (0.630.69 m), and
band 2 (0.520.60 m) were computed for the entire image using the gray-scale
method with a gliding cube of size 3 and a moving widow size of 29 29. The
lacunarity-transformed bands were then stacked as additional bands for maximum
likelihood classification.
The results show that adding three lacunarity transformed bands to the original three spectral bands (band 4, 3, 2) increased the overall classification accuracy
from 55% to 92%, a drastic improvement. If only the three lacunarity transformed
bands were used for classification, the overall accuracy still increased, but not substantially, to 68%. This study further confirmed the study by Emerson et al. (2005)
356
N.S.-N. Lam
discussed above and demonstrated clearly that an integrated textural and spectral approach is needed for more accurate land cover/land use classification and mapping.
13.4 Land Cover/Land Use Change Analysis

Obviously, extending the mapping and modeling of land cover/land use at one time
period to multiple time periods to analyze change adds lot more complexity and
challenges. We outline in this section the inherent difficulties of change detection,
summarize the existing methods into a framework, and then argue that the textural
approach has potential for rapid change detection.
13.4.1 Change Detection Issues

There are inherent difficulties involved in using time-series remote sensing data for
land cover/land use change detection. Ideally, same type of images that have the
same spectral, radiometric, spatial, and temporal resolutions should be used. However, this may not be possible especially for change studies that involve a longer
time span. For example, Landsat-MSS with a pixel resolution of 80 m started in
1972, whereas Landsat-TM with a 30 m pixel resolution became available in 1982.
Using digital imagery for change analysis before 1972 would be difficult. Often
times, old aerial photographs before these dates are the only image sources to be
used. The imminent danger of discontinuing global coverage due to sensor malfunction or budget constraints, such as Landsat-7 ETM+, will definitely hamper
land cover/land use research and make long-term change analysis impossible.
When two very different types of images are involved, the only viable approach
to change detection is to conduct detailed image classification of individual images
and then overlay the two classified images to assess the changes. However, even with
the same sensor, images taken in different dates may be affected by several factors,
causing them to be different even if there is no real land cover change. Therefore,
the following factors need to be considered for more accurate change detection:
1. The dates of the two images should be approximately the same to avoid seasonal
difference in vegetation, soil moisture, sun-angle, and other response that are not
real land cover change.
2. Even with the same date but in different years, individual atmospheric condition
can obscure our ability to uncover real changes, such as cloud cover or rainfall.
An image taken right after rainfall in a desert will look very different from an
image of the same desert before rainfall. Atmospheric correction of each image
may need to be applied before the analysis. For coastal landscape, tidal stage
is a crucial factor in conducting change detection. Significant changes can be
observed simply because the images were taken at different tidal stages (Jensen,
2005).
357
3. Pre-image processing steps may also contribute errors. Extra caution is needed
to ensure no pixel mis-registration between the two images. A single pixel shift
will shift the entire image and that could lead to substantial error in assessing
change. Another point that has seldom been mentioned in the literature refers to
the algorithm used to convert pixel values from analog to digital scale. Assuming an 8-bit scale (0255) is used, some algorithms will convert the continuous
signal using the images minimum and maximum values as the limit, whereas
others use the 99% or 95% interval. The result is that the same digital number
in different images may have very different actual radiometric value, and the
value is only true relative to the rest of the pixel values in its own image. Hence,
change detection methods that involve direct pixel-by-pixel spectral comparison
could be misleading, whereas change detection methods that are based on ratios
among bands within its own image are more reliable. By the same token, it is
expected that comparing the textural difference between two images, instead of
pixel-by-pixel spectral comparison, would yield more accurate change analysis.
13.4.2 A Classification of Change Detection Methods

New methods for change detection using remote sensing imagery have been continuously reported in the literature, and it is not the scope of this chapter to exhaust
the list and provide evaluation of each approach. However, the following framework
may be useful to classify the various change detection methods. Figure 13.4 shows
how the various commonly used methods are placed in this framework.
Fig. 13.4 A framework for classifying change detection methods
358
N.S.-N. Lam
Change detection methods can be differentiated into two main groups, depending on whether the method requires classification before or after the changes are
detected. As shown in Fig. 13.4, the first group of change detection methods, which
is also the most traditional approach to change detection, will first classify individual images of two dates using a statistical maximum likelihood classifier and then
compare the classified images to provide an assessment of change. This traditional
approach generally requires extensive human supervision for classifying the images.
However, new image classification methods, other than the traditional maximum
likelihood classifier, can be applied to increase accuracy and efficiency. These methods include, for example, fuzzy classification, artificial intelligence based classifier,
Bayesian approach, and even the textural approach (Moller-Jensen, 1990; Gopal
and Woodcock, 1996; Jensen, 2005; Gong, 2006). Recently, object-based image
segmentation and classification has gained increasing attention, with new software
such as eCognition (Definiens, 2004; Benz et al., 2004) and Feature Analyst (Visual
Learning Systems, Inc.) made available to general users. These methods use both
the spectral (or color) information and various spatial metrics to define homogenous
areas (called objects). Despite its potential, this group of change detection methods
is not germane to rapid change detection, as extensive human supervision is needed
to pre-classify the images.
The second group of change detection methods does not require images to be
pre-classified. Image differencing, change vector method, and multidate comparison methods (Fig. 13.4 box d) can be applied directly to original pixel values or
indirectly to modified values from the spectral bands (e.g., band ratios, principal
components, chi-squared transformed, and texture transformed) (Fig. 13.4 boxes
b and c). The main advantage of this group of methods is that pre-classification is
not necessary until significant changes are detected, hence avoiding the tedious classification process at the beginning. The problem remains to be that of determining
the threshold value at which the difference between the two images is considered
significant.
Continuous monitoring of land cover/land use and rapid identification of their
changes is crucial to providing timely decision support and risk assessment especially during extreme events (e.g., hurricanes, earthquakes, forest fires, terrorist
attacks, disease spread). There is a need to develop efficient and reliable change detection methods that can be automated, easy to use, and applicable to different land
covers observed by different sensors at different scales, times, and places. Although
it is difficult to achieve fully automated change detection, we expect that an integrated approach that incorporates both textural and spectral indices could alleviate
some of the existing change detection problems for two reasons (which have also
been elaborated in Section 13.2.2). First, the texture measures that have property
5 and 6 (e.g., fractals, lacunarity, wavelets, and spatial autocorrelation statistics) can
be applied directly to pre-classified images without the need to go through the image
classification process, thereby reducing the need for extensive human supervision
upfront. Second, since the spatial/texture methods measure the spatial variations
across the image instead of comparing brightness values on a pixel by pixel basis,
they are more likely to reflect dominant changes rather than spurious changes that
359
occur due to noise, clouds, or illumination differences. However, whether a combination of these methods can be successfully applied to reliably characterize land
covers and identify changes remains to be studied and is part of our ongoing research.
13.5 Change Detection Using the Textural Approach:

An Example
In this section, we demonstrate the use of textural measures only (not including
spectral information) for change detection in New Orleans before and after Hurricane
Katrina. The utility of fractal dimension and spatial autocorrelation statistics in this
capacity is compared. Hurricane Katrina hit New Orleans on August 29, 2005. Two
Landsat-TM images, dated November 7, 2004 and September 7, 2005, were obtained
from the US Geological Survey/National Wetlands Research Center at Louisiana
State University and the LSU FEMA GIS Store project (http://www.cadgis.lsu.edu).
Although not exactly the same anniversary dates, these two images are the best
available so far for change detection. Both images have already been registered and
geometrically rectified with a pixel resolution of 28.5 m prior to this study. For this
study, we created a subset of 512 512 pixels from both images. Figure 13.5a and
b display the subsets using band 4 (near-infrared band), which is the most effective band in discriminating water and non-water features. The subsets were mainly
confined to the Orleans Parish, with a small part of the Lower Ninth Ward in St.
Bernard Parish shown at the east edge of the image (east of the Industrial Canal
breach). The post-Katrina image (Fig. 13.5b) clearly shows that most of New Orleans was flooded 9 days after Katrina, except in the natural levee area along the
Mississippi River. The northwest corner of the image is Lake Pontchartrain, where
significant storm surge has destroyed properties around the Lake. The three levee
breaches at Industrial Canal, the 17th Street Canal, and the London Avenue Canal
are also marked in Fig. 13.5b.
Both Landsat-TM images used in this example have previously been normalized to minimize sensor calibration offsets and differences in atmospheric effects.
However, other factors may change the pixel values even though there are no real
changes on the ground. A random check on the pixels at the northwest corner (Lake
Pontchartrain) shows that a fair amount of difference in pixel values occurred in
the same location between the two time periods (e.g., 32 vs 72 in pre- and postKatrina image, respectively), even though no significant real change is expected
at this location. This shows that change detection using the original spectral value
pixel-by-pixel comparison approach alone could be problematic, especially regarding the determination of the threshold value to identify real changes.
Using ICAMS, we computed the local fractal dimension using the modified triangular prism algorithm and the local Morans I for each image. For the local fractal
dimension method, the following parameters were used: moving window size of
17 17, step size of 5, stretch option, and arithmetic progression. For Morans I,
360
N.S.-N. Lam
Fig. 13.5 (a and b) Display of pre-Katrina (November 7, 2004) and post-Katrina (September 7,
2005) Landsat-TM images using band 4. (c) and (d) are Morans I transformed pre- and postimages; (e) and (f) are fractal-transformed pre- and post-images
the only parameter needed to be input was the moving window size, which was also
set to 17 17. The 17 17 window was chosen because a previous study on the impacts of Hurricane Hugo along the South Carolinas coast by Kulkarni (2004) found
that this window size was the best in representing land cover features. Figure 13.5cf
361
Table 13.2 Summary statistics of band 4 for pre- and post-Katrina Landsat TM images
Min
Max
Mean
SD
CV
Pre
Original
Post
24.00
255.00
161.50
52.28
0.32
1.00
255.00
45.02
25.51
0.57
Diff.
250.00
96.00
116.49
44.95
0.38
Fractal-transformed
Pre
Post
Diff.
1.86
4.11
2.77
0.22
0.08
1.75
4.13
2.74
0.21
0.08
1.23
1.15
0.03
0.15
5.00
Morans I-transformed
Pre
Post
Diff.
0.07
0.97
0.61
0.17
0.28
0.05 0.63
0.97
0.63
0.66
0.05
0.14
0.11
0.21
2.20
SD standard deviation; CV coefficient of variation (= SD/mean); difference image =

(post pre)
show the Morans I-transformed and the fractal-transformed images. Brighter pixels
refer to higher values in fractal dimension or Morans I. It should be stressed that
since fractal dimension and Morans I have an inverse relationship, features with
low fractal dimension, such as the Mississippi River (darker pixels in the fractaltransformed images), will be shown as brighter pixels in the Morans I transformed
images.
The difference images were computed by subtracting the pre-Katrina image
from the post-Katrina image, and the summary statistics of all images are listed in
Table 13.2. In general, the post-image had lower spectral values than the pre-image,
and the mean difference between the two images (band 4) was as high as 116.49.
This is expected as most of the study area was flooded after Katrina, resulting in
lower spectral reflectance value in the near-infrared band. The fractal-transformed
summary statistics show that the mean spatial complexity, as represented by fractal
dimension, slightly decreased from 2.77 to 2.74. Conversely, the mean Morans I
increased from 0.61 to 0.66, which also indicates that the overall spatial complexity
decreased slightly for the post-image.
The fractal and Moran difference images were first mapped in a continuous mode
(with a two-standard deviation stretch) using ICAMS. The fractal difference image
(Fig. 13.6a) shows that increases in fractal dimension (positive changes), as represented by brighter pixels, occurred in areas close to the Industrial Canal (east
side of the image) and the areas between the 17th Street Canal and London Avenue
Canal (middle part of the image). Areas with decrease in fractal dimension (negative
changes) are represented by darker pixels and they scattered over the image. With a
17 17 window size, the general features of the study area, such as the Mississippi
River, can still be recognized. The Morans I difference image (Fig. 13.6b) shows
the same pattern; the darkest pixels represented the highest decrease in spatial autocorrelation, implying an increase in spatial complexity. It can be observed from the
two difference images that the location of the darkest pixels in the Morans I difference image generally coincided with the brightest pixels in the fractal different
image, and vice versa.
The difference images can also be mapped using standard deviation unit as class
intervals in ICAMS. One of the display options is to map the changes in three
class intervals using two-standard deviations as class boundaries. The first interval, which contains pixels that have highest positive changes in spatial complexity
362
N.S.-N. Lam
Fig. 13.6 Display of the fractal-difference and Moran-difference images in a continuous mode (a)
and (b) and three-class mode (c) and (d). In (c) and (d), the brightest pixels indicate the highest
positive changes in spatial complexity (>2 standard deviations), the darkest pixels indicate highest negative changes (<2 standard deviations), and the gray pixels are values in between. Both
brightest and darkest pixels should be of interest, which may point to areas that are most affected
(>2 SD), is shaded as the brightest. The second class, which is in middle gray,
is for pixels that have difference values falling between 2 standard deviations.
The third class, which has the darkest shade, is for pixels that have the greatest negative changes in spatial complexity (<2SD). Using this mapping method,
both the brightest and the darkest pixels in the images (Fig. 13c and d) signal the
greatest changes and hence attention is most warranted for these pixels and their
surrounding pixels. This method should guide resources to the most affected
areas, and in this case, greatest change in spatial complexity in both positive and
negative directions.
Figure 13.6d, the fractal difference image, shows only a few concentrated spots
belonging to the first and third classes (the brightest and darkest pixels), and they
were generally located close to the three canal breach areas. The rest of the secondclass pixels were scattered throughout the image. For the Moran difference image
(Fig. 13.6c), because of its inverse relationship with fractal dimension (i.e., the
higher the fractal dimension, the lower the Morans I value), one should expect that
363
the brightest spot in the fractal difference image would coincide with the darkest
spot in the Morans I difference image. A visual comparison between the two images
(Fig. 13.6c and d) shows that this is generally true, with the Morans I difference
image portraying a wider area of brightest and darkest spots than the fractal difference image. Based on the Morans I difference image (Fig. 13.6c), the greatest
decrease in Morans I values (greatest increase in spatial complexity darkest pixels) were also found near the three canal breach areas. Areas that showed greatest
increase in Morans I values (greatest decrease in spatial complexity brightest
pixels) were scattered in the mid city and the area surrounding Lake Pontchartrain.
In summary, this example shows that the textural approach alone could be useful
in pinpointing the areas that need the most attention. With additional information
layers, these maps could serve as a useful guide to focus our efforts in detecting
largest and meaningful changes in a rapid and reliable manner. It is expected that
combining spectral and spatial layers, as well as combining different textural measures, will increase the accuracy of this approach. Other mapping methods could
also be employed to further enhance the visualization of these changes.
13.6 Conclusions
Efficient methods for rapid monitoring of land cover/land use and their changes
through remote sensing imagery are urgently needed to provide timely decision
support and risk assessment especially during extreme events (e.g., terrorist attacks,
hurricanes, forest fires, earth quakes, disease spread). Although there is a huge literature on land cover classification and change detection, we are still far from being
able to automate these tasks via remote sensing and GIS. The high variability of
ground conditions as manifested in individual as well as time-series imagery makes
it very difficult to generalize and automate. The search for useful approaches and
methods for rapid land cover identification and change detection remains a very
challenging task.
This chapter introduced the use of textural and spatial metrics as a promising
approach to automated land cover/land use classification and change detection. We
identified in this chapter the major criteria for selecting textural measures and then
illustrated through several examples from previous studies how textural metrics, in
combination with the original spectral bands, have greatly improved the classification accuracy. For change detection analysis, we developed a framework for classifying the numerous change detection approaches. Then, using a recent example
of evaluating the impacts of Hurricane Katrina on New Orleans land cover change,
we illustrated the use of local fractal dimension and local Morans I to detect the
largest changes that might need further attention. More research is needed to determine the effectiveness of the various textural metrics with different types of remote
sensing imagery, different scales and resolutions, different land cover features, and
different environments. These issues are currently being examined in our ongoing
research.
364
N.S.-N. Lam
Acknowledgements John Barras, Geographer at the USGS/CR/BRD National Wetlands Research

Center, Coastal Restoration Field Station, Louisiana State University, provided the Landsat-TM
2004 image data. Mary Lee Eggart assisted in preparing the figures. This research is supported in
part by a NASA Intelligent Systems research grant (NCC-2-1246) and in part by a US National
Science Foundation grant (BCS-0554937).
References
Baskent EZ, Jordan GA (1955) Characterizing spatial structure of forest landscapes. Can. J. Forest
Res. 25:18301849
Benz UC, Hofmann P, Willhauck G, Lingenfelder I, Heynen M (2004) Multi-resolution, objectoriented fuzzy analysis of remote sensing data for GIS-ready information. ISPRS J. Photogramm. Remote Sens. 58:239258
Bian L (2003) Retrieving urban objects using a wavelet transform approach. Photogramm. Eng.
Briggs JM, Nellis JM (1991) Seasonal variation of heterogeneity in the tallgrass prairie: A quantitative measure using remote sensing. Photogramm. Eng. Remote Sens. 57:407411
Cao C, Lam NSN (1997) Understanding the scale and resolution effects in remote sensing and GIS.
In: DA Quattrochi, MF Goodchild (eds), Scale in remote sensing and GIS. Lewis Publishers,
Boca Raton, FL, pp 5772
Carr JR (1999) Classification of digital image texture using variograms. In: PM Atkinson, NJ Tate
(eds), Advances in remote sensing and GIS analysis. Wiley, London, pp 135146
Carr JR, de Miranda FP (1998) The semivariogram in comparison to the co-occurrence matrix for
classification of image texture. IEEE Trans. Geosci. Remote Sens. 36(6):19451952
Clarke KC (1986) Computation of the fractal dimension of topographic surfaces using the triangular prism surface area method. Comput. Geosci. 12(5):713722
Clausi DA, Jobanputra R (2006) Preserving boundaries for image texture segmentation using grey
level co-occurring probabilities. Pattern Recog. 39(2):234245
Cliff AD, Ord JK (1973) Spatial autocorrelation. Methuen, New York
Coppin PR, Bauer ME (1966) Digital change detection in forest ecosystems with remote sensing
imagery. Remote Sensing Reviews 13:207234
Crews-Meyer KA (2002) Characterizing landscape dynamism using paneled-pattern metrics.
Photogramm. Eng. Remote Sens. 68(10):10311040
Dale MRT (2000). Lacunarity analysis of spatial pattern: a comparison. Landscape Ecol.
15(5):467478
Daubechies I (1990) The wavelet transform, time/frequency localization and signal analysis. IEEE
Trans. Inf. Theory 36:9611005
De Pietri DE (1995) The spatial configuration of vegetation as an indicator of landscape degradation due to livestock enterprises in Argentina. J. Appl. Ecol. 32:857865
Definiens AG (2004) eCognition User Guide (accessed May 2006)
Dong P (2000) Test of a new lacunarity estimation method for image texture analysis. Int.
J. Remote Sens. 21(17):33693373
Dunn CP, Sharpe DM, Guntenspergen GR, Stearns F, Yang Z (1991) Methods of analyzing temporal changes in landscape pattern. In: MG Turner, RH Gardner (eds), Quantitative methods
in landscape ecology. The analysis and interpretation of landscape heterogeneity. Springer,
New York
Emerson CW, Lam NSN, Quattrochi DA (1999) Multiscale fractal analysis of image texture and
pattern. Photogramm. Eng. Remote Sens. 65(1):5161
Emerson CW, Lam NSN, Quattrochi DA (2005) A comparison of local variance, fractal dimension,
and Morans I as aids to multispectral image classification. Int. J. Remote Sens. 26(8):1575
1588
365
Estreguil C, Lambin E (1996) Mapping forest disturbances in Papua New Guinea with AVHRR
data. J. Biogeogr. 23:757773
Falconer K (1988) Fractal geometry: mathematical foundations and applications. Wiley, New York
Frank TD (1984) The effect of change in vegetation cover and erosion patterns on albedo and
texture of Landsat images in a semiarid environment. Ann. Assoc. Am. Geogr. 74:393407
Franklin SE, Hall RJ, Moskal LM, Maudie AJ, Lavigne, MB (2000) Incorporating texture into
classification of forest species composition form airborne multispectral images. Int. J. Remote
Sens. 21(1):6179
Gong P (2006) Information extraction. In: M Ridd, JD Hipple (eds), Remote sensing of human
settlements. ASPRS, Bethesda, MD, pp 275334
Gong P, Marceau DJ, Howarth PJ (1992) A comparison of spatial feature extraction algorithms for
land use classification with SPOT HRV data. Remote Sens. Environ. 40:137151
Goodchild MF (1980) Fractals and the accuracy of geographical measures. Mathematical Geology
12:8598
Goodchild MF (1986) Spatial Autocorrelation. CATMOG (Concepts and Techniques in Modern
Geography) No. 47. Geo Books, Norwich, England
Gopal S, Woodcock C (1996) Remote sensing of forest change using artificial neural networks.
Haralick RM (1979) Statistical and structural approaches to texture. Proc. IEEE 67(5):786804
Haralick RM, Shanmugan K, Dinstein J (1973) Textural features for image classification. IEEE
Trans. Syst. Man Cybern. 3(6):610621
Jaggi S, Quattrochi D, Lam NSN (1993). Implementation and operation of three fractal measurement algorithms for analysis of remote sensing data. Comput. Geosci. 19(6):745767
Jensen J, Cowen D, Althausen J, Narumalani S, Weatherbee O (1993) An evaluation of the CoastWatch change detection protocol in South Carolina. Photogrammetric Engineering and Remote
Sensing 59(6):10391046
Jensen J, Cowen D, Narumalani S, Halls J (1997) Principles of change detection using digital
remote sensor data. In: JL Star, JE Estes, KC McGwire (eds), Integration of geographic information systems and remote sensing. Cambridge University Press, Cambridge, pp 3754
Jensen J (2005) Introductory digital image processing: a remote sensing perspective, 3rd edn.
Prentice-Hall, New Jersey
Jupp DLB, Walker J, Pendridge LK (1986) Interpretation of vegetation structure in Landsat MSS
imagery: a case study in disturbed semi-arid eucalypt woodland. Part 2. model-based analysis.
J. Environ. Manag. 23:3557
Kulkarni A (2004) Evaluation of the Impacts of Hurricane Hugo on the Land Cover of Francis
Marion National Forest, South Carolina Using Remote Sensing. M.S. thesis, Louisiana State
University, Baton Rouge, Louisiana
Lam NSN (1990) Description and measurement of Landsat TM images using fractals. Photogramm. Eng. Remote Sens. 56(2):187195
Lam NSN (2004) Fractals and scale in environmental assessment and monitoring. In: E Sheppard,
R McMaster R (eds), Scale and Geographic Inquiry: Nature, Society, and Method. Blackwell,
Oxford, pp 2340
Lam NSN, Quattrochi DA (1992) On the issues of scale, resolution, and fractal analysis in the
mapping sciences. Prof. Geogr. 44(1):8999
Lam NSN, De Cola L (eds) (1993) Fractals in geography. Prentice-Hall, Englewood Cliffs,
NJ, 308p
Lam NSN, Quattrochi DA, Qiu HL, Zhao W (1998) Environmental assessment and monitoring
with image characterization and modeling system using multiscale remote sensing data. Appl.
Geogr. Stud. 2(2):7793
Lam NSN, Qiu HL, Quattrochi DA, Emerson CW (2002) An evaluation of fractal methods for
measuring image complexity. Cartogr. Geogr. Inform. Sci. 29:2535
Lam NSN, Catts C, Quattrochi DA, Brown D, McMaster R (2004) Scale. In: R McMaster, L Usery
(eds), A Research Agenda for Geographic Information Science. CRC Press, Bacon Raton, FL,
Chapter 4, pp 93128
366
N.S.-N. Lam
Lambin EF (1996) Change detection at multiple temporal scales: seasonal and annual variations in
landscape variables. Photogramm. Eng. Remote Sens. 62:931938
Lambin EF, Strahler AH (1994) Indicators of land-cover change for change-vector analysis in
multitemporal space at coarse spatial scales. Int. J. Remote Sens. 15:20992119
Lo CP, Quattrochi DA, Luvall JC (1997) Application of high-resolution thermal infrared remote
sensing and GIS to assess the urban heat island effect. International Journal of Remote Sensing
18(2):287304
Lu D, Mausel P, Brondizio E, Moran E (2005) Land-cover binary change detection methods for
use in the moist tropical region of the Amazon: a comparative study. Int. J. Remote Sens.
26(1):101114
Lunetta R, Elvidge C (1998) Remote sensing change detection: environmental monitoring methods
and applications. Sleeping Bear Press, Ann Arbor, MI
Mallat SG (1989) A theory for multi-resolution signal decomposition: the wavelet representation.
IEEE Trans. Pattern Anal. Machine Intell. 11:674693
Mandelbrot B (1982) The fractal geometry of nature. Freeman, New York
Manjunath BS, Ma WY (1996) Texture features for browsing and retrieval of image data. IEEE
Trans. Pattern Anal. Machine Intell. 18(8):837842
Mark DM, Aronson PB (1984) Scale-dependent fractal dimensions of topographic surfaces: an
empirical investigation, with applications in geomorphology and computer mapping. Math.
Geol. 11:671684
Mas JF (1999) Monitoring land-cover changes: a comparison of change detection techniques. Int.
McGarigal K (2002) Landscape pattern metrics. In AH El-Shaarawi, WW Piegorsch (eds), Encyclopedia of environmentrics, vol 2. Wiley, Sussex, England, pp 11351142
McGarigal K, Mark BJ (1995) FRAGSTATS: spatial pattern analysis program for quantifying landscape structure. USDA Forest Service General Technical Report PNW-351, Portland, Oregon
Mallat, S (1989) A theory of multi-resolution signal decomposition: the wavelet representation.
IEEE Trans. Pattern Anal. Machine Intel. 11:674693
Moller-Jensen L (1990) Knowledge-based classification of an urban area using texture and context
information in Landsat-TM imagery. Photogramm. Eng. Remote Sens. 56(6):899904
Muneeswaran K, Ganesan L, Arumugam S, Sounda KR (2005) Texture classification with combined rotation and scale invariant wavelet features. Pattern Recogn. 38(10):14951506
Myint S, Lam NSN (2005a) A study of lacunarity based texture analysis approaches to improve
urban image classification. Comput. Environ. Urban Syst. 29:501523
Myint S, Lam NSN (2005b) Examining lacunarity approaches in comparison with fractal and spatial autocorrelation techniques for urban mapping. Photogramm. Eng. Remote Sens. 71(8):927
937
Myint S, Lam NSN, Tyler J (2002) An evaluation of four different wavelet decomposition procedures for spatial feature discrimination in urban areas. Trans. GIS 6(4):403429
Myint S, Lam NSN, Tyler J (2004) Wavelets for urban spatial feature discrimination: Comparisons
with fractals, spatial autocorrelation, and spatial co-occurrence approaches. Photogramm. Eng.
Nackaerts K, Vaesen K, Muys B, Coppin P (2005) Comparative performance of a modified change
vector analysis in forest change detection. Int. J. Remote Sens. 26(5):839852
ONeill RV, Johnson AR, King AW (1989). A hierarchical framework for the analysis of scale.
Landscape Ecol. 7(1):5561
ONeill RV, Krummel JR, Gardner RH, Sugihara G, Jackson B, DeAngelis DL, Milne BT, Turner
MG, et al. (1998) Indices of landscape pattern. Landscape Ecol. 1:153162
Openshaw S (1989) Automating the search for cancer clusters. Prof. Statistician 8:78
Pickup G, Foran BD (1987) The use of spectral and spatial variability to monitor cover change on
inert landscapes. Remote Sens. Environ. 23:351363
Plotnick RE, Gardner RH, ONeill RV (1993) Lacunarity indices as measures of landscape texture.
Landscape Ecol. 8:201211
367
Quattrochi DA, Goodchild MF (1997) Scale in remote sensing and GIS. CRC/Lewis Publishers,
Boca Raton, FL
Quattrochi DA, Lam NSN, Qiu HL, Zhao W (1997) Image characterization and modeling system
(ICAMS): a geographic information system for the characterization and modeling of multiscale
remote sensing data. In: D Quattrochi, M Goodchild (eds), Scaling in remote sensing and GIS.
CRC/Lewis Publishers, Boca Raton, FL, pp 295307
Read JM, Lam NSN (2002) Spatial methods for characterizing land-cover changes for the tropics.
Int. J. Remote Sens. 23(12):24572474
Smits PC, Annoni A (2000) Towards specification-driven change detection. IEEE Trans. Geosci.
Tate N, Atkinsons P (eds) (2001) Modelling scale in geographical information science. Wiley,
New York
Turner MG, Dale VH, Gardner RH (1989) Predicting across scales: theory development and testing. Landscape Ecol. 3(3/4):245252
Voss R (1986) Random fractals: characterization and measurement. In: R Pynn, A Skjeltorp (eds),
Scaling phenomena in disordered systems. Plenum, New York
Woodcock CE, Strahler (1987) The factor of scale in remote sensing. Remote Sens. Environ.
21:311332
Yuan D, Elvidge C, Lunetta R (1998) Survey of multispectral methods for land cover change
analysis. In: R Lunetta, D Elvidge (eds), Remote sensing change detection. Sleeping Bear Press,
Ann Arbor, MI
Zhu C, Yang X (1998) Study of remote sensing image texture analysis and classification using
wavelet. Int. J. Remote Sens. 19(16):31973203
Chapter 14
Methodologies for Mapping Plant Functional

Types
Wanxiao Sun and Shunlin Liang
Abstract Plant functional type (PFT) is a crucial variable needed in studies of

global climate, carbon cycle and ecosystem change. Using remote sensing techniques to extract PFTs is a relatively recent field of research. To date, only a very few
methods for mapping PFTs have been reported. This chapter provides an overview
of recent developments in this evolving field and discusses future research needs.
A brief survey of existing methods for mapping PFTs is presented, followed by a
discussion of several methodological issues pertaining to the development of robust
remote sensing techniques for mapping of PFTs at regional to global scales. The
chapter also outlines a multisource data fusion framework for improved mapping of
PFTs from satellite observations.
14.1 Introduction
Vegetation plays a vital role in the exchange of energy, carbon, water, and
momentum between the land surface and the atmosphere (Sellers and Schimel,
1993; Bonan, 1996; Sellers et al., 1997; Lawton et al., 2001; Marland et al., 2003;
Nair et al., 2003; Gamon et al., 2004; Feddema et al., 2005). Reliable information
about the geographic distribution and extent of major types of vegetation around the
globe is required to initiate and validate various land surface models that provide
the boundary conditions for the simulation of climate, carbon cycle and ecosystem
change. Traditionally, land surface models represent vegetation as discrete biomes
such as evergreen broadleaf forest, shrub, grass, and savanna. These biomes then set
surface biogeophysical variables such as albedo, LAI, f PAR, canopy roughness and
stomatal physiology for each grid cell (Sellers et al., 1986; Running and Coughlan,
1988; Bonan, 1993; Prince and Goward, 1995).
Wanxiao Sun
Department of Geography and Planning, Grand Valley State University, USA
Shunlin Liang
Department of Geography, University of Maryland, USA
369
370
W. Sun, S. Liang
A major problem with the biome-based land classification approach is that biomes are not natural vegetation units but are products of classification. Many of the
vegetation parameters needed by land models are leaf-level and whole-plant parameters that are difficult to parameterize in the case of mixed life-form biomes such
as mixed forests and savannas (Bonan et al., 2002). For example, how does one
obtain the necessary leaf physiological and whole-plant allocation parameters for a
savanna, which consists of physiologically distinct grasses and trees? Land models
are expanding beyond their traditional biogeophysical roots to include biogeochemistry, especially photosynthesis and the carbon cycle (Bonan, 1995; Foley et al.,
1996; Dickinson et al., 1998; Kucharik et al., 2000). Inclusion of photosynthesis
and the carbon cycle in land models makes the mixed life-form problem even more
acute. To address this problem, the land modeling community has started using plant
functional types to represent land surface.
Plant functional types (PFT) are groups of plant species that share similar functioning at the organismic level, similar responses to environmental factors and/or
similar effects on ecosystems (Smith et al., 1997). Deciduous broadleaf trees, evergreen needleleaf trees, grasses and broadleaf crops are examples of PFTs, whereas
savannas, mixed forests and cropland/natural vegetation mosaics are not PFTs but
biomes. Representing land surface in terms of PFTs offers several important advantages over the biome approach (Smith et al., 1997; Bonan et al., 2002). First, PFT
provides a direct link to leaf-level physiological measurements, making it possible
to more accurately set ecological parameters in land models. Second, PFT allows
modelers to more accurately represent the land surface by separately altering the
vegetation composition (i.e., the number of PFTs and their abundance) and structure (e.g., LAI, canopy height) within a grid cell. Third, representing landscapes as
patches of PFTs also allows land surface models to better interface with ecosystem models, because the latter typically simulate vegetation change in terms of the
abundance of PFTs.
Reliable PFT information is increasingly needed by the global change research
community, especially the carbon, climate and ecosystem modeling community. For
example, the carbon models used to scale carbon fluxes typically require specification of PFTs (Denning et al., 1996; Sellers et al., 1997). The National Center
for Atmospheric Research land surface model (NCAR LSM) has recently shifted
from using land cover information to using satellite-derived PFT maps (Bonan
et al., 2002; Tian et al., 2004). Plant functional types have also been advocated
in dynamic global vegetation models (DGVM) to predict the composition and functioning of ecosystems in a changing environment (Running and Coughlan, 1988;
Prentice et al., 1992; Woodward and Cramer, 1996; Smith et al., 1997; Kucharik
et al., 2000). As such, accurate mapping of PFTs over large areas can contribute to
improved predictive capabilities of global and regional carbon cycle, climate and
ecosystem models.
Remote sensing is the only practical means by which land surface biogeophysical variables can be obtained at the temporal and spatial scales required by global
change research (Townshend et al., 1991; Roughgarden et al., 1991; Sellers et al.,
1995; Myneni et al., 1997; Liang, 2004). In fact, land surface parameter estimation is
14 Methodologies for Mapping Plant Functional Types
371
a major direction in quantitative remote sensing (Liang, 2004). There is a wealth of

literature on how to use remotely sensed data to extract land cover information and
certain biogeophysical variables (e.g., Townshend and Justice, 1981; Townshend
et al., 1987; Lloyd, 1990; Kimes et al., 1991; Myneni et al., 1995, 1997; DeFries
et al., 1998, 1999; Loveland et al., 1991, 1995, 2000; Liang, 2001, 2003; Sun, 2004;
Fang and Liang, 2005). However, using remote sensing techniques to extract PFTs
is a relatively recent field of research. A review of the literature shows that to date
only a very few methods for mapping PFTs have been reported (Strahler et al., 1999;
Bonan et al., 2002; Sun et al., 2005, 2007). Furthermore, the accuracies of the PFT
maps generated with existing methods have yet to be validated. Studies have demonstrated that the use of different PFT data sets generated with different methods has a
significant effect on global climate modeling results (e.g., Oleson and Bonan, 2000;
Bonan et al., 2002; Tian et al., 2004). These results highlight the importance of the
quality of PFT data in global change research.
Landscape can be thought of as a mixture of PFTs. As such, spectral unmixing
techniques can be used to separate different types of vegetation within a geographic
area. Substantial progress has been made in this context such that prototype data
are now available. For example, a linear mixture model was employed to estimate a
global land cover with continuous fields of vegetation characteristics (DeFries et al.,
1999, 2000a). This global data set of continuous fields was able to capture the heterogeneity of vegetation within a grid cell at subpixel level. The main advantage of
using continuous distributions of vegetation in land models is to avoid abrupt boundaries and unrealistic homogeneity that can be introduced into parameter estimates
as is the case with discrete biome-based data sets.
Linear unmixing was also used to derive a pure PFT NDVI data set using every
1 km AVHRR pixel with its corresponding PFT abundances (Bonan et al., 2002). In
addition to the linear unmixing, another approach was also utilized to generate the
pure PFT NDVI (Bonan et al., 2002). In that approach, the pure NDVI was extracted
by averaging the NDVI over 1 km pixels in which the abundance of the PFT was
greater than 60%. In fact, the pure PFT NDVI obtained in this way is not actually
pure but rather a mix of co-occurring PFTs. Although the linear unmixing method
is theoretically preferable and this method was expected to give a more accurate
representation of the pure PFT NDVI, it was unstable (Bonan et al., 2002). Possible reasons may include calibration problems of the AVHRR data. When applied to
EOS era data such as higher spatial and spectral resolution MODIS data, spectral
unmixing techniques hold promises for exacting subpixel PFT fractions with improved accuracy and for estimating the individual PFT LAI of mixed pixels (Oleson
and Bonan, 2000). Moreover, more sophisticated methods to allow for multiple endmembers and nonlinear statistical unmixing will potentially improve the accuracy
of vegetation characteristics (DeFries et al., 2000a).
The objective of this chapter is to provide an overview of recent developments
in this evolving field and discuss future research needs in developing new methodologies for mapping of PFTs. The next section provides a brief summary of the existing methods for mapping PFTs reported in the literature. Section 14.3 discusses
several methodological issues pertaining to the development of effective and robust
372
W. Sun, S. Liang
remote sensing techniques for extracting PFTs over large areas (e.g., continental to
global scales). Section 14.4 outlines a multisource data fusion framework for inferring PFTs from MODIS. We conclude the chapter with some general remarks.
14.2 Existing Methods for Mapping PFTs: State of the Art

14.2.1 Extracting PFTs from Existing Land Cover Data Sets
The land modeling community pioneered in the development of PFT data sets over
large areas for use in land models (e.g., Bonan, 1996; Oleson and Bonan, 2000;
Bonan et al., 2002). Typically, the methods used by modelers to extract PFT information rely on preexisting land cover maps. An early example of representing
vegetation as PFTs is the National Center for Atmospheric Research land surface
model (NCAR LSM) (Bonan, 1996). In this model, the vegetation in a grid cell is
represented as mixtures of 12 PFTs. Up to three PFTs can form distinct patches in a
grid cell, with lakes and wetlands forming additional patches. In the standard LSM,
the PFT composition for each grid cell is obtained by classifying grid cells as one of
28 possible biomes. In other words, each grid cell is assigned a biome type, which
in turn determines the patch fractions for each PFT. For example, savanna grid cells
are represented as consisting of 70% C4 grasses and 30% tropical trees. All needleleaf evergreen forests are 25% bare ground and 75% needleleaf evergreen tree. The
types of biomes and their geographic distributions used in the standard LSM are
based on Olson et al. (1983).
In their study of boreal forest surface fluxes, Oleson and Bonan (2000) prepared
a PFT data set for the BOREAS region. This PFT data set was derived from the
1 km AVHRR BOREAS land cover map of Steyaert et al. (1997). The land cover
classes defined at 1 km resolution were used to calculate the number of PFTs and
their abundance (i.e., fractional areas) of each grid cell. Up to six PFTs can coexist
within each grid cell. Fractions of each grid cell occupied by water and wetlands
were also determined.
Using a methodology similar to Oleson and Bonan (2000), Bonan et al. (2002)
developed a global PFT map from 1 km satellite data. In this method, each model
grid cell is divided into four primary land cover types: glacier, lake, wetland, and
vegetation. The vegetated portion of a grid cell is further divided into patches of up
to four of the models 15 PFTs. The geographic distribution and abundance of these
15 PFTs were derived from the 1 km IGBP DISCover data set (Loveland et al., 2000)
and the 1 km University of Maryland tree cover data set (DeFries et al., 1999, 2000a,
2000b). Specifically, each 1 km pixel was assigned the percentage of needleaf evergreen, needleleaf deciduous, broadleaf evergreen, and broadleaf deciduous trees as
given in the tree cover data. The non-tree-covered portion of the 1 km pixel was determined from the IGBP DISCover data. Figure 14.1 is an example of the PFT map
for North America prepared by Bonan et al. (2002).
373
Bare ground
Needleleaf evergreen temperate trees
Needleleaf evergreen boreal trees
Needleleaf deciduous trees
Broadleaf evergreen tropical trees
Broadleaf evergreen temperate trees
Broadleaf deciduous tropical trees
Broadleaf deciduous temperate trees
Broadleaf deciduous boreal trees
Broadleaf evergreen temperate shrubs
Broadleaf deciduous temperate shrubs
Broadleaf deciduous boreal shrubs
C3 Arctic grasses
C3 Non-arctic grasses
C4 Grasses
Crops
Fig. 14.1 PFT map for North America from Bonan et al. (2002)
Studies done by the land research community have demonstrated that satellitederived PFT data has the potential to allow for more accurate representation of
land surface processes and properties. However, the approach used by modelers
to derive PFTs has several limitations. First, the accuracies of the existing land
cover maps used to extract PFTs are generally unknown and, in many cases, appear poor (Townshend et al., 1991). Second, virtually all land cover maps produced
in the past represent the land surface in terms of biomes. As a result, the number of
PFTs and their abundance within each grid had to be prescribed or estimated from
the land cover types depicted on the maps. The accuracy of the PFT data generated this way may be a concern. Third, this approach often involves using several
separate and not necessary compatible land cover data sets. Consequently, modelers had to make lots of assumptions in places where existing land cover maps
conflict with each other. Fourth, existing land cover data sets often lack detailed
information needed for accurate parameterization of the land surface. Investigators
were frequently forced to make assumptions about missing information. For example, Bonan et al. (2002) noted that, because consistent information on nonvegetated
cover was not available, they had to assume that non-tree-covered land in forests,
savannas, and grasslands was covered by grasses, in shrub lands by shrubs, in croplands by crops. Finally, as Fig. 14.1 illustrates, the PFT maps prepared by modelers
are often of very coarse resolutions (e.g., 0.5 0.5 ). Such PFT data sets are also
hard to update.
374
W. Sun, S. Liang
14.2.2 MODIS PFT: Supervised Decision Tree Classifiers

