Documente Academic
Documente Profesional
Documente Cultură
Editor
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
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
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
University of Maryland, USA
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
xiii
xiv
Contents
10
11
12
13
Methodologies for Mapping Land Cover/Land Use and its Change . . 341
Nina Siu-Ngan Lam
14
16
17
18
19
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
Institute of Agricultural Resources & Regional Planning, Chinese Academy
of Agricultural Sciences, Beijing 100081, China
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
Institute of Agricultural Resources & Regional Planning, Chinese Academy
of Agricultural Sciences, Beijing 100081, China
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
Department of Geography and Environment, Boston University,
675 Commonwealth Ave., Boston, MA 02215, USA
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
Department of Geography and Environment, Boston University,
675 Commonwealth Ave., Boston, MA 02215, USA
alan@bu.edu
Wanxiao Sun
Department of Geography and Planning, Grand Valley State University, USA
sunwa@gvsu.edu
Johannes R. Sveinsson
Department of Electrical and Computer Engineering, University of Iceland,
Hjardarhaga 2-6, 107 Reykjavik, Iceland
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
Department of Electrical and Computer Engineering, University of Iceland,
Hjardarhaga 2-6, 107 Reykjavik, Iceland
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
Institute of Agricultural Resources & Regional Planning, Chinese Academy
of Agricultural Sciences, Beijing 100081, China
Mingwei Zhang
Institute of Agricultural Resources & Regional Planning, Chinese Academy
of Agricultural Sciences, Beijing 100081, China
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
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,
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.
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
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
Chapter 2
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
S. Liang (ed.), Advances in Land Remote Sensing, 918.
c Springer Science + Business Media B.V., 2008
10
T.J. Jackson
Frequency
Polarization
Viewing geometry
Spatial resolution
Temporal coverage
Signal to Noise
Variable
Atmosphere
Oceans
Sea ice
Snow
Land
SMOS
Low
AMSR-E Meteorological
Satellites
Bare
Sensitivity
11
Vegetated
3
5
10
20 30
Frequency (GHz)
50
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
13
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.
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
Sens. Environ. 36:203212
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
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
S. Liang (ed.), Advances in Land Remote Sensing, 1949.
c Springer Science + Business Media B.V., 2008
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).
21
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
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 =
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.
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.
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
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.
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.
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 =
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
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.
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).
(3.5)
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.
(3.6)
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
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
g
hh
g
vv
+ e(r , k1 ) log10
g
hh
g
vv
2
(3.7)
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
+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
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.
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
43
(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
44
J. Shi
2.0
Estimated
1.5
1.0
0.5
0.5
1.0
1.5
2.0
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
(3.15)
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):
(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)
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.
47
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.
IEEE Trans. Geosci. Remote Sens. 31(5):10361043
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
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
53
54
M. Menenti et al.
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
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.
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)
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.
20
Frequency (%)
0
10
15
20
25
30
35
Frequency (%)
15
10
5
0
10
15
20
25
30
35
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
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).
59
Horizontal
Divergence
Canopy Height
NET
SENSIBLE
RADIATION
HEAT
WATER
VAPOR
Water Vapor
Water Vapor
Sensible Heat
Horizontal
Divergence
dz
dy
dx
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.
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
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
0-180 go
0-180 back
90-270 go
90-270 back
0
60
40
20
20
40
60
20
4
11 April
14:00
60
11 April
13:00
60
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
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
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
11 April
10:30
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.)
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.
66
M. Menenti et al.
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
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.)
68
M. Menenti et al.
Central wavelength
(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.
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.
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.)
(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
(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
76
M. Menenti et al.
77
Broadband emissivity
of leaf and soil
Other inputs
Cupid model
Sensor geometry
Atmospheric
profiles
MODTRAN
Canopy RT
model
Trad( ,, )
Tb0( ,, )TOC
Atmospheric RT
model
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).
79
Cloud
Atmospheric correction
for VIS / NIR / SWIR channels
Atmospheric correction
for TIR channels
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.)
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:
(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):
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.
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)
(4.9)
j
i
= R ji with R ji =
i
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).