In response to the needs of the land research community, the Moderate Resolution Imaging Spectroradiometer (MODIS) Land Team is producing a global PFT
map (i.e., MODIS Land Cover 5) for use in the Community Land Model (CLM).
This PFT product is generated as one of the five MODIS Land Cover products
(MOD12Q1). Technical details of the methodology for producing MODIS Land
Cover products are given in Strahler et al. (1999). We provide here a brief description of the key components of the methodology pertaining to the production of the
MODIS PFT map.
MODIS Land Cover is a global database of land cover types. The Land Cover
product is prepared at 1 km spatial resolution and includes five different sets of land
cover labels (Table 14.1).
Land Cover Type 1 includes the 17 classes of land cover defined in the International Geosphere-Biosphere Programme (IGBP) global vegetation classification
scheme.
Land Cover Type 2 uses the University of Maryland land cover classification
scheme, which is a version of the IGBP scheme modified to exclude wetlands,
vegetation mosaic, and snow information.
Land Cover Type 3 uses the MODIS LAI/fPAR scheme, which is input to
the MODIS LAI/fPAR product (i.e., MODIS15A2) and emphasizes vegetation
classes.
Land Cover Type 4 uses the MODIS Net Primary Production scheme (Biome
BGC Model classes).
Land Cover Type 5 uses the Plant Functional Types (PFT) scheme.
The IGBP land cover classification scheme provides the primary layer of MODIS
Land Cover data sets. This scheme includes 11 natural vegetation classes broken
down by life form, three classes of developed and mosaic lands, and three classes
of nonvegetated lands (Table 14.1). Strahler et al. (1999) note that the IGBP classes
can be re-labeled or cross-walked to provide compatibility with current and future land cover classification systems used by the modeling community. Detailed
information on how the IGBP land cover classes are cross-walked to the PFT
classes currently produced by the MODIS Land Team is not released. A comparison of the IGBP scheme and the PFT scheme shows that for 10 of the 11 PFT
classes, there is a direct mapping of one or more IGBP classes to their PFT equivalents. However, the PFT classes of cereal crop and broadleaf crop do not have
their equivalents in the IGBP scheme, because in the latter scheme, crops are either lumped together as croplands or as part of cropland/natural vegetation mosaic. Information regarding what procedures are used to separate cereal crop from
broadleaf crop in the production of the current MODIS PFT map is not available.
Figure 14.2 shows a MODIS PFT map for North America produced by the MODIS
Land Team.
MODIS land cover classes, and by extension the PFT classes, are distinguished
with a decision tree classification method. A decision tree is a supervised classifier
375
Table 14.1 The five classification schemes used in MODIS land cover
Land Cover Types
Class
Type 1
IGBP
Type 2
UMD
Type 3
LAI/FPAR
Type 4
NPP
Type 5
PFT
0
1
Water
Evergreen
needleleaf forest
Water
Evergreen
needleleaf forest
Water
Grasses/cereal
crops
Water
Evergreen
needleleaf trees
Evergreen
broadleaf forest
Evergreen
broadleaf forest
Shrubs
Deciduous
needleleaf forest
Broadleaf crops
Deciduous
broadleaf forest
Deciduous
needleleaf
forests
Deciduous
broadleaf forest
Mixed forests
Mixed forests
Broadleaf forest
6
7
Closed
shrublands
Open shrublands
Closed
shrublands
Open shrublands
Needleleaf
forest
Unvegetated
8
9
Woody savannas
Savannas
Woody savannas
Savannas
Urban
Water
Evergreen
needleleaf
vegetation
Evergreen
broadleaf
vegetation
Deciduous
needleleaf
vegetation
Deciduous
broadleaf
vegetation
Annual
broadleaf
vegetation
Annual grass
vegetation
Non-vegetated
land
Urban
10
11
Grasslands
Permanent
wetlands
Croplands
Urban and
built-up
Cropland/natural
vegetation
mosaic
Permanent snow
and ice
Barren or
sparsely
vegetated
Grasslands
12
13
14
15
16
Savanna
Evergreen
broadleaf trees
Deciduous
needleleaf trees
Deciduous
broadleaf trees
Shrub
Grass
Cereal crop
Broadleaf crop
Urban and built
up
Snow and Ice
Barren or sparse
vegetation
Croplands
Urban and
built-up
Barren or
sparsely
vegetated
that recursively partitions a data set into smaller sub-divisions via binary rules and a
heterogeneity-minimization function (Breiman et al., 1984). The tree is composed of
a root node, a set of intermediate notes (splits), and a set of terminal nodes (leaves).
In this framework, a data set is classified by sequentially sub-dividing it according to the decision framework defined by the tree, and class labels are assigned to
each observation according to the leaf node into which the observation falls. Several
studies have demonstrated the utility of decision trees in land cover classification at
376
W. Sun, S. Liang
Evergreen needleleaf tree
Evergreen broadleaf tree
Deciduous needleleaf tree
Deciduous broadleaf tree
Shrub
Grass
Cereal crop
Broadleaf crop
Urban and built-up
Snow and ice
Barren or sparsely vegetated
Fig. 14.2 An example of MODIS PFT map for North America
regional to global scales (e.g., Hansen et al., 1996, Friedl and Brodley, 1997; DeFries
et al., 1998; Friedl et al., 1999). Decision tree classifiers have several advantages
over traditional supervised classification procedures used in remote sensing such as
maximum likelihood classification (Hansen et al., 1996; Friedl and Brodley, 1997).
First, decision trees are nonparametric and do not require any assumptions regarding the distributions of the input data. Second, decision trees can handle noisy or
missing features and capture nonlinear and hierarchical relationships between the
input variables. Third, decision trees have significant intuitive appeal because the
classification structure is explicit and therefore easily interpretable.
The MODIS land classification method exploits spectral and temporal information from MODIS. Key inputs include Nadir BRDF-Adjusted Reflectances
(NBARs) derived from the MODIS BRDF/Albedo product (MOD43B4) in the
MODIS Land Bands (17), MODIS Enhanced Vegetation Index (EVI) (MOD13),
etc. (Strahler et al., 1999; Friedl et al., 2002). These data are composited over a
32-day period to produce a globally consistent, multitemporal database on a 1 km
grid as input to the classification algorithm. Land cover classes are assigned by
processing 12 (annual) 32-day composites using a decision tree classifier trained
by site data. The success of decision tree classifiers requires extensive, high quality training site data base. The System for Terrestrial Ecosystem Parameterization
(STEP) database is used to train the MODIS decision tree classifier. The STEP database is based on the information interpreted from Landsat and ancillary data. Key
STEP parameters include vegetation life form, cover fraction, leaf types phenology,
elevation, moisture regime, disturbance as well as descriptions of site and type.
A more detailed discussion of the decision tree algorithm and its implementation in
the production of MODIS Land Cover product is beyond the scope of this chapter.
377
The reader is referred to Strahler et al. (1999), Hansen et al. (1996), Friedl and
Brodley (1997), and Friedl et al. (1999) for more technical details.
14.2.3 Multisource Evidential Reasoning

Sun et al. (2005, 2007) recently presented a multisource evidential reasoning method
for mapping PFTs from MODIS data. Their method first utilizes a suite of improved
and standard MODIS products to generate evidence measures for each PFT class.
The multiple lines of evidence computed from input data are then combined using
DempsterShafer theory of evidence (Shafer, 1976) to make classification decisions.
In the Sun et al. (2007) study, the evidential reasoning method is implemented in
three steps.
Step 1: Generating evidence measures for each PFT class from each input data
source. The PFT classification scheme used in the Sun et al. (2007) study is the same
as the one used in the MODIS PFT product, which consists of 12 PFTs (Table 14.1).
In DempsterShafer theory of evidence, a set of these 12 PFT classes constitute the
frame of discernment, denoted by . The degree of belief in the evidence from an
input data source (e.g., LAI data) in support of a PFT class (e.g., grass) is expressed
in terms of a mass function (m). A mass function has the following property:
m(X) = 1
(14.1)
m() = 0
(14.2)
XP()
where P() is the power set consisting of 2C subsets for a set of size C; is the
empty set. The size of a set is the number of singleton classes in the frame of discernment. In the Sun et al. (2006) study, they used C instead of 2C because their
study focused only on singleton hypotheses.
Due to its generality, DempsterShafer theory of evidence does not specify how
to compute evidence measures. Sun et al. (2007) developed a three-step procedure
to derive masses of evidence. First, a mean vector over a whole year or the growing
seasoning (AprilOctober) is computed for each PFT class using each input data
source. An example of the computed mean values of EVI for 10 PFT classes over
the year 2001 for the state of Illinois, USA is given in Fig. 14.3. Second, the spectral
distances of each pixel to the mean vector of each PFT class are then calculated for
each input data source. Finally, these spectral distances are converted to probabilities
of class membership (i.e., mass functions) using the sigmoidal fuzzy set membership
function or a weighing function (Liang, 2004). Figure 14.4 is an example showing
the masses or probabilities of each pixel belonging to three PFT classes computed
from EVI data for Illinois, USA. Note that the higher the grayscale value of a pixel,
which appears darker in the image, the higher the probability of that pixel belonging
to a certain PFT.
378
W. Sun, S. Liang
0.7
Evergreen Needleaf Tree
0.6
Evergreen Broadleaf Tree
0.5
Deciduous Needleleaf Tree

Deciduous Broadleaf Tree
0.4
Shrub
0.3
Grass
0.2
Cereal Crop
0.1
Broadleaf Crop
Urban and Build-up
Barren or Sparsely Vegetated
35
32
28
25
22
3
19
16
9
12
97
65
0.1
33
Fig. 14.3 Mean values of EVI for each PFT class over the year 2001 for Illinois, USA (X axis =
day of the year, Y axis = EVI value)
Fig. 14.4 Evidence measures computed from EVI data for Illinois, USA showing the masses
(or probabilities) of each pixel belonging to (a) deciduous broadleaf trees, (b) grass, and
(c) broadleaf crop
Step 2: Combining evidence from all sources. Once the masses of evidence from
all sources for each PFT have been determined, they are combined using Dempsters
rule of combination (Dempster, 1967). The equation for computing the orthogonal
sum () of source 1 (with mass m1 over a set of labels X) and source 2 (with m2
over a set of labels Y ) is as follows:
m1 m2 (Z) =
where:
k=
XY =Z
379
m1 (X)m2 (Y )
1k
m1 (X)m2 (Y )
(14.3)
(14.4)
XY =
k indicates the extent of conflict between the two sources considered (Shafer, 1976).
Orthogonal summation of additional sources is achieved by repeated application of
Eqs. (14.3) and (14.4).
Step 3: Making classification decisions. To classify a pixel into one of the PFT
classes, a decision rule is applied to the measure of support and/or plausibility. Support or belief function (Bel) is the total belief of a set and all its subsets. It is defined
in terms of the mass:
(14.5)
Bel(X) = m(H)
HX
where H represents any subset of a set X.

Plausibility (Pls) is defined as the degree to which a proposition, X, cannot be
disbelieved or refuted. It is calculated as one minus the support for all other propositions (Shafer, 1976).
(14.6)
Pls(X) = 1 Bel(X )
where (X ) is not (X).
In the context of remote sensing image classification, Bel defines the lower
boundary of the support committed to a class labeling proposition, Pls defines an
upper boundary, and the range [Bel, Pls] is referred to as evidential interval. In the
Sun et al. (2007) study, a maximum support decision rule is used, that is, the class
with the highest support is assigned as the pixel label. The reader is directed to Sun
et al. (2007) for a more detailed discussion of the theoretical background and key
steps in the implementation of the evidential reasoning algorithm.
The method of Sun et al. (2007) was tested over the states of Illinois, Indiana,
Iowa and North Dakota, USA. They supplied the evidential reasoning algorithm
with nine MODIS data sets, that is, improved Leaf Area Index (LAI) product obtained from a NASA funded data assimilation project, MODIS Enhanced Vegetation Index (EVI) (MOD13A2), and seven bands of black-sky spectral albedos
(MOD43B3). Sun et al. (2007) reported that for several major PFTs in the study
areas, the PFT maps generated with the evidential reasoning classifier represent significant improvements over the MODIS PFT product.
Figure 14.5c shows an example of the classification results from evidential reasoning for the state of Illinois, USA. As a first step in the validation of their results
and the MODIS PFT products, Sun et al. (2007) used the Cropland Data Layer
(CDL) data provided by the National Agricultural Statistics Services (NASS) of
the United States Department of Agriculture (USDA) as reference data (Fig. 14.5b).
The NASS data was derived from high-resolution Landsat ETM+ data and its spatial
resolution is 30m. As can be seen from Fig. 14.5, the PFT map generated with evidential reasoning gives much more spatial details of the PFTs in Illinois compared
380
W. Sun, S. Liang
Fig. 14.5 Comparison of MODIS PFT, USDA NASS data, and PFT classification results generated with evidential reasoning for Illinois, USA: (a) MODIS PFT map, (b) USDA NASS data
used as ground truth, and (c) evidential reasoning classification results using evidence measures calculated from Illinois state growing season means plus weighing factors (weights used:
LAI = 0.001, EVI = 0.6, albedo 1, 2, 3, 4 = 0.2, albedo 5, 6, 7 = 0.1)
Table 14.2 Percent area of each PFT class identified in USDA NASS data (NASS), MODIS PFT
(MODIS), and evidential reasoning classifications (EV) for Illinois, USA
Class
Trees
Grass and shrub
Crop
Urban and built-up
Water and wetland
Clouds
Snow and ice
NASS
11.63
19.54
59.04
5.64
1.61
2.53
0
MODIS
2.63
0.87
92.11
3.13
1.25
0
0
MODIS NASS
9
18.67
33.07
2.51
0.36
2.53
0
EV
11.61
12.01
72.88
2.65
0.86
0
0
EV NASS
0.02
7.53
13.84
2.99
0.75
2.53
0
to the MODIS PFT map (Fig. 14.5a). In terms of percent area of each PFT class,
the results obtained from evidential reasoning are much closer to the NASS data
(ground truths), whereas large discrepancies exist between the MODIS PFT map
and the NASS data (Table 14.2). It appears that the MODIS PFT map grossly overestimated the crop class, while it underestimated other important PFT classes such
as trees and grass and shrub. This observation applies to all of the four states examined in their study.
381
Sun et al. (2007) also experimented the evidential reasoning algorithm with various combinations of input data to examine the sensitivity of classification results to
input data. A main conclusion from their experiment is that careful selection of input
data is critical to obtaining satisfactory results from the evidential reasoning method.
They also argued that despite the encouraging results from their study, more work
is needed to validate their method over other regions, at other geographic scales
(e.g., continents to global) and using additional sources of information including
ancillary data such as climate and terrain. Overall, the work of Sun et al. (2005,
2007) demonstrates that multisource evidential reasoning is a promising approach
to improved mapping of PFTs from MODIS data.
14.3 Key Issues In Mapping of PFTs Over Large Areas

The MODIS PFT product (MODIS Land Cover 5) is the only global PFT data
set produced on a regular basis at the present time. The current MODIS PFT is a
land cover-derived product in that it is produced by cross-walking the land cover
classes in the IGBP scheme to MODIS PFT classes. This means that, for the most
part, the procedures currently used to produce the MODIS PFT map are not developed specifically for extracting PFT information from MODIS. Rather, the overall
methodology employed by the MODIS Land Team is geared toward distinguishing land cover classes based on the IGBP scheme. Although the IGBP scheme has
gained wide acceptance by the remote sensing community, it is fundamentally a
biome-based land cover classification system. It seems arguable that methodologies
effective for extracting biome-based land cover classes may not be optimal for mapping PFTs. Thus, at the present time it appears that no satisfactory methodology
exists for the extraction of PFT information from satellite observations.
An important reason why the land research community has requested for a PFT
product is that reliable PFT data would give modelers more confidence in determining land surface processes and parameters for use in carbon cycle, climate and
ecosystem models. Validation of the current MODIS PFT product is a demanding
task and science quality information about the accuracy of the MODIS PFT data
appears scarce. Errors and uncertainties in PFT data can propagate and compromise
the credibility of global change research. As such, it seems desirable to develop
and test new approaches to generate new PFT products. Such new products can be
used as an independent validation method for the MODIS PFT, in addition to the
validation activities undertaken by the MODIS Land Team.
Due to the enormous diversity of terrestrial plant species and the spatial and
temporal variability in the morphological and spectral characteristics of PFTs
across regions, climates and years, accurate mapping of major PFTs around the
globe is a difficulty task (Prentice et al., 1992; Box, 1996; Smith et al., 1997;
Semenova and van der Maarel, 2000). The amount and quality of the data being
generated by EOS sensors and other sensors are truly unprecedented (Justice et al.,
2002). The increased availability and information content of remotely sensed data
382
W. Sun, S. Liang
has provided considerable potential for generating global PFT data sets. To date,
however, research on how to effectively utilize multi-sensor data to accurately
map PFTs remained quite limited. We provide below a brief discussion of several
methodological issues that appear to be critical in the design and implementation of
robust methodologies for extracting PFTs over large areas.
(1) PFT classification scheme. The search for plant functional types represents a
long standing desire of ecologists to seek simplified explanatory variables for
understanding patterns in the richness of plants and the complexity of ecosystems. A review of the literature shows that there is a noticeable lack of consistency regarding how to define and group the major PFTs around the globe
(Box, 1996; Woodward and Cramer, 1996; Smith et al., 1997). The PFT scheme
used in the production of MODIS PFT is adopted from a scheme described in
Bonan et al. (2002). The PFT scheme of Bonan et al. (2002) itself is adapted
from the logic of Running et al. (1995) and Nemani and Running (1996).
Briefly, the PFT scheme of Bonan et al. (2002) consists of seven primary PFTs
(i.e., needleleaf evergreen or deciduous tree, broadleaf evergreen or deciduous
tree, shrub, grass, and crop) and these primary PFTs are expanded to 15 physiological variants based on climate rules.
McIntyre et al. (1999) call for simplified functional species classifications on the
premise that the fewer groups used the greater the chance that broad patterns can be
discerned. However, it is likely that the utility of PFT classifications will be scaledependent. Broad categories are likely to be most useful when considering patterns
over large geographic areas such as continents or global, while more narrowly defined types may be necessary at more local scales. In order to accommodate the
needs of environmental research at varying geographic scales, it seems desirable to
develop a system of PFT classifications that can be used in remote sensing of PFTs
at a hierarchy of scales.
(2) How to characterize PFTs. PFTs can and should be characterized by a variety of variables in domains such as plant physiognomy, vegetation structure,
phenology, and environmental conditions (Running et al., 1995). However, the
synthesis we have done of the research on PFTs to date shows that the question
of what variables are most crucial to distinguishing major PFTs using remote
sensing techniques remains unanswered. As such, there is an urgent need to
identify a set of key variables that together can effectively characterize major
PFTs and therefore should be used to extract PFTs using remote sensing techniques.
(3) Limitation of remote sensing instruments. While some of the characteristics exhibited by individual PFTs such as their phenologies and stand structure are observable by remote sensing, others may not. For example, certain site-specific
environmental and ecological conditions such as climate and terrain appear to
be less observable, but they are among the most important factors determining
the geographic distribution of PFTs. As such, it seems that the use of remotely
sensed data alone is inadequate to distinguish PFTs. This is especially true of
mapping PFTs over large areas and using moderate resolution satellite data.
383
(4) Limitation of spectral information. Remote sensing scientists have long recognized that the use of spectral information alone to interpret remote sensing data
is seriously inadequate (Campbell, 1978; Townshend and Justice, 1981; Sun
et al., 2003). Many studies, though not directly related to mapping of PFTs, have
demonstrated that, when mapping land cover types over large areas and using
moderate resolution satellite data, ancillary data such as climate and elevation
contributed essential evidence for postclassification refinement and/or labeling
of land cover classes where differing types exhibited similar spectral-temporal
signatures (e.g., DeFries et al., 1998; Loveland et al., 2000; Hansen et al., 2000).
The classification method used in the production of MODIS Land Cover products relies primarily on spectral and temporal information in MODIS. As such,
how to integrate a variety of ancillary data into PFT classification procedures is
an issue that deserves further research.
(5) Environmental complexity of large geographic areas. When mapping PFTs over
large areas (e.g., continents and global), the huge differences in climate and terrain across vast landmasses will almost certainly complicate image interpretation (Brown et al., 1993). This again suggests that integration of both satellite
data and other relevant spatial data describing the environmental conditions of
major PFTs will be critical to reliable mapping of PFTs, especially at regional
to global scales. Ideally it would be useful for the remote sensing community
to synthesize and encode a system of environmental and ecological constraints
in some sort of hierarchy from major constraints operating at the global scale
down to constraints operating at the regional and local level. Such knowledge
may prove extremely valuable in improving the mapping of PFTs.
(6) Need to integrate a priori knowledge. Although remote sensing of PFTs is a
relatively recent field, there are multiple sources of knowledge that can be used
to infer PFTs. For example, knowledge exists on what climatic conditions or
thresholds (e.g., minimum temperature, occurrence of frost or freezing, and water moisture) are required for a plant species to occur on the earths surface.
Such knowledge can be encoded in the form of a system of climatic envelopes
and utilized, for example, to differentiate tropical, subtropical, temperate and
boreal varieties of PFTs. There is also considerable knowledge about the spatial
associations of major PFTs. For example, evergreen needleleaf trees and evergreen broadleaf trees rarely intermix geographically and, therefore, they could
be separated by simple climate rules. Models of elevation-plant species relationships (altitudinal zonation) can also be incorporated into PFT classification
procedures.
Knowledge of the spectral characteristics of major PFTs can be generated from
training sites. For example, the mean and variance of albedo, LAI, EVI, etc. can be
calculated for each PFT class. The phenologies of each PFT class can be modeled
using statistical techniques and the parameters derived from phenological models,
such as maximum EVI, minimum EVI, greenup, maturity, senescence and dormancy
onset dates, duration of growing seasons, etc., can be used to distinguish PFTs. Certain statistical rules can also be established, such as the EVI of broadleaf evergreen
384
W. Sun, S. Liang
trees has to be larger than a threshold value, or shrub albedo values have to be larger
than certain threshold values, or the LAI/albedo values of grass are within an envelop for a given time.
The above discussion suggests that considerable work is still needed before PFTs
can be reliably mapped over large areas and across years. An important direction in
the development of new methodologies for extracting PFTs appears to be the need
to bring much more information, including both satellite observations and ancillary
data, to bear on the PFT identification problem than is currently used in the MODIS
product.
14.4 Developing a Multisource Data Fusion Approach

to Extracting PFTs
In this section, we outline a methodology that we are currently developing to extract
PFTs from MODIS data. A key feature of our approach is that it utilizes not only
spectral data but also ancillary information to infer PFT from satellite observations.
It should be noted that our methodology is not specific to PFT mapping. It can
also be applied to mapping of traditional land cover types.
14.4.1 Overall Framework

The methodology described here is a fully automated PFT classification system capable of fusing multiple sources of data and information using a DempsterShafer
belief system. Our methodology is a hybrid approach in that it combines both physically based methods and rule-based methods. Past research has demonstrated that
purely physically based methods using spectral information alone often lead to poor
results (e.g., Campbell, 1978; Townshend and Justice, 1981; DeFries et al., 1998;
Loveland et al., 2000; Hansen et al., 2000). At the other end of the spectrum, purely
empirical rule-based approaches have also been tried in the past with marginal success. As such, integrating physically based and rule-based methods into a single
classification system appears to be a sound strategy for mapping PFTs over large
areas.
Our multisource data fusion framework consists of three main components:
a knowledge base, evidence generation, and evidential reasoning algorithms (see
Fig. 14.6). Each of these components is described in more details in the following
sections.
1. The knowledge database contains a set of key parameters and rules essential to
discerning major PFTs.
2. The evidence generation module is used to generate masses of evidence for PFT
classification from each input data source using a variety of techniques.
MODIS data &

products
Evidence
generation
385
Other spatial
data
Knowledge
base
Evidential reasoning
algorithm
PFT product
Validation
Fig. 14.6 Flowchart of the multisource data fusion approach
3. The evidential reasoning classifier collects masses of evidence and combines

them to compute the belief and plausibility of a pixel belonging to a particular
PFT class.
14.4.2 Developing a Knowledge Database

Our methodology relies greatly on the quality and amount of evidence and
knowledge that we can provide to the evidential reasoning classifier. As discussed
above, PFTs can be characterized in multiple domains including plant physiognomy, vegetation structure, phenology, and environmental conditions (Box, 1995;
Running et al., 1995). An effective approach to extracting PFTs should therefore
take into account the manifestations of PFTs in all of the above domains. To encode existing knowledge about the physiological, phenological, and ecological
properties of each PFT and to organize various parameters and decision rules
in a structured manner, we are currently constructing a comprehensive knowledge base.
Our knowledge database contains a large number of parameters and rules that
are essential to discerning major PFTs using remote sensing techniques. The main
areas from which knowledge is being extracted include:
386
W. Sun, S. Liang
Table 14.3 Examples of the key items in the knowledge database

Parameters/rules
Plant physiognomy
Growth and life form
Leaf longevity
Canopy structure/architecture
Vegetation structure
Fractional vegetation cover
Phenology
Intra-annual change
Inter-annual change
Environmental conditions
Temperature
Precipitation
Terrain
Spatial
Texture
Spatial association
Used to separate
Input data/knowledge
Perennial versus annual

Evergreen versus deciduous
Broadleaf versus needleleaf
MODIS albedo, LAI, etc.

MODIS albedo, LAI, etc.
MODIS EVI, NDVI, LAI, etc.
Woody versus nonwoody
MODIS LAI, etc.
Growing activity
Vegetation change, climatic
impacts
EVI, NDVI, LAI

AVHRR NDVI, MODIS
NDVI, etc.
Limit to vegetation growth
MOD LST, long-range global

monthly temperature data
Limit to vegetation growth
Long-rang global monthly
precipitation data
Spatial context, specially
Elevation-vegetation models,
elevation zonation of vegetation Global digital elevation models
(DEM)
Different forms of plants
MOD albedo, etc.
Spatial vegetation associations Meta-analysis of known
associations published in the
literature
Physiological and phenological characteristics of each PFT class

Environmental conditions (e.g., climatic envelopes and elevation-PFT models)
determining the geographic distribution of PFTs
Spatial associations between PFTs
Image characteristics of each PFT class in terms of its textures, etc.
Examples of the key items used in the knowledge database are given in
Table 14.3. It should be noted that the knowledge database can be easily updated and refined when new knowledge becomes available. This will allow us to
continuously improve the accuracies of PFT product.
14.4.3 Generating Evidence for PFT Classification

Integrating a wide array of data and information from multiple sources in a carefully structured manner is a key feature of the methodology described here. In our
framework, we ingest a variety of both remotely sensed and other spatial data into
the data system for generation of evidence to support PFT classification decisions.
Besides standard MODIS products and the data acquired by other sensors, we use
a suite of improved higher-level MODIS products (e.g., albedo and LAI) generated
from a NASA funded data assimilation project. A range of ancillary data pertaining
to the environmental conditions associated with each PFT class is also fused into
the data system.
387
Technically, the evidence generation module involves transforming the input data
into measures of evidence for each PFT class with the aid of the knowledge database
described above. The measures of evidence generated in this step will be input to
the evidential reasoning algorithm. In our framework, we employ three techniques
to accomplish this transformation.
1. Distances of each pixel to the mean vectors of each PFT class are first calculated
for each source of information used in the classification system. These distances
are then converted into probability images, with a higher probability indicating a
higher level of belief of a pixel belonging to a particular PFT class. A variety
of weighing functions and fuzzy set membership functions are used to convert
distances into measures of evidence. Below are examples of the three weighing
functions currently used in our work (Liang, 2004):
.
R2 di,2 j
(14.7)
w(ri , r j ) = max 0, 2
R + di,2 j
.
di,2 j
w(ri , r j ) = exp 2
(14.8)
2R

di, j
di, j
w(ri , r j ) = 1 +
exp
(14.9)
R
R
where: w(ri , r j ) is weighing function dependent on the distance di, j between point
ri and r j ; R defines the spatial location in the one-, two-, or three-dimensional
domain.
2. Measures of evidence computed from training data within a supervised classification framework. In this approach, evidential support is computed with respect
to the frequency of occurrence of pixel values within training samples (Peddle,
1995a). This technique is considered more objective and involves identifying representative areas within an image for each PFT class. The underlying premises to
this method are that training data contain evidence for a set of classes, and that
the frequency of occurrence of a given value in the training set represents the
magnitudes of support for those classes.
3. Rule-based measurements. Our knowledge database also contains a set of rules
that are encoded to generate evidence from certain data sources (climate, elevation, etc.) for use in the evidential reasoning algorithm. Each rule has an antecedent, a consequent, and a weight. The antecedent consists of certain attribute
values describing a pixel in the image. The consequent is the resultant assertion
that the pixel belongs to a certain PFT class. The weight is the confidence in
the assertion. Briefly, the rule-based measurements are obtained in the following way. Given an image, the method computes a set of attribute values for each
pixel. Next, the classifier feeds pixels with the corresponding attribute values
into a rule-based system. A pixel may trigger the firing of multiple rules, asserting complimentary or conflicting PFT classifications with a confidence value or
weight.
388
W. Sun, S. Liang
14.4.4 Evidential Reasoning Classifiers

For each pixel, the evidence generation module described above generates a set of
basic probability assignments (or masses of evidence) from all sources of information used. A basic probability assignment ranges between 0 and 1 for positive
belief, and between 1 and 0 for negative belief. Basic probability assignment of
1.0 and 1.0 indicate absolute certainty that a feature belongs or does not belong
to a class, respectively. To determine which PFT class (if any) is supported by the
strongest evidence, we need to combine these masses of evidence in a consistent
manner. In our framework, this is done with an evidential reasoning algorithm.
In the evidential reasoning method, the masses of evidence from all data sources
for each PFT are combined using Dempsters rule of combination (Eqs. (14.3) and
(14.4)). Once the evidence from each source of information is combined by repeated application of the orthogonal summation rule, a decision rule is applied
to the mass function to classify the pixel into one of the classes within the frame
of discernment. In our classification system, the decision rule is based on maximum support, where the class with the highest support is assigned as the pixel
label.
A number of studies have demonstrated that evidential reasoning is a powerful
approach to data fusion (Peddle, 1995a, b; Lein, 2003; Le Hegarat-Mascle et al.,
2003; Soh et al., 2004). The method is built on the mathematical theory of evidence originally developed by Dempster (1967) and extended and refined by Shafer
(1976). Evidential reasoning is based on the recognition that the knowledge and
information we use in making decisions such as image classification is often uncertain, incomplete, and imprecise. As such, the method mimics human decisionmaking processes and utilizes many pieces of evidence to support a decision. As
evidence increases, the algorithm will narrow the hypothesis set down to a smaller
number of possibilities and improve its confidence in making a decision. Because
the algorithm bases its decision on many pieces of evidence, it can usually generate
better results compared to those classification procedures that rely on a single data
source.
A fuller discussion of the evidential reasoning method is beyond the scope of this
chapter. It suffices here to note that evidential reasoning offers several important
advantages over traditional image classifiers (Peddle, 1995a, b). First, it is a nonparametric classifier and therefore can handle data which may violate the Gaussian
assumption of parametric classifiers. Second, it can handle data from any number of
sources at any measurement scale. Third, the method has an explicit mechanism for
handling information uncertainty. In situations of missing or incomplete information, evidential reasoning provides a more accurate representation of the available
knowledge compared with the arbitrary assignment of probability used in Bayesian
theory. Fourth, it can provide several interpretive measures such as support, ignorance and conflict that can be used to assess uncertainties in classification results.
For example, the evidential reasoning method possesses the ability to express ignorance, which describes the incompleteness of ones knowledge as a measure of the
degree to which we cannot distinguish between any of the PFT classes. As such,
389
ignorance can be viewed as a measure of uncertainty in the resultant classification.

From a users point of view, information about the degree of ignorance (class confusion) may prove valuable.
14.5 Conclusion
The production of an improved global PFT product at the MODIS scale should be of
great interest to the global change research community. Improved mapping of PFTs
can contribute to improved ability to simulate carbon cycle, climate and ecosystem
change at regional to global scales. Using remote sensing techniques to map PFTs
over large areas is a relatively recent and therefore evolving field of research. In this
chapter, we briefly reviewed the methodologies developed by MODIS Land Team,
Sun et al. (2005, 2007), and the land modeling community to extract PFTs from
satellite observations. It seems clear that to date a limited number of methods for
mapping PFTs have been reported in the literature. As such, developing and testing
a variety of data analysis methods and extraction techniques for inferring PFTs from
remotely sensed data constitutes a fertile field of research for the remote sensing
community.
In this chapter, we also discussed a number of methodological issues that deserve special attention in the development of new approaches to extracting PFTs
over large areas. An important conclusion that we have drawn from this discussion is that incorporation of a wide array of information including both satellite
observations and ancillary data into PFT classification procedures is indispensable
to reliable mapping of PFTs. Integrating both satellite and other digital spatial data
into a single classification procedure is indeed very challenging, because data from
different sources often differ in spatial and temporal resolution, scale of measurement, accuracy, completeness, etc. Our ongoing research has shown that evidential
reasoning is a promising method capable of fusing a large number of data sets and
is effective in extracting and utilizing information from various data sources to generate reliable results. It is our hope that the multisource data fusion framework outlined in this chapter could spark further interest in the development of innovative
and robust extraction techniques for inferring PFTs from remotely sensed data.
References
Benediktsson JA, Swain PH, Ersoy OK (1990) Neural network approaches versus statistical methods in classification of multisource remote sensing data. IEEE Trans. Geosci. Remote Sens.
28:540552
Bonan GB (1993) Importance of leaf area index and forest type when estimating photosynthesis in
boreal forests. Remote Sens. Environ. 43:303314
Bonan GB (1995) Land-atmosphere CO2 exchange simulated by a land surface process model
coupled to an atmospheric general circulation model. J. Geophys. Res. 100:28172831
390
W. Sun, S. Liang
Bonan GB (1996) A land surface model (LSM version 1.0) for ecological, hydrological, and atmospheric studies: technical description and users guide. NCAR Technical Note NCAR/TN417 + STR. National Center for Atmospheric Research, Boulder, Colorado, 150pp
Bonan GB, Levis S, Kergoat L, Oleson KW (2002) Landscapes as patches of plant functional
types: an integrating concept for climate and ecosystem models. Global Biogeochem. Cycles
16(2), doi:10.1029/2000GB001360
Box EO (1995) Factors determining distributions of tree species and plant functional types. Vegetatio 121(1/2):101116
Box EO (1996) Plant functional types and climate at the global scale. J. Veg. Sci. 7:309320
Breiman L, Friedman JH, Olshen RA, Stone CJ (1984) Classification and regression trees.
Wadsworth, Belmont, CA.
Brown JF, Loveland TR, Merchant JW, Reed BC (1993) Using multisource data in global land
cover characterization: concepts, requirements and methods. Photogramm. Eng. Remote Sens.
59:977987
Campbell JB (1978) A geographical analysis of image interpretation methods. Prof. Geogr.
30:264269
DeFries RS, Hansen M, Townshend JGR, Sohlberg R (1998) Global land cover classification at
8 km spatial resolution: the use of training data derived from Landsat imagery in decision tree
classifiers. Int. J. Remote Sens. 19:31413168
DeFries RS, Townshend JRG, Hansen MC (1999) Continuous fields of vegetation characteristics
at the global scale at 1-km resolution. J. Geophys. Res. 104D:1691116923
DeFries RS, Hansen MC, Townshend JRG (2000a) Global continuous fields of vegetation characteristics: a linear mixture model applied to multi-year 8 km AVHRR data. Int. J. Remote Sens.
21:13891414
DeFries RS, Hansen MC, Townshend JRG, Janetos AC, Loveland TR (2000b) A new global 1-km
dataset of percentage tree cover derived from remote sensing. Global Change Biol. 6:247254
Dempster AP (1967) Upper and lower probabilities induced by multivalued mappings. Ann. Math.
Stat. 38:325329
Denning AS, Collatz GJ, Zhang C, Randall DA, Berry JA, Sellers PJ, Colello GD, Dazlich DA
(1996) Simulations of terrestrial carbon metabolism and atmospheric CO2 in a general circulation model. Part 1: surface carbon fluxes, Tellus 48B:521542
Dickinson RE, Shaikh M, Bryant R, Graumlich L (1998) Interactive canopies for a climate model.
J. Climate 11:28232836
94:405424
Feddema JJ, Oleson KW, Bonan GB, Mearns LO, Buja LE, Meehl GA, Washington WM (2005)
The importance of land-cover change in simulating future climates. Science 310:16741678
Foley JA, Prentice CI, Ramankutty N, Levis S, Pollard D, Sitch S, Haxeltine A (1996) An integrated biosphere model of land surface processes, terrestrial carbon balance, and vegetation
dynamics. Global Biogeochem. Cycles 10:603628
Friedl MA, Brodley CE (1997) Decision tree classification of land cover from remotely sensed
Friedl MA, Brodley CE, Strahler AH (1999) Maximizing land cover classification accuracies produced by decision trees at continental to global scales. IEEE Trans. Geosci. Remote Sens.
37:969977
Friedl MA, McIver DK, Hodges JCF, Zhang X, Muchoney D, Strahler AH, Woodcock CE, Gopal
S, Schnieder A, Cooper A, Baccini A, Gao F, Schaaf C (2002) Global land cover mapping from
MODIS: algorithms and early results. Remote Sens. Environ. 83(12):287302
Gamon JA, Huemmrich KF, Peddle DR, Chen J, Fuentes D, Hall FG, Kimball JS, Goetz S, Gu J,
McDonald KC, Miller JR, Moghaddam M, Rahman AF, Roujean JL, Smith EA, Walthall CL,
Zarco-Tejada P, Fernandes R, Cihlar J (2004) Remote sensing in BOREAS: lessens learned.
391
Hansen M, Dubayah R, DeFries R (1996) Classification trees: an alternative to traditional land

cover classifiers. Int. J. Remote Sens. 17:10751081
Hansen MC, DeFries RS, Townshend JRG, Sohlberg R (2000) Global land cover classification at
1 km spatial resolution using a classification tree approach. Int. J. Remote Sens. 26:13311364
Justice CO, Townshend JRG, Vermote E, Masuoka E, Wolfe R, Saleous N, Roy D, Morisette J
(2002) An overview of MODIS land data processing and product status. Remote Sens. Environ.
83:315
Kimes DS, Harrison PR, Ratclife PA (1991) A knowledge-based expert system for inferring vegetation characteristics. Int. J. Remote Sens. 12:19872020
Kucharik CJ, Foley JA, Delire C, Fisher VA, Coe MT, Lenters JD, Young-Molling C,
Ramankutty N, Norman JM, Gower ST (2000) Testing the performance of a dynamic global
ecosystem model: water balance, carbon balance, and vegetation structure. Global Biogeochem.
Cycles 14:795825
Lawton RO, Nair US, Pielke Sr. RA, Welch RM (2001) Climatic impact of tropical lowland deforestation on nearby montane cloud forest. Science 294:584587
Le Hegarat-Mascle S, Richard D, Ottle C (2003) Multi-scale data fusion using Dempster-Shafer
evidence theory. Int. Comput.-Aid. Eng. 10:922
Lein JK (2003) Applying evidential reasoning methods to agriculture land cover classification. Int.
Liang S (2001) Land cover classification methods for multiyear AVHRR data. Int. J. Remote Sens.
22:14791493
Liang S (2003) A direct algorithm for estimating land surface broadband albedos from MODIS
Imagery. IEEE Trans. Geosci. Remote Sens. 41:136145
Liang S (2004) Quantitative remote sensing of land surfaces. Wiley, New York, 534pp
Lloyd D (1990) A phenological classification of terrestrial vegetation cover using shortwave vegetation index imagery. Int. J. Remote Sens. 11:22692279
Loveland TR, Merchant JW, Ohlen DO, Brown JF (1991) Development of a land cover characteristics database for the counterminous US. Photogramm. Eng. Remote Sens. 57:14531463
Loveland TR, Merchant JW, Reed BC, Brown JF, Ohlen DO, Olson P, Hutchinson J (1995) Seasonal land cover regions of the United States. Ann. Assoc. Am. Geogr. 85:339355
Loveland TR, Reed BC, Brown JF, Ohlen DO, Zhu Z, Yang L, Merchant JW (2000) Development
of a global land cover characteristics database and IGBP DISCover from 1 km AVHRR data.
Int. J. Remote Sens. 21:13031330
Marland G, Pielke Sr. RA, Apps M, Avissar R, Betts RA, Davis KJ, Frumhoff PC, Jackson ST,
Joyce LA, Kauppi P, Katzenberger J, MacDicken KG, Neilson RP, Niles JO, Niyogi DS, Norby
RJ, Pena N, Sampson N, Xue Y (2003) The climatic impacts of land surface change and carbon management, and the implications for climate-change mitigation policy. Climate Policy
3:149157
McIntyre S, Diaz S, Lavorel S, Wolfgang C (1999) Plant functional types and disturbance
dynamics introduction. J. Veg. Sci. 10:604608
Myneni RB, Maggion S, Iaquinta J, Privette JL, Gobron N, Pinty B, Verstraete MM, Kimes DS,
Williams DL (1995) Optical remote sensing of vegetation: modeling, caveats and algorithms.
northern high latitudes from 19811991. Nature 386:698702
Nair US, Lawton RO, Welch RM, Pielke Sr. RA (2003) Impact of land use on Costa Rican tropical
montane cloud forests: sensitivity of cumulus cloud field characteristics to lowland deforestation. J. Geophys. Res. 108:D7, 4206, doi:10.1029/2001JD001135
Nemani R, Running SW (1996) Implementation of a hierarchical global vegetation classification
in ecosystem function models. J. Veg. Sci. 7:337346
Oleson KW, Bonan GB (2000) The effects of remotely sensed plant functional type and leaf
area index on simulations of boreal forest surface fluxes by the NCAR land surface model.
J. Hydrometeorol. 1:431446
392
W. Sun, S. Liang
Olson JS, Watts JA, Allison LJ (1983) Carbon in live vegetation of major world ecosystems.
ORNL-5862, Oak Ridge National Laboratory, Oak Ridge, TN
Peddle DR (1995a) Knowledge formulation for supervised evidential classification. Photogramm.
Eng. Remote Sens. 61:409417
Peddle DR (1995b) MERCURY: an evidential reasoning image classifier. Comput. Geosci.
21:11631176
Prentice IC, Cramer W, Harrison SP, Leemans R, Monserud RA, Solomon AM (1992) A global
biome model based on plant physiology and dominance, soil properties, and climate. J. Biogeogr. 19:117134
Prince SD, Goward SN (1995) Global primary production: a remote sensing approach. J. Biogeogr.
22:815835
Roughgarden J, Running SW, Matson PA (1991) What does remote sensing do for ecology?
Ecology 72:19181922
Running SW, Coughlan JC (1988) A general model of forest ecosystem processes for regional
applications. I. Hydrological balance, canopy gas exchange and primary production processes.
Ecol. Model. 42:125154
Running SW, Loveland TR, Pierce LL, Nemani RR, Hunt ER Jr. (1995) A remote sensing based
vegetation classification logic for global land cover analysis. Remote Sens. Environ. 51:3948
Sellers PJ, Schimel DS (1993) Remote sensing of the land biosphere and biogeochemistry in the
EOS era: science priorities, methods and implementation EOS land biosphere and biogeochemical cycles panels. Global Planet. Change 7:279297
Sellers PJ, Mintz Y, Sud YC, Dalcher A (1986) A simple biosphere model (SIB) for use within
general circulation models. J. Atmos. Sci. 43:505531
Sellers PJ, Meeson BW, Hall FG, Asrar G, Murphy RE, Schiffer RA, Bretherton FP, Dickinson
RE, Ellingson RG, Field CB, Huemmrich KF, Justice CO, Melack JM, Roulet NT, Schimel
DS, Try PD (1995) Remote sensing of the land surface for studies of global change: models
algorithms experiments. Remote Sens. Environ. 51:326
Sellers PJ, Dickinson RE, Randall DA, Betts AK, Hall FG, Berry JA, Collatz GJ, Denning
AS, Mooney HA, Nobre CA, Sato N, Field CB, Henderson-Sellers A (1997) Modeling the
exchanges of energy, water, and carbon between continents and the atmosphere. Science
275:502509
Semenova GV, van der Maarel E (2000) Plant functional types a strategic perspective. J. Veg.
Sci. 11:917922
Shafer G (1976) A Mathematical theory of evidence. Princeton University Press, Princeton, NJ
Smith TM, Shugart HH, Woodward FI (Eds.) (1997) Plant functional types: their relevance to
ecosystem properties and global change. Cambridge University Press, New York, 369 pp
Strahler A, Coauthors (1999) MODIS Land Cover Product Algorithm Theoretical Basis Document
(ATBD). Version 5.0. Boston University
Steyaert LT, Hall FG, Loveland TR (1997) Land cover mapping, fire regeneration, and scaling
studies in the Canadian boreal forest with 1 km AVHRR and Landsat TM data. J. Geophys.
Res. 102:2958129598
Soh L, Tsatsoulis C, Gineris D, Bertoia C (2004) ARKTOS: an intelligent system for SAR sea ice
image classification. IEEE Trans. Geosci. Remote Sens. 42(1):229247
Sun W, Heidt V, Gong P, Xu G (2003) Information fusion for rural land-use classification with high
resolution satellite imagery. IEEE Trans. Geosci. Remote Sens. 41:883890
Sun, W. (2004) Land-Use Classification Using High Resolution Satellite Imagery: A New Information Fusion Method An Application in Landau, Germany. University of Mainz Press, 95pp
Sun W, Liang S, Xu G, Fang H (2005) Improving MODIS PFT product using multisource evidential reasoning. Proceedings of the 9th International Symposium on Physical Measurements and
Signatures in Remote Sensing 1, pp. 225227
Sun W, Liang S, Xu G, Fang H, Townshend, JR, Dickinson R (2007) Mapping plant functional
types from MODIS data using multisource evidential reasoning. Remote Sensing of Environment (in press)
393
Tian Y, Dickinson RE, Zhou L, Myneni RB, Friedl M, Schaaf CB, Carroll M, Gao F (2004) Land
Geophys. Res. Lett. 31:L05504, doi:10.1029/2003GL019104
Townshend JRG, Justice CO (1981) Information extraction from remotely sensed data a user
view. Int. J. Remote Sens. 2:313329
Townshend JRG, Justice CO, Kalb V (1987) Characterization and classification of South American
land cover types. Int. J. Remote Sens. 8:11891207
Townshend JRG, Justice CO, Li W, Gurney C, McManus J (1991) Global land cover classification by remote sensing: present capabilities and future possibilities. Remote Sens. of Environ.
35:243255
Woodward FI, Cramer W (1996) Plant functional types and climatic changes: Introduction. J. Veg.
Sci. 7:306308
Part III
Remote Sensing Applications
Chapter 15
Monitoring and Management of Agriculture

with Remote Sensing
Zhongxin Chen, Sen Li, Jianqiang Ren, Pan Gong, Mingwei Zhang,
Limin Wang, Shenliang Xiao, and Daohui Jiang
The intrinsic characteristics of agriculture make remote sensing an ideal technique

for its monitoring and management (Chen et al., 2004). These characteristics include: (a) Agricultural activities are usually carried out in large spatial regions,
which makes the conventional field survey or census time-consuming and usually
costly; (b) the per-unit-area economic output from agriculture is not so significant
in comparison with other industries; (c) most of the crops are annual herbs having
different growth and development stages in different seasons which means that agricultural activities have obvious phenological rhythms and the intra-annual change
may be very drastic; (d) agriculture is strongly affected by human activities and
management where timely and accurate monitoring information is required. These
intrinsic characteristics of agriculture demand novel techniques in the monitoring of
crop growth and agricultural productions. Remote sensing technology meets these
requirements by its rapidness, accuracy, economy, timing, dynamic and repetitive
monitoring capacity. Remote sensing technology has been applied in agriculture
extensively since its early stage in the 1960s. Now several global and national operational systems of monitoring agriculture with remote sensing have been operated. The number of similar operational systems at regional scale is much more.
These systems provide timely and valuable information for agricultural production,
management and policy-making. On the other hand, the demands arising from the
applications in agricultural sectors have also enhanced the progress and innovation
in remote sensing technology. The main applications of remote sensing in agriculture management and monitoring include: crop identification and cropland mapping,
Zhongxin Chen, Jianqiang Ren, and Limin Wang
Key Laboratory of Resource Remote Sensing & Digital Agriculture, Ministry of Agriculture,
Beijing, China
chzhx@yahoo.com
Zhongxin Chen, Sen Li, Jianqiang Ren, Pan Gong, Mingwei Zhang, Limin Wang, Shenliang Xiao,
and Daohui Jiang
Institute of Agricultural Resources & Regional Planning, Chinese Academy of Agricultural
Sciences, Beijing, China
397
398
Z. Chen et al.
crop growth monitoring and yield estimation/prediction, inversion of key biophysical, biochemical and environmental parameters, crop damage/disaster monitoring,
precision farming, etc. In this paper, crop mapping, yield prediction, soil moisture
monitoring and crop phenology monitoring with remote sensing are reviewed.
15.1 Crop Identification and Crop Mapping

Crop distribution and acreage, as well as yield, is the basic information necessary
for agricultural management and policy-making (Pradhan, 2001). Remote sensing of
the extent and distribution of individual crop types has proven useful to a wide range
of end-users, including governments, farmers, and scientists (Allen et al., 2002).
The traditional method of crop identification is supervised classification of multiple
land resource satellite (including Landsat MSS, TM, ETM+, SPOT-XS, IRS-LISS,
CBERS, etc.) remotely sensed images throughout the growing season. This approach usually requires a lot of manual interpretation and cloud-free imagery for
critical phenological stages which could be barriers for operational implementation
over large areas and in multiple years (Broge and Mortensen, 2002; Doraiswamy
et al., 2005; Kressler and Steinnocher, 1999). The Advanced Very High Resolution Radiometer (AVHRR) series of sensors onboard the National Oceanic and
Atmospheric Administration (NOAA) satellites and the Moderate Resolution Imaging Spectroradiometer (MODIS) offer a unique combination of spectral and temporal resolutions, making them alternatives for large scale crop type mapping with
novel classification techniques such as time series analysis and so on (Broge and
Mortensen, 2002; Doraiswamy et al., 2005). The former is more accurate, but may
be expensive and time-consuming when applied to a large area. The latter has the
characters of low-cost and rapidness, but is low in accuracy because of the restriction of spatial resolution. Furthermore, SAR data is used in the discrimination of
agricultural land use classes which can eliminate the influence of cloud and time
(Soares et al., 1997; Chakraborty et al., 1997; Shao et al., 2001).
Confusion between natural vegetation and cropland is a major source of error in
remote sensing-based crop mapping with low-resolution imagery. Sometimes this
is true even with high-resolution imagery. This kind of confusion is also obvious in
regions with very complicated crop-planting patterns which can be generated from
complicated topography or land ownership. For example, Loveland et al. (1999) reported that nearly 60% of the problems addressed in the post classification process
for the International Geosphere Biosphere Programmes Data and Information System (IGBP-DIS) global land cover data set arose from confusion between natural
vegetation and cropland. This problem was also noted in the MODIS classification
of agricultural land use. The most obvious confusion in this regard arose because
of seasonal variation in the Normalized Differentiated Vegetation Index (NDVI)
signals caused by seasonal variation in illumination geometry that mimicked a phenological cycle (Spanner et al., 1990; McIver and Friedl, 2002). The accuracy of
a method to estimate the variability of cropland is affected by lots of factors, such
15 Monitoring and Management of Agriculture with Remote Sensing
399
as the data used, scale, crop type, etc. There are lots of researches on it in the past
years. Agricultural land use has as specific characteristic that the surface reflectance
changes regularly in time with the growth of a crop. This may cause it difficult to
calculate accurately the total sown area of a specific crop in case of different types
of cropping systems. Satellite data must cover the key phenological phase of the
cropping system (Thenkabail et al., 2000).
15.1.1 Traditional Classification Methods

The frequently adopted crop classification methods include unsupervised classification, supervised classification, and decision tree classifier. In the case where there is
less information for a study area, only the image characteristics are used. Multiple
groups, from randomly sampled data, will be automatically divided into homogeneous spectral classes using a clustering technique. The clustered classes are then
used for estimating the population statistics. This method is only used under special conditions. The accuracy of this case is usually lower than that of supervised
classification for shortage of knowledge on the study area. In order to determine a
decision rule for the classification, it is necessary to know the spectral characteristics or features with respect to the population of each class. The spectral features
can be measured using ground-based spectrometers or by sampling training data
from clearly identified training areas, corresponding to defined classes that are usually made for estimating the population statistics. Statistically unbiased sampling of
training data should be made in order to represent the population correctly. Maximum likelihood classification has been the most common method used for supervised classification of remotely sensed data (Wessel et al., 2004). The premise of
this case is enough field samples in the study area. The date of collecting samples
should be close to the date of satellite images, because the surface reflectance is
changing with crop growth. The accuracy of crop identification is dependent on the
quality of the training data. Many authors report classification accuracies exceeding
85% for crop mapping (Van Niel and McVicar, 2004; Baban and Luke, 2000).
15.1.2 Non-Parametric Classification Methods

Increasingly, nonparametric classification algorithms are being used, which make no
assumptions regarding the distribution of the data being classified, because the frequently adopted normal distribution hypothesis is usually not true for the remotely
sensed data sets. The decision tree classifier is a hierarchically based classifier which
compares the data with a range of properly selected features. The selection of features is determined by an assessment of the spectral distributions or separability of
the classes (Pal and Mather, 2003). There is no generally established procedure and
each decision tree or set of rules should be designed by an expert. When a decision
400
Z. Chen et al.
tree provides only two outcomes at each stage, the classifier is called a binary decision tree classifier (BDT). This method was used in many cases for its flexible
characteristic. The rules of a decision tree are obtained through analyzing the particular attributes of different crop types. Usually auxiliary factors are also added to
distinguish crop types, including climate, topography, spectral characteristics, etc.
McIver and Friedl (2002) present a method for incorporating prior probabilities in
remote sensing-based land cover classification using a supervised decision tree classification algorithm.
Conventional classification has limited success for farm lands with high fragmentation and high spatial or/and temporal ecological heterogeneity. A knowledgebased approach combined with imagery and geographical data within a framework
of an intelligent recognition system can improve the accuracy of crop identification. Experts interpret remote sensing images with knowledge based on experiences.
However, computer assisted classification utilizes only very limited expert knowledge. The following two types of knowledge are required for an expert system in
remote sensing. (a) Knowledge about image analysis: a feedback system should be
introduced for checking and evaluating the objectives and the results. (b) Knowledge about the objects to be analyzed should be introduced in addition to the ordinary classification method. The fact that cropland does not exist over 60 slope, is
one example of the type of knowledge that can be introduced. Cohen and Shoshany
(2002) utilize the split-and-merge rules derived from the entire data set of imagery and auxiliary data to enable the formalization of different interpretation keys
for each crop in Israel. This research area contains eight crop types that represent
70% of Israeli agricultural production. Multi-date Landsat TM images representing
seasonal vegetation cover variations were converted to NDVI layers. Field boundaries were delineated by merging Landsat data with SPOT-panchromatic images.
The difference of this method is based on auxiliary geographical and expert knowledge in the post-classification phase.
Crop classification mapping with time series data was developed in the 1980s,
and applied in remote sensing analysis. The basic theory of this method is that the
spectral reflectance or vegetation index from satellite data is changing with crop
growth (Lobell and Asner, 2004). Badhwar (1984) identified corn, soybean, and
other land cover classes with a multi-temporal classification approach based on the
greenness profile derived from Landsat MSS spectral bands. Buttner and Csillag
(1989) mapped crop and soil inhomogeneities for a complex area through analyzing
spring, summer, and autumn SPOT images. Chakraborty et al. (1997) classified the
crop types using ERS-1 SAR data due to its independence from cloud cover, and
results showed more than 90% classification accuracy for all types of wetland rice
using three-date SAR data.
Low spatial resolution data was extensively used in case of a large study area,
e.g., NOAA-AVHRR and MODIS time series data, because of their advantages of
high-temporal resolutions. Jakubauskas et al. (2002) identified crop types based on
temporal changes in NDVI values of AVHRR. Harmonic analysis, or Fourier analysis, decomposed a time-dependent periodic phenomenon into a series of constituent
sinusoidal functions, or terms, which were defined by a unique amplitude and phase
401
value. Amplitude and phase angle images were produced by analysis of the timeseries NDVI data and used within a discriminant analysis to develop a methodology
for crop type identification. Doraiswamy et al. (2005) separated the cropland area
for crop yield estimation with MODIS 8-day surface reflectance images of bands
1 and 2. Areas of Interest (AOI) were created from the Landsat TM classification
and standard procedures were followed to develop a land use classification using the
parallelepiped method as the non-parametric rule, and the Mahalanobis Distance as
the parametric rule. Xiao et al. (2005) developed a paddy rice mapping algorithm
that uses time series of three vegetation indices (including NDVI, Enhanced Vegetation Index, EVI and the Land Surface Water Index, LSWI) derived from MODIS
images to identify the initial period of flooding and rice transplanting in China.
This research indicated that MODIS-based paddy rice mapping could potentially
be applied at large spatial scales to monitor paddy rice agriculture on a timely and
frequent basis.
Because the pixel size of the space-born remotely sensed data is usually larger
than the small parcel of cropland, the mixel problem is nearly ubiquitous in cropland mapping. Mixels can be the main source of errors in agricultural monitoring
and agricultural statistics by remote sensing. A spectral mixture model method was
used in cropland mapping research from microcosmic to macroscopic scale. Spectral mixture analysis (SMA) has become the basic tool for land cover analysis with
remote sensing (Broge and Mortensen, 2002; Doraiswamy et al., 2005; Tompkins
et al., 1997; Fitzgerald et al., 2005; Theau et al., 2005). SMA methods are typically
utilized to calculate the fraction of each endmember in a mixed pixel using an inverse least square devolution method and endmember spectra. One basic assumption
of SMA models is that the spectrum for each pixel is a linear or nonlinear combination of endmembers spectra dependent on the significance of multiple scattering of
light of land cover types (Shimabukuro and Smith, 1991; Wu, 2004).
Neural networks have proven to be the most significant improvement in information extraction in remote sensing in the last 15 years (Del Frate et al., 2003).
This method compares favorably with an optimal Bayesian classifier. The classification performance of the method is proven superior compared with other statistical
and neural classifiers (Moshou et al., 2001). The classification of remote sensing
data using artificial neural networks appeared in remote sensing interpretation about
10 years ago (Park et al., 2005). Afterwards, this method was used in crop type identification and other agricultural research fields (Del Frate et al., 2003).
In crop identification and cropland mapping, the accuracy has been improved by
innovation of new classification algorithms, integrating various classification methods, data fusion of remotely sensed data from different types of sensors, incorporating other auxiliary data and expert knowledge, etc. To guarantee the precision of
crop area monitoring in a large area, it is necessary to develop the technology of field
sampling. There is a trend to integrate high-resolution and low-resolution remotely
sensed imagery in large-scale crop area estimation. That is, the low-resolution imagery can be used to estimate the distribution of crops and this information can be
used as sampling frame. Then high-resolution imagery can be sampled for accurate
crop mapping.
402
Z. Chen et al.
15.2 Crop Yield Estimation and Prediction

Crop yield and production data are key indicators for national food security and sustainable development of society. Early and accurately gathering knowledge of crop
yield and production or the changes of production is very important. Regional and
national crop yield estimation and prediction with remote sensing are of great interest for scientists, policy-makers and the general public. Crop yield estimation with
remote sensing is not only for the staple crops such as wheat, maize, rice, cotton,
soybean, but also for some marginal crops. The techniques and methods to estimate
crop yield and production include statistical sampling methods, agro-climate models, crop growth models, remote sensing, and some integrated methods. Because of
the specific character of remote sensing, for example, the field of view and swath
width is wide and the period of detecting the earth surface is short, remote sensing has been used to estimate crop yield and production at a large scale in many
countries (Fuller, 1998; Huang et al., 2002; Hochheim and Barber, 1998; Jiao et al.,
2005; Potdar, 1993; Groten, 1993; Weissteiner et al., 2004). Crop yield estimation
with space-born remotely sensed data can be traced back to mid-20th century. The
National Aeronautics and Space Administration (NASA), NOAA and the United
States Department of Agriculture (USDA) implemented a large area crop inventory experiment (LACIE) project in the years 19741978 and agricultural and a resources inventory surveys though aerospace remote sensing (AgRISTARS) project
in the years 19801985 (Wang, 1996; Allen et al., 2002). Important information and
many experiences about crop monitoring with remote sensing were gained from
these tasks. From then on, European and other countries also developed crop yield
estimation and crop monitoring systems.
Remote sensing-based crop yield estimation is a prominent example of the
macro-research of remote sensing of vegetation (Boogaard et al., 2002). Because of
the occurrence of a seasonal rhythm of vegetation, the micro-structure of plant cells
and the macro-structure of vegetation canopies changes accordingly and the spectral
response of individual vegetation types or of a population also changes periodically.
So human can carry out many researches such as vegetation growth monitoring and
biomass estimation depending on the multi-spectral response of vegetation to derive
them and learn their changing information.
With the development of remote sensing, the researches of remote sensing of
vegetation have developed practicably. Scientists have put forward Vegetation Index
(VI) models which are indicators to vegetation and are linear or non-linear combinations of multi-spectral data (Dadhwal and Ray, 2000). The different combinations
of measured reflectance in the visible (Rv ) and near-infrared (Rn ) parts of the spectrum compose the core of VI (Carlson and Ripley, 1997; Gitelson and Kaufman,
1998).
Remote sensing-based crop yield estimation is a series of techniques and methods to forecast crop yield before harvesting of the target crops. Based on the theories of biology and spectroscopy and cognition of crop foliage and canopy spectral
response, the crop classes are identified and the spectral data of different crops in
different spectral band are acquired through the sensors. Then using the spectra data,
403
we can monitor the crop growth and establish various models for crop yield forecasting. Generally, there are three categories of models based on remote sensing:
empirical models, physiological models and crop growth models.
15.2.1 Empirical Models

The empirical approach to estimate crop yield is based on the hypothesis that biological/physical crop parameters or ecological environmental parameters of the land
surface such as biomass, LAI, temperature are correlative to final crop yield or production and remote sensing data are obviously correlative to these above parameters
of critical growth stages (Dadhwal and Ray, 2000; Mkhabela et al., 2005). Of course,
the empirical relationships may be simple linear or non-linear and may be multiple
(Tucker et al., 1980; Sridhar et al., 1994). The crop parameters may be in one critical growth stage or may be in more (Barnett and Thompson, 1983). In some cases,
there may be some non-remote sensing-based parameters such as weather data used
in crop yield monitoring and prediction (Kalubarme et al., 1995).
There are some advantages using empirical relationships for yield estimation.
Firstly, it is simple to estimate crop yield. A few ground-truth measurements are
sufficient for the model validation. Single or several images can meet the need
of yield estimation. Secondly, the accuracy of estimation is higher especially in a
homogenous region and it is enough to meet the information need of agricultural
management.
There are also some disadvantages for this method. For example, it is based on
a link between crop biomass and remote sensing data, which sometimes may be
occasional but not intrinsic. It is lack of flexibility and stability, so the results at
different stages vary drastically.
15.2.2 Physiology-Based Models

A physiology-based model is mainly based on crop physiological functions and
assumes that crop production results from photosynthesis through which a fraction
of incident solar radiation is converted into biomass. The most popular formulation
of this approach is the efficiency equation of Monteith (1972), which can be applied
to remote sensing data. The simple model for crop yield estimation can be written
as follows:
DM =
& t2
t1
fPAR PAR dt
Yield = DM HI
(15.1)
(15.2)
404
Z. Chen et al.
where DM is dry matter production in a time period t2 t1 , is light use efficiency, PAR (MJ m2 ) is the incoming photosynthetically active radiation for the
wave bands between 0.4 and 0.7 m. PAR is part of the short wave solar radiation
(0.33 m) and is absorbed by chlorophyll for photosynthesis in the crop. fPAR
is the fraction of the photosynthetical radiation absorbed by the canopy. HI is the
harvest index which means the ratio of grain mass to aboveground biomass.
Judging from Eqs. (15.1) and (15.2), we know that using this method to estimate crop yield we only need to learn the parameters of , HI, fPAR and PAR.
Researchers have done a lot of work on these parameters. (g MJ1 m2 ) differs
between C3 and C4 plants and is affected by climate factors such as temperature and
rainfall (Hanan et al., 1995; Bastiaanssen and Ali, 2003; Field et al., 1995). As to the
maximum for each kind of crop, different researchers often get different results.
Many studies have shown that is a relative constant property of plants and only
varies over a relatively narrow range for crop ecosystems but over a wider range for
natural ecosystems.
fPAR is a key variable in the assessment of vegetation production. The fPAR
can be calculated depending on the relationships between LAI, VI and fPAR which
has been studied by many authors (Hu et al., 2003; Shabanov et al., 2003; Myneni
et al., 2002). PAR can be obtained from observation data of weather stations or can
be derived from remote sensing such as TOMS (Total Ozone Mapping Satellite)
reflectance (Goldberg and Klein, 1980; Eck and Dye, 1991).
As to the strong points of this method, it is more universal and suitable for more
crop types and it only needs a relatively simple data set such as solar radiation and
fPAR. But light use efficiency depends on the phenological stage and environment
conditions such as temperature and rainfall.
15.2.3 Crop Growth Models

A crop growth model describes the primary physiological mechanisms of crop
growth such as phenological development, photosynthesis and dry matter partitioning, and their interactions with the underlying environmental factors using mechanistic and sometimes empirical equations (Delecolle et al., 1992).
Using crop growth models to estimate crop yield requires a lot of inputs that
are specific to the crop, soil characteristics, management practices and local climate
conditions. So it has limitations to use this kind of models in large regions because
fewer inputs are generally available at this scale. Remote sensing has shown to be
capable of providing certain crop characteristics and some other parameters. So a
crop growth model can be combined with remote sensing and use input parameters
which are derived from remote sensing for a larger region (Guerif et al., 1993; Maas,
1988).
When integrating remotely sensed data with a crop growth model, we can use
several methods (Doraiswamy et al., 2003). Firstly, model initialization can be done
by the parameters estimated from remote sensing and/or taking these parameters
405
directly as input data. This kind of parameters is used successfully including LAI
and crop canopy cover. The other method is data assimilation of data derived from
remote sensing that is used to calibrate the crop growth model. There are many successful researches on mechanistic crop growth model incorporating with remote
sensing (Clevers and van Leeuwen, 1996; Guerif and Duke, 2000; Doraiswamy
et al., 2003). Clevers and van Leeuwen (1996) used optical and microwave remote
sensing data to estimate LAI for calibrating the SUCROS (simplified and universal
crop growth simulator) model and the improved the accuracy of sugar beet yield estimation. Guerif and Duke (2000) studied the method of coupling the radiative transfer model SAIL (Scattered by Arbitrary Inclined Leaves) and the crop growth model
SUCROS. Doraiswamy et al. (2003) derived LAI data from Landsat TM and NOAA
AVHRR to calibrate the EPIC (Erosion Productivity-Impact Calculator) model depending on the link provided by the radiative transfer model SAIL and applied this
method to simulate spring wheat yield for the state of North Dakota of America. Recently, Yang (2005) integrated LAI data from Landsat TM with the EPIC model and
calibrated the crop growth model including cropping system, planting and harvest
data, and applied this method to estimate winter wheat yield in North China.
There have been great progresses in crop yield monitoring and prediction because
of novel remote sensing sensors, improved algorithms and models. The progress in
quantitative remote sensing made the accurate inversion of land surface and crop
parameters possible, which strengthened the data input for various crop monitoring models and systems. Different kinds of crop models have been applied for different purposes. Empirical relationships for crop yield monitoring and prediction
are widely used for its simplicity. But its shortcoming is also obvious, the unstableness and sometimes site-specific relationship between yield and remote sensing
data. Physiology-based models are mainly based on crop physiological functions,
which is its strong point. But some parameters are not consistent over a large region
and/or for different crops and sometimes not easily acquired by remote sensing or
an in situ survey. Crop growth models have a long history and extensive use for
crop growth monitoring, yield prediction and farm management around the world.
Huge work of data collection and preparation hampered its good performance at a
regional scale. During recent years, crop growth models with remote sensing data
assimilation have been improved greatly at regional scales for better estimation of
crop parameters.
15.3 Crop Phenology Monitoring

Crop phenology is important in crop monitoring because it can have great impact on
the monitoring accuracies of crop yield and acreage change. Accurate monitoring of
crop development patterns (i.e., phenology and growth) is an important component
of farm management since it enables us to assess crop growth under various regional
weather conditions. Monitoring seasonal changes in vegetation activity and crop
phenology over wide areas is essential for many applications, such as estimation of
406
Z. Chen et al.
net primary production (Kimball et al., 2004), deciding on the time period for crop
yield modeling (Bouman et al., 2002) and supporting decisions about water supply
(Digkuhn and Gal, 1996).
Because of the synoptic coverage and repeated temporal sampling that satellite
observations afford, remotely sensed data possess significant potential for monitoring vegetation dynamics at regional to global scales (Myneni et al., 1997). Various methods have been developed to monitor seasonal vegetation changes using
time series of normalized difference vegetation index (NDVI) data. These methods have employed a variety of different approaches including the use of specific
threshold-based methods (Viovy et al., 1992; White et al., 1997), Fourier-based
fitting methods (Moody and Johnson, 2001), asymmetric function fitting methods
(Jonsson and Eklundh, 2002; Zhang et al., 2003), backward-looking moving averages method (Reed et al., 1994), and so on.
The data collected by MODIS onboard NASAs Terra/Aqua spacecrafts are useful for monitoring crop phenology (Zhang et al., 2003; Sakamoto et al., 2005). Satellite vegetation index (VI) data such as the NDVI are correlated with green leaf area
index (LAI), green biomass, and vegetation cover (Baret and Guyot, 1991). In most
agricultural remote sensing applications, monitoring spectral characteristics of the
crops at one particular stage is more popular than those over the entire growing season. So, the spectral characteristics of vegetation and the background surface (soil
or water) should be well understood. Monitoring crop phenologies can serve the
purpose of forecasting grain harvest (yield prediction), collecting crop production
statistics, facilitating crop rotation records, mapping soil productivity, identification
of factors influencing crop stress, assessment of crop damage due to storms and
drought, and monitoring farming activity.
Crop phenology controls the temporal changes observed from satellite data. Crop
development is shown in remote sensing by integrating space and time. Therefore,
incorporating crop development stages into remote sensing analyses is conducive
to crop type identification and area measurement. Different crops (winter wheat,
maize, soybeans, alfalfa, etc.) exhibit a distinctive seasonal pattern and period of
maximal greenness. This information (phenology) may be used in the classification
process to accurately discriminate vegetation types. Vegetation indices (VIs) have
been extensively used for identifying crop types. For instance, NDVI time series
analysis was applied to identify several common crop types occurring within the
Western Great Plains (Jakubauskas et al., 2002).
Overall, the temporal NDVI images provide the crop cycle and the cropping system pattern of agricultural land use. Observed changes in the NDVI through time
are generally thought to reflect vegetation type, phenology and local environmental conditions. Therefore, compared to land cover classification using single date
data, multi-temporal datasets are often found to improve the accuracy of classification. With the use of NDVI/EVI time-series we can detect and monitor vegetation
cycles and timings over large areas. This offers many opportunities for more complete vegetation descriptions than could be achieved with only single-date images.
The discrimination of NDVI/EVI time-series is based on their characterization of
seasonal dynamics of vegetation growth (phenology).
407
Phenological knowledge plays a critical role in identifying optimal acquisition

dates for selection of satellite imagery for agricultural monitoring. For example,
greenup is a particularly important vegetative stage (greenup is the date of onset
of photosynthetic activity). Greenup of winter wheat occurs earlier than for other
crops. At this stage, winter wheat can be easily identified from other vegetation
because of its greater greenness. In general, paddy fields are plowed and flooded
before rice planting. Planting date of rice is when rice is often distinctive from other
crops since the water background of rice can easily be distinguished from both soil
and green vegetation (Xiao et al., 2002).
Crop phenology is generally divided into vegetative and reproductive stages. The
vegetative stage is largely defined by the part of the growth cycle where the crop
develops and grows, starting emergence to tasseling. The reproductive starts at anthesis and ends after maturity. For dryland crops, several transitions are important
in terms of management: emergence, tasseling, and initiation of senescence. For
example, unfavorable conditions occurring between corn emergence and leaf development will limit the size of the leaves and thus the amount of photosynthetic
biomass (Vina et al., 2004). Harmful conditions at the beginning of the reproductive cycle will result in impairment of pollination and decrease of the number of
fertilized kernels that are destined to be filled. The water stress or temperature or
disease stress during the grain-filling period directly affects the crop yield. Therefore, identifying stress-induce during the key transitions is important for influencing
farm management practice.
For different locations, the timing of the growth cycle of rice varies depending on
local climate, management, and cultivar planted. Traditional methods of obtaining
information on crop phenology are census and ground surveying. When planning
a monitoring campaign that extends over large areas it is often difficult to obtain
precise crop calendars. In general these crop calendars only record specific dates (or
ranges) of activities (i.e., field preparation, sowing, emergence, and harvest). These
calendars may be applied to the selection of remotely sensed data to discriminate
crop types, but they have limitations in supporting decisions about farm management practice. Remote sensing can provide relevant data collection and information
extraction strategies.
15.4 Soil Moisture Monitoring

Soil moisture is an important variable controlling biogeochemical cycles, heat exchange and infiltration rates at the land/atmosphere boundary (Li and Islam, 1999).
Soil moisture information is critical in agricultural water management and drought
monitoring. However, soil moisture observations are often inadequate, due to an inability to economically monitor spatial variation in soil moisture from traditional
in situ measurement techniques such as weighting and TDR (Time Domain Reflectometry) method. In contrast, the development of remote sensing technology makes
it possible to monitor soil water in nearly real time at large scale.
408
Z. Chen et al.
Monitoring soil moisture and drought with remote sensing is based on two basic
principles. The first is that soil water content change leads to variations of soil spectral reflectance; the other is that soil water content change can cause a physiological
change of a plant, thus changes the spectral characteristics of a leaf and affects
the reflectance of the canopy. There are three categories of remote sensing systems
used in soil moisture monitoring: optical (including reflective near-infrared), thermal and microwave systems. Optical systems acquire data from visible wavelengths
up to, and including, near infrared (4001, 000 nm). They measure the amount and
wavelength of the suns radiant energy reflected from a target object. The inclusion
of wavelengths visible to the human eye often allows an intuitive interpretation of
images, especially when combined with expert knowledge of soil properties. Applications for the determination of soil moisture have been investigated but were
limited by the need for bare soil surfaces and also by the fact that soil moisture
and soil organic carbon properties give similar signals when they change. Thermal systems refer, in general, to sensors that measure energy emitted within the
wavelength range from 3 to 13 m. According to different approaches, microwave
systems include passive systems, i.e. radiometers such as SMMR (Scanning Multichannel Microwave Radiometer), SSM/I (Special Sensor Microwave Imager), TMI
(TRMM Microwave Imager), AMSR /AMSR-E (Advanced Microwave Scanning
Radiometer); and active systems, i.e., radar systems such as ERS-1/2, RADARSAT,
ENVISAT ASAR, and so on.
Research work on soil moisture remote sensing was initiated in the late 1960s to
early 1970s. Watson et al. (1971) used a thermal model for the geologic interpretation of infrared (IR) images. Bijleveld (1978) did further research based on Watsons
work and raised the model of thermal inertia and daily evaporation calculation. Pratt
and Ellyett (1979) presented the thermal inertia approach for mapping soil moisture.
All this initiative work founded a base for multi-method research on soil moisture
remote sensing. Great achievements were made in the field of soil moisture remote
sensing in the 1980s. Soil moisture data can be obtained by means of ground, space
and satellite remote sensing. The spectrum covers visible light, near-, middle- and
far-infrared bands, thermal infrared band and L, C and X microwave bands. Methods
and indices used for soil moisture mapping were developed step by step such as regional evaporation estimation, crop surface temperature, soil heat capacity, drought
indices, crop water stress and leaf water content, and so on. Pratt and Ellyett (1979)
and Price (1985) developed the thermal inertia approach and the mechanism of thermal infrared imagery. Attempts for large scale soil moisture remote sensing using
meteorological satellite data were starting at the same time. Carlson (1986) used
NOAA /AVHRR data to estimate surface moisture and thermal inertia at regional
scale. Owe (1988) used NOAA /AVHRR data and satellite microwave data to derive
vegetation indices and estimate surface soil moisture.
Although visible, near-infrared and thermal infrared bands can be used to monitor soil moisture, they are not effective in case of cloud cover. Microwaves have a
good penetration ability of clouds, so it has an advantage over the other methods in
soil moisture monitoring, especially in cloudy regions. Soil moisture remote sensing by microwaves was initiated in the 1970s (Schmugge et al., 1974; Choudhury
409
et al., 1979) and developed after the 1980s (Schmugge et al., 1986; Njoku and
ONeill, 1982; Wang, 1985). With the launch of a series of microwave sensors
such as ERS-1& 2, RADARSAT, ENVISAT, TMI, AMSR and AMSR-E, soil moisture monitoring by microwaves developed rapidly in recent years (Gu et al., 2002;
Wickel and Jackson, 2001; Blumberg and Freilikher, 2001; Glenna and Carr, 2003).
Approaches or indices used in soil moisture monitoring mainly include thermal
inertia approaches, crop water stress index (CWSI), transformed normalized difference vegetation index (TNDVI), vegetation temperature condition index (VTCI),
thermal infrared approaches and microwave remote sensing.
15.4.1 Thermal Inertia Approaches

Water has a large heat capacity and heat conductivity, so the thermal inertia of
wet soil is larger than that of dry soil. Thermal inertia can be observed from
the monitoring of ground surface temperature by optical remote sensing, such as
NOAA /AVHRR or EOS / MODIS data. The thermal inertia method is commonly
used in soil water monitoring research.
Soil thermal inertia is a heat characteristic of soil, it is the internal factor which
causes the change of soil surface temperature. Thermal inertia has very close relation
with soil water content, and it affects the diurnal temperature range of soils at the
same time. The thermal inertia method is suitable in case of bare soil and the early
stages of crop growth. Thermal inertia can be written as:

(15.3)
P = c
where P is thermal inertia (J m2 K1 s1/2 ), is heat conductivity (J m1
s1 K1 ), is the density of soil (kg m3 ), c is specific heat (J kg1 K1 ).
The classical thermal inertia method raised by Price in 1985 needs input of a
large number of in situ data, thus leads to poor practicability of this method. So,
thermal inertia is usually used in practice as apparent thermal inertia (ATI). ATI can
be obtained from satellite data, which greatly enhances the applications of thermal
inertia.
Apparent thermal inertia is written as
AT I =
1A
Tmax Tmin
(15.4)
where Tmax is the maximum temperature and Tmin is the minimum temperature on
the same day, A is the full spectrum albedo.
Soil water content is usually estimated from an empirical linear equation. The
equation is written as
W = aAT I + b
(15.5)
where W is soil water content, a and b are coefficients.
410
Z. Chen et al.
15.4.2 Crop Water Stress Index

The crop water stress index was proposed by Jackson et al. (1981). This index is
based on the energy balance theory and calculated from the difference between
canopy and air temperatures.
The energy balance equation can be expressed as
Rn G = E + H
(15.6)
where Rn , G, E and H(W m2 ) are net radiation, soil heat flux, latent heat flux
and sensible heat flux, respectively. The sensible heat flux can be expressed in terms
of a temperature difference as
H = Cp
Tc Ta
a
(15.7)
where (kg m3 ) is the air density, Cp (J kg1 K1 ) the specific heat of the air,
a (s m1 ) the aerodynamic resistance and Tc and Ta (K) the canopy temperature
and air temperature at the reference height, respectively.
The CWSI is expressed as
CW SI =
(Tc Ta ) (Tc Ta )ll

(Tc Ta )ul (Tc Ta )ll
(15.8)
where the subscripts ll and ul denote lower (well-watered crop) and upper (nontranspiring crop) limits, respectively.
Substituting Eq. (15.7) into Eq. (15.6) and solving for Tc Ta ,
Tc Ta =
a
[(Rn G) E]
Cp
(15.9)
In a well-watered crop, E is equal to the potential crop evaporation (expressed

in g m2 s1 ) times the latent heat of vaporization (, J g1 ). In a completely
stressed crop, E is zero, thus Eq. (15.9) reduces to:
(Tc Ta )ul =
a (Rn G)
Cp
(15.10)
The CWSI has been commonly applied to the detection of water stress of plants, but
difficulties in measuring canopy temperature of crops with less than 100% vegetation cover has limited its operational application.
411
15.4.3 Transformed Normalized Difference Vegetation Index

Soil water is the key factor which affects crop growth when the radiation and temperature condition show little change. The crop grows better under good soil water
condition than with water stress conditions. This method calculates the long-term
decadal average vegetation index from remote sensed data, then it determines the
growth condition of crops and soil water condition from the difference of the decadal
average vegetation index and the long-term average.
Kogan (1997) proposed two useful indices VCI and TCI to monitor crop water
and temperature stress when drought occurs. VCI is the Vegetation Condition Index;
it reflects the growth difference of crops among different years. VCI is expressed as
VCI =
(NDV I NDV Imin )

(NDV Imax NDV Imin )
(15.11)
TCI is the Temperature Condition Index, which reflects the different responds of
crops to temperature. TCI is expressed as
TCI = (BTmax BT ) (BTmax BTmin )
(15.12)
in Eqs. (15.11) and (15.12), NDVImin (BTmin) and NDVImax (BTmax) refer to
the absolute minimum and maximum NDVI (BT) measured for a given decade (or
month) over a multi-year series of image data. NDVI (BT) refers to the current year
NDVI (BT) for the same decade. BT is the brightness temperatures derived from
channel 4 of NOAA /AVHRR.
The VCI is an indicator of the status of the vegetation cover as a function of the
NDVI minima and maxima encountered for a given ecosystem over many years. It
normalizes the NDVI (or any other vegetation index) and allows for a comparison
of different ecosystems. It is an attempt to separate the short-term climatic signal
from the long-term ecological signal and in this sense it is a better indicator of water
stress conditions than the NDVI.
The TCI is an equivalent indicator based on the surface skin temperature derived
from NOAA AVHRR data. Both, the VCI and the TCI, are dimensionless and vary
between the values of 0 and 1. Zero indicates the worst condition ever encountered
over the period of available images, one indicates the best condition encountered
during the same period of time. If the period covered includes dry and wet years and
under the assumption that the vegetation condition is mainly related to the water
availability, these indicators have a high potential for monitoring water stress.
15.4.4 Vegetation Temperature Condition Index

The vegetation temperature condition index is defined as:
V TCI =
LSTNDVIi. max LSTNDVIi