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
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
ff (v ) = a0 (s , v ) + ai (s , v )i (s , v , )
(4.16)
i=1
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).
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.
IEEE Trans. Geosci. Remote Sens. 35:13161337
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, Ottl`e 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.
Geosci. Remote Sens. 34(2):457470
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.
IEEE Trans. Geosci. Remote Sens. 30(5):871884
Kimes DS (1980) Effects of vegetation canopy structure on remotely sensed canopy temperatures.
Remote Sens. Environ. 10:165174
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.
Remote Sens. Environ. 33:4153
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.
IEEE Trans. Geosci. Remote Sens. 32(5):10601066
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.
Geosci. Remote Sens. 37(3):14251434
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.
IEEE Trans. Geosci. Remote Sens. 41(3):719724
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, Pragn`ere 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
Mark J. Chopping
Department of Earth and Environmental Studies, Montclair State University, USA
chopping@pegasus.montclair.edu
S. Liang (ed.), Advances in Land Remote Sensing, 95144.
c Springer Science + Business Media B.V., 2008
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
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
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).
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).
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
Radiometric
Radiometric, geometric
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
Launch Date
Angular coverage
Multiangle pushbroom
imager
Description
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
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.
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.)
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.
(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.)
106
M.J. Chopping
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
25
20
15
10
5
10
15
20
25
75
100
b
Predicted Undisturbed Stand
True Undisturbed Stand
Predicted Selective Cut
True Selective Cut
20
15
10
0
0
25
50
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).
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.
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.
(5.1)
(5.2)
(5.3)
(5.4)
where:
(5.5)
(5.6)
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.)
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).
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.)
116
M.J. Chopping
(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.
117
5
4
3
2
1
0
1
0
0.2
0.4
0.6
0.8
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.
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)
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.)
122
M.J. Chopping
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
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
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/).
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
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
(b)
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
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
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
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.
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
0.8
0.6
0.4
0.2
0
0
0.2
0.4
0.6
0.8
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
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).
133
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
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
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.
IEEE Trans. Geosci. Remote Sens. 42(7):15121520
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
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
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
Chapter 6
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
S. Liang (ed.), Advances in Land Remote Sensing, 147171.
c Springer Science + Business Media B.V., 2008
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.
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.
Illumination
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
152
P. Stenberg et al.
153
154
P. Stenberg et al.
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.
Needles
Shoots
Tree crowns,
internal shoot
distribution patttern
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.
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)
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.
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.
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.
166
P. Stenberg et al.
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.
IEEE Trans. Geosci. Remote Sens. 39:10611071
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
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
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 (1992) Geometric-optical bidirectional reflectance modeling of the discrete
crown vegetation canopy: effect of crown shape and mutual shadowing. IEEE Trans. Geosci.
Remote Sens. 30:276292
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
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
Sens. Environ. 110:332343
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.
Geosci. Remote Sens. 38:23872401
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, Pragn`erea 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
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
S. Liang (ed.), Advances in Land Remote Sensing, 173201.
c Springer Science + Business Media B.V., 2008
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.
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).
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)
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)
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.
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.
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
n=1
N
J = RnRnRn
n=1
N
J=
n=1
N
Rn R n
Rn +R n
2
(Gobron et al., 1997)
J = (Rn R n )2
n=1
N
J=
n=1
N
Rn R n
Rn
J = n
n=1
2
Rn R n
Rn
2
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
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)
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.
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
186
F. Baret, S. Buis
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
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
NDVI
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
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).
190
F. Baret, S. Buis
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
Hapke
3 typical +
understorey
DISORD
6 biomes
TOC
LUT
measurements
prescribed at
20%
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
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
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
193
N(p, d) = d nA + p nC + d p nS
(7.10)
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
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.
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.
196
F. Baret, S. Buis
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 param`etres biophysiques des couverts vegetaux
par inversion de mod`eles 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.
Remote Sens. Environ. 35:161173
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, Pragn`ere 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
Sens. Environ. 34:7591
Jacquemoud S, Baret F, Hanocq JF (1992) Modeling spectral and directional soil reflectance.
Remote Sens. Environ. 41:123132
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
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):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.
Remote Sens. Environ. 27:157167
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
Sens. Environ. 55:95107
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.