LSTNDVIi. max LSTNDVIi. min
(15.13)
412
Z. Chen et al.
where
LSTNDVIi. max = a + bNDVIi
LSTNDVIi. min = a + b NDVIi
(15.14)
where LSTNDV Ii. max and LSTNDV Ii. min are maximum and minimum land surface
temperature (LSTs) of pixels which have same NDV Ii value in a study region, respectively, and LSTNDV Ii denotes LST of one pixel whose NDVI value is NDV Ii .
Coefficient a, b, a and b can be estimated from an area large enough where soil
moisture at the surface layer should span from wilting point to maximum field
water-holding capacity at pixel level. In general, the coefficients are estimated from
the scatter plot of LST and NDVI for the study area. The shape of the scatter plot is
normally triangular at a regional scale (Gillies et al., 1997; Wang et al., 2001) if the
study area is large enough to provide a wide range of NDVI and surface moisture
conditions.
VTCI is not only related to NDVI changes in the region, but also related to LST
changes of pixels with a specific NDVI value. It can be physically explained as the
ratio of temperature differences among the pixels. The numerator of Eq. (15.13) is
the difference between maximum LST of the pixels and LST of one pixel, while the
denominator of Eq. (15.13) is the difference between maximum and minimum LSTs
of the pixels. LSTmax can be regarded as the warm edge where there is less soil
moisture availability and plants are under dry conditions; LSTmin can be regarded as
the cold edge where there is no water restriction for plant growth. The value of
VTCI ranges from 0 to 1. The lower the value of VTCI, the higher the occurrence
of drought.
15.4.5 Thermal Infrared Method

There is an inherent relationship between soil moisture and soil temperature. The
spatial distribution can be deduced through thermal images obtained by thermal infrared remote sensing. Experiments indicated that soil moisture of bare soil can be
estimated by surface temperature, and can reach 50 cm depth (Myers and Heilman,
1969). Bartholic et al. (1972) found that the maximum temperature of the bare soil
surface during the day decreases according to the increase of surface soil moisture.
The spatial distribution of soil moisture can be estimated by the daytime temperature of the underlying surface on a sunny and windless day. It contains two main elements for soil moisture monitoring using infrared methods, one is the measurement
of land surface temperature, and the other is the establishment of the relationship
between soil moisture and the land surface temperature.
15.4.6 Microwave Remote Sensing

The soil dielectric constant at microwave frequencies exhibits a strong dependence
on the soil water content. For example, at L band, the real part of the dielectric
constant ranges from 3 for dry soil to about 25 for saturated soil. This variation can
413
result in a change in the order of 10 dB in the magnitude of the radar-backscatter coefficient (Oh et al., 1992), and of 100 K in the magnitude of the brightness temperature. According to the operation method, there are two kinds of microwave remote
sensing systems, the passive system and the active system.
15.4.6.1 Passive Systems

Research on the soil moisture content with microwave radiometers has been performed since the late 1970s, and has recently been revitalized by new missions:
the already in orbit AMSR-E and AMSR, and the planned SMOS and AQUARIUS
(Kerr et al., 2001; Le Vine et al., 2001).
Various experimental researches on soil moisture with passive systems have been
carried out and different algorithms were proposed. For example, large airborne
experiments, called the Southern Great Plains Hydrology Experiments, were conducted in the USA in 1997 (SGP97) and 1999 (SGP99) to address significant gaps
in our knowledge, and to validate retrieval algorithms designed for the AMSR and
the AMSR-E. These approaches differ primarily in the methods used to correct for
the effects of soil texture, roughness, vegetation, and surface temperature. Soil moisture content retrieval approaches that have been investigated in previous studies include: (a) Single-channel retrieval with sequential corrections using ancillary data
(Jackson et al., 1993; Vinnikov et al., 1999); (b) iterative forward model corrections
using multi-channel brightness temperatures (Njoku and Li, 1999); (c) correction
using multi-frequency polarization indices (Paloscia et al., 2001); and (d) variations or combinations of the above methods (Owe et al., 2001; Jackson and Hsu,
2001).
The thickness of the soil layer through which moisture can be directly estimated
by means of a microwave radiometer has been investigated by many experimental
studies. Most researchers have come to the conclusion that at L band this layer is
about 510 cm. This result matches well with the requirements of those processes
such as infiltration and evapotranspiration that take place within this first layer of
the soil medium. In other applications, where soil moisture profiles down to several
decimeters are necessary, microwaves must be coupled to appropriate hydrological
models. Galantowicz et al. (1999) pointed out the effectiveness of the Kalman filter
for retrieving such a quantity by using field observations and a simulation study. The
usefulness of assimilating remotely sensed measurements into land-surface models
was discussed in (Houser et al., 1998; Burke et al., 2001). Burke and Simmonds
(2003) explored the potential for using low-resolution passive microwaves in a twodimensional Land Data Assimilation System (LDAS) for estimating deep-soil moisture from surface-soil moisture. Houser et al. (1998) investigated four-dimensional
(4D) soil-moisture assimilation using in situ and remote sensing observations.
414
Z. Chen et al.
15.4.6.2 Active Systems

The possibility of monitoring soil moisture changes using SAR data has stimulated
a large number of studies focused on establishing a relationship between the observed SAR response and surface soil moisture content. For a homogeneous soil
with a perfectly smooth surface, the scattering of electromagnetic waves is totally
forward, and depends on permittivity of the medium. For a rough surface, radiation
is scattered in various directions, and also generates backscattering. Thus, two basic properties determine the backscatter response observed by the SAR system: the
permittivity of the medium and the roughness characteristics of the surface. Both
parameters are, in turn, related to different geophysical parameters of the soil. With
the advent of the polarimetric SAR, radar remote sensing of soil moisture has attained significant prominence in the past two decades. Initially, extensive experimental studies using polarimetric scatterometers were carried out to establish a
relationship between radar response and the surface roughness and soil moisture
(Sarabandi et al., 1992). Extensive field experiments have also been conducted to
examine retrieval algorithms ranging from simple analytical to regression/empirical
models (Oh et al., 1992; Dubois et al., 1995; Kim and van Zyl, 2002).
Apart from surface-roughness parameters, the existence of short vegetation on
the surface makes the retrieval of soil moisture content very complicated. Vegetation
cover and its temporal variations are believed to be the major stumbling blocks in
monitoring soil moisture content variations using microwaves. Chiu and Sarabandi
(2000) proposed a very complicated coherent-scattering model, which accounts for
scattering from rough surfaces, vegetation cover, and their near-field interaction.
The inverse of this model was then used to demonstrate its ability for estimating the
physical parameters of a soybean field, including soil moisture from a polarimetric
set of AIRSAR images.
The remote sensing monitoring of soil moisture is a complicated indirect process:
many factors, such as the choice of bands and characteristics of the sensor are critical
for the monitoring of soil moisture; the properties of the soil, the different landforms,
climate change and soil management by human beings all affect the soil moisture
to different degrees, and the characteristics of the land surface are correlative to soil
moisture in some degree. So, the improved thermal inertia and CWSI (Crop water
stress index) method are the mature methods of remote sensing-based soil moisture
monitoring methods. The extensive application of GIS technology, the optimization
of models for the remote sensing monitoring of soil moisture and the application of
microwaves are the key directions for future study.
Currently, theories and methods on soil moisture remote sensing monitoring are
relative mature. Most approaches such as crop indices, CWSI, thermal infrared
and microwaves can get good results in soil moisture retrieval. But different approaches have their own applicability and limits. So, when using a single method
in soil moisture retrieval there may be large errors. For example, the thermal inertia
method is used in case the soil is bare; CWSI is used on crop covered surfaces, and
for partly covered soil the improved thermal inertia method is applicable. To avoid
large errors in soil moisture estimation, we should select different approaches to get
good results.
415
15.5 Summary
The rapid development of remote sensing technology, including novel sensors and
remote sensing data sets, quantitative inversion algorithms, and so on, has greatly
benefited the application of remote sensing for agricultural monitoring and management. The intrinsic characteristics of agriculture make remote sensing an ideal technique for its monitoring and management. Meanwhile, the demands arising from the
applications in agricultural sectors have also enhanced the progress and innovation
in remote sensing technology.
For crop identification and crop mapping, traditional unsupervised and supervised classification methods are still widely used. Increasingly, nonparametric classification algorithms are being used, which make no assumptions regarding the
distribution of the data to be classified, because the frequently adopted normal distribution hypothesis is usually not true for remotely sensed data sets. Auxiliary data
and expert knowledge are added in the process of cropland classification, which improve the overall classification accuracy. Multi-sensor data fusion and classification
of time series data are applied in cropland monitoring more and more.
For crop yield monitoring and prediction, different kinds of models are applied
for different purposes. Empirical relationships for crop yield monitoring and prediction are widely used for their simplicity. But its shortcomings are also obvious:
the unstableness and site-specific relationship between yield and remote sensing
data. Physiology-based models are mainly based on the crop physiological functions, which is its strong point. But some parameters are not consistent over a large
region and/or for different crops and sometimes not easily acquired by remote sensing or in situ survey. Crop growth models have a long history and are extensively
used for crop growth monitoring, yield prediction and farm management around the
world. The huge work of data collection and preparation hampered its good performance at a regional scale. In recent years, crop growth models with remote sensing
data assimilation have been improved greatly at regional scales for better estimation
of crop parameters.
Crop phenology is important in crop monitoring because it can have great impact
on the monitoring accuracies of crop yield and acreage change. Accurate monitoring
of crop development patterns (i.e., phenology and growth) is an important component of farm management since it enables us to assess crop growth under various
regional weather conditions.
Soil moisture is an important parameter for crop monitoring and management.
A lot of efforts have been made for accurate monitoring soil moisture or drought.
Monitoring soil moisture and drought with remote sensing is based on two basic
principles. The first one is that soil water content change leads to variations of soil
spectral reflectance; the other is that soil water content change can cause physiological changes of a plant, thus changes the spectral characteristics of a leaf and affects
the reflectance of a canopy. There are three categories of remote sensing systems
used in soil moisture monitoring: optical (including reflective near-infrared), thermal and microwave systems. Optical systems acquire data from visible wavelengths
up to, and including, near infrared. They measure the amount and wavelength of the
416
Z. Chen et al.
suns radiant energy reflected from a target object. The inclusion of wavelengths visible to the human eye often allows an intuitive interpretation of images, especially
when combined with expert knowledge of soil properties. According to different
approaches, microwave systems include passive systems, i.e., radiometers; and active systems, i.e., radar systems. Currently, theories and methods on soil moisture
remote sensing monitoring are relative mature. Most approaches such as crop indices, CWSI, thermal infrared and microwaves can get good results in soil moisture
retrieval. But different approaches have their own applicability and limits. So, if we
use a single method in soil moisture retrieval there may be large errors. For example, the thermal inertia method is used in case the soil is bare; CWSI is used on crop
covered surfaces, and for partly covered soil the improved thermal inertia method
is applicable. To avoid large errors in soil moisture estimation, we should select
different approaches to get good results.
Acknowledgements This work is funded by a National Key Technologies R&D Program of China
(No. 2006BAD10A06) and a National High Technology Research and Development Program of
China (863 Program No. 2006AA12Z103).
References
Allen R, Hanuschak G, Craig M (2002) History of remote sensing for crop acreage in USDAs
National Agricultural statistics Service. http://www.usda.gov/nass/nassinfo/remotehistory.htm
Baban SMJ, Luke C (2000) Mapping agricultural land use using retrospective ground referenced
data, Satellite sensor imagery and GIS. Int. J. Remote Sens. 21:17571762
Badhwar GD (1984) Classification of corn and soybeans using multitemporal thematic mapper
Baret F, Guyot G (1991) Potentials and limits of vegetation indices for LAI and APAR assessment.
Barnett TL, Thompson DR (1983) Large area relation of Landsat MSS and NOAA-6 AVHRR
spectral data to wheat yields. Remote Sens. Environ. 13:277290
Bartholic JE, Namken LN, Wiegand CL (1972) Aerial thermal scanner to determine temperatures
of soils and crop canopies differing in water stress. Agron. J. 64:603608
Bastiaanssen WGM, Ali S (2003) A new crop yield forecasting model based on satellite measurements applied across the Indus Basin, Pakistan. Agric. Ecosyst. Environ. 94:321340
Bijleveld RA (1978) A combined surface temperature, soil moisture and evaporation mapping approach. 12th International Symposium on Remote Sensing of Environment, Manila,
Philippines, 2026
Blumberg DG, Freilikher V (2001) Soil water-content and surface roughness retrieval using ERS-2
SAR data in the Negev Desert, Israel. J. Arid Environ. 49:449464
Boogaard HL, Eerens H, Supit I, van Diepen CA, Piccard I, Kempenseers P (2002) Description
of the MARS Crop Yield Forecasting System (MCYFS). METAMP-1/3. Alterra and VITO,
Wageningen and Mol
Bouman B, Kropff M, Tuong T, Wopereis M, Berge H, Laar H (2002) ORYZA2000: modeling
lowland rice. IRRI, Manila, Philippines
Broge NH, Mortensen JV (2002) Deriving green crop area index and canopy chlorophyll density
of winter wheat from spectral reflectance data. Remote Sens. Environ. 81:4557
Burke EJ, Simmonds LP (2003) Effects of sub-pixel heterogeneity on the retrieval of soil moisture
from passive microwave radiometry. Int. J. Remote Sens. 24:20852104
417
Burke EJ, Shuttleworth WJ, Lee K, Bastidas LA (2001) Using area-average remotely sensed surface soil moisture in multipatch land data assimilation systems. IEEE Trans. Geosci. Remote
Sens. 39:20912100
Buttner G, Csillag F (1989) Comparative study of crop and soil mapping using multitemporal and
multispectral SPOT and landsat thematic mapper data. Remote Sens. Environ. 29:241249
Carlson TN, Ripley DA (1997) On the relation between NDVI, fractional vegetation cover, and
leaf area index. Remote Sens. Environ. 62:241252
Carlson TN (1986) Regional-scale estimates of surface moisture availability and thermal inertia
using remote thermal measurements. Remote Sens. Environ. 1:197247
Chakraborty M, Panigrahy S, Sharma SA (1997) Discrimination of rice crop grown under different
cultural practices using temporal ERS-1 synthetic aperture radar data. ISPRS J. Photogramm.
Chen Z, Zhou Q, Tang H (2004) The future agricultural applications demanding on space-born
remote sensing in China. In: H Tang (ed) Research on the Agricultural Resource Utilization
and Regional Sustainable Development. China Outlook Press, Beijing, pp 236250
Chiu TC, Sarabandi K (2000) Electromagnetic scattering from short branching vegetation. IEEE
Choudhury BJ, Schmugge TJ, Newton RW, Chang A (1979) Effect of surface roughness on the
microwave emission from soils. J. Geophys. Res. 84:56995706
Clevers JGPW, van Leeuwen HJC (1996) Combined use of optical and microwave remote sensing
data for crop growth monitoring. Remote Sens. Environ. 56:4251
Cohen Y, Shoshany M (2002) A national knowledge-based crop recognition in Mediterranean environment. Int. J. Appl. Earth Observ. Geoinform. 4:7587
Dadhwal VK, Ray SS (2000) Crop assessment using remote sensing-Part II: crop condition and
yield assessment. Indian J. Agric. Econ. 55:5567
Del Frate F, Ferrazzoli P, Schiavon G (2003) Retrieving soil moisture and agricultural variables by
microwave radiometry using neural networks. Remote Sens. Environ. 84:174183
Delecolle R, Maas SJ, Guerif M, Baret F (1992) Remote sensing and crop production models:
present trends. ISPRS J. Photogramm. Remote Sens. 47:145161
Digkuhn M, Gal P (1996) Effect of drainage date on yield and dry matter partitioning in irrigated
rice. Field Crops Res. 46:117126
Doraiswamy PC, Moulin S, Cook PW, Stern A (2003) Crop yield assessment from remote sensing.
Photogramm. Eng. Remote Sens. 69:665674
Doraiswamy PC, Sinclair TR, Hollinger S, Akhmedov B, Stern A, Prueger J (2005) Application of MODIS derived parameters for regional crop yield assessment. Remote Sens. Environ.
97:192202
Dubois PC, van Zyl J, Engman T (1995) Measuring soil moisture with imaging radars. IEEE Trans.
Eck TF, Dye DG (1991) Sattellite estimation of incident photosynthetically active radiation using
ultraviolet reflectance. Remote Sens. Environ. 38:135146
Field CB, Randerson JT, Malmstrom CM (1995) Global net primary production: combining
ecology and remote sensing. Remote Sens. Environ. 51:7488
Fitzgerald GJ, Pinter J, Hunsaker DJ, Clarke TR (2005) Multiple shadow fractions in spectral
mixture analysis of a cotton canopy. Remote Sens. Environ. 97:526539
Fuller DO (1998) Trends in NDVI time series and their relation to rangeland and crop production
in Senegal, 19871993. Int. J. Remote Sens. 19:20132018
Galantowicz JF, Entekhabi D, Njoku EG (1999) Tests of sequential data assimilation for retrieving profile soil moisture and temperature from observed L-Band radiobrightness. IEEE Trans.
Gillies RR, Carlson TN, Cui J, Kustas WP, Humes KS (1997) A verification of the triangle
method for obtaining surface soil water content and energy fluxes from remote measurement
of the Normalized Difference Vegetation Index (NDVI) and surface radiant temperature. Int. J.
Gitelson AA, Kaufman YJ (1998) MODIS NDVI optimization to fit the AVHRR data series
spectral considerations. Remote Sens. Environ. 66:343350
418
Z. Chen et al.
Glenna NF, Carr JR (2003) The use of geostatistics in relating soil moisture to RADARSAT-1 SAR
data obtained over the Great Basin, Nevada, USA. Computers and Geosciences 29:577586
Goldberg B, Klein WH (1980) A model for determination the spectral quality of daylight on a
horizontal surface at any geographical location. Solar Energy 24:351357
Groten SME (1993) NDVI crop monitoring and early yield assessment of Burkina Faso. Int. J.
Gu S, Gao H, Zhu Y, Zhao B, Lu N, Zhang W (2002) Remote sensing land surface wetness by use
of TRMM/TMI microwave data. Meteorol. Atmos. Phys. 80:5963
Guerif M, Duke CL (2000) Adjustment procedures of a crop model to the site specific characteristics of soil and crop using remote sensing data assimilation. Agric. Ecosyst. Environ. 81:5769
Guerif M, de Brisis S, Seguin B (1993) Combined NOAA-AVHRR and SPOT-HRV data for assessing crop yields of semiarid environments. EARsel Adv. Remote Sens. 2:110123
Hanan NP, Prince SD, Begue A (1995) Estimation of absorbed photosynthetically active radiation and vegetation net production efficiency using satellite data. Agric. Forest Meteorol.
76:259276
Hochheim KP, Barber DG (1998) Spring wheat yield estimation for Western Canada using NOAA
NDVI data. Can. J. Remote Sens. 24:1727
Houser PR, Shuttleworth WJ, Famiglietti JS, Gupta HV, Syed KH, Goodrich DC (1998) Integration of soil moisture remote sensing and hydrological modeling using data assimilation. Water
Resour. Res. 34:34053420
Hu J, Tan B, Shabanov N, Crean KA, Martonchik JV, Diner DJ, Knyazikhin Y, Myneni RB (2003)
Performance of the MISR LAI and FPAR algorithm: a case study in Africa. Remote Sens.
Environ. 88:324340
Huang J, Tang S, Ousama AI, Wang R (2002) Rice yield estimation using remote sensing and
simulation model. J. Zhejinag Univ. (Science Edition) 3:461466
Jackson RD, Idso SB, Reginato RJ, Pinter PJ (1981) Canopy temperature as a crop water stress
indicator. Water Resour. Res. 17:11331138
Jackson TJ, Hsu AY (2001) Soil moisture and TRMM microwave imager relationships in
the Southern Great Plains 1999 (SGP99) Experiment. IEEE Trans. Geosci. Remote Sens.
39:16321642
Jackson TJ, Le Vine DM, Griffis AJ, Goodrich DC, Schmugge TJ, Swift CT, ONeill PE (1993)
Soil moisture and rainfall estimation over a semiarid environment with the ESTAR microwave
radiometer. IEEE Trans. Geosci. Remote Sens. 31:836841
Jakubauskas ME, Legates DR, Kastens JH (2002) Crop identification using harmonic analysis of
time-series AVHRR NDVI data. Comput. Electron. Agric. 37:127139
Jiao X, Yang B, Pei Z, Wang F (2005) Monitoring crop yield using NOAA/AVHRR-based vegetation indices. Trans. Chin. Soc. Agric. Eng. 21:104108
Jonsson P, Eklundh L (2002) Seasonality extraction by function fitting to time-series of satellite
sensor data. IEEE Trans. Geosci. Remote Sens. 40:18241832
Kalubarme MH, Mahey RK, Dhaliwal SS, Sidhu SS, Singh R, Mahajan A, Sharma PK (1995)
Agromet-Spectral wheat yield modeling in Punjab. Proceedings of National Symposium on
Remote Sensing of Environment with special emphasis on green revolution, pp 1117
Kerr YH, Waldteufel P, Wigneron JP, Martinuzzi JM, Font J, Berger M (2001) Soil moisture retrieval from space: the Soil Moisture and Ocean Salinity (SMOS) mission. IEEE Trans. Geosci.
Kim Y, van Zyl J (2002) On the relationship between polarimetric parameters and soil moisture. Proceedings of the IEEE 2002 International Geoscience and Remote Sensing Symposium
(IGARSS 2002), Toronto, June 2002
Kimball J, McDonald K, Running S, Frolking S (2004) Satellite radar remote sensing of seasonal
growing seasons for boreal and subalpine evergreen forests. Remote Sens. Environ. 90:243258
Kogan FN (1997) Global drought watch from space. Bull. Am. Meteorol. Soc. 78:621636
Kressler FP, Steinnocher K (1999) Detecting land cover changes from NOAA-AVHRR data using
spectral mixture analysis. Int. J. Appl. Earth Observ. Geoinform. 1:2126
419
Le Vine D, Koblinsky C, Pellerano F (2001) The measurement of salinity from space: sensor
concept, IEEE International Geoscience and Remote Sensing Symposium, Sidney, Australia,
July 913, 2001, pp 10101012
Li J, Islam S (1999) On the estimation of soil moisture profile and surface fluxes partitioning from
sequential assimilation of surface layer soil moisture. J. Hydrol. 220:86103
Lobell DB, Asner GP (2004) Cropland distributions from temporal unmixing of MODIS data.
Loveland TR, Zhu Z, Ohlen DO, Brown JF, Reed BC, Yang L (1999) An analysis of the IGBP
global land-cover characterization process. Photogramm. Eng. Remote Sens. 65:10211032
Maas SJ (1988) Using satellite data to improve model estimates of crop yield. Agron. J.
80:655662
McIver DK, Friedl MA (2002) Using prior probabilities in decision-tree classification of remotely
sensed data. Remote Sens. Environ. 81:253261
Mkhabela MS, Mashinini NN (2005) Early maize yield forecasting in the four agro-ecological
regions of Swaziland using NDVI data derived from NOAAs-AVHRR. Agric. Forest Meteorol.
129:19
Monteith JL (1972) Solar radiation and productivity in tropical ecosystems. J. Appl. Ecol.
9:747766
Moody A, Johnson DM (2001) Land-surface phonologies from AVHRR using the discrete fourier
transform. Remote Sens. Environ. 75:305323
Moshou D, Vrindts E, De Ketelaere B, De Baerdemaeker J, Ramon H (2001) A neural network
based plant classifier. Comput. Electron. Agric. 31:516
Myers VI, Heilman HD (1969) Thermal IR for soil temperature studies. Photogramm. Eng. Remote
Sens. 35:10241032
Myneni RB, Knyazikhin Y, Privette JL, Glassy J (2002) Global products of vegetation leaf area
and fraction absorbed PAR from year one of MODIS data. Remote Sens. Environ. 83:214231
Njoku EG, Li L (1999) Retrieval of land surface parameters using passive microwave measurements at 618 GHz. IEEE Trans. Geosci. Remote Sens. 37:7993
Njoku EG, ONeill PE (1982) Multifrequency microwave radiometer measurements of soil moisture. IEEE Trans. Geosci. Remote Sens. 20:468475
Oh Y, Sarabandi K, Ulaby FT (1992) An empirical model and an inversion technique for radar
scattering from bare soil surfaces. IEEE Trans. Geosci. Remote Sens. 30:370381
Owe M (1988) Estimating surface soil moisture from satellite microwave measurement and a satellite derived vegetation index. Remote Sens. Environ. 24:331345
Owe M, de Jeu R, Walker J (2001) A methodology for surface soil moisture and vegetation optical
depth retrieval using the microwave polarization difference index. IEEE Trans. Geosci. Remote
Sens. 39:16431654
Pal M, Mather PM (2003) An assessment of the effectiveness of decision tree methods for land
cover classification. Remote Sens. Environ. 86:554565
Paloscia S, Macelloni G, Santi E, Koike T (2001) A multi-frequency algorithm for the retrieval of
soil moisture on a large scale using microwave data from SMMR and SSM/I satellites. IEEE
Park SJ, Hwang CS, Vlek PLG (2005) Comparison of adaptive techniques to predict crop yield
response under varying soil and land management conditions. Agric. Syst. 85:5981
Potdar MB (1993) Sorghum yield modelling based on crop growth parameters determined from
visible and near-IR channel NOAA AVHRR data. Int. J. Remote Sens. 14:895905
Pradhan S (2001) Crop area estimation using GIS, remote sensing and area frame sampling. International J. Appl. Earth Observ. Geoinform. 3:8692
Pratt A, Ellyett CD (1979) The thermal inertia approach to mapping of soil moisture and geology.
Price JC (1985) On the analysis of thermal infrared imagery. The limited utility of apparent thermal
inertia. Remote Sens. Environ. 18:5973
420
Z. Chen et al.
Reed BC, Brown JF, van der Zee D, Loveland TL, Merchant JW, Ohlen DO (1994) Measuring
phenological variability from satellite imagery. J. Veg. Sci. 5:703714
Sakamoto S, Yokozawa M, Toritani H, Shibayama M, Ishitsuka N, Ohno H (2005) A crop phenology detecting method using time-series MODIS data. Remote Sens. Environ. 96:366374
Sarabandi K, Oh Y, Ulaby FT (1992) Measurement and calibration of differential mueller matrix
of distributed targets. IEEE Trans. Antennas Propag. 40(12):15241532
Schmugge TJ, Gloersen P, Wilheit T, Geiger F (1974) Remote sensing of soil moisture with microwave radiometers. J. Geophys. Res. 79:317323
Schmugge TJ, ONeill PE, Wang JR (1986) Passive microwave soil moisture research. IEEE Trans.
Shabanov NV, Wang Y, Buermann W, Dong J, Hoffman S, Smith GR, Tian Y, Knyazikhin Y,
Myneni RB (2003) Effect of foliage spatial heterogeneity in the MODIS LAI and FPAR algorithm over broadleaf forests. Remote Sens. Environ. 85:410423
Shao Y, Fan X, Liu H, Xiao J, Ross S, Brisco B, Brown R, Staples G (2001) Rice monitoring and
production estimation using multitemporal RADARSAT. Remote Sens. Environ. 76:310325
Shimabukuro YE, Smith JA (1991) The least squares mixing models to generate fraction images
derived from remote sensing multispectral data. IEEE Trans. Geosci. Remote Sens. 29:1620
Soares JV, Renno CD, Formaggio AR, da Costa Freitas Yanasse C, Frery AC (1997) An investigation of the selection of texture features for crop discrimination using SAR imagery. Remote
Spanner MA, Pierce LL, Peterson DL, Running SW (1990) Remote sensing of temperate coniferous forest leaf area index. The influence of canopy closure, understory vegetation and background reflectance. Int. J. Remote Sens. 11:95111
Sridhar VN, Dadhwal VK, Chaudhary KN, Sharma R, Bairagi GD, Sharma AK (1994) Wheat
production forecasting for a predominantly unirrigated region of Madhya Pradesh (India). Int.
Theau J, Peddle DR, Duguay CR (2005) Mapping lichen in a caribou habitat of Northern Quebec,
Canada, using an enhancement of classification method and spectral mixture analysis. Remote
Thenkabail PS, Smith RB, De Pauw E (2000) Hyperspectral vegetation indices and their relationships with agricultural crop characteristics. Remote Sens. Environ. 71:158182
Tompkins S, Mustard JF, Pieters CM, Forsyth DW (1997) Optimization of endmembers for spectral
mixture analysis. Remote Sens. Environ. 59:472489
Tucker CJ, Holden BN, Elgin GH Jr, Mcmurtrey E (1980) Remote sensing of total dry matter
accumulation in winter wheat. Remote Sens. Environ. 11:267277
Van Niel TG, McVicar TR (2004) Current and potential uses of optical remote sensing in rice-based
irrigating systems: a review. Aus. J. Agric. Res. 55:155185
Vina A, Gitelson AA, Rundqist DC, Keydan G, Leavit B, Schepers J (2004) Monitoring maize
(Zea mays L.) phenology with remote sensing. Agron. J. 96:11391147
Vinnikov K, Robock A, Qiu S, Entin J, Owe M, Choudhury B, Hollinger S, Njoku E (1999) Satellite remote sensing of soil moisture in Illinois, USA. J. Geophys. Res. 104:41454168
Viovy N, Arino O, Belward A (1992) The Best Index Slope Extraction (BISE): a method for
reducing noise in NDVI time-series. Int. J. Remote Sens. 13:15851590
Wang JR (1985) Effect of vegetation on soil moisture sensing observed from orbiting microwave
radiometer. Remote Sens. Environ. 17:141151
Wang N (1996) Wheat area and yield monitoring with remote sensing in China. China Science
Press, Beijing
Wang P, Li X, Gong J, Song C (2001) Vegetation temperature condition index and its application
for drought monitoring. Proceedings of International Geoscience and Remote Sensing Symposium, July 914, 2001, Sydney, Australia. IEEE, Piscataway, NJ, pp 141143
Watson K, Rowen LC, Offield TW (1971) Application of thermal modeling in the geologic interpretation of IR images. Remote Sens. Environ. 3:20172041
Weissteiner CJ, Braun M, Kuhbauch W (2004) Regional yield predictions of malting barley by
remote sensing and ancillary data. Proceedings of SPIE The International Society for Optical
Engineering, Barcelona, Spain, vol 5232, pp 528539
421
Wessel KJ, De Fries RS, Dempewolf J, Anderson LO, Hansen AJ, Powell SL, Moran EF (2004)
Mapping regional land cover with MODIS data for biological conservation: examples from the
Greater Yellowstone Ecosystem, USA and Para State, Brazil. Remote Sens. Environ. 92:6783
White MA, Thornton PE, Running SW (1997) A continental phenology model for monitoring vegetation responses to interannual climatic variability. Global Biogeochem. Cycles 11:217234
Wickel AJ, Jackson TJ (2001) Multitemporal monitoring of soil moisture with RADARSAT
SAR during the 1997 Southern Great Plains hydrology experiment. Int. J. Remote Sens.
22:15711583
Wu C (2004) Normalized spectral mixture analysis for monitoring urban composition using ETM
+ imagery. Remote Sens. Environ. 93:480492
Xiao X, Boles S, Frolking S, Li C, Salas W, Berrien Moore III (2005) Mapping paddy rice agriculture in South and Southeast Asia using multi-temporal MODIS images. Remote Sens. Environ.
100:95113
Xiao X, Boles S, Frolking S, Salas W, Moore BI, Li C (2002) Observation of flooding and rice
transplanting of paddy rice fields at the site to landscape scales in China using VEGETATION
sensor data. Int. J. Remote Sens. 23:30093022
Yang P (2005) Linking Multi-temporal Remotely Sensed Data, Field Observations and GIS-based
Crop Growth Model to Estimate Winter Wheat Yield in North China Plain. Doctoral Degree
Thesis of University of Tokyo, Tokyo. 104pp
Zhang X, Friedl MA, Schaaf CB, Strahler AH, Hodges JCF, Gao F, Reed BC, Huete A (2003)
Monitoring vegetation phenology using MODIS. Remote Sens. Environ. 84:471475
Chapter 16
Remote Sensing of Terrestrial Primary

Production and Carbon Cycle
Abstract The objective of this chapter is to review the historical development of and
the recent advances in the application of satellite remote sensing data for estimating
terrestrial gross and net primary production (GPP and NPP), while also monitoring
carbon cycle related ecosystem dynamics and changes. We achieve this objective by
separating the topic into five sections:
1. A review of the history of using satellite data to study the carbon cycle, concentrating on using the Normalized Difference Vegetation Index (NDVI) and its
derived Fraction of Photosynthetically Active Radiation (FPAR) and Leaf Area
Index (LAI) for biomass and NPP estimations
2. A description of recent advances in the application of Moderate Resolution Imaging Spectroradiometer (MODIS) data to estimates of GPP and NPP, along with
related findings using MODIS Land Surface Temperature (LST) and the Enhanced Vegetation Index (EVI)
3. A discussion of the application of long-term satellite data to the study of terrestrial ecosystems, including phenology monitoring, changes in regional carbon
storage resulting from land use change, carbon flux changes induced by climate
change, disturbance detection, and validation of ecosystem models
4. A proposed general scheme for applying satellite data to terrestrial ecosystem
studies, highlighting the role of modeling
5. A summary that emphasizes the continuity of vegetation monitoring with satellites
The use of remote sensing information for studying terrestrial primary production
and the global carbon cycle is significant both for an increased understanding of the
earth system and improved management of land and natural resources.

Numerical Terradynamic Simulation Group, Department of Ecosystem and Conservation Science,
University of Montana, Missoula, USA
zhao@ntsg.umt.edu
423
424
M. Zhao, S.W. Running
16.1 Introduction
Terrestrial Net Primary Production (NPP), the difference between Gross Primary
Production (GPP) and plant autotrophic respiration (Ra), is the net carbon fixed by
vegetation through photosynthesis. Quantification of NPP has socioeconomic significance, since NPP can directly measure the quantity of goods (e.g., food, fuel and
fiber) provided to human beings by ecosystems (Imhoff et al. 2004). NPP provides
the carbon required for maintenance of the structure and functions of an ecosystem.
The current climate change caused primarily by increasing anthropogenic greenhouse gas emissions, especially CO2 , is the largest global environmental issue facing
the world (IPCC, 2001), and there is now ample evidence of the ecological impacts
of recent climate change (Walther et al., 2002). Climate and terrestrial ecosystems
interact with and influence each other. On one hand, climate change and increasing
CO2 can cause changes in NPP and carbon storage in ecosystems (Nemani et al.,
2003a; Prentice et al., 2001), impacting the well-being of humans (Milesi et al.,
2005). On the other hand, terrestrial ecosystems can affect climate through carbon,
water, and energy exchange. For example, terrestrial ecosystems and oceans equally
absorbed nearly half of CO2 emission by human activities in the 1990s (Prentice
et al., 2001). Understanding the response of terrestrial NPP and the carbon cycle to
climate change is therefore critical for predicting future environmental change and
mitigating the impacts.
Satellite remote sensing data provide us with invaluable continuous temporal and
spatial information, which help us understand the processes, dynamics, and disturbances (such as land use change, wildfires, and insect outbreaks) in the biosphere,
and the impacts of environmental changes on terrestrial ecosystems. Since the application of AVHRR data to the study of vegetation in the early 1980s, great progress
has been made in the study of terrestrial carbon storage and fluxes, especially from
the NASA Earth Observing System (EOS) program. The NASA EOS program has
been planning and executing satellite-based earth monitoring for 15 years, and is the
heart of global change science for the USA. The central sensor on board the Terra
Satellite Platform is the Moderate Resolution Imaging Spectroradiometer (MODIS).
Terra was successfully launched on December 18, 1999, and the second MODISbased satellite, Aqua, was launched on May 4, 2002. For the first time in history, we
are able to obtain near-real time global vegetation growth status, including primary
production, at an 8-day time interval with 1 km spatial resolution (Justice et al.,
2002; Running et al., 2004).
This chapter concentrates on applying satellite data to terrestrial primary production and carbon cycle studies. We (1) review the history of using satellite data to
study the carbon cycle; (2) describe recent advances in the applications of MODIS
data; (3) discuss monitoring terrestrial ecosystems with the long-term satellite data
records; (4) propose a general scheme of applying the satellite data to terrestrial
ecosystem studies, highlighting the role of modeling; and (5) summarize the chapter by emphasizing the continuity of vegetation monitoring with satellites. We focus on the optical sensors on-board polar-orbiting satellites, especially on those
with medium to coarse resolution instruments. Radar and Lidar (light detection and
16 Remote Sensing of Terrestrial Primary Production and Carbon Cycle
425
ranging), which can also provide valuable vegetation information from microwaves,
are not covered in this chapter. In addition, land cover classification, an important
aspect of carbon relevant application, is detailed elsewhere (Chapter 13).
16.2 History of the Development of Using NDVI

16.2.1 Related NDVI to Green Biomass for Vegetation Monitoring
The basis for remote sensing of vegetation is the sharp contrast in reflectivity of
visible (0.40.7 m) and near-infrared (0.71.3 m, NIR) spectra caused by the optical properties of chlorophyll and the internal structure of green leaf cells. The
most widely used mathematic formula combining red and NIR reflectance (RED
and NIR ) is the Normalized Difference Vegetation Index (NDVI),
NDV I =
NIR RED
NIR + RED
(16.1)
Tucker (1979) evaluated several proposed vegetation indices and found that NDVI
is strongly related to biomass, and in 1980, his lab experiments also revealed that
NDVI is most strongly related to the percentage of green/brown herbage, suggesting the potential of using satellite data to estimate dry green biomass (Tucker, 1980).
The first reported study for using NDVI from satellites (AVHRR / NOAA) to quantify grass production at the regional scale was published in 1983 by Tucker et al.
(1983). They first constructed the relationship between AVHRR NDVI and clipped
grass biomass on the ground for a limited number of sites, extrapolating the relationship to the study region using NDVI to estimate spatial patterns of biomass.
Today, similar studies still employ this method, sometimes using relatively higher
resolution satellite data (e.g., 30 m TM/ETM+, or 15 m ASTER) as an intermediate
to scale up from field observations to coarse resolution satellite data to enhance the
accuracy of estimations (e.g., Reeves et al., 2006).
The first use of multi-temporal AVHRR NDVI to monitor the dynamics of vegetation at the continental and global scales was in 1985. Tucker et al. (1985) found the
differences in temporal dynamic of vegetation reflected by NDVI are associated with
variations in climate and dependent on biome types, and that the integrated NDVI
over a given time interval is related to NPP. Justice et al. (1985) expanded this study
to the global scale, suggesting that AVHRR NDVI is effective for monitoring phenology of global vegetation. Townshend and Justice (1986) further analyzed NDVI
from different years in Africa and found inter-annual variation in NDVI can reveal
the response of vegetation to climate anomalies.
426
16.2.2 Related NDVI to FPAR, LAI and NPP for Carbon

Cycle Study
Goward et al. (1985) first demonstrated the linear relationship between the growing
season integrated NDVI and ground-based observations of NPP for different biomes
over North America. Tucker et al. (1986) found a strong relationship between seasonal variations in atmospheric CO2 and NDVI. Both studies suggested that NDVI
can be used to estimate terrestrial photosynthesis, and therefore NPP. In the following year, Fung et al. (1987) was the first to relate NDVI to annual NPP at the
global scale to study atmosphere-biosphere exchange of CO2 . Fung et al.s (1987)
study simply distributed annual NPP into monthly level based on the monotonic
function of monthly NDVI, without considering solar radiation or environmental
stresses. However, Running and Nemani (1988) examined the seasonal relationship
between photosynthesis estimated by a process-based model, FOREST-BGC, and
NDVI/AVHRR for seven sites in North America representing a wide range of annual climates, and found that NDVI can not reflect the drawdown of photosynthesis induced by summer drought for water stressed sites, implying that NDVI alone
could not fully represent seasonal photosynthesis.
Along with the above studies to directly relate NDVI to NPP and the carbon cycle, additional studies were conducted to develop the physiological linkage between
NDVI and NPP. Professor Monteith was the pioneering scientist who proposed the
concept of photosynthesis efficiency logic. Monteith (1972, 1977) found that crop
production under non-stressed conditions is linearly related to the amount of photosynthetically active radiation solar radiation (PAR) that is absorbed by green leaf
(APAR). Kumar and Monteith (1982) decomposed the linear model into independent parameters such as incoming solar radiation, radiation absorption efficiency,
and conversion efficiency of APAR, while also demonstrating how the fraction of
APAR (FPAR) is related to the ratio of red reflectance to NIR. Asrar et al. (1984)
found that NDVI can be simultaneously used to estimate both FPAR and leaf area
index (LAI), FPAR is a linear function of NDVI, and LAI is a curvilinear function of NDVI, while Sellers (1985, 1987) also found a linear relationship between
NDVI and FPAR. Running et al. (1989) first demonstrated a prototype for incorporation of the LAI derived from AVHRR NDVI into a process-based ecosystem model
to simulate regional forest evapotranspiration and photosynthesis. While numerous
empirical studies have found a strong relationship between NDVI and ground LAI
and FPAR, the first theoretical interpretation of this was provided by Myneni et al.
(1995), stating that vegetation indices are a measure of chlorophyll abundance and
energy absorption. Further, Myneni et al. (1997a) proposed using a canopy radiation
transfer model to derive FPAR and LAI.
Models calculating regional NPP based on Monteiths photosynthesis efficiency
logic combined with FPAR estimated from remotely sensed NDVI were proposed
after 1990. Prince (1991) first proposed a model of regional primary production for
use with coarse resolution satellite data, accounting for the physiological costs of
maintenance and growth respirations, and the environmental constraints tending to
427
reduce maximum light use efficiency ( ). Prince and Goward (1995) later named
the model the Global Production Efficiency model (GLO-PEM), and simulate
global GPP and NPP at 8 km resolution. The first global NPP images estimated
using satellite NDVI were generated with the CarnegieAmesStandford approach
(CASA) model by Potter et al. (1993), which had a spatial resolution of one degree. Running and Hunt (1993) proposed a NDVI-based NPP model, also exploring the range and variability of using a mechanistic model. Ruimy et al. (1994,
1996) proposed similar models of NPP from AVHRR NDVI, but their models had
no constraints on potential maximum resulting from environmental stresses. In
addition, a forest stand growth model, 3-PG (Use of Physiological Principles in Prediction Growth) can use satellite FPAR as an input, and was developed by Landsberg
and Waring (1997). Though there are some differences among the different models,
Photosynthetic Efficiency Models (PEMs) can be expressed generally as,
P = PAR FPAR m f (T ) F(W )
(16.2)
where P is GPP or NPP over a given time interval t, FPAR is derived using vegetation indices, m is the maximum , and f (T ) and f (W ) are the constraints resulting
from temperature and water stress. In general, water stress results from soil moisture
and air vapor pressure deficit (VPD).
16.3 Advances for the MODIS Sensor