IEEE Trans. Geosci. Remote Sens. 34(5):12541265
Weiss M, Baret F, Myneni R, Pragn`ere 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
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
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).
206
J. Wang, X. Li
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).
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
209
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.
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).
212
J. Wang, X. Li
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.
214
J. Wang, X. Li
Ancillary
Parameter space
Data
(8.3)
215
(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
References
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(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
physically-based models. Remote Sens. Rev. 18:381439
Kuusk A (1991) Determination of vegetation canopy parameters from optical measurements.
Remote Sens. Environ. 47:194202
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
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).
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.
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
0.72
0.63
Surface Albedo
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
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.
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
0.20
0.15
0.10
0.05
80
60
40
20
20
40
60
80
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
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
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).
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.
Remote Sens. Environ. 65:367375
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 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. 30:276292
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
Sens. Environ. 76:213238
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.
IEEE Trans. Geosci. Remote Sens. 36:12121227
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.
IEEE Trans. Geosci. Remote Sens. 40:15201531
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.
Geosci. Remote Sens. 43:144158
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
surface for the correction of remote sensing data. J. Geophys. Res. 97:2045520468
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
bidirectional reflectance. J. Geophys. Res. 100:2107721090
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
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
S. Liang (ed.), Advances in Land Remote Sensing, 245291.
c Springer Science + Business Media B.V., 2008
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
247
248
F. Jacob et al.
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.
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
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
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
Close nadir
VZA
TIMS
(multispectral)
15 m
Across track
VZA: 38
DAIS
(multispectral)
15 m
Across track
VZA: 26
MAS / MASTER
(multispectral)
15 m
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
Nadir VZA
Hyperspectral FTIR
BOMEM suite
Few cm
Goniometric
systems
Few cm
Airborne
Ground-based
VZA [090 ]
VAA [0360 ]
251
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
Eq. 10.6
[81, 178]
Atmosphere
Brightness temperature
(Atmospheric corrections)
Composite surface
radiative transfer
Eq. 10.7
[12, 158]
Land surface
Eq. 10.9
[125, 126]
Eq. 10.10
[68, 127]
Land surface
Kernel-driven
radiative transfer
Eq. 10.11
[128, 129]
Land surface
Ensemble emissivity
Soil and vegetation temperatures
Radiative transfer
MODTRAN [134]
Atmosphere
Brightness temperature
(Atmospheric corrections)
Radiative transfer
Prevots [139]
SAIL [74, 137]
Land surface
Brightness temperature
Ensemble emissivity
Soil and vegetation temperatures
with sunlit and shaded components
Geometric-optics
Kimess [141]
Caselless [143]
Yus [73]
Land surface
Brightness temperature
Soil and vegetation temperatures
with sunlit and shaded components
Geometric-optics
radiative transfer
CUPID [147]
Thermo [148]
Jias [149]
DART [76, 77]
Land surface
Brightness temperature
Canopy temperature profile
Monte Carlo
ray tracing
Land surface
Brightness temperature
Ensemble emissivity
Soil and vegetation emissivities
Simulation models
253
with
(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].
254
F. Jacob et al.
C1 5
exp TC2 1
(10.4)
(10.5)
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.
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].)
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
1 + 2
+ D(1 2 ) + E (10.9)
2
(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
(10.11)
i=1
258
F. Jacob et al.
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].
M1
M1
P1
P1
..
.
.
1 .
. = F .. , [K1 Kk ] .. = F .. , [K1 Kk ](10.12)
Mm
Pp
Pp
Mm
260
F. Jacob et al.
261
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.
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.)
264
F. Jacob et al.
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].
266
F. Jacob et al.
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].
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
270
F. Jacob et al.
functionally equivalent to aerodynamic temperature (Eq. 10.3). Second, they provide information for monitoring vegetation photosynthesis and soil respiration.