16.3.1 MODIS GPP and NPP Products
MODIS may be so far the most complex instrument built and flown on a spacecraft
for civilian research purposes (Guenther et al., 2002). The MODIS sensor provides
higher quality data for monitoring terrestrial vegetation and other land processes
than previous AVHRR, not only because of its narrower spectral bands that enhance
the information derived from vegetation (Justice et al., 2002), onboard calibration
to guarantee the consistent time-series reflectance (Guenther et al., 2002), and orbit
and altitude satellite maneuvers to ensure sub-pixel geolocation accuracy (Wolfe
et al., 2002), but also because leading scientists are working as a team to improve the
accuracy of the data from low level reflectance data, to high level data, such as land
cover, fire, land surface temperature, vegetation indices (NDVI and EVI; EVI is the
enhanced vegetation index), FAPR / LAI and GPP / NPP (Justice et al., 2002). The
other important feature of MODIS data is that all of the land products have quality
flags, denoting any negative atmospheric impacts (e.g., cloudiness and aerosol) to
help users screen the data for their purposes (Justice et al., 2002). Combined with the
MODIS atmosphere and ocean products, MODIS vegetation data provide invaluable
information for earth system study.
428
The algorithm for the MODIS 1 km 8-day GPP and annual NPP employs 1 km
land cover and 1 km 8-day FPAR/LAI in addition to daily spatial coarse resolution
meteorological data. The LAI is used to estimate the biomass of leaf, fine root and
live wood for plant maintenance respiration calculations in the algorithm (Running
et al., 2000; Heinsch et al., 2003; Running et al., 2004; Zhao et al., 2005). Because of
the near real-time processing of MODIS GPP/NPP products and the required suite
of large datasets, water stress is represented solely by air VPD rather than a full
water stress from both VPD and soil moisture, thereby avoiding extremely intensive
computation and the creation of the additional physical and biophysical datasets
required for water balance calculation. Our study has shown that VPD alone can
capture the inter-annual variability of the full water stress, although it may fail to
capture the seasonality of water stresses for dry areas dominated by strong monsoons (Mu et al., 2007). The quality flags in these MODIS land data products allow
us to fill the gaps in the time series of FPAR and LAI resulting from contamination by unfavorable atmospheric conditions, generating more accurate estimations
of GPP and NPP (Zhao et al., 2005).
For the first time, we have more than 6 years of consistent global 1 km GPP and
NPP data products estimated from MODIS. Figure 16.1 shows the averaged seasonality of MODIS GPP from 6-year (20002005) results, in which we have aggregated
8-day values to 3 month averages (Fig. 16.1a). Spatial seasonal variations clearly
Fig. 16.1 Spatial patterns of the seasonality of MODIS GPP (a), and the mean annual cycle of
GPP at an 8-day interval for four latitude bands and the globe (b). (From Zhao et al., 2006b.)
429
demonstrate the expected peak GPP in June, July, and August and the low values in
December, January, and February over the mid- and high-latitudes of the northern
hemisphere (NH). The rainforests of the Amazon Basin have higher GPP during the
dry season from July to November than during the wet season, which agrees with
the studies by Huete et al. (2006) and Xiao et al. (2006). Monthly precipitation in
the Amazon Basin can reach approximately 100 mm in the dry season, making solar radiation, not water, the leading limiting factor in this region, limiting growth in
the wet season. Figure 16.1b shows the annual cycle of total GPP for four latitudinal bands and for the entire globe. Relatively strong seasonal signals occur for the
mid- and high-latitudes of the NH (i.e., north of 22.5 N). The areas south of 22.5 S
have the opposite seasonal profile relative to the mid- and high-latitudes of NH, and
the seasonality for the southern hemisphere is much weaker because there is significantly less land mass. For the entire tropical region (22.5 S22.5 N), there is
almost no discernible seasonality, and total GPP is always the highest among the
four latitudinal bands. Therefore, at the global scale, the magnitudes of annual GPP
cycle can be mostly attributed to the tropical region, while the seasonality in the
global cycle is largely determined by areas north of 22.5 N.
Figure 16.2 reveals the spatial pattern of the 6-year mean annual total NPP. As
expected, high MODIS NPP occurs in areas covered by forests and woody savannas, especially in tropical regions. Low NPP occurs in areas dominated by harsh
environments, such as high latitudes with short growing seasons constrained by low
temperatures and daylength, and dry areas with limited water availability. At the
global scale, from 2000 to 2005, MODIS estimated a total terrestrial annual GPP
of 109.07 Pg C (std. 1.66), and an annual NPP of 52.03 Pg C (std. 1.17), ignoring
Fig. 16.2 Spatial patterns of global terrestrial MODIS NPP averaged over 6 years (20002005).
(From Zhao et al., 2006b.)
430
Fig. 16.3 Interannual variations in global total C4.8 MOD17 NPP driven by NCEP and GMAO,
respectively, as compared with the inverted atmospheric CO2 growth rate. A Multivariate ENSO
Index (MEI) is shown in gray scale, where darker shades represent higher MEI values. (From Zhao
et al., 2006b.)
barren land cover as defined by the MODIS land cover product. For vegetated areas,
the mean annual GPP and NPP are 996.03 (std. 823.67) and 475.19 (std. 375.44)
g C m2 year1 , respectively.
Interannual anomalies in the MODIS global NPP record correlate well with
the inverted atmospheric CO2 growth rate, corresponding to results (correlation =
0.70, P < 0.001) found by Nemani et al. (2003a) for the AVHRR period of record
(19821999). Figure 16.3 shows the relationship between MODIS NPP anomalies
and CO2 growth rates. To account for the uncertainties from inputs from different
meteorological datasets, Zhao et al. (2006a) used both GMAO and NCEP meteorology to drive the MODIS GPP and NPP algorithm. For the 6-year MODIS record, the
correlations are 0.85 (P < 0.016) and 0.91 (P < 0.006) using GMAO and NCEP,
respectively (Zhao et al., 2006b), implying that NPP is the primary driver of the
atmospheric CO2 growth rate.
16.3.2 Recent Findings Using MODIS Vegetation Products

There were no standard global land surface products (temperature (LST), EVI, fire,
GPP and NPP) prior to 2000 when MODIS began providing data, but information
from these 1 km or sub 1 km data products help us understand the dynamics of terrestrial ecosystems and detect the underlying mechanisms at different levels ranging
from local to regional scales. MODIS LST and EVI data can provide information on
surface resistance and disturbance (Nemani and Running, 1989; Mildrexler et al.,
431
2007), and EVI has been found to be strongly related to the GPP derived from
eddy covariance flux tower measurements (Xiao et al., 2004; Rahman et al., 2005;
Huete et al., 2006), indicating the potentially superior utility of EVI in comparison
to NDVI. However, Daniel A. Sims (personal communication on 3/2/06) has found
that EVI models break down for sites subjected to summer drought as indicated by
high summer VPD. This finding is consistent with the similar conclusion by Running and Nemani (1988) using NDVI, implying that neither EVI nor NDVI can
reflect the total reduction in carbon uptake resulting from water stress, especially
for evergreen ecosystems in dry areas. Hence, water stresses must be incorporated
into global remote sensing primary production models. For Amazon rainforests,
MODIS EVI has revealed enhanced vegetation growth during dry seasons (Huete
et al., 2006; Xiao et al., 2006), and the MODIS GPP, incorporating such stresses,
also shows this enhanced production during dry seasons (Fig. 16.1a).
16.4 Monitoring the Long-Term Dynamics of Ecosystems

and Carbon Cycle
The advantages of using satellites to monitor land are numerous. Not only can satellite data provide detailed spatial patterns and variations in ecosystem processes, but
they also provide information on the temporal changes. With the accumulated longterm satellite data from Landsat since 1972 and AVHRR/NOAA since 1981, there
are more than three decades of satellite data, enabling us to study dynamics of and
changes in terrestrial ecosystems. The following subsections present several important applications of long-term satellite data.
16.4.1 Long Term Phenology Monitoring

Vegetation activities at mid- and high-latitudes are largely controlled by short growing seasons resulting from temperature and daylength constraints. Recent warming
(19762000) has been greater over the continents of the northern hemisphere (NH)
than elsewhere (Folland et al., 2001), lengthening the growing season and stimulating vegetation growth. CO2 measurements have shown an advance of upto 7 days
in the timing of the drawdown of CO2 in spring and early summer since the 1960s
(Keeling et al., 1996). However, CO2 data alone can not depict detailed spatial information on the response of vegetation to the warming temperature. Though 3D
transfer models can retrieve some spatial information on carbon sinks and sources,
the spatial resolution is very coarse and the results contain relatively large uncertainties (Fan et al., 1998). Long-term AVHRR NDVI data allow us to detect vegetation
responses to the warming climate spatio-temporally. Myneni et al. (1997b) utilized
the AVHRR NDVI from 1981 to 1991 and found that biospheric activity increased
remarkably in the northern high latitudes as a result of warming. Zhou et al. (2001)
2.5
2
1.5
1
0.5
0
0.5
1
1.5
2
2.5

NORTH AMERICA (40N-70N)
TEMPERATURE
1.5
1
0.5
0
0.5
GREENNESS
EURASIA (40N-70N)
2
TEMPERATURE
GREENNESS
ANOMALY
ANOMALY
432
82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99
YEAR
1
1.5
2
2.5
82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99
YEAR
Fig. 16.4 Anomaly of standardized NDVI and air temperature from 1982 to 1999 for North
America and Eurasia (40 N70 N). (Redrawn from Zhou et al., 2001.)
continued the study by using a longer term NDVI record from 1981 to 1999, comparing that data to anomalies in air temperature (Fig. 16.4) for North America and
Eurasia from 1982 to 1999. They found that Eurasia had larger increases in both the
magnitude and duration of the seasonal cycle of NDVI than did North America. The
length of active growing season increased by more than 2 weeks (18 4 days) for
Eurasia and nearly 2 weeks for North America (12 5 days). This lengthening of
the growing season has implications for many aspects of ecosystem processes, especially the carbon cycle. It may indicate an increase in NPP, but it may also enhance
processes that release carbon to the atmosphere, such as decomposition, wildfires
and insect outbreaks.
16.4.2 Estimation of Regional Carbon Storage Changes

from Land Use Change
Year-to-year variability in the growth rates of atmospheric CO2 concentrations is
large, and is principally induced by inter-annual variations in terrestrial metabolism
(Kindermann et al., 1996; Prentice et al., 2000; Keeling et al., 2001; McGuire et al.,
2001; Nemani et al., 2003a). Over a relatively long period (e.g., decades), the mean
carbon fluxes between atmosphere and land are controlled mainly by changes in
biomass density, soil carbon and land use (Prentice et al., 2001; Schimel et al.,
2001). Historical records and national inventories together with bookkeeping models have been used to estimate the contribution of carbon flux from land use change
(e.g., Houghton et al., 1983). There are, however, large uncertainties in quantifying
the role of changes in biomass density and land use in the contemporary global carbon cycle (Prentice et al., 2001; Schimel et al., 2001), largely because of varying
definitions of forest cover among countries and differing time intervals (Matthews,
2001). Satellite data offer the possibility of providing spatially and temporally consistent estimates of forest cover to complement national reports. Using AVHRR
data, for example, DeFries et al. (2002) estimated the changes in forest cover at the
sub-pixel level over tropical regions for the 1980s and 1990s, and found that for the
433
1980s, the carbon emissions from land use change were 0.6 (0.30.8) Pg C year1 ,
much less than 1.7 (0.62.5) Pg C year1 , a value from IPCC report (Prentice et al.,
2001), which was estimated mainly based on the statistics from United Nation Food
and Agriculture Organization (FAO) and deforestation rates from the FAO Forest Resource Assessment (FRA). The measurements from satellite suggested less
missing carbon in the global carbon budget than the previous estimates. DeFries
et al. (2002) also found that, compared with the 1980s, clearing of tropical forests
increased by about 10% as revealed by satellite data, as opposed to the 11% reduction reported by FRA, implying that there are increasing carbon emissions from
changes in tropical land use. For forests in mid- and high-latitudes of NH, Myneni
et al. (2001) simply used the fitted empirical equations between growing season
cumulative NDVI and total biomass derived from inventory data of stem wood volume and long-term AVHRR NDVI records to discover an overall carbon sink for
these forests, especially for forests in Eurasia and east North America. Compared
with statistics and inventory data, satellite data can provide more detailed spatial
information, and the results are generally consistent and accurate.
16.4.3 Estimation of Regional Carbon Fluxes Changes

With a long-term AVHRR dataset and the PEMs mentioned previously, it is possible
to detect the trends and interannual variability in global NPP. At the regional level,
there have been several reports of increased NPP. For example, Randerson et al.
(1999) used the CASA model to study the relationships among NPP, Net Ecosystem Production (NEP) and seasonal cycle of atmospheric CO2 at high latitudes and
found that the enhanced photosynthesis activities in the early spring due to warming
temperature are responsible for the changing trend in the seasonal cycle of CO2 .
Also using the CASA model, Hicke et al. (2002a, b) reported an increase in NPP
over most of North America from 1982 to 1999; Fang et al. (2003) found increased
NPP from 1982 to 1999 in China.
Studies have also found increased NPP at the global scale. Ichii et al. (2001)
found that NPP increased during the 1980s with AVHRR NDVI. Cao et al. (2004),
using the GLO-PEM model, found that NPP increased from 1981 to 2000, and the
authors attribute the increased NPP to the effects of both increased CO2 fertilization effects and the favorable change in climate during the periods. Potter et al.
(2003a) used CASA and AVHRR data to calculate changes in NPP and NEP during
19821998, determining that NEP increased at the global scale, and further, that
Eurasia was a larger carbon sink than North America, contradicting the atmospheric
inversion study by Fan et al. (1998). These studies are concentrated, using only one
(solar radiation; Ichii et al., 2001) or two climate variables (temperature and precipitation; Potter et al., 2003a) or the combined effects of climate and CO2 (Cao et al.,
2004). Nemani et al. (2003a) used a PEM developed for the global MODIS GPP
and NPP products (Running et al., 2000) to study the effects of all three primary
climate controls (incoming solar radiation, temperature and water) on NPP for the
434
Fig. 16.5 Spatial distribution of linear trends in estimated NPP (% change per year) using FPAR
and LAI derived from AVHRR data for 19821999. (From Nemani et al., 2003a.)
period 19821999, and found that global NPP increased 6%, with tropical regions,
especially Amazon rain forests (accounting for 42% of the global increase), being
the largest contributors (Fig. 16.5). The increased NPP in the Amazon is caused by
increased solar radiation resulting from reduced cloudiness during the dry season.
In addition, Nemani and colleagues found that climate alone was responsible for
more than 40% of the increase in the global NPP. Other factors, such as CO2 fertilization and nitrogen deposition may also played a role in enhancing NPP. Nemani
et al.s (2003a) study emphasizes the role of radiation, generally a dominant limiting
climate factor in humid, highly productive rain forests covering large areas of the
tropics.
16.4.4 Monitoring Disturbances

Disturbances from fire, insect outbreaks, logging, etc., constitute the major components to ecosystem change and have great impact on the global carbon cycle, but
many terrestrial biogeochemistry models do not account for disturbances (Canadell
et al., 2000). The chief reason for omitting disturbances in these models is that there
is no reliable record of the timing, distribution, spatial extent, or severity of these
disturbances at the regional or the global scale. Only medium- to coarse-resolution
satellites can provide the timing of major disturbances at the global scale. Potter
et al. (2003b) retrieved major disturbances at the global scale with the AVHRR
FPAR data for the period 19821999 and combined it with the NASA-CASA model
to estimate the above-ground biomass carbon lost to the atmosphere during that
435
period. Van der Werf et al. (2004) used fire activity obtained from several satellites together with a biogeochemical model and an inverse analysis of atmospheric
CO anomalies to estimate the CH4 and CO2 emissions from fire during the 1997
2001 El Nino/La Nina period. They found that CO2 released from fire accounts for
66 24% of the observed atmospheric CO2 annual growth rates during El Nino.
Randerson et al. (2005) further separated the contributions C3 and C4 vegetation
to fire emissions and found that C3 vegetation is largely responsible for interannual
variations in global fire emissions.
MODIS provides real-time fire burned area products (Roy et al., 2005). However,
other disturbances besides fires, including insect outbreaks, flood, irrigation etc., are
also important contributors to the carbon cycle. With MODIS LST and EVI datasets,
Mildrexler et al. (2007) have proposed a powerful Disturbance Index (DI),
DI = (LSTmax /EV Imax )/(LSTmax /EV Imax )
(16.3)
to detect all disturbance, regardless of origin. In Eq. 16.3, LSTmax (EV Imax ) is the
annual maximum composite MODIS LST (EVI), and LSTmax /EV Imax is the multiyear mean of the ratio of LSTmax to EV Imax . Compared with the field data, the DI
can effectively detect fire scars and other disturbances at regional and global scales
on an annual basis (Fig. 16.6). As mentioned previously, MODIS NPP results show
that NPP anomalies can explain 7283% of annual CO2 growth rate (Fig. 16.3),
while 66 24% was found by Van der Werf et al. (2004) from fire emission. Incorporating DI into MODIS NPP algorithm will help to separate the roles of NPP and
disturbance in the global carbon cycle.
16.4.5 Validation of Process-Based Ecosystem Model

with Remote Sensing Data
Process-based ecosystem models are now playing an indispensable role in simulating the dynamics of ecosystems. These models all experience improvements
through validation and advances in knowledge. However, unlike General Circulation
Models (GCMs), which can be validated with observations from several thousand
weather stations covering nearly all climatic zones, there is a paucity of ecological
observations to validate ecosystem models. Remote sensing data, because of their
spatio-temporal consistency, can be used as independent observations to validate
such models. For example, Zhao et al. (2002) used AVHRR NDVI to test improvements in a biogeography model for simulation of leaf longevity (deciduous or evergreen) over China relative to the original method. Zhuang et al. (2003) compared
the trend in carbon storage simulated by a process-based ecosystem model with that
derived from NDVI data by Myneni et al. (2001) over the extra-tropical NH, determining that there is similar spatial pattern between the modeled and satellite data,
though the modeled results had a smaller rate of change than the satellite estimates.
436
Fig. 16.6 Correspondence between the DI results, the MODIS active fire-detection data (black
dots) and fire perimeter maps (black outlines) for (a, b) the 2003 wildfires near Missoula, Montana,
and (c, d) the 2003 southern California wildfires. The southern California fires occurred in savanna
and shrublands, vegetation types with the highest frequency of major disturbance at the global
scale. (From Mildrexler et al., 2006.)
More importantly, results from models and satellite observations can support
each other and strengthen our confidence in understanding the mechanisms behind
ecosystem processes. Lucht et al. (2002) provide an excellent example of this support in their study of LAI following the Mt. Pinatubo eruption. The trend in vegetation activities derived from the long-term AVHRR dataset has long been suspect
because of the need for data corrections resulting from instrument and navigational
drift, intercalibration of successive instruments, and consideration of aerosol effects
(Cihlar et al., 1998). In addition, there have been several different explanations for
the reduced atmospheric CO2 growth rates after the effects of Mt. Pinatubo have
been taken into account (Arora, 2003). Lucht et al. (2002) used a biogeochemical
model of vegetation and observed climate data to discover that there is a drawdown
of maximum LAI simulated by the model in the northern high-latitude growing seasons of 19921993, consistent with the reduced LAI in the same period derived from
437
AVHRR. This consistency between two independent datasets lead to the discover
that the lower AVHRR-estimated LAI following the 1991 Mt. Pinatubo eruption
was not caused by contamination of the satellite record from high aerosol loading,
but by the volcanic-induced climate anomaly, especially low temperatures. Moreover, the model results demonstrated that the unbalanced effect of cooling on NPP
and heterotrophic respiration provides a much simpler explanation for the additional
high-latitude CO2 uptake than the proposed mechanism of increased NPP due to the
increased diffuse sky light by Gu et al. (2003). Krakauer and Randerson (2003) also
found reduced tree-ring width following Pinatubo eruption, confirming the reduced
NPP due to volcanic effects.
16.5 Towards an Integrated Study of the Terrestrial Carbon

Dynamics
Remote sensing data are able to provide spatio-temporal dynamics of vegetation at
regional and global scales. To quantify the role of terrestrial vegetation in the carbon
cycle, however, requires ecosystem models. These models are validated, calibrated
and refined based on the relationship between remote sensing data and the groundbased data related to carbon at local level (Heinsch et al., 2006; Turner et al., 2006;
Zhao et al., 2005). Remote sensing data are then used to scale up from the local to the
regional and the global scale both temporally and spatially (e.g., Zhao et al., 2005).
The ecosystem models mentioned here cover a large complexity, ranging from very
simple regression models such as the empirical relationship between NDVI and biomass or LAI to complex ecosystem process models. Figure 16.7 shows a scheme of
quantification of the terrestrial carbon cycle using remote sensing data in an integrated system. Ground data, such as these measured at the eddy flux towers, and
other observations, such as land cover, NPP, biomass, are the source for developing
linkages between field-based and remote sensing data. Then the ecosystem models
use remote sensing data to quantify the regional or global scale carbon exchange
(Running et al., 1999).
In reality, the interactions among ground data, remote sensing data, ecosystem
models and carbon cycle are more complex. As depicted by Fig. 16.7, remote sensing data sometimes can also detect some mechanisms not observable with field measurements and remote sensing data may even assist in selecting appropriate sites
for field observations. The results from ecosystem models may help us explain remote sensing data, and vice versa, while remote sensing data also can be used to
validate or drive ecosystems models. Ecosystem models can quantify carbon exchange between atmosphere and terrestrial ecosystems, but carbon cycle data can
also constrain and validate ecosystem models. Ecosystem models are the centerpiece
of terrestrial carbon cycle studies, because modeling is the only method for understanding ecosystem processes in an interactive manner and predicting the impacts of
biotic and abiotic factors on the carbon cycle, providing important information for
policy makers and stockholders. With the improved quality and increasing spatial
438
Fig. 16.7 An interactive scheme for studying the terrestrial carbon cycle using remote sensing data
coverage of ground-based data, improved quality of satellite data, and the increasing knowledge gained from these data, the performance of the ecosystem models
will continue to be enhanced, thereby enhancing our ability to study the earth as a
system.
16.6 Summary
Human beings have never before had such a great impact on the earth environment in
recorded history. Our activities have changed land cover, water and nutrient cycling,
the chemistry of the atmosphere, and therefore, climate systems and the structure
and function of ecosystems. In turn, the anthropogenically induced environmental
changes have influenced our well-being. More importantly, many current human-
439
induced environment issues are not restricted to a given region or country, but are
having global impacts. Regular global measurements from satellites play a crucial
role in monitoring the earth, which can provide advanced warning to allow for favorable environmental change (Running et al., 2006). The atmospheric CO2 concentration measurements at Mauna Loa since 1958 (the famous Keeling curve),
has given us the advanced warning for global climate change, and led to the 1997
Kyoto Protocol, the international treaty on climate change, assigning mandatory
targets for the reduction of greenhouse gas emissions to signatory nations. Ironically, Keelings measurements have been nearly halted several times due to the reluctance of the US government to provide continued support (Keeling, 1998). With
the accumulated long-term time-series data, we gain more knowledge and find new
discoveries (Keeling, 1998). Likewise, EOS has provided valuable scientific knowledge, and we hope that NASA continues this mission in some EOS-like project.
The other lesson we learned from the use of satellite data to study the carbon
cycle is that it is vital to have basic standard datasets for use in land science. We
have made a lot of discoveries using only NDVI. However, the quality of NDVI
is very critical for the science research, and there should be some MODIS-like land
products generated continuously, with similar sensor spatial resolution, quality flags,
and easy access data formats. As discussed above, for example, without standard
LST and EVI datasets, it is impossible for the ecologists to effectively use satellite
data (e.g., Mildrexler et al., 2007); without FPAR and LAI derived from satellite
data, it is impossible for us to have a deeper understanding of the terrestrial carbon
cycle following the Mt. Pinatubo eruption (e.g., Lucht et al., 2002), and the longterm trend in global NPP (e.g., Nemani et al., 2003a).
With advancements in ecosystem modeling based on remotely sensing data, as
well as the enhancement of computer performance and Internet technology, it is now
becoming possible for land managers and policy makers to make decisions based on
near real-time satellite data (Running et al., 2004), or even to take preventative action based on the information from near future forecasting using the same satellite
data (Nemani et al., 2003b). Thus, the advancements in the study of terrestrial primary production and carbon cycle using satellite data are significant, not only for
understanding the global carbon cycle, but also for application of satellite data to
carbon-related natural resource management and land management.
Acknowledgements The work is funded by the NASA/EOS Natural Resource/Education Training
Center (grant NAG5-12540) and NASA MODIS Project (NNG04HZ19C). We thank Dr. Faith Ann
Heinsch for comments and Dr. Liming Zhou for providing Fig. 16.4 for the chapter.
References
Arora VK (2003) Decreased heterotrophic respiration reduced growth in atmospheric CO2 concentration. IGBP Global Change Newsletter 54:2122
Asrar G, Fuchs M, Kanemasu ET, Hatfield JL (1984) Estimating absorbed photosynthetic radiation
and leaf area index from spectral reflectance in wheat. Agron. J. 76:300306
440
Cao M, Prince SD, Small J, Goetz SJ (2004) Remotely sensed interannual variations and trends in
terrestrial net primary productivity 19812000. Ecosystems 7:233242
Canadell JG, Mooney HA, Baldocchi DD, Berry JA, Ehleringer JR, Field CB, Gower ST,
Hollinger DY, Hunt JE, Jackson RB, Running SW, Shaver GR, Steffen W, Trumbore SE,
Valentini R, Bond BY (2000) Carbon metabolism of the terrestrial biosphere: a multi-technique
approach for improved understanding. Ecosystems 3:115130
Cihlar J, Chen JM, Li Z, Huang F, Latifovic R, Dixon R (1998) Can interannual land surface
signal be discerned in composite AVHRR data? J. Geophys. Res. 103(D18):2316323172, doi:
10.1029/98JD00050
DeFries RS, Houghton RA, Hansen M, Field CB, Skole DL, Townshend J (2002) Carbon emissions
from tropical deforestation and regrowth based on satellite observations for the 1980s and 90s.
Proc. Natl. Acad. Sci. USA 99(22):1425614261
Fan S, Gloor M, Mahlman J, Pacala S, Sarmiento J, Takahashi T, Tans P (1998) A large terrestrial
carbon sink in North America implied by atmospheric and oceanic carbon dioxide data and
models. Science 282:442446
Fang J, Piao S, Field CB, Pan Y, Guo Q, Zhou L, Peng C, Tao S (2003) Increasing net primary
production in China from 1982 to 1999. Front. Ecol. Environ. 1:293297
Folland CK, Karl TR, Christy JR, Clarke RA, Gruza GV, Jouzel J, Mann ME, Oerlemans J,
Salinger MJ, Wang S-W (2001) Observed climate variability and change. In: JT
Houghton,Y Ding, DJ Griggs, M Noguer, PJ van der Linden, X Dai, K Maskell, CA Johnson
(eds), Climate Change 2001: The Scientific Basis. Contribution of Working Group I to the Third
Assessment Report of the Intergovernmental Panel on Climate Change. Cambridge University
Press, Cambridge, United Kingdom and New York, USA, pp 182237
Fung I, Tucker CJ, Prentice K (1987) Application of advanced very high resolution radiometer
vegetation index to study atmosphere-biosphere exchange of CO2 . J. Geophys. Res. 92(D3):
29993015
Goward SN, Tucker CJ, Dye DG (1985) North American vegetation patterns observed with the
NOAA-7 Advanced Very High Resolution Radiometer. Vegetatio 64:314
Gu L, Baldocchi DD, Wofsy SC, Munger JW, Michalsky JJ, Urbanski SP, Boden TA (2003) Response of a deciduous forest to the Mount Pinatubo eruption: enhanced photosynthesis. Science
299:20352038
Guenther B, Xiong X, Salomonson VV, Barnes WL, Young J (2002) On-orbit performance of
the earth observing system moderate resolution imaging spectroradiometer; first year of data.
Heinsch FA, Reeves M, Votava P, et al. (2003) Users Guide: GPP and NPP (MOD17A2/A3)
Products, NASA MODIS Land Algorithm, pp 157
Heinsch FA, Zhao M, Running SW, et al. (2006) Evaluation of remote sensing based terrestrial
productivity from MODIS using regional tower eddy flux network observations. IEEE Trans.
Hicke JA, Asner GP, Randerson JT, Tucker C, Los S, Birdsey R, Jenkins JC, Field CB (2002a)
Trends in North American net primary productivity derived from satellite observations, 1982
1998. Global Biogeochem. Cycles 16(2), doi:10.1029/2001GB001550
Hicke JA, Randerson J, Asner G, Randerson J, Tucker C, Los S, Birdsey R, Jenkins J, Field C,
Holland E (2002b) Satellite-derived increases in net primary productivity across North
America, 19821998, Geophys. Res. Lett. 29(10), doi:10.1029/2001GL013578
Houghton RA, Hobbie JE, Melillo JM, et al. (1983) Changes in the carbon content of terrestrial
biota and soils between 1860 and 1980: a net release of CO2 to the atmosphere. Ecol. Monogr.
53:235262
Huete AR, Didan K, Shimabukuro YE, Ratana R, Saleska SR, Hutyra LR, Yang W, Nemani RR,
Myneni R (2006) Amazon rainforests green-up with sunlight in dry season. Geophys. Res. Lett.
33, L06405, doi: 10.1029/2005GL025583
Ichii K, Matsui Y, Yamaguchi Y, Ogawa K (2001) Comparison of global net primary production
trends obtained from satellite-based normalized difference vegetation index and carbon cycle
model. Global Biogeochem. Cycles 15(2):351364, doi: 10.1029/2000GB001296
441
Imhoff ML, Bounoua L, Richetts T, Loucks C, Harriss R, Lawrence WT (2004) Global pattern in
human consumption of net primary production. Nature 429:870873
IPCC (Intergovernmental Panel on Climate Change) (2001) IPCC (Intergovernmental Panel on
Climate Change). In: JT Houghton, Y Ding, DJ Griggs, M Noguer, PJ van der Linden, X Dai X
(eds), Climate Change 2001: The Scientific Basis Contribution of Working Group I to the Third
Assessment Report of the IPCC. Cambridge University Press, Cambridge, United Kingdom
Justice CO, Townshend JRG, Holben BN, Tucker CJ (1985) Analysis of the phenology of global
vegetation using meteorological satellite data. Int. J. Remote Sens. 6:12711318
Justice CO, Townshend JRG, Vermote EF, Masuoka E, Wolfe RE, Saleous N, Roy DP, Morisette JT
(2002) An overview of MODIS land data processing and product status. Remote Sens. Environ.
83:315
Keeling CD, Chin JFS, Whorf TP (1996) Increased activity of northern vegetation inferred from
atmospheric CO2 measurements. Nature 382:146149
Keeling CD (1998) Rewards and penalties of monitoring the earth. Annu. Rev. Energy Environ.
23:2582
Keeling CD, Piper SC, Bacastow RB, Wahlen M, Whorf TP, Heimann M, Meijer HA (2001) Exchange of atmospheric CO2 and 13 CO2 with the terrestrial biosphere and oceans from 1978 to
2000, I. Global aspects. SIO Ref. Ser. 0106, Scripps Institute of Oceanography, La Jolla, CA
Kindermann J, Wurth G, Kohlmaier GH, Badeck F-W (1996) Interannual variation of carbon
exchange fluxes in terrestrial ecosystems. Global Biogeochem. Cycles 10(4):737756, doi:
10.1029/96GB02349
Krakauer NY, Randerson JT (2003) Do volcanic eruptions enhance or diminish net primary production? Evidence from tree rings. Global Biogeochem. Cycles 17(4):1118,
doi:10.1029/2003GB002076
Kumar M, Monteith JL (1982) Remote sensing of crop growth. In H Smith (ed), Plants and daylight
spectrum. Academic Press, London, pp 133144.
Landsberg JJ, Waring RH (1997) A generalized model of forest productivity using simplified concepts of radiation use efficiency, carbon balance and partitioning. Forest Ecol. Manag. 95:
209228
Lucht W, Prentice IC, Myneni RB, Sitch S, Friedlingstein P, Cramer W, Bousquet P, Buermann W,
Smith B (2002) Climatic control of the high-latitude vegetation greening trend and Pinatubo
effect. Science 296:16871689
Matthews E (2001) Understanding the FRA 2000: Forest Briefing No. 1. World Resource Institute,
Washington, DC, p 12
McGuire AD III, Prentice IC, Ramankutty N, Reichenau T, Schloss A, Tian H, Williams LJ,
Wittenberg U (2001) Carbon balance of the terrestrial biosphere in the twentieth century: analyses of CO2 , climate and land use effects with four process-based ecosystem models. Global
Biogeochem. Cycles 15(1):183206
Mildrexler D, Zhao M, Heinsch FA, Running SW (2007) A new satellite based methodology for
continental scale disturbance detection. Ecol. Appl. 17(1):235250
Milesi C, Hashimoto H, Running SW, Nemani RR (2005) Climate variability, vegetation productivity and people at risk. Global Planet Change 47:221231
Monteith JL (1972) Solar radiation and productivity in tropical ecosystems. J. Appl. Ecol. 9:
747766
Monteith JL (1977) Climate and efficiency of crop production in Britain. Phil. Trans. Royal Soc.
Lond. B 281:277294
Mu Q, Zhao M, Heinsch FA, Liu M, Tian H, Running SW (2007) Evaluating water stress controls
on primary production in biogeochemical and remote sensing based models. J. Geophys. Res.
112,G01012
Myneni RB, Hall FG, Sellers PJ, Marshak AL (1995) The interpretation of spectral vegetation
indexes. IEEE Trans. Geosci. Remote Sens. 33:481486
Myneni RB, Nemani RR, Running SW (1997a) Estimation of global leaf area index and absorbed
par using radiative transfer models. IEEE Trans. Geosci. Remote Sens. 35:13801393
442
Myneni RB, Keeling CD, Tucker CJ, Asrar G, Nemani RR (1997b) Increase plant growth in the
Myneni RB, Dong J, Tucker CJ, Kaufmann RK, Kauppi PE, Liski J, Zhou L, Alexeyev V,
Hughes MK (2001) A large carbon sink in the woody biomass of northern forests. Proc. Natl.
Acad. Sci. USA 98(26):1478414789
Nemani RR, Running SW (1989) Estimation of regional surface resistance to evapotranspiration
from NDVI and thermal-IR AVHRR data. J. Appl. Meteorol. 28:276284
Nemani RR, Keeling CD, Hashimoto H, Jolly WM, Piper SC, Tucker CJ, Myneni RB, Running SW
(2003a) Climate-driven increases in global terrestrial net primary production from 1982 to
1999. Science 300:15601563
Nemani RR, White MA, Pierce L, Votava P, Coughlan J, Running SW (2003b) Biospheric monitoring and ecological forecasting. Earth Observ. Mag. 12(2):68
Potter CS, Randerson JT, Field CB, Matson PA, Vitousek PM, Mooney HA, Klooster SA (1993)
Terrestrial ecosystem production: a process model based on global satellite and surface data.
Global Biogeochem. Cycles 7:811841
Potter CS, Klooster S, Myneni R, Genovese V, Tan P, Kumar V (2003a) Continental scale comparisons of terrestrial carbon sinks estimated from satellite data and ecosystem modeling 198298.
Global Planet. Change 39:201213
Potter CS, Tan P, Steinbach M, Klooster S, Kumar V, Myneni R, Genovese V (2003b) Major
disturbance events in terrestrial ecosystems detected using global satellite data sets. Global
Change Biol. 9(7):10051021
Prentice IC, Heimann M, Sitch S (2000) The carbon balance of the terrestrial biosphere: ecosystem
models and atmospheric observations. Ecol. Appl. 10:15531573
Prentice IC, Farquhar GD, Fasham MJR, Goulden ML, Heimann M, Jaramillo VJ, Kheshgi HS,
Le Quere C, Scholes RJ, Wallace DWR (2001) The carbon cycle and atmospheric carbon dioxide. In: JT Houghton,Y Ding, DJ Griggs, M Noguer, PJ van der Linden, X Dai, K Maskell,
CA Johnson (eds), Climate Change 2001: The Scientific Basis. Contribution of Working Group
I to the Third Assessment Report of the Intergovernmental Panel on Climate Change. Cambridge University Press, Cambridge, United Kingdom and New York, USA, pp 182237
Prince SD (1991) A model of regional primary production for use with coarse resolution satellite
data. Int. J. Remote Sens. 12:13131330
Prince SD, Goward SN (1995) Global primary production: a remote sensing approach. J. Biogeogr.
22:815835
Rahman AF, Sims DA, Cordova VD, El-Masri BZ (2005) Potential of MODIS EVI and surface temperature for directly estimating per-pixel ecosystem C fluxes. Geophys. Res. Lett. 32,
L19404, doi:10.1029/2005GL024127
Randerson JT, Field CB, Fung IY, Tans PP (1999) Increases in early season ecosystem uptake
explain recent changes in the seasonal cycle of atmospheric CO2 at high northern latitudes.
Geophys. Res. Lett. 26:27652768, doi: 10.1029/1999GL900500
Randerson JT, van der Werf GR, Collatz GJ, Giglio L, Still CJ, Kasibhatla P, Miller JB, White JWC,
DeFries RS, Kasischke ES (2005) Fire emissions from C3 and C4 vegetation and their influence on interannual variability of atmospheric CO2 and d13 CO2 . Global Biogeochem. Cycles
19:GB2019
Reeves MC, Zhao M, Running SW (2006) Applying improved estimates of MODIS productivity
to characterize grassland vegetation dynamics. Rangeland Ecol. Manag. 59:110
Roy DP, Jin Y, Lewis PE, Justice CO (2005) Prototyping a global algorithm for systematic fireaffected area mapping using MODIS time series data. Remote Sens. Environ. 97:137162
Ruimy A, Saugier B, Dedieu G (1994) Methodology for the estimation of terrestrial net primary
production from remotely sensed data. J. Geophys. Res. 99:52635283
Ruimy A, Dedieu G, Saugier B (1996) TURC: a diagnostic model of continental gross primary
productivity and net primary productivity. Global Biogeochem. Cycles 10: 269286, doi:
10.1029/96GB00349
443
Running SW, Nemani RR (1988) Relating seasonal patterns of the AVHRR vegetation index to
simulated photosynthesis and transpiration of forests in different climates. Remote Sens. Environ. 24:347367
Running SW, Nemani RR, Peterson DL, Band LE, Potts DF, Pierce LL, Spanner MA (1989)
Mapping regional forest evapotranspiration and photosynthesis by coupling satellite data with
ecosystem simulation. Ecology 70:10901101
Running SW, Hunt ER (1993) Generalization of a forest ecosystem process model for other biomes, BIOME-BGC, and an application for global-scale models. In JR Ehleringer, CB Field
(ed), Scaling physiological processes: leaf to globe. Academic, San Diego, CA, pp 141158
Running SW, Baldocchi DD, Turner DP, Gower ST, Bakwin PS, Hibbard KA (1999) A global
terrestrial monitoring network integrating tower fluxes, flask sampling, ecosystem modeling
and EOS satellite data. Remote Sens. Environ. 70:108128
Running SW, Thornton PE, Nemani RR, Glassy JM (2000) Global terrestrial gross and net primary
productivity from the earth observing system. In O Sala, R Jackson, H Mooney (eds), Methods
in ecosystem science. Springer, New York, pp 4457
Running SW, Nemani RR, Heinsch FA, Zhao M, Reeves M, Hashimoto H (2004) A continuous
satellite-derived measure of global terrestrial primary productivity: future science and applications. Bioscience 54:547560
Running SW, Nemani RR, Townshend J, Baldocchi D (2006) Next generation terrestrial carbon
monitoring. American Geophysical Union Monography XX. A tribute to the career of Charles
David Keeling
Schimel DS, House JI, Hibbard KA, Bousquet P, Ciais P, Peylin P, Braswell BH, Apps MJ, et al.
(2001) Recent patterns and mechanisms of carbon exchange by terrestrial ecosystems. Nature
414:169172
Sellers PJ (1985) Canopy reflectance, photosynthesis and transpiration. Int. J. Remote Sens.
6:13351372
Sellers PJ (1987) Canopy reflectance, photosynthesis and transpiration. II. The role of biophysics
in the linearity of their interdependence. Remote Sens. Environ. 21:143183
Townshend JRG, Justice CO (1986) Analysis of the dynamics of African vegetation using the
normalized difference vegetation index. Int. J. Remote Sens. 7:14351445
Tucker CJ (1979) Red and photographic infrared linear combinations for monitoring vegetation.
Tucker CJ (1980) A spectral method for determining the percentage of green herbage material in
clipped samples. Remote Sens. Environ. 9:175181
Tucker CJ, Vanpraet C, Boerwinkel E, Gaston A (1983) Satellite remote sensing of total dry matter
production in the Senegalese Sahel. Remote Sens. Environ. 13:461474
Tucker CJ, Townshend JRG, Goff TE (1985) African land-cover classification using satellite data.
Science 227:369375
Tucker CJ, Fung I, Keeling C, Gammon R (1986) Relationship between atmospheric CO2 variations and satellite-derived vegetation index. Nature 319:195199
Turner DP, Ritts WD, Cohen WB, Gower ST, Running SW, Zhao M, Costa MH, Kirschbaum A,
Ham J, Saleska S, Ahl D (2006) Evaluation of MODIS NPP and GPP products across multiple
biomes. Remote Sens. Environ. 102:282292
Van der Werf GR, Randersonm JT, Collatz GJ, Giglio L, Kasibhatla P, Arellano A, Olsen S,
Kasischke ES (2004) Continental-scale partitioning of fire emissions during the 1997 to 2001
El Nino/La Nina period. Science 303:7376
Walther G-R, Post E, Convey P, Menzel A, Parmesan C, Beebee TJC, Fromentin J-M,
Guldberg OH, Bairlein F (2002) Ecological responses to recent climate change. Nature
416:389395
Wolfe RE, Nishihama M, Fleig AJ, Kuyper JA, Roy DP, Storey JC, Patt FS (2002) Achieving subpixel geolocation accuracy in support of MODIS land science. Remote Sens. Environ. 83:3149
Xiao X, Hollinger D, Aber J, Goltz M, Davidson EA, Zhang Q, Moore III B (2004) Satellite-based
modeling of gross primary production in an evergreen needleleaf forest. Remote Sens. Environ.
89:519534
444
Xiao X, Hagen S, Zhang Q, Keller M, Moore III B (2006) Detecting leaf phenology of seasonally
moist tropical forests in South America with multi-temporal MODIS images. Remote Sens.
Environ. 103:465473
Zhao M, Neilson R, Yan X, Dong W (2002) Modelling the vegetation of China under changing
climate. Acta Geographica Sinica 57(1):2838
Zhao M, Heinsch FA, Nemani RR, Running SW (2005) Improvements of the MODIS terrestrial
gross and net primary production global dataset. Remote Sens. Environ. 95:164176
Zhao M, Running SW, Nemani RR (2006a) Sensitivity of Moderate Resolution Imaging Spectroradiometer (MODIS) terrestrial primary production to the accuracy of meteorological reanalyses.
J. Geophys. Res. 111:G01002, doi:10.1029/2004JG000004
Zhao M, Running SW, Heinsh FA, Nemani RR (2006b) Variations of global terrestrial primary
production observed by Moderate Resolution Imaging Spectroradiometer (MODIS) from 2000
to 2005. (in preparation)
Zhou L, Tucker CJ, Kaufmann RK, Slayback D, Shabanov NV, Myneni RB (2001) Variations
in northern vegetation activity inferred from satellite data of vegetation index during 1981 to
1999. J. Geophys. Res. 106:2006920083
Zhuang Q, McGuire AD, Melillo JM, Clein JS, Dargaville RJ, Kicklighter DW, Myneni RB,
Dong J, Romanovsky VE, Harden J, Hobbie JE (2003) Carbon cycling in extratropical terrestrial ecosystems of the Northern Hemisphere during the 20th Century: a modeling analysis
of the influences of soil thermal dynamics. Tellus B 55:751776
Chapter 17
Applications of Terrestrial Remote Sensing

to Climate Modeling
Robert E. Dickinson
Abstract Terrestrial processes are an important component of climate. Climate can

be viewed as a nonlinear dynamical system which generates statistics to be compared with observational statistics. The surface is forced by net radiation balanced
by sensible and latent fluxes, and by precipitation balanced by evapotranspiration,
soil moisture storage, and runoff. These balances depend on detailed geographic descriptions of parameters required by the modeling. These details are constrained by
satellite remote sensing (as demonstrated by various recent studies) with consequent
substantial improvement in climate models. Some of these parameters, especially
those involving vegetation, may be evolved with the climate system.
When climate models characterize their radiative processes consistent with the
remote sensing algorithms useful for their detection, they become physically more
realistic and provide a suitable modeling framework for forward modeling data assimilation. In particular, the 3D nature of canopy radiation needs to be represented
in climate models and how this connects trees and bushes to various underlying
surfaces.
17.1 Introduction to the Formulation of Climate Models

Climate dynamical models solve various conservation equations that describe how
the climate system evolves in time, and consequently generate numerical data for its
statistical characterization. These climate statistics may include, e.g., monthly mean
temperature and humidity of near surface air, temperature of leaves and profiles of
soil temperature and water content. Additional important statistics involve diurnal
variations, annual cycles and year to year variability of such quantities. The terrestrial component of climate consists of various state variables related to the terms
Robert E. Dickinson
Georgia Institute of Technology
robted@eas.gatech.edu
445
446
R.E. Dickinson
just mentioned, linked together by process descriptions, and constrained by appropriate observational information. Statistics of the process descriptions are also often
examined as part of the study of the climatology of a model.
Exactly what state variables, process descriptions, and constraining observations
should be used is determined by the multiple purposes for which the terrestrial system is included in a climate model (e.g., for their contributions to weather prediction,
inter-annual climate prediction, projection of long term change from greenhouse
gases, study of hydrology and water resources, crop production, ecology of various
natural systems, biogeochemical cycling, especially that of carbon, as an ancillary
source of information for remote sensing studies, or for forward modeling of remote
sensing radiances).
The primary role of the terrestrial system as an interactive component of a climate model is to determine the near surface and surface variables that are used to
characterize climate observationally and to control the dynamic and thermodynamic
structure of the atmosphere as mentioned above. Major ingredients of this control
are the surface exchanges of energy, moisture, and momentum.
Land surface models have evolved from quite simple treatments of these exchanges to increasingly complex descriptions. For earlier reviews, cf. Dickinson
(1983, 1984, 1989, 1992, 1995a, b), Dickinson et al. (1991), Sellers et al. (1997),
Pitman (2003), Yang (2003). Land is most simply treated as a single reservoir for
whatever quantities it is supposed to exchange with the atmosphere. Earliest climate models also averaged over the diurnal and annual cycles so that climate was
a steady state system. The land component of such early models required a statement as to what fraction of the incident solar radiation it would absorb; that is, its
albedo. It also required a specification of heat capacity and water holding capacity
and a rule relating its evaporation to that which would come from a wet surface,
depending on how much water was stored.
Later models have become much more complex. Dai et al. (2003) describes one
example of current models. Such models include descriptions of the global distribution of vegetation and various mechanisms changing with model time-step by
which vegetation returns water to the atmosphere. They may also describe overall
energy and water balances and fluxes from lakes, wetlands, glaciers; river-flows are
determined as an input to ocean dynamics. Rather than describe the details of any
particular such approach, we frame the problem somewhat abstractly. All appropriate approaches to modeling the land surface in a climate model, although they may
look quite different, are implementations of essentially the same ideas.
17.2 What Does the Land Surface of a Climate Model Consist of?
The terrestrial surface in a climate model consists of either soils covered by vegetation or bare soil. It includes wetlands and lakes and a description of land elevations
at the resolution of the atmospheric model. An initial question might be: what soils
and what vegetation should the model have. This question could be responded to by
17 Applications of Terrestrial Remote Sensing to Climate Modeling
447
the various descriptive names that observed such vegetation is given: e.g., a given
grid-square in a climate model may be covered by an open forest with an understory
of grass, overlying a soil classified as an Alfisol. Such names, alone, are useless
for quantitative physical modeling. However, they may be used through their correlation with needed properties to infer what is needed. One application of terrestrial
remote sensing has been to extrapolate over wider areas names that have been given
to particular kinds of vegetation at the surface. This assignment of a geographic distribution of land cover is useful for other purposes but may be losing some of the
information needed by climate modelers. If so, this is regrettable since much of the
quantitative information measured by a satellite sensor, especially such as provided
by reflectance imagery, is closely connected to the information directly needed by
climate modelers.
Initial geographic characterization of land cover was provided by Wilson (1984),
Matthews (1984). More recent land cover data sets have come from AVHRR (e.g.,
Strugnell et al., 2001) and now MODIS (Friedl et al., 2002).
17.3 Water and Energy Balance Requirements

Energy exchange at the terrestrial surface is dictated by its net absorption of solar energy. The solar flux incident at the top of the atmosphere is determined by location,
time of day, and time of year. In passing through the atmosphere, some radiation
is absorbed by atmospheric molecular constituents and some scattered, i.e., Raleigh
scattering by molecules, and Mie scattering by aerosols and cloud droplets. Overall,
on average, about 20% is absorbed and about the same lost by upward scattering.
More than half the incoming solar radiation remains to reach the terrestrial surface.
This radiation incident at the surface is approximately half downward scattered from
the atmosphere (much less on clear days and nearly 100% on cloudy days). It can in
the short term be stored, but on longer time scales must be balanced by net upward
long-wave thermal radiation and turbulent fluxes carrying internal energy from atmospheric temperature (sensible fluxes), and the energy stored in evaporated water
(latent fluxes).
Different surfaces act in different ways to achieve a balance between their absorption of solar radiation and the thermal radiation and turbulent flux exchanges
that collectively provide this balance. Because the net thermal emission and storage
variation (e.g., thermal conduction into soil) are relatively regular, during the daytime these are commonly lumped for diagnostic purposes with the absorbed solar
radiation to get a net forcing of the sensible and latent fluxes. As by construction,
this net radiation plus storage must balance the turbulent fluxes, the primary issue is
then what determines the relative amount of each. This partitioning between the turbulent flux terms is commonly characterized by the ratio of sensible to latent fluxes,
referred to as the Bowen ratio.
According to the above generalities, state variables should be connected to components of the system that store water and energy and change in time as this storage
448
R.E. Dickinson
is modified. Additional diagnostic variables are used to relate the state variables
to fluxes. The amount of energy stored in a given material is proportional to its
temperature with this proportionality factor, a heat capacity. Furthermore, sensible
energy exchanges are driven by the differences in temperature between two reservoirs and moisture exchange by some difference in moisture potentials. In sum, the
temperatures and moisture of various system elements must be distinguished. Models address how these elements are heated by the sun and how they lose energy to
each other or to the atmosphere.
The most significant contributors to these elements to be modeled are the vegetation, surface water in the form of snow or free water (e.g., lakes), and soil, where
soil refers to mineral or dead organic matter stratified vertically. Vegetation generally consists of individual surfaces such as leaves or branches with spatial scales
from a few mm to meters. These surfaces have an area in contact with air that is
large compared to the flat area of underlying soil (thus, acting in some ways as a
porous medium, but on a relatively large spatial scale). They respond to solar heating by increasing their temperatures until their heating is balanced by turbulent and
convective air motions carrying thermal energy and water. Soils and snow are also
porous but their spatial scales are only on a nanometer to micrometer scale so that
water moves through them as a very viscous fluid and thermal energy by conduction. Soil by volume is about half mineral or organic materials, and the rest some
combination of air and water. Snow consists of crystalline ice with quite a bit of air.
In simple terms, the modeled atmosphere delivers solar radiation and precipitation to the modeled terrestrial surface. How this energy and water is returned to the
atmosphere depends on quite a few modeling details. From the climate viewpoint,
the details of the return fluxes in amount and timing can be important for temperatures and precipitation. Precipitation may either be lost at nearly the same time it
falls through evaporation from canopy (referred to as interception loss) or can infiltrate the soil and be extracted weeks later as part of the transpiration fluxes (e.g.,
Dickinson et al., 2003). Solar heating of springtime boreal vegetation may either return directly to the atmosphere or be used to melt the snow pack. The latter is often
invoked as a vegetation feedback on snow melting. More concrete descriptions of
particular processes will be addressed in the following discussion of current issues
that emerge in the context of remote sensing data requirements.
17.4 Role of Solar Radiation in the Climate Model Terrestrial

System
Climate models determine the solar radiation absorbed by the various components of
the terrestrial system to determine separate energy and water balances. For example,
the solar radiation incident at the surface may heat a canopy, its understory, and the
underlying soil. If all these terms are lumped together into a single complex surface,
the solar radiation absorbed is simply the complement of (one minus the) albedo,
which is quantified by remote sensing data products such as from MODIS (e.g., Gao
449
et al., 2005). However, surfaces treated in more detail will determine total solar
absorbed as the sum over the absorption by individual terms, and hence observed
albedo becomes a constraint rather than a simple boundary condition.
Climate models use leaves, green or brown, and the underlying woody stems to
determine the canopy contribution to absorption of solar radiation. This absorbed
solar radiation raises the vapor pressure of water inside green leaves to higher values than that of the air outside the leaves and consequently forces the leaves to
transpire. Since the storage of water inside leaves is commonly small, this water
loss must be compensated by extraction of water stored in the soil by plant roots.
The contribution of leaves is quantified in the model in terms of the leaf-area-index
(LAI). This quantity is the one-sided area of leaves (flattened if necessary, to be a
spatial projection) per area of soil surface. MODIS (Myneni et al., 2002) provides
remote sensing estimates of the global distribution of leaves. This term has been determined over periods of 8 days from the beginning of the year 2000 to the present.
It is accompanied by a product FPAR which provides that fraction of the absorbed
visible solar radiation that is taken up by the canopy. The characterization of dead
leaves and stems is more problematic as not yet supported by remote sensing data.
The most common approach to address sub-grid variability in vegetation is to
portion the climate model grid square into various sub-grid tiles. Some fraction of
the grid square must be taken as vegetated, and can be subdivided in terms of plant
functional types (pfts) (Bonan and Levis, 2002). For example, a savanna might be
covered 80% by grass and 20% by trees, and these are put on separate areas and their
radiative exchange determined by one-dimensional models. The estimated area of
bare surfaces (i.e., whatever is not directly under a tree or grass canopy) is similarly
treated as a separate tile. The area directly under canopies has been treated as bare
soil. Such models assume that trees and grass are in separate clumps; consequently,
if there are trees overlying grass or soil, the light environment is treated very unrealistically in terms of the effects of the 3D shading of the grass or soil by the trees.
For open canopies, this treatment can underestimate the visible radiation absorbed by the canopy by at least a factor of 2. However, many canopies are closed
enough, and reflections from the underlying surface sufficiently compensate, that
much smaller errors are seen in comparing the modeled with observed fraction of
absorbed visible radiation.
Models of transpiration and photosynthesis have a nonlinear dependence on the
intensity of incident light. That is, their dependence on light is linear at low light
levels but asymptotes to some constant value (saturation) at high light levels. Direct
sunlight either strikes a leaf and is greatly attenuated or does not hit any leaves and
continues at the same intensity it started with. Thus, the appropriate statistic for the
effect of direct solar radiation on photosynthesis is not average light intensity but
rather what is the relative area of leaves that receive the direct sun. Diffuse light, on
the other hand is idealized as coming from all sky directions with the same intensity
(cf. Pinty et al., 2005, for a relaxation of this assumption.) and consequently strikes
all leaf surfaces with the average intensity of the diffuse light at a given level in the
canopy. Climate models commonly use an average value of the diffuse radiation,
an assumption which may lead to a factor of 2 overestimation of the average shade
450
R.E. Dickinson
leaf intensity compared to that from a better scaling (e.g., Dai et al., 2004) and
further serious error if light intensities are too strong to be in the light limited regime
of transpiration/photosynthesis. Such may happen, e.g., for a midday summer sun,
passing through a thin cloud or thick aerosol layer.
Light of average intensity locally attenuates by Beers Law either in fractional
area of sunlit leaves (direct sun) or its intensity (diffuse radiation) and thus according to
dI
+GI = S
(17.1)
dL
where I is the incident light intensity, L is path length measured in leaf area increment, S is an internal source term, accounting for scattering if needed, and G is
a factor for the projection of leaves into the direction of the light (commonly assumed to be a factor of 0.5 approximating the effect of a uniform distribution of
leaf orientations). For a canopy idealized as a uniform spatial distribution of leaves,
Eq. (17.1) generalizes to a global expression and the path length is the vertical path
length divided by a cosine projection factor. However, leaves, especially those in
forest canopies, are far from uniformly distributed. The most obvious spatial heterogeneities that must be addressed are the organization of plant canopies as discrete
objects, organization of leaves into clusters and the variety of path lengths (i.e.,
number of leaves) that a light ray must pass by.
Although this issue of canopy heterogeneity has been recognized by climate
modelers for a long time (e.g., Dickinson, 1983), practical approaches to address
it are only now being developed (e.g., Pinty et al., 2006). Ideally, this part of a climate model should not be computationally much more intensive than the evaluation
of few exponentials (as in analytic two-stream solutions), thus probably precluding
the use of fine layering approaches with multi-scatter iteration or even the direct inversion of a large matrix as might arise out of various idealizations of the situation.
The physically most realistic and complete description using Monte Carlo is slow
by factors of at least thousands compared to what is needed but can be very useful
for validating approaches of low computational cost (e.g., Pinty et al., 2006).
Where climate model treatments of canopy radiation are weakest, they have also
been most limited by lack of data. These are the situations where sparse canopies
have large openings through which light can directly reach the underlying surface.
Thus, however bad the treatments of radiation by climate models may appear from a
conceptual viewpoint, they have done the required job within the context of the very
little available information. With the archiving of several years of MODIS data, the
time has arrived to evaluate and improve the treatments of radiation used in climate
models.
Jin et al. (2002) reported a very large range of albedos over snow covered
areas depending on the masking by vegetation of the underlying snow surface.
Figure 17.1, from Gao et al. (2005) compares observed albedos (visible light) for
evergreen forest and grassland covers and for varying degrees of snow cover. It illustrates the increase of albedo with increasing snow but with a much greater darkening
of the forested region and less sensitivity to snow. It also shows that an extremely
large spatial variability of albedo especially with the largest snow covers (factor of 2
451
Fig. 17.1 Shows how albedos vary with snow cover fractions (Gao et al., 2005). Points in the figure
were extracted from global high-quality retrievals during the whole year of 2001 (23 production
periods). Each point in the figure represents a CMG resolution (0.05 ) pixel. Grassland albedos
are seen to be four times as sensitive to snow fractions as that of an evergreen forest
or more). The masking of snow by forests has been included in at least some climate
models since the early 1980s but not with the correct geometric considerations as
discussed above. Furthermore, climate modelers have had little basis for incorporating in their global models from first principles or surface observation descriptions of
how open canopies might be. Prior remote sensing data such as from AVHRR has
been of limited value in developing climate model radiation schemes that adequately
address such details.
Gao et al. (2005) show that the albedos measured by MODIS have the spatial and
temporal patterns appropriate to their underlying land cover classes. The seasonal
variations of these albedos are consistent with the phenology of leaf cover (e.g.,
outside the growing season, deciduous broadleaf forests become much darker in the
near-infrared and a bit brighter in the visible). The albedos over areas of sparser
vegetation also show influences of underlying soil, e.g., grasslands in the band of
1020 N and presumably mostly in North Africa, show a much higher albedo than
elsewhere.
Tsvetsinskaya et al. (2002) has shown in the context of North Africa how measured MODIS albedos can be applied to characterize the albedos of an arid region.
They find over this region, considerable spatial variability in its surface albedos,
apparently related to soil mineral composition and geographical characteristics. For
example, MODIS shortwave albedos vary by a factor of about 2.5 from the darkest
volcanic terrains to the brightest sand over the Sahara. Overall, the Sahara desert
has much higher albedos than deserts elsewhere. Climate models have previously
assumed a single albedo for desert.
452
R.E. Dickinson
17.5 Climate Model Evaluation Studies

Recent studies have begun to evaluate in the context of the new satellite data the
performance of the treatments of radiation in climate models and their various ancillary assumptions such as the leaf area used to produce their radiation. Two studies
have evaluated the climate model described by Zeng et al. (2002). This model utilized land surface data developed from AVHRR (e.g., Zeng et al., 2000; Strugnell
et al., 2001) and implemented in the CLM model as described by Dai et al. (2003).
Zhou et al. (2003) found that largest discrepancies between observed and modeled
albedos occur over snow covered regions and in arid regions. They emphasized a
lowering of winter albedos compared to that observed by MODIS, resulting from
an apparent excess of tree and bush stem areas included in the model. Tian et al.
(2004a) compared the LAI assumed in the model with that observed by MODIS and
the contributions of differences between model verus observed LAI to model error.
They also emphasized that whereas the Zeng et al. (2002) climate model may have
overstated the absorption of light by the canopy stems, that this absorption should
still be included but with more appropriate optical properties and that the MODIS
algorithms for LAI and FPAR have left out their effect entirely.
Zhang et al. (2005a) have further demonstrated that a significant fraction of solar
radiation is absorbed by non-photosynthesizing canopy surfaces.
Other studies have evaluated the climate model version described by Bonan et al.
(2002). It differs in its modeling of the land surface from that used by Zeng et al.
(2002) in replacing some aspects of the previous version of CLM (common land
model) with the canopy radiation description developed by Bonan (1996), and with
land data sets developed by Bonan and Levis (2002), also from AVHRR. Although
the AVHRR satellite data used by Zeng et al. (2002) may have been developed more
thoroughly than that of Bonan and Levis (2002), it was expressed in terms of the
IGBP classification and Bonan et al. (2002) required the land data to be formulated
in terms of plant functional type tiles (pfts) as described in Bonan and Levis (2002).
In after-sight, the treatment of canopy radiation by Zeng et al. (2002) suffered some
numerical flaws (e.g., as mentioned by Pinty et al., 2006), and other inadequacies
so its replacement by the full two-stream computation of Bonan (1996), essentially
the same as used in the Sib model (Dickinson, 1983; Sellers, 1985), was likely an
improvement. To make a long story short, the CLM took two steps forward and
one step back as a result of different individual ideas as to what was currently the
best way to do canopy radiation and derive land boundary conditions from AVHRR.
This CLM2 model (community land model version 2) described by Bonan et al.
(2002) was assessed by Oleson et al. (2003), Wang et al. (2004), in the context of
its simulation of albedos. They reported a large negative bias in the albedos over
the Sahara in contrast to the large positive bias reported by Zhou et al. (2003).
This negative bias was a consequence of the Sahara albedo constrained to fit data
from the ERBE instrument, whereas the positive bias was a consequence of data
constrained to reproduce albedos inferred from AVHRR. In sum, prior to MODIS,
previous less reliable satellite estimates of the albedo of the Sahara may have provided no real information beyond that already available to climate modelers from
453
surface and aircraft observations. For the latter, soils, sands, and rock had only been
distinguished by models in terms of their soil hydrological properties and albedo
inferred from an original; light, medium and dark color classification. The
ratio of near-infrared to visible albedo ratio has been fixed at 2 and any dependence
on solar zenith angle neglected. Much more detailed spatial and spectral information, especially over semi-arid regions is now available from MODIS. For example,
the ratio of near-infrared to visible albedos observed in MODIS over the deserts
of Sahara, Taklimakan, and Australia varies from 1.6 to 2.7 (Zhou et al., 2003) and
albedo in the Sahara has been shown to increase significantly with solar zenith angle
(Wang et al., 2005). Zhou et al. (2005) has proposed a method using principal component analysis to economically represent a high quality soil albedo dataset over
non-vegetated North Africa. An extension of their approach can be used to separate
the MODIS albedos into soil and canopy contributions.
Oleson et al. (2003) (as Zhou et al., 2003) reported that the land surface model,
CLM2, overall reasonably simulated vegetation albedos but that the two-stream
treatment of radiation appeared to have too strong a dependence on solar zenith
angle so that the CLM black-sky albedo with sun at local noon was in better agreement with MODIS than the white sky diffuse radiation. Wang et al. (2004) further
documented the substantial differences between model and observed dependences
on solar zenith angle. Oleson et al. (2003) also emphasized an overestimate of albedos by the model over snow covered regions in contrast to the emphasis in Zhou
et al. (2003) of an apparent underestimate. Evidently, it is easy for climate models
to make large errors in regions of snow cover and the shading/masking effects of
the vegetation can be either under or over emphasized in different approaches to the
canopy radiation that are too oversimplified and unconstrained by observations to
be expected to give the right answer.
Tian et al. (2004b) illustrate the usefulness of new land surface datasets developed from MODIS as model boundary conditions. They related LAI, vegetation
continuous field, and the land cover maps to the pft formulation of Bonan and Levis
(2002), using data collection 4 for the period of September, 2000August, 2002.
This new dataset showed large differences from the old dataset of Bonan and Levis
(2002). The new LAI was larger than the older version by at least 1.5 over the
Amazon, central Africa, southeastern Asia, and north Europe, and by about 0.51.0
over most areas beyond 60 N in both seasons (Fig. 17.2). These increases in LAI
are likely a result of the AVHRR inversion saturating at a lower LAI than that of
MODIS. The new LAI was also found to be smaller over many regions (Fig. 17.2).
Their use of MODIS data decreased the amount of crops and grass by 2040% globally, and increased the bare category by a large amount (Fig. 17.3). The previous
version of the data assumed: that non-tree covered land in forest, savannas, and
grasslands was covered by grasses, in shrub lands by shrubs, in croplands by crops
(Bonan and Levis, 2002). In semi-arid regions, there was no information (Bonan
and Levis, 2002) as to the fraction of area covered by bushes so leaves were spread
uniformly. This assumption has been found to be incompatible with the existing parameterization of under canopy energy fluxes and a revised approach for the latter
was implemented (Zeng et al., 2005).
454
R.E. Dickinson
Fig. 17.2 Spatial pattern of LAI difference between the old and new land surface datasets (newold as derived respectively from MODIS and AVHRR LAI products) in winter (DJF) and summer
(JJA) (Tian et al., 2004c)
In the newer MODIS data for fractional tree cover (Hansen et al., 2003) fractional
regions are characterized either as trees or shrubs or bare. Apparently, the bare
category in the continuous field data refers to all understory components, but for lack
of further information, Tian et al. (2004b) used literally the bare classification.
Presumably, in semi-arid systems, understories are mostly bare soil, but by contrast,
in moist forests they should mostly be dead leaves, mosses, or herbaceous small
plants.
Tian et al. (2004b, 2004c) used CLM2 coupled with the Community Atmospheric
Model (CAM2) to investigate how the modeled surface variables such as temperature and albedo were modified by the new dataset. For snow-free regions, the
increased LAI and changes in the percent cover from grass/crop to tree or shrub
decreased albedo, but also decreased surface air temperatures. Increases of canopy
evapotranspiration and decreases of ground evaporation over tropical regions improved the modeled surface temperature. MODIS albedo data can be used to adjust
model parameters controlling absorption of solar radiation to provide a better fit to
the albedo observations (e.g., Liang et al., 2005).
455
Fig. 17.3 Spatial pattern of percent cover difference (new-old) in grass/crop, tree, shrub, and bare
soil at the model spatial resolution as inferred from MODIS land cover classification versus use of
AVHRR (Tian et al., 2004c)
17.6 How can Terrestrial Remote Sensing Best Support Climate

Models?
Observational characterizations of the terrestrial surface may advance climate models more by providing information that climate models should be using rather than
that required by the current formulation of the climate models. For example, we
have discussed in Section 17.4, that climate models have oversimplified treatments
of the interaction of solar radiation with the terrestrial surface. What observational
information could support a more correct treatment of solar radiation? Observations
456
R.E. Dickinson
should characterize, to the extent possible, the 3D structure of the surface. A description of fractional tree cover (e.g., Hansen et al., 2003) is a first step in that
direction. Besides the fraction of tree cover, other tree statistics needed for climate
model details are their size, i.e., height and aspect ratio, and a characterization of
their average distance to nearest neighbors (e.g., Widlowski et al., 2001, 2004). In
addition, a climate model needs all the smaller scale parameters that contribute to the
characterization of surface shading, in particular the relative areas of leaf and stems.
The surfaces beneath a canopy that are affected by the canopy radiatively must
be connected to the description of the canopy. Thus, the fractional areas of other
surfaces besides trees should be established at an appropriate level of detail, in particular, the fraction that is radiatively connected versus unconnected and the composition of the underlying surface. The concept of radiative connectivity is somewhat
vague, but it can be quantified by some simple rules: e.g., it might include all underlying surfaces within a horizontal distance 3 H from the canopy, where H is
the crown height. A description of the underlying surfaces so connected can be relatively simple. What are the characteristics of the dominant material shading the
mineral soil, what fraction is photosynthesizing, and what fraction is open to the
underlying soil?
As discussed in the preceding sections, understory has been variously characterized as grass or bare. Either choice of surface may lead to erroneous results in
absorption of radiation when, as now done, its shading by the overlying canopy is
neglected except for surfaces directly under the canopy (which are currently always
taken to be bare soil). Other potentially important under-canopy-cover can be, e.g.,
dead leaves, moss, lichen, or wetland. In general, if fc denotes the fractional cover
of trees, then the sum of all understory components should be 1, not fc as currently
assumed by climate modelers. The difference between surfaces shaded for overhead
sun and those shaded at other times of day is not discontinuous in nature and should
not be so modeled.
17.7 Terrestrial Remote Sensing as a Component of Climate

Prediction
In the climate context, the use of information from terrestrial remote sensing is still
relatively immature. The initial MODIS products involve individual estimates of
parameters at a given pixel and composited over enough scene views to obtain a
variation of view angles. Views are rejected for cloud contamination and quality
flags are set. What is still lacking is utilization of the expected strong space and
time correlation of the data. A pixel with, e.g., a pine forest, will have neighboring
pixels also classified as pine forest and for the next several years at least, they will
mostly remain pine forests. An observation of neighboring pixels can be viewed
as all having the same expected value plus a random noise element. There will be
real differences in their albedo, LAI, etc. The random noise characterizations simply
mean that these differences are too small and irrelevant to be of interest in detail.
457
Likewise, pixels should look the same from year to year except for random changes.
Such changes could be as large as those caused by forest fires, which will certainly
be of great interest for some issues, but may not be significant for, e.g., a climate
models albedo over a continent.
The state variables computed by a climate model are necessarily continuous in
time and space whereas the pixel by pixel remote sensing products have many data
holes caused by clouds, or more rarely by instrument failures. These holes must
be filled before statistics can be obtained appropriate for use by climate modelers.
Ideally, this filling should be done with no loss of real information. Moody et al.
(2005) reports an approach to such filling for albedos measured by MODIS.
What might be the optimum characterization of, e.g., pine forest albedos in July?
First, what additional parameters besides the month and land cover description do
we expect these albedos to depend on? Probably latitude, and fraction of tree cover
need to be considered. After these correlations are quantified, there may still be significant variability that can be ascribed to characteristics of different understories. In
principle, any significant variability can be inverted from its cause, provided the radiative model used is realistic enough to include this source of variability. Whatever
remains is uninterpreted noise. Characterizing the amplitude of this residual noise
is an important aspect of the data analysis.
The state variables of vegetation change in time as a consequence of their
interactions with climate. Various approaches have been developed to model this
dynamical system. Ecosystems change their structure as a response to competition between different plant functional types (e.g., Lu et al., 2001; Sitch et al.,
2003; Bonan et al., 2003; Woodward and Lomas, 2004; Krinner et al., 2005). These
models have generated their leaves according to prescribed phenologies based on
accumulated degree days, apparently by itself not on adequate constraint (e.g.,
Arora and Boer, 2005). The onset of leaves in Northern forests appears to have
as much correlation with mid-summer temperature as degree days (Jenkins et al.,
2002). In semi-arid or tropical systems, the phenology is largely controlled by the
onset of rainy and dry seasons. Their dependences are shown in Fig. 17.4 (Zhang
et al., 2005b). Detailed day by day dynamics of leaf growth can be attempted (e.g.,
Dickinson et al., 1992; 2002), but the underlying principles for this may be inadequately understood. Including dynamic growth of roots is another difficult modeling
issue (e.g., Arora and Boer, 2003).
The terrestrial state variable X moves forward in time, formally as a multivariate
differential equation:
dX
+ F(X, ) = Q( )
(17.2)
dt
where are various fixed parameters, and Q is a forcing term depending on other
parameters . A satellite measures various reflectance quantities, denoted Yo , which
consists of the real Y plus a measurement error term. The model can use its model
value of X = Xm to provide an estimate of Y = Ym . For example, X can be some combination of the model soil moisture, or LAI or fc ; the latter two would be changed
by dynamic vegetation versions of the model.
458
R.E. Dickinson
210
Date (DOY)
180
150
Forests
Shrublands
Savannas
Grasslands
120
90
60
30
5
11
Latitude (N)
13
15
17
210
Date (DOY)
180
150
2001
120
2002
90
2003
60
30
5
11
Latitude (N)
30
25
20
15
10
5
0
15
25
Sahelian and subSahelian region
Southern Africa
Frequency distribution (%)
Frequency distribution (%)
35
13
80 60 40 20
20
40
60
Sahelian and subSahelian region

Southern Africa
20
15
10
5
0
80
Differnce in greenup onset and ORS (days)
17
20
40
60
80 100 120 140 160
Differnce in domancy onset and ORS (days)
Fig. 17.4 The phenology of vegetation over Africa as time of greenup from beginning of year
(a, b) and (c, d) its timing relative to the onset (o) and end (e) of the rainy season (Zhang et al.,
2005)
Assume
Ym = g(Xm )
(17.3)
The term Ym will differ from Y either because Xm differs from the real X or from
structural error in use of g. We can require that X and Y be adjusted to be as close
as possible to their real values in an rms sense, i.e., find an optimum estimate X as
the minimum value of
2
2