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
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]
Accounts for
multiple scattering
P4
P5
[71]
Accounts for
multiple scattering
P6
[119]
P7
P8
[71]
Accounts for
multiple scattering
and cavity effect
272
F. Jacob et al.
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
274
F. Jacob et al.
275
Glossary
AATSR
ANEM
ASTER
Advanced ATSR
Adjusted NEM
Advanced Spaceborne Thermal Emission and Reflection Radiometer
276
F. Jacob et al.
ATSR
AVHRR
DA
DAIS
DAISEX
DART
DART-EB
EFEDA
ENVISAT
ERS
ETM
FIFE
FLUXNET
FTIR
GLI
GLobal Imager
HAPEX
HIRS
IASI
IRSUTE
ISLSCP
JORNEX
JORNada EXperiment
LAI
LIDF
MAS
MASTER
METOP
MIR
MODIS
MODTRAN
MSG
MTI
NDVI
NEM
NOAA
NPOESS
NTB
ReSeDA
SAIL
SALSA
SVAT
SW
SWVCR
TES
TIMS
TIR
TIROS
TISIE
TM
TOVS
TTM
VIIRS
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, Lacarr`ere 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.
Remote Sens. Environ. 61:96109
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.
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.
Geosci. Remote Sens. 42:22902304
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.
99.
100.
101.
102.
103.
104.
105.
106.
107.
108.
109.
110.
111.
112.
113.
114.
115.
116.
117.
118.
283
284
F. Jacob et al.
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
Remote Sens. 42(3):664672
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
physically-based models. Remote Sens. Rev. 18:381439
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.
Remote Sens. Environ. 83:118
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
Sens. Environ. 92:425426
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, Nicl`os 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, S`oria 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, Nicl`os R, Sanchez J, Rivas R (2005) Ground measurements for the validation of land surface temperatures derived from AATSR and MODIS
data. Remote Sens. Environ. 97:288300
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.
Remote Sens. Environ. 91:377389
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
Trans. Geosci. Remote Sens. 38:25572569
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.
Forest Meteorol. 112:2330
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.
Forest Meteorol. 94:1329
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, Bussi`ere F, Cabidoche Y, Cellier P, Debaeke P, Gaudill`ere 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
Sens. Environ. 92:535547
260. Lyons T, Halldin S (2004) Surface heterogeneity and the spatial variation of fluxes. Agric.
Forest Meteorol. 121:153165
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
Sens. Environ. 60:195216
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
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.
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.
295
296
A. Vesteinsson et al.
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
(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
A. Vesteinsson et al.
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.
299
q*F
*
A
w
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.
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)
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
A. Vesteinsson et al.
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
SP1 + SP2
.
2
(11.9)
302
A. Vesteinsson et al.
R1 = R + (P1 PL ) cos( )
R2 = R + (P2 PL ) cos( ),
(11.10)
FP , FR
= arccos
|FP | |FR |
'
FP FR dw
= arccos
.
|FP | |FR |
(11.11)
(11.12)
Vc ( f ).
cC
(11.14)
303
aspectral
Pspectral
asmooth
1/T
Psmooth
aIP
PIP
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
A. Vesteinsson et al.
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)
306
A. Vesteinsson et al.
Fig. 11.4 (a) The multispectral image. (b) The panchromatic image
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
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
A. Vesteinsson et al.
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
IHS
0.955
0.930
0.886
309
Table 11.3 Quality metrics for the image using smoothness, spectral and IP constraints using a
fixed non-zero temperature
Metric
Result
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
A. Vesteinsson et al.
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
Chapter 12
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
S. Liang (ed.), Advances in Land Remote Sensing, 313339.
c Springer Science + Business Media B.V., 2008
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
315
(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 ().
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.
(12.3)
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.
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.
318
S. Liang, J. Qin
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)
320
S. Liang, J. Qin
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.
322
S. Liang, J. Qin
(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.
(12.6)
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
x0:t
(12.7)
i = 1, . . . ,t
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
(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.
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)
(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 ) =
&
(12.13)
(12.14)
(12.15)
&
(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
(12.17)
(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).
327
(i)
(i)
(i)
(i)
(i)
(i)
p(yt+1 |xt+1 )
N
(i)
p(yt+1 |xt+1 )
i=1
(i)
(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.
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.
329
330
S. Liang, J. Qin
331
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.
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
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.
Geosci. Remote Sens. 43(1):144158
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
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.