(17.4)
J = X X + Y gX .
459
is close enough
where all parameters have been normalized by error estimates. If X
to Xm , Eq. (17.4) can be linearized and solved by matrix computations. In this case,
it reverts to linear least squares fitting, as appropriate for a statistical model with
Gaussian error terms. The standard technologies of data assimilation in atmospheric
models are summarized in Kalnay (2003).
Detailed implementation should make use of known spatial correlations and error
characterizations. Some aspects of this issue have already been addressed by hydrologists in the context of the fusion of microwave data and soil moisture modeling
(e.g., Entekhabi et al., 2004; Reichle and Koster, 2005). For statistics that change a
lot in time, i.e., depend on X, it has been found useful to do ensemble integrations of
the system Eq. (17.2). A future target for such a data assimilation approach will be
terrestrial carbon (e.g., Hese et al., 2005). Various satellites are under development
to measure the variability of atmospheric CO2 with enough accuracy to infer terrestrial and oceanic sources and sinks. Such estimation can be substantially improved
with the inclusion of these measurements in atmospheric data assimilation schemes
along with terrestrial and oceanic process modeling.
Assimilation of terrestrial reflectance imaging into models for the time evolution
of canopy structure and LAI will become another major component of this activity.
The implementation of such assimilation will require more advanced algorithmic
approaches to canopy radiation so that the climate model simulations are realistic
enough to reproduce the reflectances seen by remote sensing data.
17.8 Conclusions
Climate models have become increasingly realistic in their descriptions of land
surface processes. They absorb solar radiation and emit long wave radiation
depending on structural details such as arrangement of leaves and soil moisture.
Remote sensing instruments measure the same radiation as reproduced by climate
models. Various aspects of vegetation change in time in response to variations in
climate. Climate models have begun to include such changes in terms of dynamic
vegetation. However, the climate models still use much less realistic treatments of
terrestrial radiation than have been achieved by the remote sensing community.
By simplification and improved efficiencies, the treatments of radiation used in
remote sensing can be adopted into climate models. With such, many terrestrial
remote sensing products should be derivable through forward calculations in data assimilating climate models. The logic, as used today, for atmospheric observations is
that the climate model provides a first estimate, which in principle has incorporated
in it all past observational information. Current observations then provide a correction to this a priori estimate. The resulting optimal estimate is essentially a weighted
average of two estimates, where the weighting is determined by an understanding
of correlated error statistics. Although assimilating terrestrial data should involve
these general principles, the details of approaches needed will be vastly different
460
R.E. Dickinson
because of very different time and space scales of the dynamics and observations.
In particular, vegetation is only spatially coupled on climate scales, has very fine
spatial sampling structure, and its temporal sampling by satellite is sparse.
Acknowledgements Yuhong Tian and Liming Zhou are thanked for their suggestions on the
manuscript, including some content, Qing Liu for editing and Janet McGraw for her help
with the manuscript. Also thanks to an anonymous reviewer. Support has been provided by
the authors NASA Grant NNG04GO61G; NASA Grant NG04GK87G, and S.L. Liangs NASA
Grant NNG04GL85G.
References
Arora VK, Boer GJ (2003) A representation of variable root distribution in dynamic vegetation
models. Earth Interactions 7(6): 119
Arora VK, Boer GJ (2005) A parameterization of leaf phenology for the terrestrial ecosystem
component of climate models. Global Change Biol. 11(1): 3959
Bonan GB, Levis S (2002) Landscapes as patches of plant functional types: an integrating concept for climate and ecosystem models. Global Biogeochem. Cycles, 16 (2),
doi:10.1029/2000GB001360
Bonan GB, Oleson KW, Vertenstein M, Levis S, Zeng X, Dai Y, Dickinson RE, Yang, Z-L (2002)
The land surface climatology of the community land model coupled to the NCAR community
climate model. J. Climate 15: 31233149
Bonan GB, Levis S, Sitch S, Vertenstein M, Oleson KW (2003) A dynamic global vegetation model
for use with climate models: concepts and description of simulated vegetation dynamics. Global
Change Biol. 9:15431566
Dai Y, Zeng X, Dickinson RE, Baker I, Bonan GB, Bosilovich MG, Denning AS, Dirmeyer PA,
Houser PR, Niu G-Y, Oleson KW, Schlosser CA, Yang Z-L (2003) The common land model
(CLM). Bull. Amer. Meteor. Soc. 84(8):10131023
Dai Y, Dickinson RE, Wang Y-P (2004) A two-big-leaf model for canopy temperature, photosynthesis and stomatal conductance. J. Clim. 17(12):22812299
Dickinson RE (1983) Land surface processes and climate Surface albedos and energy balance.
In: B Saltzman (ed) Theory of climate, advances in geophysics. Academic Press, New York,
vol 25, pp 305353
Dickinson RE (1984) Modeling evapotranspiration for three-dimensional global climate models.
Climate Processes and Climate Sensitivity. JE Hansen, K Takahashi (eds), American Geophysical Union Geophysical Monograph 29, Maurice Ewing, vol 5, pp 5872
Dickinson RE (1989) Water and energy exchange. Chapter 6, Remote Sensing of Biosphere Functioning. R. Hobbs, H. Mooney (eds), Springer, New York, pp 105133
Dickinson RE, Henderson-Sellers A, Rosenzweig C, Sellers PJ (1991) Evapotranspiration models
with canopy resistance for use in climate models, a review, Agric. Forest Meteor. 54:373388
Dickinson RE (1992) Land surfaces. Chapter 5, Climate Systems Modeling. K. Trenberth (ed),
Cambridge University Press, Cambridge, pp 149171
Dickinson RE (1995a) Satellite systems and Models for Future Climate Change. Chapter 7, World
Survey of Climatology, 16, Future Climates of the World. A. Henderson-Sellers (ed), Elsevier,
Amsterdam, pp 245280
Dickinson RE (1995b) Land processes in climate models, International Satellite Land Surface
Climatology Project special issue of Remote Sensing of Environment 51:2738
Dickinson RE, Shaikh M, Bryant R, Graumlich L (1998) Interactive canopies for a climate model.
J. Clim. 11:28232836
461
Dickinson RE, Berry JA, Bonan GB, Collatz GJ, Field CB, Fung IY, Goulden M, Hoffman WA,
Jackson RB, Myneni R, Sellers PJ, Shaikh M (2002) Nitrogen controls on climate model evapotranspiration. J Clim. 15 (3):278295
Dickinson RE, Wang G, Zeng X, Zeng Q-C (2003) How does the partitioning of evapotranspiration
and runoff between different processes affect the variability and predictability of soil moisture
and precipitation? Adv. Atmos. Sci. 20(3):475478
Dickinson RE, Oleson KW, Bonan GB, Hoffman F, Thornton P, Vertenstein M, Yang Z-L, Zeng X
(2006) The community land model and its climate statistics as a component of the community
climate system model. J. Clim. 19:23022324
Entekhabi D, Njoku EG, Houser P, Spencer M, Doiron T, Kim Y, Smith J, Girard R, Belair S,
Crow W, Jackson TJ, Kerr YH, Kimball JS, Koster R, McDonald KC, ONeill PE, Pultz T,
Running SW, Shi J, Wood E, Van Zyl J (2004) The Hydrosphere state (Hydros) satellite mission: an earth system pathfinder for global mapping of soil moisture and land freeze/Thaw.
IEEE Trans. Geosci. Remote Sens. 42 (10):21842195
Friedl M, McIver D, Hodges J, Zhang X, Muchoney D, Strahler A, Woodcock C, Gopal S,
Schneider A, Cooper A, Baccini A, Gao F, Schaaf C (2002) Global land cover mapping from
MODIS: Algorithms and early results. Remote Sens. Environ. 83:287302
Gao F, Schaaf CB, Strahler AH, Roesch A, Lucht W, Dickinson R (2005) MODIS bidirectional
reflectance distribution function and albedo Climate Modeling Grid products and the variability
of albedo for major global vegetation types. JGR 110 (D01104l), doi:l0.1029/2004-JD005190
Hansen MC, DeFries RS, Townsend JRG, Carroll M, Dimicelli C, Sohlberg RA (2003) Global
percent tree cover at a spatial resolution of 500 metres: first results of the MODIS vegetation
continuous fields algorithm. Earth Interactions 7, paper no. 10
Hese S, Lucht W, Schmullius C, Barnsley M, Dubayah R, Knorr D, Neuman K, Reidel T,
Schroter K (2005) Global biomass mapping for an improved understanding of the CO2 balance the Earth observation mission Carbon-3D. Remote Sens. Environ. 94:94104
Jenkins JP, Braswell BH, Frolking SE, Aber JD (2002) Detecting and predicting spatial and interannual patterns of temperate forest springtime phenology in the eastern US. Geophys. Res.
Lett. 29(24): 5456, doi:10.1029/2001GL014008
Jin Y, Schaaf CB, Gao F, Li X, Strahler AH, Zeng X, Dickinson RE (2002) How does snow impact
the albedo of vegetated land surface as analyzed with MODIS data. Geophys. Res. Lett. 29,
doi:10.1029/2001GL014132
Kalnay E (2003) Atmospheric modeling, data assimilation and predictability. Cambridge University Press, Cambridge, 341pp
Krinner G, Viovy N, de Noblet-Ducoudre N, Ogee J, Polcher J, Friedlingstein P, Ciais P, Sitch S,
Prentice IC (2005) A dynamic global vegetation model for studies of the coupled atmospherebiosphere system. Global Biochem. Cycles 19 (GB1015), doi:10.1029/2003GB002199
Liang X-Z, Xu M, Gao W, Kunkel K, Slusser J, Dai Y, Min Q, Houser PR, Rodell M,
Schaaf CB, Gao F (2005) Development of land surface albedo parameterisation based on Moderate Resolution Imaging Spectroradiometer (MODIS) data. J. Geophys. Res. 110 (D11107),
doi:10.1029/2004JD005579
Lu L, Pielke RA Sr, Liston GE, Parton WJ, Ojima D, Hartman M (2001) Implementation of a
two-way interactive atmospheric and ecological model and its application to the central united
states. J. Clim. 14:900919
Matthews E (1984) Prescription of land-surface boundary conditions in GISS GCMII and Vegetation, land-use and seasonal albedo data sets: documentation of archived data tape. NASA
Technical Memos 860096 and 86107, NASA, Goddard Institute for Space Studies, New York,
20pp and 9pp
surface albedos: value-added datasets derived from Terra MODIS Land Products. IEEE Trans.
Geosci. Remote Sens. 43 (l):144158
Myneni RB, Hoffman S, Knyazikhin Y, Privette JL, Glassy J, Tian Y, Wang Y, Song X, Zhang Y,
Smith GR, Lotsch A, Friedl M, Morisette JT, Votava P, Nemani RR, Running SW(2002) Global
products of vegetation leaf area and fraction absorbed PAR from year one of MODIS data.
462
R.E. Dickinson
Oleson KW, Bonan GB, Schaaf C, Gao, F, Jin Y, Strahler A (2003) Assessment of global
climate model land surface albedo using MODIS data. Geophys. Res. Lett. 30 (8):1443,
doi:1029/2002GL016749
Pinty B, Lattanzio A, Martonchik JV, Verstraete MM, Gobron N, Taberner M, Widlowski J-L,
Dickinson RE, Govaerts Y (2005) Coupling diffuse sky radiation and surface albedo. J. Atmos.
Sci. 62:25802591
Pinty B, Lavergne T, Dickinson RE, Widlowski J-L, Gobron N, Verstraete MM (2006) Simplifying
the interaction of land surfaces with radiation for relating remote sensing products to climate
models. J. Geophys. Res. 111 (D02116), doi:10.1029/2005JD005952
Pitman AJ (2003) The evolution of, and revolution in, land surface schemes designed for climate
models. Int. J. Climatol. 23:479510
Reichle RH, Koster RD (2005) Global assimilation of satellite surface soil moisture retrievals into the NASA Catchment land surface model. Geophys. Res. Lett. 32 (L02404),
doi:10.1029/2004GL021700
Sellers PJ (1985) Canopy reflectance, photosynthesis and transpiration. Int. J. Remote Sens.
6:13351372
Sellers PJ, Dickinson RE, Randall DA, Betts AK, Hall FG, Berry JA, Collatz GJ, Denning AS,
Mooney HA, Nobre CA, Sato N, Field CB, Henderson-Sellers A (1997) Modeling the
exchanges of energy, water, and carbon between continents and the atmosphere. Science
275:502509
Sitch S, Smith B, Prentice IC, Arneth A, Bondeau A, Cramer W, Kaplan JO, Levis S, Lucht W,
Sykes MT, Thonicke K, Venevsky S (2003): Evaluation of ecosystem dynamics, plant geography and terrestrial carbon cycling in the LPJ dynamic global vegetation model. Global Change
Biol. 9:161185
Strugnell NC, Lucht W, Schaaf C (2001) A global albedo data set derived from AVHRR data for
use in climate simulations. Geophys. Res. Lett. 28(1):191194
Tian Y, Dickinson RE, Zhou L, Zeng X, Dai Y, Myneni RB, Knyazikhin Y, Zhang X, Friedl M,
Yu H, Wu W, Shaikh M (2004a) Comparison of seasonal and spatial variations of leaf area
index and fraction of absorbed photo synthetically active radiation from Moderate Resolution Imaging Spectroradiometer (MODIS) and common land model. J. Geophys. Res. 109,
doi:10.1029/2003JD003777
Tian Y, Dickinson RE, Zhou L, Myneni RB, Friedl M, Schaaf CB, Carroll M, Gao F (2004b) Land
Geophys. Res. Lett. 31(5), L05504, 10.1029/2003GL019104
Tian Y, Dickinson RE, Zhou L, Shaikh M (2004c) Impact of new land boundary conditions from
Moderate Resolution Imaging Spectroradiometer (MODIS) data on the climatology of land
surface variables. J. Geophys. Res. 109 (D20115), doi:10.1029/2003JD004499
Tsvetsinskaya EA, Schaaf CB, Gao F, Strahler AH, Dickinson RE, Zeng X, Lucht W (2002) Relating MODIS derived surface albedo to soil and landforms over Northern Africa and the Arabian
peninsula. Geophys. Res. Lett. 29, doi:10.1029/2001GL014096
Wang Z, Zeng X, Barlage M, Dickinson RE, Gao F, Schaaf CB (2004) Using MODIS BRDF and
albedo data to evaluate global model land surface albedo. J. Hydrometeorol. 5:314
Wang Z, Barlage M, Zeng X, Dickinson RE, Schaaf CB (2005) The solar zenith angle dependence
of desert albedo. Geophys. Res. Lett. 32 (L05403), doi:10.1029/2004GL021835
Widlowski JL, Pinty B, Gobron N, Verstraete MM (2001) Detection and characterization of boreal
coniferous forests from remote sensing data. J. Geophys. Res. 106 (D24): 3340533419
Widlowski JL, Pinty B, Gobron N, Verstraete MM, Diner DJ, Davis AB (2004) Canopy structure parameters derived from multi-angular remote sensing data for terrestrial carbon studies.
Climatic Change 67 (23):403415
Wilson MF (1984) The construction and use of land surface information in a general circulation
climate model. Unpublished Ph.D. thesis, University of Liverpool, United Kingdom, 346 pp
Woodward FI, Lomas MR (2004) Vegetation dynamics simulating responses to climatic change.
Biol. Rev. 79:643670, doi:10.1017/S1464793103006419
463
Yang Z-L (2003) Modeling land surface processes in short-term weather and climate studies. In:
X Zu, X Li, M Cai, S Zhou, Y Zhu, F-F Jin, X Zou, M. Zhang (eds), Theory and modeling of
atmospheric variability. World Scientific Series on Meteorology of East Asia, World Scientific,
New Jersey, pp 288313
Zeng X, Dickinson RE, Walker A, Shaikh M, DeFries RS, Qi J (2000) Derivation and evaluation
of global 1-km fractional vegetation cover data for land modeling. J. Appl. Meteor. 39:826839
Zeng X, Shaikh M, Dai Y, Dickinson RE, Myneni R (2002) Coupling of the common land model
to the NCAR community climate model. J. Clim. 15:18321854
Zeng XB, Dickinson RE, Barlage M, et al. (2005) Treatment of undercanopy turbulence in land
models. J. Clim. 18(23):50865094
Zhang Q, Xiao X, Braswell B, Linder E, Baret F, Moore B III (2005a) Estimating light absorption by chlorophyll, leaf and canopy in a deciduous broadleaf forest using MODIS data and a
radiative transfer model. Remote Sens. Environ. 99:357371
Zhang X, Friedl MA, Schaaf CB, Strahler AH (2005b) Monitoring the response of vegetation
phenology to precipitation in Africa by coupling MODIS and TRMM instruments. J. Geophys.
Res. 110 (D12103), doi:10.1029/2004JK005263
Zhou L, Dickinson RE, Tian Y, Zeng X, Dai Y, Yang Z-L, Schaaf CB, Gao F, Jin Y, Strahler A,
Myneni RB, Yu H, Wu W, and Shaikh M (2003) Comparison of seasonal and spatial variations of albedos from Moderate-Resolution Imaging Spectroradiometer (MODIS) and Common Land Model. J. Geophys. Res. 108(D15):4488, doi:101029/2002JD003326
Zhou L, Dickinson RE, Tian Y (2005) Derivation of a soil albedo dataset from MODIS using
principal component analysis: Northern Africa and the Arabian Peninsula. Geophys. Res. Lett.
32 (L21407), doi:10.1029/2005GL024448
Chapter 18
Improving the Utilization of Remotely

Sensed Data
John R. Townshend, Stephen Briggs, Roy Gibson, Michael Hales,
Paul Menzel, Brent Smith, Yukio Haruyama, Chu Ishida, John Latham,
Jeff Tschirley, Deren Li, Mengxue Li, Liangming Liu, and Gilles Sommeria
Abstract This chapter reports on the work of a team set up by the Committee on
Earth Observation Satellites to examine the factors encouraging the increased use
of space data. The main emphasis of the study was on situations where the desired
outcome is a sustained capability rather than increasing the number of one-off users.
Twenty-five case studies were compiled by the team. Examination of the successful
case studies indicates that more of the factors favoring adoption of remote sensing products by users are within the province of the users themselves, rather than
John R. Townshend
University of Maryland, College Park, USA
jtownshe@umd.edu
Stephen Briggs
European Space Agency
Roy Gibson
EUMETSAT
Michael Hales, Paul Menzel and Brent Smith
NOAA
Yukio Haruyama and Chu Ishida
JAXA
John Latham and Jeff Tschirley
Food and Agriculture Organization
Deren Li
Wuhan University
Mengxue Li
NRSCC
Liangming Liu
National Remote Sensing Center of China
Gilles Sommeria
WCRP

465
466
J.R. Townshend et al.
with the space agencies, but successful adoption always involves balanced cooperation between space agencies and users. In this cooperation space agencies can
and should do more to ensure the greater uptake of their data and services. Critical
factors from the space agency point of view are provision of user-specific products;
making resources available speedily and with a minimum of bureaucracy; willingness to listen over an extended period to user needs and practical difficulties and to
accommodate them to the maximum extent practicable. From the user standpoint,
success appears to be very dependent on (a) the user having an understanding, and
preferably first-hand experience, of the use of remote sensing data and products, and
its integration into geospatial information systems; (b) a willingness to make appropriate staff available before and during the project to work with the space agency;
(c) the user having a core need, such as a statutory need, for the information; (d)
an ability to define rather precisely what the needs are; and (e) the user being willing and able to help develop a realistic financial scenario for an operational system.
On both sides, the case studies indicate the need for innovation and flexibility in
response to new opportunities for data products.
18.1 Introduction
A wide range of sensors regularly supply observations of the Earths environment
and ever improved capabilities are provided to meet increasingly stringent demands
are frequently made to satisfy users (e.g., Duchossois and Sommeria, 2003). Achieving enhanced use of observations depends on many factors including the successful
transition from research to operational observations (NRC, 2000, 2003a). Another
major factor is data policy (Harris, 2002) which can strongly impact the availability
of data through charging policies. Increasing the utilization of these data represents
an improved return on the considerable investments in the space and ground segments (NRC, 2003b). There is a widely held belief that such data are often underutilized. To facilitate and encourage the increased use of space data and the products
derived from them, the Committee on Earth Observation Satellites (CEOS) set up a
Utilization Team to advise space agencies about their roles in achieving these goals
(CEOS, 2003).
The approach adopted by the Utilization Team was first to identify case studies,
where adoption of remote sensing was both rapid and successful and then to examine
what can be inferred from an analysis of them. The advantage of this approach was
to ground the Teams work and advice on concrete examples. Second the case
studies were used to identify impediments, the necessary preconditions for adoption
of remote sensing and the ways to facilitate utilization. Finally using the previous
considerations specific recommendations were made to space agencies to increase
utilization.
In this chapter we next describe the procedures adopted, including the assembly
of case studies as well as insights from interviews, which authors conducted with
experienced users of satellite data. We then identify and analyse the factors, which
18 Improving the Utilization of Remotely Sensed Data
467
either facilitate or hamper the increased use of satellite data, products and services.
These considerations are used to derive a set of principles, which we believe will
lead to enhanced adoption of space-based data and products. Finally we summarize
the recommendations which were addressed to CEOS as a body and to individual
members and associates.
18.2 Approach and Procedures

Basic to this study is that increased utilization of space-derived data, products and
services is a good thing in its own right. While all space agencies undoubtedly wish
to see the widest use of data acquired by the satellite systems for which they are
responsible, it should be recognised, that some space agencies do not have a specific
remit to be proactive in promoting the use of their data for users for whom the
Earth observation activity was intended, nor do they necessarily have a mandate to
undertake the regular re-processing necessary to ensure consistent term data sets
over the longer term, crucial to many applications. We also consulted a number of
experts on these issues, who are recognized in the acknowledgments.
The main emphasis of the study was on situations where the desired outcome is
a sustained capability rather than increasing the number of one-off users. We examined increased use of existing data, as well as increasing the potential use of
future satellite missions. We considered both those situations where immediate increases in utilization are realisable as well as those where sustained collaboration
between space agencies and users may be needed before operational applications
can be developed.
For each case study a template was completed, which provided the following
information:
Characterizing the activity
Providing a chronology
Identification of the factors encouraging adoption
The principal obstacles that were overcome and how this was achieved
Related spin-off applications
The lessons learnt by the authors from this case study
The templates were completed by the authors through interviews with both the users
and the responsible space agencies and their partners. In Fig. 18.1 there is the completed template for one of the case studies concerned with rapid response of fires.
Table 18.1 contains a listing of all of the 22 case studies compiled by the Team with
the names of space agencies and partners as well as the users of the data.
Discussions of potential users imply that there are always many more waiting
to be found. This is not the case for all types of users: some markets may already
be close to saturation. For example, although weather forecasting centres may need
improved observations, the number of such centres is unlikely to grow and may even
contract. Hence there may be situations where there may only be the prospect of an
improved use of remote sensing, without any actual increase in the number of users.
468
CEOS ad hoc team on utilization: implementation case study template

Name of activity
MODIS Fire Rapid Response System
Further information (URL)
http://rapidresponse.umd.edu/
Name of space agency/ies and partners
NASA, University of Maryland, USFS
Name of adopting agency/ies
USFS and the National Inter-Agency Fire Center
Brief description of activity
1. Images from MODIS are down-loaded, processed and placed in a GIS/map format for use by NIFC and the
USFS on a daily basis for the whole of the conterminous US plus Alaska.
2. Areas of active fires and cumulative burn scars are identified on a topographic map base showing relief,
main roads the urban interface, and administrative and state boundaries.
3. The products are used by NIFC to make strategic resource allocation decisions (i.e., where to redeploy firefighters and major equipment such as aircraft as opposed to scale local tactical fire fighting decision-making)
Chronological description of process of adoption
2000 USFS learnt of the availability of fire products from MODIS and examined examples prepared on an
ad hoc basis.
2001 Fully implemented prototype boot-strapped system developed, using NASA, NOAA and UMD and
USFS facilities to provide products on a daily basis to NIFC in time for the 2001 fire season (June 1).
Algorithms simplified, main EOS-DIS system by-passed and completely automated implementation to
allow sufficiently rapid delivery of products. In practice products are available 24 h after imaging.
2002 USFS takes over complete operations through use of own direct receiving system for western US
and transfer to them of all needed software. The NASA system now acts as a back-up. NASA and UMD
involvement moved to joint research in developing new products (e.g., burn intensity in addition to providing
eastern US and Alaska data beyond USFS station coverage).
Factors favoring adoption
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
Adopting agency had existing experience in using remote sensing.

Adopting agency had clear, strong need for products.
Pre-existing scientific expertise in algorithm development for fire detection.
Close working relationship developed between algorithm/system developers and adopting agency
Funds quickly deployed by space agency and adopting agency allowing prototype system to be
implemented.
Highly competent technical environment (at NASA, USFS and UMD) allowing information system components to be built and integrated from COTS and reuse of hardware.
Funds quickly deployed to allow USFS to purchase dish allowing them take over as operational agency.
Products delivered in form specified by users - paper maps with information kept to the minimum required
for decision-making. New products compatible with heritage formats and procedures.
Those developing the system passed over operations as quickly as possible to adopting agency.
Developers interacted directly with users (e.g., RR staff visited active wildfires and spoke with range of
users including management, direct attack, and rehabilitation team members.
Principal obstacles that were overcome (indicate whether and how they were overcome)
1. Routine EOS-DIS data processing system for science had to be by-passed to allow sufficiently rapid
delivery.
2. Need for non-operational prototype was avoided. Strictly speaking no true prototype.
Other related spin-off applications
1. Rapid response system implemented for whole world with reflectance background and fire locations identified (http://rapidfire.sci.gsfc.nasa.gov/).
2. Products distributed through UN system.
3. Products made available regionally through GOFC-GOLD mechanism.
Lessons learned from case study
1.
2.
3.
4.
5.
This example worked exceedingly well and quickly.

Technical expertise played a major part.
But so did the receptivity of the adopting agency.
The quickness of the funders to back a winner when they saw one.
Perhaps this example may be atypical in that everything was in its favor.
Fig. 18.1 Example of completed case study template
469
Table 18.1 Case studies where remote sensing was adopted successfully and rapidly
Name of case study
Responsible space agency

(and partners)
User agency
MODIS Fire Rapid

Response System
NASA
Application of the HY-1

satellite for sea ice
monitoring and forecasting
(the sea ice retrieval
system)
Meteorological Satellite
Fire Monitoring System
Chinese National Oceanic

Satellite Application Center
US forest Service and

National Inter-agency Fire
Center
National Marine
Environment Forecasting
Center
Integrated operational
system of flood monitoring
and assessment
Monitoring system of ice

conditions for the Yellow
River
Crop monitoring system
using remote sensing
10
11
The application of
CBERS-1 data in forest
resources investigation
Application of satellite
remote sensing to fishing
industry
Collaborative research on
sea ice observation by Earth
Observation data
Collaborative research on
development of land
utilization distribution by
Earth Observation satellite
data
MODIS polar winds
National Satellite
Meteorological Center of
China
Remote Sensing
Technology Application
Center, Ministry of Water
Resources Institute of
Geography, CAS
National Remote Sensing
Center of China and Wuhan
University
Institute of Remote Sensing
Applications, CAS
China Center for Resources
Satellite Data and
Application (CRESDA)
JAXA
JAXA
JAXA
NASA, NOAA
12
Atmospheric Sounder data

utilization
NOAA
13
Assimilation of satellite
data in Numerical Weather
Prediction
NOAA, Eumetsat
National Satellite
Meteorological Center of
China, Ministry of
Agriculture, China Forestry
Administration
Chinese Ministry of Water
Resources
Information Center of
Yellow River Conservancy
Commission
State Food Bureau,
Committee for State
Planning Development
Guizhou Forestry and
Planning Institute
Japan Fisheries Information
Service Center
Hydrographic and
Oceanographic Department,
Japanese Coast Guard
Kumamoto City
Environmental Research
Institute
European center for

Medium Range Weather
Forecasting
US National Centers for
Environmental Protection,
European Center for
Forecasting, UK Met.
Office
European Center for
Forecasting
(Continued)
470
Table 18.1 Continued

Name of case study
Responsible space agency

(and partners)
User agency
14
Regional Burned Forest

Mapping
Italian Department of Civil

Protection (DPC)
15
Kyoto Inventory

(ESA) with Telespazio,
Intecs, Vitrociset, Cesia and
Mestor (Italy)
ESA, Intecs (IT), Planetek
(IT), NEO (NL), Telespazio
(IT), Dataspazio (IT),
Bosdata (NL), Gamma
Remote Sensing (CH),
University of Trento (IT) +
precursor study: VTT
(FIN), European Forest
Institute, Stora Enso Forest
Consulting (FIN)
16
Urban expansion and

monitoring (URBEX)
17
Oil Spill Monitoring Center
18
Ecosystem (environmental)
monitoring
NASA, ESA, CNES
19
Ozone alert
ESA, KNMI
20
Surface albedo
21
Sea level and ocean

circulation real time
monitoring
EUMETSAT and Joint

Research Center, Ispra
CLS, CNES, EUMETSAT,
JPL NASA, NOAA, US
Navy
22
EnviroCast (remote sensing

products for the media)
NASA, NOAA, EPA,

USDA Forest Service

(ESA), Advance Computer
Systems (ACS),
Norwegian Space Centre,
Kongsberg Satellite
Services (KSAT)
Italian Ministry of
Environment and Territory;
Swiss Agency for the
Environment, Forest and
Landscape; Finnish
Ministry of
Environment/METLA;
Norwegian Ministry of
Agriculture/NIJOS; Dutch
Ministry of Agriculture,
Nature Management and
Fishery
World Wildlife Fund
(WWF)
State Pollution Authorities
until 2002. Norwegian
National Coastal
Administration from 2003
Various NGOs including
The Nature Conservancy,
Conservation International,
and World Resources
Institute
ECMWF, Met. Offices,
Education Institutes,
Scientific Measurements
Campaign Planners,
Environmental Agencies,
LOreal
EUMETSAT initially for
distribution
ECMWF, FNMOC, FOAM,
MERCATOR, MFS,
NAVOCEANO, NCEP,
NOAA, NRL
Private company
(StormCenter
Communications, Inc.) for
distribution to media
471
It is worth noting that, compared with even 10 years ago, the costs of many
elements of information systems are much lower, including costs of local ground
receiving stations and computer hardware and software. These all combine to make
it less expensive to integrate remote sensing products into environmental information systems, and hence potentially increases the number of users.
18.3 What Factors Favor Adoption of Remote Sensing?

18.3.1 Analysis of Case Studies
Table 18.2 lists the factors identified in the case studies as favoring the adoption of
remote sensing by users and the frequency of their occurrence in the case studies.
Those completing the studies were asked to identify only the most important factors
but were not given any fixed number of factors that had to be selected.
Examination of the successful case studies indicates that more of the factors
favouring adoption of remote sensing products by users are within the province of
the users themselves, rather than with the space agencies, but it needs to be stressed
that successful adoption always involves balanced cooperation between space agencies and users. In this co-operation space agencies can and should do more to ensure
the greater uptake of their data and services.
Table 18.2 Occurrence of factors important in case studies where remote sensing technology was
rapidly and successfully adopted
Types of factors encouraging adoption
Space agencys and partners factors
Additional resources provided for new uses
Provision of technical capability
Provision of ready access to data
Provision of data at affordable costs
Products in user-specified forms (incl. timeliness)
Inter-agency factors
Strong inter-agency cooperation
Rapid transfer of capability to users
External linking expertise linking users and space
agencies
User factors
Information needs strong and well quantified
Additional resources allocated
Resources to install ground receiving capability
In-house scientific and remote sensing expertise
Existing operational use of remote sensing
Existing use of geospatial technologies
Commitment to use information operationally
Provision of field and validation information
Demonstrated cost savings
Number of occurrences
7
6
3
4
10
11
2
6
19
7
1
13
6
6
6
3
2
472
Critical from the space agency standpoint is (a) provision of user-specific products; (b) making resources available speedily and with a minimum of bureaucracy;
(c) willingness to listen over an extended period to user needs and practical difficulties; and (d) to accommodate these difficulties to the maximum extent practicable going so far sometimes as to incorporate user representatives into a joint
project team.
Added to these factors, several case studies stressed the importance of keeping
user development projects simple, and building further on reliable, successful initial
projects. However, there is evidence to suggest that turning prototypes into sustained
operations is frequently unsuccessful if the funding from users does not continue
beyond the prototype phase. Space agencies therefore need at a relatively early stage
to seek assurance that the user has sufficient interest and capacity to continue the
project beyond the prototype stage.
From the user standpoint, success appears to be very dependent on (a) the user
having strong needs for the information and that these needs are well characterized;
(b) the user having an understanding, and preferably first hand-experience, of the
use of remote sensing data and products, and their integration into geospatial information systems; (c) a willingness to make appropriate staff available before and
during the project to work with the space agency; and (d) the user being willing and
able to help develop a realistic financial scenario for an operational system.
On both sides, the case studies indicate the need for innovation and flexibility in
response to new opportunities for data products. Equally important is the role that
individuals can play. Sometimes, these are senior champions in either space agencies or less frequently in user organisations. In other cases it may be an individual
facilitator who has experience in new remote sensing technologies and the development of algorithms and who also has close contacts with user communities; this can
lead to speedy adoption by new users. Not surprisingly strong cooperation between
space and user agencies was frequently recognized as being key. The development
of partnerships between users, space agencies and research organizations including
universities often figured prominently. In several case studies industry also played a
significant role in increasing utilization.
18.3.2 Technical Factors Encouraging the Adoption of Remote

Sensing
Analysis of case studies showed that the principal obstacles faced, and to varying
extents overcome, were overwhelmingly technical (Table 18.3) and fell into the following categories:
Need to provide adequate bandwidth for data access
Devising adequate distribution networks
Need to provide new or improved algorithms
473
Table 18.3 Occurrence of significant obstacles overcome in case studies, where adoption was rapid
and successful
Obstacles overcome
Bandwidth
Inadequate characterization of observations
Need for improved algorithms
Need for improved information system
Gaining sufficient resources
Improving timeliness
Lack of trained personnel
Lengthy prototyping
Lack of understanding of potential of
Remote Sensing
Proving impact
Number of occurrences
5
1
10
4
3
4
1
2
2
2
Development of faster data processing techniques (e.g., calibration techniques

for SAR images)
Need to improve the basic classification systems to which the remote sensing
project is designed to contribute
Need to demonstrate that continuity of space observations can be assured
However, the importance of technical obstacles relative to other kinds might have
been smaller had the study also considered unsuccessful case studies.
Where major improvements in observational capabilities occur there is often immediate potential for improved use. For example, the availability of moderate spatial resolution MODIS (Moderate resolution Imaging Spectrometer) data with its
much higher capabilities compared with the AVHRR (Advanced Very High Resolution Radiometer) has led to a number of rapid improvements in utilization. The
use of MODIS data to assist fire-fighting in the United States or the use of MODIS
polar winds in ECMWF (European Center for Medium Range Weather Forecasting) assimilations may be mentioned. Adoption of MODIS is encouraged by the
planned follow-on of the VIIRS (Visible Infrared Imaging Radiometer Suite) sensor
of NPOESS (National Polar Orbiting Environmental Satellite System) and its predecessor NPP (NPOESS Preparatory Project), with similar capabilities to MODIS.
Such new sensing capabilities offer both the possibility of enhancing systems already using remote sensing and the possibility of significant new uses of the data.
One notable example of a space agency positioning itself for new uses of its data belongs to EUMETSAT who has set up seven Satellite Application Facilities (SAFs)
to exploit the increased capabilities of Meteosat Second Generation for non-weather
and climate applications. This offers a potentially valuable model for other agencies
concerned with increased exploitation of new observational capabilities. Enhancement in observational capabilities can also arise from improved acquisition strategies. The use of Landsat data increased substantially with the introduction of a much
improved acquisition strategy for Landsat 7 greatly increasing coverage; increased
use also stemmed from a sharp reduction in costs.
474
Where there will be major changes in observational capabilities in the foreseeable

future it is normally incumbent on the responsible agencies to ensure that users
are aware of these. Agencies should also consider the desirability of ensuring the
provision of appropriate training and contributing to capacity building especially
where it is possible to identify existing operational users who are likely to benefit
from these improvements.
18.3.3 Education and Capacity Building Encouraging

the Adoption of Remote Sensing
Lack of trained staff and modern hardware and software were, not surprisingly, also
encountered. Remote sensing can in some respects be regarded as a relatively mature
technology and, as a result, in some sectors many potential users are well aware of
its ability to meet their needs. However, this does not always apply in developing
countries, and even users in the developed word may need additional training (as
well as updated hardware and software) if they are to make best use of new data
sets, which often have much higher data volumes. The availability of the technical
capability to improve algorithms is clearly important in many of the case studies.
To increase utilization, it appears preferable that education and capacity building
is carried out in conjunction with a practical project, rather than performed generically. Where remote sensing products are being introduced it is essential at an early
stage remote sensing realistically to assess the amount of education and training
which will be needed in order to ensure that the project will successfully develop
into an operational capacity. The understandable tendency to underestimate the cost
of this element will certainly militate against success.
18.3.4 Financial Factors Affecting the Adoption of Remote Sensing

There is no substitute for users being willing to spend their own time and resources
in the adoption of remote sensing. Such resources will be used on staff time being
allocated in the prototyping phase and in expenditures on information systems and
outreach to the decision-makers and other end-users of the resultant products.
Alongside space agencies and new users, third parties can also play an important
role. This may be in a funding role, as in the case of the European Commission
in Europe or the World Bank. Such financers rightly insist on the project being
designed realistically and economically to meet user needs. In Europe, private companies often also play a major role in developing user capabilities and in linking
users with space agencies.
475
18.3.5 Policy Factors Affecting the Adoption of Remote Sensing

Two other obstacles, of a somewhat different nature, are apparent: restrictions on
data access because of local data policies, and the regular complaint of the high cost
of satellite data and products. It was beyond the remit of the authors to deal in depth
with these two factors, but it would be unrealistic not to identify them. Perhaps
the most that can be said is that space agencies should be encouraged to release
data expeditiously and at the lowest cost, whenever they are needed in connection
with an approved project or for a use consistent with the work which space agencies
promote such as uses recognized by the World Summit on Sustainable Development
and with the international environmental conventions.
18.4 Principles to Encourage the Adoption of Remote Sensing

Based on the case studies we can identify three main ways in which the use of
remote sensing can be enhanced.
First there are those situations where users already have information or forecast
systems using remotely sensed data. Improved observations become available
and these are introduced to the existing operational system. Even in these circumstances increasing utilization, may not be a trivial exercise and usually requires
active cooperation between user and space agencies.
Second there are situations where users have not normally used remote sensing.
This will usually require a significant effort for the needed skills and facilities to
be transferred. Many of the case studies include such examples.
Third, there are situations where space agencies or contracted intermediaries are
made responsible for base-lined standard products, with well-characterized properties in terms of Calibration and Validation and Quality Assessment. In these
situations many users could adopt products without the need greatly to add to
their technical knowledge and facilities for remote sensing.
18.4.1 Assurance of Continuity

For users to adopt remote sensing there is a need for systems to have ensured continuity. Security of data and product supply is essential and operational users have
to receive firm assurances to be sure that there will be long-term continuity. Operational agencies need to be assured that the significant initial costs incurred in
changes in their business practices to accommodate the use of satellite data will be
recovered through guaranteed long-term access to satellite data.
476
18.4.2 Need for Adequate Information Systems and Spatial Data

Infrastructure
In many parts of the world successful information systems may exist and the benefits of remote sensing may appear marginal compared with existing in situ sampling
schemes. But in other parts of the world adequate environmental information systems often do not exist, especially in the developing world. Where aid is provided
it is vital that there is not merely investment in receiving station capability (in some
cases there are too many receiving stations which seem to be regarded as status
symbols), but also in end-to-end capabilities from acquisition to product generation
and use. Although there may be considerable potential for application of remote
sensing, sustaining such capabilities in the poorest countries of the world may be
difficult without long-term aid.
18.4.3 Space Agencies Responsibilities for Information Systems

Although in some cases it is not included expressly within their mandates, space
agencies have key long-term stewardship responsibilities in ensuring data delivery,
data archiving and reprocessing for climate-quality data sets. Users have to be sure
that these will be executed regularly and to the highest possible standards
18.4.4 Enhancing Data Delivery Systems

In some situations data systems may be designed for scientific researchers and be inadequate for other applications. For example delivery of data for scientific uses may
be quite acceptable even if days or weeks after acquisition. Providing a capability to
access and process data almost instantaneously significantly increases the range of
possible applications. This helps explain the increasing popularity of local ground
receiving stations. One noteworthy example of this is the Advanced TIROS Operational Vertical Sounder (ATOVS), where direct broadcast has been maintained and a
preprocessing package was created and distributed with international collaboration.
18.4.5 Ensuring that Consistent Long-Term Data Sets are Created

Many uses of remotely sensed data rely on internally consistent long-term data sets
especially in relation to understanding climate change and assessing its impacts
on socio-economic activities. From time to time major new uses of environmental records may arise and it would be prudent for space agencies to be prepared to
respond to these. Creation of such data sets usually requires very substantive efforts
normally involving a considerable amount of reprocessing and analysis.
477
18.4.6 Access to Integrated Assets of Multiple Agencies

There is a need for users to be able to use the integrated assets of multiple agencies; essential to ensure increased usage. This will often involve the merging and
integration of data sets by data fusion. There is a key role for universities and research efforts outside of space agencies. We recognize that individual space agencies have to promote their own EO assets, and there is a continuing need for such
promotion to be realistic in terms of capabilities. There is a need for independent
mechanisms to promote integration and to carry out independent assessments of
capabilities.
18.4.7 Role of Research in Stimulating Innovative Uses of Remote

Sensing
It is important to highlight the importance of research in general. Examples of areas
where significant recent strides have been made include the development of SAR
Interferometry and research in Numerical Weather Prediction. Development of applications often arises on the basis of research. Also there is a need to recognize
the importance of integrated research activities between the industrial and research
communities. The key roles of research institutions must be recognized in addition
to the roles of space agencies and users.
18.4.8 Improved Cooperation Between Space Agencies

There is a need for mechanisms to avoid potentially unhelpful duplication between space agencies. Proposals for a more integrated approach between NASAs
Pathfinder and ESAs Explorer program are to be welcomed. Cooperation will
needed as early as possible, including the definition and selection phases. The optimal situation in terms of international cooperation may often be where users pose
key needs which receive a coordinated response; an exemplary case of this is of
which Jason-2. Where there are international jointly funded programs such as the
Global Precipitation Mission (GPM) or other missions with multiple international
components, consideration should be given to joint science teams for the above
activities.
18.4.9 Sustained Capacity Building

With improvements in observational capabilities there will be a continuing need for
capacity building in developing and developed and countries. Such outreach needs
478
to be embedded within continuing projects to secure the long-term benefits available

from increased capability. This must be supported by a continuing stream of high
quality technical and educational publications and outreach to users.
18.4.10 Need for Strategies to Cope with the Long Time Periods
Between Problem Identification and the Arrival
of an Operational System
Strategies are needed to deal with very long times between problem identification and provision of comprehensive observational solutions. The timescale of response to the relatively simple problem of global stratospheric ozone, ideally suited
to remote sensing techniques, is indicative of the long time constants inherent in
addressing such problems. The long time period needed and gradual evolution of
the system to introduce the use of operational remote sensing satellites for weather
forecasting in Europe is illustrated in Fig. 18.2.
18.4.11 Recognition of New Partners and the Distinctive Needs

of New Users
During the last few years major new types of users are beginning to adopt remote
sensing. These include NGOs who require observations for purposes such as conservation and humanitarian relief. Others include media organizations requiring very
timely information. Space agencies will need to recognize and adapt to their special
needs if adoption is going to be successfully achieved by these new sets of users.
In developing partnerships with the private sector this should not automatically be
interpreted as meaning the aerospace industry, but new partners, such as insurance
companies, already providing services on a large scale through non-space means.
18.4.12 Political Ownership of Benefits

There is a need to promote uses that have a high political impact. The benefits
for political leaders have to be included, since political ownership of the benefits
of remote sensing will encourage investment in the adoption of remote sensing. It
is also important that legislators are kept aware of major advances in capabilities,
which could influence the introduction of legislation, e.g., reliable monitoring of
tropospheric ozone could conceivably lead to the introduction of legislation based
on this capability.

A)
479
Weather Forecasting 40 years ago in Europe

Data ....to .. Information
Feedback
USERS
Forecasts
Global
System
design
In situ Data
Govt
Public
Private
D
S
Modeling,
OPS
Data
Assimilation
RTD
EC-Funded
Research in
Universities
B)
Weather Forecasting 20 years ago in Europe

Feedback
USERS
Modeling,
Global
System
design
In situ Data
Forecasts
Data
Assimilation
Govt
Public
Private
OPS
RTD
Satellite Data
ESA
EC - Funded
Research in
Universities
C)
The Example of Weather Forecasting today in Europe

Feedback
USERS
N
E
E
D
In situ Data
Global
System
design
Modeling,
Data
Assimilation
Satellite Data
Forecasts
Govt
Public
Private
EUMETSAT
OPS
RTD
ESA,
National
agencies
FP6
Universities
Fig. 18.2 The evolution of the utilization of remotely sensed data in European weather forecasting.
(Courtesy of D. Williams, Eumetsat.)
480
18.5 Recommendations to Space Agencies

The preceding sections contain information that the authors hope will be of interest
and use to at least some readers. In conclusion we emphasize a few major recommendations made to space agencies through CEOS.
1. Real-time access to data streams is leading to increasing use of remotely sensed
data, especially where preprocessing packages are made available to users. The
latter means that the initial responsibility to ensure the processing and delivering
of products rests with the agency rather than the user. It is important that agencies
make pre-processing packages available to the greatest extent possible, whether
they are developed in-house or in the research or commercial communities.
Real-time access includes but is not limited to direct broadcast. This means that
space agencies need to make data streams available in real time and whenever
possible provide any necessary preprocessing software, a benchmark data set
and a focal point for questions.
2. The quality of information about instruments, made available by space agencies,
varies considerably. Many space agencies need to reevaluate the way in which
they provide information about past, current and future systems including the
provision of high quality up-to-date information concerning instrument performance especially calibration and product quality assessments. A more standardized approach for the description of instrument performance should be adopted.
3. Best use of remotely sensed data increasingly depends on use of data from
multiple sensors often from multiple agencies, but this is often significantly hindered by the lack of coordination between information systems. Space agencies
need to develop information systems with more integrated catalog, search, ordering and retrieval mechanisms.
4. Increasing the utilization of remotely sensed data is dependent on developing
improved partnerships with users. This is, of course, very much a matter for
individual agencies, and much has already been done. Nevertheless, space agencies should commit to seeking ways of increasing and improving partnerships,
including those with NGOs and universities. These could include the following
but need not be limited to them:
Improve the representation of scientific and other users on their advisory
bodies.
Work more intensively with the scientific community to develop new applications and services and specifically work to develop improved assimilation
procedures.
Consider the possibility of selectively improving the funding of new research
and other application activities.
Consider coordination between agencies in providing funding for research and
other applications.
Where there are internationally coordinated remote sensing activities such as
the GPM, then consider setting up international science panels in conjunction
with appropriate international research organizations.
481
Establish closer links with the scientific and political arms of the international
environmental conventions and agreements.
Strengthen links with policy makers and international conventions for high
profile data utilization in keeping with the high priority given to this by space
agencies and their coordination organizations such as CEOS.
Reflecting the strong emphasis of most space agencies on planning Earth observation missions for the purpose of climate and global environment studies,
particular attention should be placed on improving the dialogue with relevant
user communities for the planning, specification, and application of these missions. This should include establishing closer links with the scientific and political arms of the climate conventions (e.g., the UN Framework Convention on
Climate Change and its Subsidiary Body for Scientific and Technical Advice)
and should take account of high priority information needs as defined by bodies
such as the Inter-government Panel on Climate Change.
5. Use of remotely sensed data is sometimes restricted by the fact that although
products have been demonstrated in a research environment and have been produced at a regional level their value at a global level has not been demonstrated
and recognized by the users. Examples include global albedo, fire products and
other high priority ones identified in the Integrated Global Observation Strategy
themes (IGOS, 2004) and more recently by the Group on Earth Observations
Table 18.4 GCOS Climate Monitoring Principles

(http://gosic.org/GCOS/GCOS climate monitoring principles.htm)
Effective monitoring systems for climate should adhere to the following principles
1. The impact of new systems or changes to existing systems should be assessed prior to
implementation
2. A suitable period of overlap for new and old observing systems is required
3. The details and history of local conditions, instruments, operating procedures, data
processing algorithms and other factors pertinent to interpreting data (i.e., metadata)
should be documented and treated with the same care as the data themselves
4. The quality and homogeneity of data should be regularly assessed as a part of routine
operations
5. Consideration of the needs for environmental and climate-monitoring products and
assessments, such as IPCC assessments, should be integrated into national, regional and
global observing priorities
6. Operation of historically uninterrupted stations and observing systems should be maintained
7. High priority for additional observations should be focused on data-poor regions,
poorly observed parameters, regions sensitive to change, and key measurements with
inadequate temporal resolution
8. Long-term requirements should be specified to network designers, operators and instrument engineers at the outset of system design and implementation
9. The conversion of research observing systems to long-term operations in a carefully
planned manner should be promoted
10. Data management systems that facilitate access, use and interpretation of data and
products should be included as essential elements of climate monitoring systems
482
(GEO) as having critical deficiencies. Space agencies members should provide

mechanisms to allow the creation of these and other key global data sets. One
possible mechanism could be for multiple agencies to make their data available
and then for one or maybe two agencies to produce them globally using different
software. Inter-comparisons and validation of the results would be carried out in
close coordination with representative user groups.
6. A restrictive data policy can often significantly limit access to data and products and hence inhibit utilization. Commercial considerations and the need to
prioritize the allocation of limited ground segment resources, for example, the
planning, acquisition and delivery of high resolution data sets are legitimate reasons for restricting access to data, but, these apart, space agencies need to make
data easily and cheaply available, especially when they are in response to a specific request consistent with efforts to promote sustainable development in its
widest sense.
7. Capacity building, education and training remain high priorities in developing
and developed countries and will remain a significant responsibility of space
agencies. Experience from the case studies shows that these are often most effective in transferring capacity to local users, when performed within the context of
specific projects, when the support of key decision makers is included and when
working in partnership from the very beginning with the end users.
8. The case studies showed that creating long-term consistent records of the environment is of the highest importance for many key users of remotely sensed data:
it is also extremely challenging for agencies responsible for their creation. As one
step in this direction, it was recommended therefore that space agencies make
best efforts to adopt the GCOS Climate Monitoring Principles (Table 18.4) and
this was subsequently agreed to by the members of CEOS at their 17th Plenary
Meeting.
Acknowledgements The study was initiated by Mr. Greg Withee, when he was chair of the Committee on Earth Observation Satellites as part of an initiative to give priority to encouraging and
facilitating the increased use of space date, products and services. He set up an ad hoc Team on Utilization comprising the authors of this chapter and they gratefully acknowledge his contributions
and encouragement during the Teams work.
The following people gave generously of their time to the case studies and in providing
more general advice: Masahiro Baba (Kumamoto City Environmental Research Institute, Japan),
Philippe Bally (Spot Image), Wu Bingfang (Institute of Remote Sensing Applications, China), Anthony Busalacchi (World Climate Research Program, Climate Variability and Predictability CLIVAR), Wei Cai (China Center for Resources Satellite Data and Application), Howard Cattle (World
Climate Research Program, Climate Variability and Predictability CLIVAR), John Eyre (Meteorological Office), United Kingdom, Yves Govaerts (EUMETSAT), Lujun Guo (National Satellite
Meteorological Center of China), Shifeng Huang (IWHR, China), Gregg Jacobs (Naval Research
Laboratory (NRL)), Anthony Janetos (The H. John Heinz III Center for Science, Economics and
the Environment), Dave Jones (StormCenter Communications, Inc.), Chris Justice (Department of
Geography, University of Maryland), K.R.S. Murthi (Indian Space Research Organization ISRO),
Jeffrey Key (NOAA, National Environmental Satellite, Data and Information Services, NESDIS),
Dieter Klaes (EUMETSAT), Pierre Yves Le Traon (Space Oceanography Division, CLS), Fing Liu
(NOSAC), Guoping Liu (Institute of Remote Sensing Applications, China), Haolu Ma (Information Center of the Yellow River Conservancy Center, China, Agostino Miozzo (Department of Civil
483
Protection, Italy), Hiromi Nakamura (Hydrographic and Oceanographic Department, Japan Coast
Guard), Marc Paganini (Science and Applications Department ESA), Bernard Pinty (Institute for
Environment and Sustainability, European Commission, Joint Research Centre), Nigel Press (Nigel
Press Associates), Diego Fernando Prieto (ESA), James Purdom (Italian Centre for Aerospace
Research (CIRA)), Mukund Rao (ISRO), Adrian Simmons (European Centre for Medium-Range
Weather Forecasts ECMWF), Per Erik Skrovseth (Norwegian Space Centre), Rob Sohlberg (University of Maryland), Hideo Tameishi (Japan Fisheries Information Service Center JAFIC), John
Trinder, Ray Harris and other Committee Chairmen of the International Society for Photogrammetry and Remote Sensing, Espen Volden (ESA), Tomohiro Watanabe (JAXA), Bob Winokur
(EarthSat Corporation) and members of the World Climate Research Program (WCRP) Satellite
Working Group.
Ms Judy Carrodeguas provided invaluable practical support for all aspects of the study and
work of the Data Utilization Working Group.
References
CEOS (Committee on Earth Observation Satellites) (2003)Utilization Team Report and Recommendations, 17th Plenary Meeting, Colorado Springs, Colorado, November 1920, 2003, Report CEOS/17/Utilization
Duchossois G, Sommeria G (2003) Update of Space Mission Requirements for WCRP, WCRP
Satellite Working Group Report, WCRP, Geneva
GCOS (2003) Second Report on the Adequacy of the Global Observing Systems for Climate in Support of the UNFCCC, Appendix 2, p. 57 GCOS-82; WMO TD 1143.
http://www.wmo.ch/web/gcos/gcoshome.html
Harris R (2002) Earth Observation Data Policy and Europe. Taylor & Francis, London
IGOS (2004) Report of the IGOS International Workshop, Tokyo International Exchange, February
46, 2004, JAXA, Japan (http://www.igospartners.org/docsIGOS.html)
NRC (National Research Council) (2000) From Research to Operations in Weather Satellites and
Numerical Weather Prediction: Crossing the Valley of Death, Commission on Geosciences,
Environment and Resources (CGER), National Academies Press, Washington, DC
NRC (National Research Council) (2003a) Satellite Observations of the Earths Environment: Accelerating the Transition of Research to Operations, Committee on NASA-NOAA Transition
from Research to Operations, National Academies Press, Washington, DC
NRC (National Research Council) (2003b) Using Remote Sensing In State And Local Government: Information for Management and Decision Making, National Academies Press,
Washington, DC
Chapter 19
Emerging Issues in Land Remote Sensing

Shunlin Liang, Michael Schaepman, Thomas Jackson, David Jupp,
Xiaowen Li, Jiyuan Liu, Ronggao Liu, Alan Strahler, John R. Townshend,
and Diane Wickland
Abstract This chapter summarizes the key questions and issues discussed by three
review panels in the 9th International Symposium on Physical Measurements and
Signatures in Remote Sensing (ISPMSRS) held in October 2005 in Beijing. The
panels focused on remote sensing systems and sensors, modeling and inversion, and
remote sensing applications. Some emerging issues in land remote sensing are presented, including sensor networks, modeling complex landscapes, machine learning
techniques for inversion, and spatial scaling.
Shunlin Liang and John Townshend

Department of Geography, University of Maryland, College Park, MD 20742, USA
Michael Schaepman
Centre for Geo-Information, Wageningen University, Wageningen, The Netherlands
Thomas Jackson
USDA ARS Hydrology and Remote Sensing Lab, Beltsville, MD 20705, USA
David Jupp
CSIRO Marine and Atmospheric Research, Canberra ACT 2601 Australia
Xiaowen Li
Center for Remote Sensing and GIS of Beijing Normal University and Institute for Remote Sensing
Applications, Chinese Academy of Sciences, Beijing, China
Jiyuan Liu and Ronggao Liu
Institute for Geographical Sciences and Natural Resource Research, Chinese Academy of Sciences,
Beijing, China
Alan Strahler
Department of Geography and Environment and Center for Remote Sensing, Boston University,
Boston, USA
Diane Wickland
NASA Headquarters, Washington, DC, USA
485
486
S. Liang et al.
19.1 Introduction
The 9th International Symposium on Physical Measurements and Signatures in Remote Sensing (ISPMSRS) in Beijing, China, from October 17 to 19, 2005, lasted
for 3 days, each day ending with a panel discussion. The panels built upon and
complemented each days papers and posters. Topics were not independent components but rather viewpoints pertinent to the science and activity of remote sensing
and its applications. Panel discussions and outcomes sometimes overlapped. In this
chapter, the summaries of the key questions and points discussed in the three panels
and more in-depth discussions on some emerging issues in land remote sensing are
provided.
19.2 Remote Sensing Sensors and Systems

Panel One drew together the work addressed during the day and the overviews prepared by the presenters. It was designed to address a series of topics, such as hardware and system software, engineering physics, platforms and sensor deployment,
and calibration and validation. It also addressed the modelling and integration of
systems and methods in a way that takes full account of their unique features. Careful attention was paid to the framework and scales of demand that drive the work
the end-user applications.
At the start, a series of questions was asked:
What have been the main areas of success and impact of the various systems
in applications?
What are the main points of issue or limitation for their operational use in applications?
What are new or emerging opportunities in systems and sensors?
What synergies arise from combinations, including data assimilation, of systems
and sensors?
Which systems and combinations hold the most promise for meaningful modelling and inversion?
Which systems and combinations will have the greatest impact on specific major
issues regarding applications?
After extensive discussions, it was recognized that attention to engineering issues
has paid off, with significant improvements seen in many indicators, such as signalto-noise ratios, resolutions, pointing accuracies, geometric and spectro-radiometric
integrity, and calibration. Many systems now incorporate multiple dimensions of
observation (i.e., angular, frequencies, polarizations) at a range of scales.
The panel identified a series of issues and pointed out the required actions and
opportunities:
For Passive Microwave systems: L-band is needed for better soil moisture estimation; pixel sizes are still too large (e.g., 40 km) for many applications but can
19 Emerging Issues in Land Remote Sensing
487
potentially be reduced by passive microwave and radar synergy; and soil moisture estimation theory in general needs further work.
For Radar systems: saturation in vegetation is still a big issue; more dimensions
will help with current problems; multi-dimensions can add value to radar operations; and combination with other sensors has great potentials.
For Lidar systems: there is no current space-borne canopy lidar except GLAS
(Geoscience Laser Altimeter System) on the Ice, Cloud and land Elevation Satellite (ICESat); airborne systems have proved the value of canopy lidar for many
applications; a space sensor would be an important step; and solving issues of
interpretation and modelling (rather than correlations) take priority.
For Thermal systems: multiple-view-angle (MVA), multispectral, and hyperspectral data are needed; emissivities are interesting yet difficult to handle; emissivity
and MVA effects have been identified and need attention (e.g., new hardware);
evaporation still is not a standard product.
For Hyperspectral systems: building upon the Hyperion experience in operational
system design, spectral coverage of wavelengths shorter than 400500 nm and
longer than 2,000 nm is essential and must be added; true spectroscopy can add
significant and valuable dimensions for land applications but demands small pixels for best results; the community is ready for a Hyperion follow-up; and the
opportunities for spectroscopy of the land environment have been established.
One particular emerging issue is the sensor-web or sensor networks. Earth environments are exceptionally dynamic and interrelated through a number of systems. To
understand Earth systems and dynamics, significant improvements in spatial and
temporal observations are required. The consensus is that fusion of data from multiple sensor systems is currently one of the optimal solutions. However, increasingly inexpensive, yet sophisticated, chips for the computer and telecommunication
industries provide a unique opportunity to develop a distributed, heterogeneous,
and adaptive observing system or sensor web in the future. The sensor networks
can be used for various applications (e.g., health, military, home) (Akyildiz et al.,
2002).
Hart and Martinez (2006) claimed that Environmental Sensor Networks (ESN)
will become a standard research tool for future Earth system and environmental science. ESNs have evolved from passive logging systems (requiring manual
downloading) into intelligent sensor networks that comprise a network of automatic
sensor nodes and communications systems. These systems actively communicate
data to a sensor network server for integration with other environmental datasets.
The sensor nodes can be fixed or mobile, and range in scale appropriate to the
environment being sensed. Over 50 representative examples with different scales
and functions were reviewed. Lemmerman et al. (2005) conducted the cost analysis by focusing on one possible approach based on an LEO (low Earth orbiting)
constellation composed of 100 spacecraft. Kung et al. (2006) described a Drought
Forecast and Alert System (DFAS), which is a four-tier system framework composed of mobile users, ecology monitoring sensors, integrated service server, and
intelligent drought decision system. DFAS combines wireless sensor networks, embedded multimedia communications, and neural network decision technologies to
488
S. Liang et al.
effectively achieve the forecast and alert of the drought. Paul (2006) described a
new initiative in the defense sector on developing the smart sensor web - an intelligent, web-centric distribution and fusion of sensor information that provides greatly
enhanced situational awareness, on demand, to fighters at lower echelons. Emphasis
is on multi-sensor fusion of large arrays of local sensors, joined with other assets,
to provide real-time imagery, weather, targeting information, mission planning, and
simulations for military operations on land, sea, and air.
Substantial progress has been made to address some of the technical issues (e.g.,
Akyildiz et al., 2002; Clark and Fearn, 2006). Although technological advances have
facilitated these changes, it is vital to continue their use and exploration.
19.3 Modeling and Inversion

The panel focused on the following questions:
What is the role of system design in making observations to invert models?
Should the issue be hardware-driven or parameter-driven?
Can better use of spatial information inherent in remote sensing be made? Are
assumptions that each pixel is independent of the next valid and required? If so,
why?
Can the estimation of error structures in retrievals be improved? How are partial
derivatives and standard errors of observations best merged?
Can error estimation lead to better validation scenarios?
Is it better to design models from the bottom up (physics first) or the top down
(parameters first)? What is the role of empirical parameters in this process?
Is a probabilistic framework the best way of incorporating ancillary information?
Is setting limits or bounds on parameters a valid approach?
Concentrating on the above questions, the panel (1) examined primarily optical and
thermal domains, with a look at microwave and optical synergy; (2) discussed two
main approaches to model-based inversion and retrieval of physical parameters:
(a) forward modeling to fit observed data, either iteratively or with look-up tables
(LUTs); and (b) using models to teach neural nets and regression models to retrieve
parameters; (3) noticed that many inversion techniques involve multiple domains:
spectral, spatial, temporal, directional; and (4) looked at how to use prior information to enhance retrievals with successful examples in the optical domain.
The panel identified a series of common themes:
Ill-posed nature of the inversion problem: models typically have more variables
than data dimensions; correlation among model parameters is common.
Information dimensions: spectral, spatial, temporal, directional; expand the problem by adding dimensions to solve it.
Ancillary Information: not all combinations of parameters are encountered in nature or in sensing scenarios; how should the weight given to ancillary information
be determined?
489
Empirical vs physical models: empirical models or parameters require training

data or calibration.
Simplicity vs complexity in models: trade-off between better description of model
physics and easier to calculate, but less accurate retrieval of parameters.
Several critical issues were extensively discussed in the symposium and are further elaborated here, including modeling, machine learning techniques for inversion,
data fusion and assimilation, and spatial scaling.
A major purpose of modeling surface radiation in remote sensing is to understand the signals observed by remote sensors, and eventually to use such models
for estimating land surface variables. Some of the models are developed mainly for
understanding the radiation field of the landscape and linking with other environmental modeling. Traditional numerical models are usually deterministic ones based
on first principle equations (Clark and Gelfand, 2006). They are formulated using relevant first principles and observational data, and are usually based on solving
deterministic equations (such as RT equations) and some secondary empirical components based on traditional statistical techniques like regression. Landscape complexity often demands sophisticated models, but inversion of land surface variables
from remotely sensed data has to rely on simple models. For example, inversion
of canopy properties often uses the plane-parallel RT models that do not describe
the actual landscape well. Quantitatively, what are the impacts of the model error
(uncertainty) on inversion of satellite data? There is growing demand for model
outputs to be associated with some sort of uncertainty estimates. Including the stochastic components in the model is one direction to pursue. Stochasticity stands
in for complexity. By making models stochastic and inferential, complexity can be
represented in ways that vastly simplify modeling (Palmer et al., 2005). One solution for developing the hybrid models based on combining deterministic modeling
and machine learning components may help to achieve this goal (Krasnopolsky and
Fox-Rabinovitz, 2006). The approaches based on the Monte-Carlo sampling of the
key parameters could be another solution to determine the model output uncertainty,
although they are computationally very expensive, and the biases due to the structure
deficiency also need to be characterized. Convincing arguments must be put forth in
support of large-scale field measurement campaigns for calibration and validation
of inversion-oriented models.
Machine learning techniques, such as artificial neural network (ANN), support
vector machine, and regression tree, have been widely used in many applications,
however, they are often considered to be black-box approaches that do not reflect
the physics of the modeled process in remote sensing (Krasnopolsky and Schiller,
2003; Loyola, 2006). Given the accumulated vast data and a priori knowledge, the
models and inversion algorithms must rely on effective learning procedures. With
the rapid development of these techniques, their potential needs further exploration.
For example, neural networks have been combined into multi-neural systems with
either redundant or modular elements. The combination of neural networks more
effectively solves complex tasks and can furthermore result in dramatic performance improvement. It fully exploits the advantages of neural methods and expands
the range of applicability. An important drawback of many ANNs is their lack of
490
S. Liang et al.
explanation capability. It is increasingly apparent that, without some form of explanation capability, the full potential of trained ANNs may not be realized. Many
studies have focused on mechanisms, procedures, and algorithms designed to insert
knowledge into ANNs (knowledge initialization), extract rules from trained ANNs
(rule extraction), and utilize ANNs to refine existing rule bases (rule refinement)
(Andrews et al., 1995; Ding and Xin, 2006; Lofstrom et al., 2004; Nunez et al.,
2006; Saad and Wunsch, 2007). The land remote sensing community should pay
close attentions to these developments.
Given the ill-posed nature of remote sensing inversion and vast expansion of the
observation data, the challenges becomes: how do we combine observations that
derive from many sources, and how do we connect observations that are specific
to location, time and setting with understanding that comes from a diverse body
of nonspecific theory? Data fusion techniques that simply register and combine data
sets together from multiple sources may not be adequate to solve our problems. Data
assimilation method allows us to use all the information available to us within a time
window to estimate various unknowns of land surface models (see Chapter 12). The
information that can be incorporated includes observational data, existing pertinent
a priori information, and, importantly, a dynamic model that describes the system
of interest and encapsulates theoretical understanding. Data assimilation has been
widely used in meteorology and oceanography, but more efforts are needed in the
land remote sensing community to explore its potential for characterizing land surface environments.
Spatial and spatiotemporal processes in the physical, environmental and biological sciences often exhibit complicated and diverse patterns across different space
time scales. Both scientific understanding and observational data vary in form and
content across scales. Remote sensing data and other observations are collected at
differing scales and resolutions, at different spatial locations, and in different dimensions. Many statistical issues are associated with combining such data for modeling
and inference. Gotway and Young (2002) gave an overview of these issues and the
approaches for integrating such disparate data, drawing on work from geography,
ecology, agriculture, geology, and statistics, with the emphasis on state-of-the-art
statistical solutions to this complex and important problem. Substantial progress
has been made recently in combining incompatible spatial data, such as multiscale
modeling using the Bayesian hierarchical framework (Wikle and Berliner, 2005), by
which the combination of such information at different spatial scales can be accomplished. Such progress has not been incorporated into remote sensing modeling and
inversion. Downscaling methods using unmixing algorithms have been extensively
reported for estimating subpixel proportions from multi-spectral and multi-temporal
remote sensing data with the traditional statistical algorithms; geostatistical methods need further exploration (Kyriakidis and Yoo, 2005; Pardo-Iguzquiza et al.,
2006).
The latest progress in modeling and inversion algorithms are recently reviewed
(Liang, 2007).
491
19.4 Remotes Sensing Applications

Panel Three identified the following factors that may lead to the disparities: data
availability (e.g., data policy, costs, inadequate information systems), need for assurance of continuity of data sets, lack of familiarity of potential users with data
sets, product quality insufficient for application requirements, terminological incongruities between remote sensing algorithms and application scientists, and lack of
investments by some space agencies in applications.
Developing successful remote sensing applications is challenging. Impediments
to success include:
Lack of user involvement early in the application development process.
Lack of technical knowledge or capability in the user.
Institutional inertia the user institution has to be interested in and willing to
change or accept new approaches to decision making. Also, the initial investment
can be difficult to secure.
Too great a gap between the underlying theory and concept for a new application
and the first recognition/acceptance of a realistic application on the part of an end
user.
Mismatches between what researchers are interested in or expect and what operational users need or can use.
Successful applications are not static but evolve as new sensors, data processing and
network technology emerge. Successful applications also require verification and
validation of products.
The scientific credibility of the observations, measurements, derived products
and model outputs must be established and maintained.
The user has the responsibility to validate the inputs to the decision-making
process, i.e., to assess the input and determine that it meets the intended need.
The panel presented some results of the CEOS (Committee on Earth Observation
Satellites) Utilization Team research based on a series of case studies from the USA,
Europe, China and Japan, and determined the critical space agency success factors
and critical user success factors. The agency success factors include:
Provision of user-specific products.
Long-term continuity of data sets.
Making resources available in a timely fashion and with a minimum of bureaucracy.
Willingness of space agencies to listen extensively to user needs, their practical
difficulties, and ultimately make accommodations.
Importance of keeping user development projects simple and building further on
reliable, successful initial projects.
Funding by space agencies needs to be linked with user funding. Turning prototypes into sustained operations is frequently unsuccessful if the external funding
from users does not continue beyond the prototype phase.
Space agencies must seek assurance that the user has sufficient interest and
capacity to maintain funding.
492
S. Liang et al.
The user success factors include:

User understanding, and preferably first-hand experience of the use of remote
sensing data and products and integration with geospatial information systems
Existence of key elements of spatial data infrastructure
Willingness to make appropriate staff available before and during the project to
work with the space agency
Having a core need (e.g., a statutory need for the information, and an ability to
define needs precisely)
Willingness and ability to help develop a realistic financial scenario for an operational system
Considering both the agency and user, case studies indicate the need for innovation
and flexibility in response to new opportunities for data products.
In stead of conducting an extensive review of specific remote sensing applications
in all areas, the application panel focused on only three areas: Land Use and Land
Cover Change (LUCC) mapping, forest monitoring in the example of Canada, and
regional agricultural monitoring and management.
After evaluating the status of LUCC mapping from remote sensing, a series of
challenging issues were identified, including:
Dynamic monitoring of LUCC with high accuracy by remote sensing

Historical retrieval of LUCC by using the data with different characteristics
Classifying not only the type but also the level and the structure of LUCC
Standardization of the LUCC products for international exchange and global
change study
Analysis of the scenarios of LUCC with coupled humannatural factors, and
modeling
Analysis of the environmental impacts of LUCC, and modeling
Various measures have been identified to monitor spatially and temporally the composition, structure and health of forests from remote sensing in the example of
Canada, including
Forest area by type (coniferous deciduous, mixed), species, and disturbance (harvesting, fire, insect damage, blow down, etc.)
Amount of forest timber volume and biomass
Forest health through measurements of forest chemistry
Remote sensing applications to agriculture were reviewed. One of the emerging
issues is the integration of remote sensing and crop growth model. Remotely sensed
data substantially improves the accuracy of the regional crop growth model, but as
yet, a number of critical agronomic parameters cannot be precisely inverted from
remotely sensed data at large scale, such as harvest index and crop management
information (e.g., applied fertilizer, irrigation, and crop variety). There are also large
gaps between the promised utility and reality, for example, inaccuracy for agroparameter inversion models, ambiguity in use of various indices vs. conventional
parameters (e.g., crop growth and soil moisture), and uncertainty of the availability
of continuous time series data.
493

Because of considerable investment, remote sensing has advanced significantly, and
the societal benefits have been tremendous. Huge amounts of data available from
NASAs EOS program and other space agencies offer promise but also present great
challenges. Developing advanced sensors and systems to address application issues
is constructive, but optimal integration of observations from various sensors, as well
as other ancillary measurements, needs urgent attention. To realize integration, new
models and algorithms, such as data fusion and data assimilation, are crucial. Developing computationally simplified surface radiation models mostly suitable for inversion of land surface variables from satellite data is also urgently required. Improved
coupling of remote sensing science and applications would be most advantageous.
Acknowledgements The authors are very grateful to the panel members who provided most materials presented in this chapter. Some of the ideas also come from numerous presentations by the
symposium participants.
References
Akyildiz IF, Su WL, Sankarasubramaniam Y, Cayirci E (2002) A survey on sensor networks. IEEE
Commun. Mag. 40(8): 102114
Andrews R, Diederich J, Tickle AB (1995) Survey and critique of techniques for extracting rules
from trained artificial neural networks. Knowledge-Based Systems 8(6): 373389
Clark JS, Gelfand AE (2006) A future for models and data in environmental science. Trends Ecol.
Evol. 21(7): 375380
Clark SD, Fearn DG (2006) The impact of advanced platform and ion propulsion technologies on
small, low-cost interplanetary spacecraft. Acta Astronautica 59(811): 899910
Ding XQ, Xin S (2006) Application research on extraction of rule from artificial neural networks
for nonlinear regression. Dyn. Contin. Discrete Impuls. Syst. Ser. A Math. Anal. 13: 565568
Gotway CA, Young LJ (2002) Combining incompatible spatial data. J. Am. Stat. Assoc. 97(458):
632648
Hart JK, Martinez K (2006) Environmental Sensor Networks: A revolution in the Earth system
science? Earth-Science Reviews, 78: 177191
Krasnopolsky VM, Fox-Rabinovitz MS (2006) A new synergetic paradigm in environmental numerical modeling: hybrid models combining deterministic and machine learning components.
Ecol. Model. 191(1): 518
Krasnopolsky VM, Schiller H (2003) Some neural network applications in environmental sciences.
Part I: forward and inverse problems in geophysical remote measurements. Neural Networks
16(34): 321334
Kung HY, Hua JS, Chen CT (2006) Drought forecast model and framework using wireless sensor
networks. J. Inform. Sci. Eng. 22(4): 751769
Kyriakidis PC, Yoo EH (2005) Geostatistical prediction and simulation of point values from areal
data. Geogr. Anal. 37(2): 124151
Lemmerman L, Raymond C, Shotwell R, Chase J, Bhasin K, Connerton R (2005) Advanced platform technologies for Earth science. Acta Astronautica 56(12): 199208
Liang S (2007) Recent developments in estimating land surface biogeophysical variables from
optical remote sensing, Prog. Phys. Geogr. 31(5): 501516
494
S. Liang et al.
Lofstrom T, Johansson U, Niklasson L (2004) Rule extraction by seeing through the model, Neural
Information Processing. Lecture Notes in Computer Science 3316, pp 555560
Loyola RDG (2006) Applications of neural network methods to the processing of Earth observation
satellite data. Neural Networks 19(2): 168177
Nunez H, Angulo C, Catala A (2006) Rule-based learning systems for support vector machines.
Neural Process. Lett. 24(1): 118
Palmer TN, Shutts GJ, Hagedorn R, Doblas-Reyes E, Jung T, Leutbecher M (2005) Representing
model uncertainty in weather and climate prediction. Annu. Rev. Earth Planet. Sci. 33: 163193
Pardo-Iguzquiza E, Chica-Olmo M, Atkinson PM (2006) Downscaling cokriging for image sharpening. Remote Sens. Environ. 102(12): 8698
Paul JL (2006) Smart Sensor Web: tactical battlefield visualization using sensor fusion. IEEE
Aerosp. Electron. Syst. Mag. 21(1): 1320
Saad EW, Wunsch DC (2007) Neural network explanation using inversion. Neural Networks
20(1): 7893
Wikle CK, Berliner LM (2005) Combining information across spatial scales. Technometrics
47(1): 8091
Index
A
AATSR, 66, 67, 85, 99, 100, 131, 249, 250
Advanced Very High Resolution Radiometer
(AVHRR), 16, 82, 98, 123, 131, 132,
137, 174, 176, 184, 215, 221, 222, 235,
249251, 264, 268, 269, 276, 331, 334,
353, 371, 372, 386, 398, 400, 405, 408,
409, 411, 424427, 430437, 447,
451455, 473
Aerodynamic temperature, 246, 248, 253,
269270, 273, 275
Aerosol optical depth, 78, 79, 82, 84, 229
Albedo, 6, 3032, 4042, 52, 83, 9698,
102, 108, 112, 115, 116, 126, 128, 131,
155, 159, 161, 162, 184, 204, 205, 207,
220228, 231237, 239, 248, 264, 274,
320
Anisotropy factor, 56
ASTER, 99, 249251, 261, 264, 266269, 275,
276, 425
Atmospheric correction, 13, 67, 68, 70, 7880,
8284, 128, 219, 221, 223, 229, 234,
248, 252, 255, 261, 263265, 334, 356
Automatic differentiation (AD), 321, 323, 324
B
Bayesian network (BN), 213214
Bidirectional reflectance factor (BRF), 55, 65,
96, 102, 113, 128, 137, 160, 177, 205,
222, 227, 228, 231
Bidirectional reflectance distribution function
(BRDF), 1, 56, 71, 96, 97, 99, 102, 106,
112, 115, 118, 119, 121, 125, 131, 132,
134, 137, 183, 191, 204207, 209, 214,
222, 224, 225, 232, 258, 376
Bihemispherical reflectance (BHR), 96, 101,
137, 222, 226228, 232, 236
Biomass, 3, 54, 106, 123, 164, 246, 318, 320,

333, 341, 402404, 406, 407, 425, 428,
432434, 437, 492
Brightness temperature, 11, 15, 51, 53, 55, 57,
60, 61, 64, 6669, 7173, 7982, 247,
248, 251, 252, 254258
C
Canopy openness, 112114, 125
Carbon cycle, 54, 148, 317, 321, 331, 332,
334, 369, 370, 381, 389, 423, 424, 426,
431, 432, 434, 435, 437439
CCD, 65, 104, 132, 137
Change detection, 26, 28, 342, 345, 346, 353,
356359, 363
CHRIS, 98, 100, 102, 103, 113, 119, 120,
131133, 136, 137
Classification, 10, 23, 26, 27, 29, 30, 38, 95,
110, 123, 125, 128, 134, 195, 208, 223
Clumping index, 71, 114, 115, 117, 137
Cost function, 176, 180182, 185, 188190,
204206, 232, 316, 317, 321, 323, 324,
329
Crop yield, 332324, 401407, 415
D
Data assimilation (DA), 196, 207, 210, 212,
213, 216, 273, 313, 331, 332, 379, 386,
405, 413, 415, 459, 479, 486, 490, 493
Data fusion, 14, 16, 234, 294, 297, 372, 384,
385, 388, 389, 401, 415, 477, 489, 490,
493
Decision tree classifier, 29, 374, 376, 399, 400
DempsterShafer theory, 377
E
Emissivity, 53, 5557, 69, 71, 74, 76, 77, 79,
80, 82, 247, 248, 251, 252, 254, 255, 257
495
496
Energy balance, 52, 59, 135, 174, 248, 270,
273, 317, 330, 331, 410, 447
Enhanced vegetation index (EVI), 376380,
383, 386, 401, 406, 423, 427, 430, 431,
435, 439
Ensemble emissivity, 248, 252, 254, 255, 257,
260, 264, 270, 273, 275
Evapotranspiration (ET), 52, 174, 247, 330,
413, 426, 445, 454
Evidential reasoning, 377, 379381, 384, 385,
387389
F
Filtering, 96, 123, 232, 294296, 322, 325, 327
Foliage temperature, 53, 54, 58, 59, 69, 71, 85
fPAR (= FAPAR = FPAR), 102, 108, 110,
119, 132, 138, 148, 163166, 177, 179,
182, 191, 196, 204, 318, 320, 374, 375,
403, 404, 423, 426428, 434, 439, 449,
452
Fractional cover, 52, 73, 74, 85, 98, 120, 122,
456
Fuzzy set, 377, 387
G
Gaussian distribution, 187, 188, 317, 324
Geometric optics, 2, 70, 163, 252, 258
Geostationary, 99, 131, 219, 220, 222,
231235, 237, 239, 249, 251, 266, 274,
294
Gibbs distribution, 302
Goniometer, 51, 6163, 86
H
Hemispherical directional reflectance factor
(HDRF), 96, 101, 138, 226, 227, 229
Hot spot, 71, 98, 104, 119, 133, 152, 156, 183,
263, 273
Hyperspectral, 97, 98, 103, 104, 107, 149, 189,
196, 211, 250, 251, 261, 265, 269, 294,
298, 487
I
Ill-posed, 157, 180, 183185, 187, 203, 216,
246, 259, 264266, 268, 313, 320, 488,
490
Image fusion, 293297
Instantaneous field of view (IFOV), 13, 58, 61,
62, 65, 72, 99, 113, 114, 134, 138
K
Kalman filter, 321, 322, 325, 326, 413
Index
L
Land cover, 10, 28, 47, 53, 76, 78, 86, 95, 115,
116, 123, 125, 131, 135, 164, 207209,
215, 220224
Land surface temperature (LST), 10, 55, 246,
249, 253, 273, 317, 318, 320, 330, 331,
386, 411, 412, 423, 427, 430, 435, 439
Land use, 76, 208, 211, 214, 215, 260, 265,
341, 342, 344346, 350, 352, 354356,
358
Latent heat flux, 52, 329, 330, 410
Leaf area index (LAI), 75, 98, 102, 109, 110,
119, 120, 138, 148, 159, 165, 174, 204,
213, 276, 318, 379, 406, 423, 426, 449
Lidar, 3, 106, 107, 109, 114, 118, 135, 136,
138, 221, 320, 424, 487
Lookup table (LUT), 109, 120, 135, 157, 180,
182, 191, 259, 275, 334
M
Maximum likelihood classifier, 125, 358
Meteosat, 219221, 231, 233237, 239
Meteosat Second Generation (MSG), 131, 138,
235, 249, 250, 276, 473
Minimization, 84, 180, 298, 302, 305, 323, 375
MISR, 98102, 107110, 112, 118, 121129,
131133, 136, 138, 182, 184
Mixture model, 52, 72, 371, 401
MODIS, 98100, 102, 105, 115, 117, 118, 122,
124, 131, 132, 136, 138, 166, 174, 177
MODTRAN, 75, 77, 190, 191, 252, 258, 276
Monte Carlo, 39, 135, 153155, 161, 180, 182,
252, 258, 327, 332, 345, 450, 489
N
National Polar Orbiting Earth Satellite System
(NPOESS), 14, 16
Net ecosystem production (NEP), 54, 332, 433
Net primary production (NPP), 106, 204, 332,
374, 375, 406, 423430, 432435, 437,
439, 473
Net radiation, 53, 59, 410, 445, 447
Neural network (NNT) (= ANN), 107, 109,
148, 177, 178, 204, 213, 259, 261, 262,
322, 326, 327, 341, 401, 487, 489
Normalized difference vegetation index
(NDVI), 117, 123, 138, 158, 177,
186188, 227228, 276, 406, 409, 411,
423, 425
O
Objective function (= cost function), 176,
180182, 185, 188190, 204206, 232,
294, 298, 302, 307, 310, 315317,
321324, 329
Index
Optimization, 121, 157, 180182, 189191,
194, 204, 259, 293, 294, 302, 306, 310,
315, 320324, 332, 334, 414
P
Phenology, 105, 226, 376, 382, 385, 386, 398,
405407, 415, 423, 425, 431, 451, 457,
458
Photosynthetically active radiation (PAR), 97,
98, 102, 110, 119, 148, 149, 204, 227,
318, 404, 423, 426
Plant functional type (PFT), 369, 370, 374,
382, 449, 452, 457
Planck function, 73
Probability distribution function (PDF),
186188, 205, 207, 232
Polarization, 10, 11, 2028, 3038, 41, 44, 46,
47, 98, 104, 221, 413, 486
POLDER, 98, 100, 104, 105, 115, 129,
131133, 136, 138, 174, 177, 178, 184,
191, 204, 205, 221, 222
A posteriori, 189, 205, 305, 316
A priori, 3, 38, 80, 98, 180, 189, 203208, 211,
212, 216, 223, 231, 259, 265, 266, 269,
298, 301, 307, 310, 313, 315, 334, 383,
459, 489, 490
PROSPECT, 75, 119, 120, 183, 191
p-theory, 160162, 166
R
Radiation Transfer Model Intercomparison
(RAMI), 110, 111, 138, 157
Radiative transfer (RT), 11, 12, 15, 39, 51,
59, 7072, 75, 78, 80, 83, 84, 86, 106,
109111
Radiometric temperature, 52, 246, 248, 252,
254, 257, 259, 260, 262, 264270,
273275
Radiosity, 71, 119, 153, 154, 258
Recollision probability, 110, 160162
Regularization, 185, 293, 301
S
Scaling, 150, 157, 161163, 203, 215, 216,
328, 351, 450, 489
497
Single scattering albedo, 155, 159, 184
Snow density, 20, 30, 32, 33, 3544
Snow water equivalence (SWE), 19, 20, 23,
3335, 39, 40, 42, 4447
Soil moisture, 916, 23, 28, 29, 53, 97, 98,
130, 150, 247, 262, 274
Soil Moisture Ocean Salinity (SMOS), 11, 15,
16, 413
Soil temperature, 5355, 58, 59, 62, 70, 71, 75,
78, 87, 272, 320, 412, 445
Soilvegetationatmosphere transfer (SVAT),
75, 248, 262, 277, 317, 330
Solar radiation, 57, 58, 63, 76, 81, 135, 147,
148, 150, 152, 403, 404, 426, 429, 433,
434, 446449, 452, 454, 455, 459
Spinning Enhanced Visible and Infra-Red
Imager (SEVIRI), 131, 138, 249251,
268, 277
Split-window method, 67, 78, 80, 82
SSM/I, 10, 13, 14, 16, 408
Stochastic inversion, 154, 274
Support vector machine (SVM), 125, 489
Synthetic aperture radar (SAR), 1935, 3739,
4146, 106, 107, 138, 398, 400, 414,
473, 477
T
Two-stream, 152, 153, 156, 450, 452, 453
V
Validation, 13, 34, 78, 86, 87, 103, 121, 150,
165, 196, 204, 207, 211, 215, 229
Vegetation continuous fields (VCF), 105, 123,
124, 138
Volumetric scattering, 117, 256
W
Water balance, 428, 446, 448
Water vapor, 10, 51, 58, 59, 72, 76, 79, 8183,
190, 234, 254, 257, 261, 262, 265, 332
Wavelet, 55, 297, 342346, 350352, 354,
355, 358

BOOK - Advances in Land Remote Sensing PDF

Încărcat de

Informații document

Titlu original

Drepturi de autor

Formate disponibile

Partajați acest document

Partajați sau inserați document

Opțiuni de partajare

Vi se pare util acest document?

Este necorespunzător acest conținut?

Drepturi de autor:

Formate disponibile

BOOK - Advances in Land Remote Sensing PDF

Încărcat de

Drepturi de autor:

Formate disponibile

Shunlin Liang

Advances in Land Remote Sensing

Advances in Land Remote

Library of Congress Control Number: 2007940919

International Scientific Committee

Recent Advances in Land Remote Sensing: An Overview . . . . . . . . . .

Part I Remote Sensing Systems

Passive Microwave Remote Sensing for Land Applications . . . . . . . . .

Active Microwave Remote Sensing Systems and Applications

Multi-angular Thermal Infrared Observations

Terrestrial Applications of Multiangle Remote Sensing . . . . . . . . . . . . 95

Part II Physical Modeling and Inversion Algorithms

Modeling the Spectral Signature of Forests: Application of Remote

Estimating Canopy Characteristics from Remote Sensing

Knowledge Database and Inversion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203

Retrieval of Surface Albedo from Satellite Sensors . . . . . . . . . . . . . . . . 219

Modeling and Inversion in Thermal Infrared Remote Sensing

Spectrally Consistent Pansharpening . . . . . . . . . . . . . . . . . . . . . . . . . . . 293

Data Assimilation Methods for Land Surface Variable Estimation . . 313

Methodologies for Mapping Plant Functional Types . . . . . . . . . . . . . . 369

Part III Remote Sensing Applications

Monitoring and Management of Agriculture with Remote Sensing . . 397

Remote Sensing of Terrestrial Primary Production

Applications of Terrestrial Remote Sensing to Climate Modeling . . . . 445

Improving the Utilization of Remotely Sensed Data . . . . . . . . . . . . . . . 465

Emerging Issues in Land Remote Sensing . . . . . . . . . . . . . . . . . . . . . . . 485

Recent Advances in Land Remote Sensing:

Earths surface is undergoing rapid changes due to urbanization, industrialization

1 Recent Advances in Land Remote Sensing: An Overview

1 Recent Advances in Land Remote Sensing: An Overview

In Chapter 17, Dickinson discusses applications of terrestrial remote sensing to

Remote Sensing Systems

Passive Microwave Remote Sensing for Land

2.2 Satellite Passive Microwave Sensor Systems

2.2.1 Frequency Selection

Table 2.1 Earth variables derived from AMSR-E

Total integrated water vapor over the ocean

2 Passive Microwave Remote Sensing for Land Applications

Fig. 2.1 Sensitivity of

function of frequency (although) there is a decrease in sensing depth with increasing

2.2.3 Viewing Geometry

2.2.4 Spatial Resolution

2.2.5 Temporal Coverage

2.2.6 Signal to Noise

2 Passive Microwave Remote Sensing for Land Applications

2.3 Selected Conical Scanning Instruments

2.4 New Directions and Missions

2 Passive Microwave Remote Sensing for Land Applications

2 Passive Microwave Remote Sensing for Land Applications

Active Microwave Remote Sensing Systems

3.1.1 Current Available Satellite SAR Data

3 Active Microwave Remote Sensing Systems and Applications to Snow Monitoring

Table 3.1 Characteristics of current available spaceborne SAR

3 Active Microwave Remote Sensing Systems and Applications to Snow Monitoring

3.1.2 SAR Measurement Properties

3.2 Snow Mapping with SAR

3 Active Microwave Remote Sensing Systems and Applications to Snow Monitoring

3.2.1 Multi-temporal Single-polarization Techniques

3 Active Microwave Remote Sensing Systems and Applications to Snow Monitoring

3.2.2 Multi-frequency and Multi-polarization Techniques

3.2.3 Repeat-pass Interferometric Technique

3 Active Microwave Remote Sensing Systems and Applications to Snow Monitoring

lake surface roughness characteristics, because of change in wind conditions (speed