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.
343
344
N.S.-N. Lam
References
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.
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
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)
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)
(13.8)
(13.9)
(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 =
(13.12)
i=1 j=1
ASM =
P(i, j)2
(13.13)
i=1 j=1
K
entropy =
|P(i, j)|2
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.
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.
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.
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.
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.
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.
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.
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
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
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.
Remote Sens. 69(2):133141
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.
IEEE Trans. Geosci. Remote Sens. 34(2):398404
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.
J. Remote Sens. 20:139152
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.
Remote Sens. 70(8):803812
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.
Remote Sens. 38:14841488
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
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
S. Liang (ed.), Advances in Land Remote Sensing, 369393.
c Springer Science + Business Media B.V., 2008
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
371
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.
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
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
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.
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
0.5
0.4
Shrub
0.3
Grass
0.2
Cereal Crop
0.1
Broadleaf Crop
Urban and Build-up
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
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.
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.
Evidence
generation
385
Other spatial
data
Knowledge
base
Evidential reasoning
algorithm
PFT product
Validation
386
W. Sun, S. Liang
Spatial
Texture
Spatial association
Used to separate
Input data/knowledge
Growing activity
Vegetation change, climatic
impacts
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
389
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
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
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
data. Remote Sens. Environ. 61:399409
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.
Remote Sens. Environ. 89:139162
391
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
boundary conditions from MODIS data and consequences for the albedo of a climate model.
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
Chapter 15
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.
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).
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.
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.
& 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.
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.
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
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.
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.
(15.6)
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 =
(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)
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
(15.11)
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.
(15.13)
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.
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.
414
Z. Chen et al.
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
data. Remote Sens. Environ. 16:175181
Baret F, Guyot G (1991) Potentials and limits of vegetation indices for LAI and APAR assessment.
Remote Sens. Environ. 35:161173
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.
Remote Sens. 52:183191
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
Trans. Geosci. Remote Sens. 38:911925
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.
Geosci. Remote Sens. 33:915926
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.
Geosci. Remote Sens. 37:18601870
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.
Remote Sens. 18:31453166
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.
Remote Sens. 14:14951515
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.
Remote Sens. 39:17291735
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.
Remote Sens. Environ. 93:412422
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
Myneni RB, Keeling CD, Tucker CJ, Asrar G, Nemani RR (1997) Increased plant growth in the
northern high latitudes from 19811991. Nature 386:698702
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
Trans. Geosci. Remote Sens. 39:16551661
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.
Remote Sens. Environ. 8:51168
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.
Geosci. Remote Sens. 24:1220
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
Sens. Environ. 59:234247
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.
J. Remote Sens. 15:13071316
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
Sens. Environ. 94:232243
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
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.
423
424
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
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).
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
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).
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.
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).
2.5
2
1.5
1
0.5
0
0.5
1
1.5
2
2.5
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.
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.
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.
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.
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.
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.
Remote Sens. Environ. 83:1630
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.
Geosci. Remote Sens. 44(7):19081925
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
northern high latitudes from 19811991. Nature 386:698702
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.
Remote Sens. Environ. 8:127150
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
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
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).
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.
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
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)
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.
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
35
13
80 60 40 20
20
40
60
20
15
10
5
0
80
17
20
40
60
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
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.
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.
Remote Sens. Environ. 83:214231
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
boundary conditions from MODIS data and consequences for the albedo of a climate model.
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
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
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
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.
468
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.
469
Table 18.1 Case studies where remote sensing was adopted successfully and rapidly
Name of case study
User agency
NASA
Integrated operational
system of flood monitoring
and assessment
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
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
470
User agency
14
15
Kyoto Inventory
16
17
18
Ecosystem (environmental)
monitoring
19
Ozone alert
ESA, KNMI
20
Surface albedo
21
22
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.
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.
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
474
475
476
477
478
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.
479
USERS
Forecasts
Global
System
design
In situ Data
Govt
Public
Private
D
S
Modeling,
OPS
Data
Assimilation
RTD
EC-Funded
Research in
Universities
B)
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)
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
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
482
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
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.
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.
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.
489
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
492
S. Liang et al.
493
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
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