This article was downloaded by: [81.180.20.

66]
On: 09 December 2014, At: 14:51
Publisher: Taylor & Francis
Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered
office: Mortimer House, 37-41 Mortimer Street, London W1T 3JH, UK

International Journal of Remote

Sensing
Publication details, including instructions for authors and
subscription information:
http://www.tandfonline.com/loi/tres20

A novel framework for urban mapping

from multispectral and hyperspectral
data
a

K. Bakos , G. Lisini , G. Trianni & P. Gamba

Dipartimento di Elettronica , Universit di Pavia , Pavia , Italy

IUSS , Pavia , Italy

Joint Research Centre, ISFEREA Action , Ispra , Italy

Published online: 02 Oct 2012.

To cite this article: K. Bakos , G. Lisini , G. Trianni & P. Gamba (2013) A novel framework for urban
mapping from multispectral and hyperspectral data, International Journal of Remote Sensing, 34:3,
759-770, DOI: 10.1080/01431161.2012.714502
To link to this article: http://dx.doi.org/10.1080/01431161.2012.714502

PLEASE SCROLL DOWN FOR ARTICLE

Taylor & Francis makes every effort to ensure the accuracy of all the information (the
Content) contained in the publications on our platform. However, Taylor & Francis,
our agents, and our licensors make no representations or warranties whatsoever as to
the accuracy, completeness, or suitability for any purpose of the Content. Any opinions
and views expressed in this publication are the opinions and views of the authors,
and are not the views of or endorsed by Taylor & Francis. The accuracy of the Content
should not be relied upon and should be independently verified with primary sources
of information. Taylor and Francis shall not be liable for any losses, actions, claims,
proceedings, demands, costs, expenses, damages, and other liabilities whatsoever or
howsoever caused arising directly or indirectly in connection with, in relation to or arising
out of the use of the Content.
This article may be used for research, teaching, and private study purposes. Any
substantial or systematic reproduction, redistribution, reselling, loan, sub-licensing,
systematic supply, or distribution in any form to anyone is expressly forbidden. Terms &

Downloaded by [81.180.20.66] at 14:51 09 December 2014

Conditions of access and use can be found at http://www.tandfonline.com/page/termsand-conditions

International Journal of Remote Sensing

Vol. 34, No. 3, 10 February 2013, 759770

A novel framework for urban mapping from multispectral

and hyperspectral data
K. Bakosa , G. Lisinib , G. Triannic , and P. Gambaa*
a

Dipartimento di Elettronica, Universit di Pavia, Pavia, Italy; b IUSS, Pavia, Italy; c Joint Research
Centre, ISFEREA Action, Ispra, Italy

Downloaded by [81.180.20.66] at 14:51 09 December 2014

(Received 25 February 2011; accepted 9 May 2011)

This work presents a first attempt to define a framework able to face the challenges
related to an efficient exploitation of multispectral and hyperspectral sensors in urban
areas for mapping applications. The difference in spatial and spectral information in
the data sets, the spatial scale of the materials, and the objects involved in different
instances of the urban mapping problem require semi-automatic approaches to select
the most useful features and techniques according to the final mapping application. The
framework proposed here is based on a flexible processing chain and is complemented
by some guidelines, aiming at helping practitioners to achieve a satisfying result from
different kinds of data sets and for different mapping applications.

1.

Introduction

The possibilities opened up by very high spatial resolution (VHR) multispectral data and
very high spectral resolution (hyperspectral) data for urban remote sensing are numerous.
The availability of more and more spatial detail allows the extraction of objects, e.g. buildings (Ehlers 2009), cars, and roads (Gautama et al. 2004), and improves the performance
of spaceborne data for cartographic applications. It also contributes to better management
of natural/anthropomorphic disasters, such as floods (Kux and Arajo 2008), earthquakes
(Yamazaki, Yano, and Matsuoka 2005), hurricanes (Friedland, Levitan, and Adams 2008),
and pollution (Nichol et al. 2006). Moreover, the availability of fine spectral detail helps
to characterize urban materials (Roessner et al. 2001), from asbestos (Marino, Panigada,
and Busetto 2001) to asphalt (Herold and Roberts 2005), and to assess the impact of
environmental agents, such as heat island effects (Nichol 1996).
All these applications are very interesting to urban managers, but the information coming from remotely sensed data sets is usually too detailed (in both the spatial and the spectral
sense) to be of immediate use to decision-makers. One of the best ways to convey information at their level is to use thematic maps, which help to identify objects and materials
of interest, as well as their spatial locations. Remote-sensing experts are thus challenged
to design procedures able to extract different maps in a way that is, as much as possible,
independent of the original sensor acquiring the image of the area under test.
Of course, any such procedure is highly dependent on the theme of the map, i.e. on the
features that the map should be able to highlight. The main reason is the different spatial
*Corresponding author. Email: paolo.gamba@unipv.it
ISSN 0143-1161 print/ISSN 1366-5901 online
2013 Taylor & Francis
http://dx.doi.org/10.1080/01431161.2012.714502
http://www.tandfonline.com

Downloaded by [81.180.20.66] at 14:51 09 December 2014

760

K. Bakos et al.

scales of the objects within a human settlement, from a few square metres for a car to
hundreds for urban parks. Using hyperspectral data, another equally important reason is
the variable spectral characteristics of the materials, dependent on its age, composition,
and geometrical positioning with respect to the sensor and the illuminating source (the
so-called bidirectional reflectance distribution function (BRDF)) (Schiefer, Hostert, and
Damm 2006). The mapping problem is therefore going to be in essence the search for a
methodology which, taking into account the different spatial scales, geometrical properties, and material characteristics of the elements of the urban scene, is able to classify as
automatically as possible the original data according to a given legend. It must be noted,
additionally, that the class legend in urban areas often refers to land-use classes, i.e. to
classes (such as medium-size residential or recreational areas) that do not directly refer
to a specific cover (i.e. spectral characteristics) or object (i.e. spatial scale). The only way
to reduce uncertainties and improve the final mapping result is therefore the combined use
of the spectral and spatial characteristics and the exploitation of the multiple scales and
radiances at more wavelengths, when these information are available. A combination with
suitable a priori knowledge about the elements of the scene is a plus for the approach and
allows definition of more efficient techniques and more targeted approaches, but must be
consistent with the legend of the final mapping product.
This article is devoted to analysis of a few different approaches developed for
multispectral and hyperspectral data classification in urban areas using both spectral and
the spatial information. As discussed in Gamba, DellAcqua, and Ferrari (2003), there are
advantages in considering the context in addition to the very fine spectral characterization of each pixel. The disadvantages are, however, the increased number of features to
be considered and the subsequent Hughes phenomenon, i.e. the reduction of classification
accuracy when more features are available. This work is meant to include these ideas into
a more general framework and to provide some guidelines to properly use them.
To this aim, a comparison between available methodologies and maps with different
legends and information coming from different sensors with different resolutions will
be used to illustrate the approach and characterize the improvements with respect to the
state-of-the-art technologies.
2.

Overall approach

The overall approach used in this work to analyse high-resolution spectral and spatial data
available on a given urban area is structured as shown in Figure 1. Generally speaking, the
procedure is organized in three steps, but not all of them are considered for any mapping
problem or spatial scale, and thus, the figure should be considered to be the more general
way to implement the framework discussed in this article.
The processing steps involved may be grouped as follows.

Pixel-based urban mapping: a set of algorithms considering each pixel as a standalone entity and trying to exploit as much as possible the spectral and spatial
information contained in the data vector corresponding to the pixel location in the
urban environment. These algorithms will be differently selected according to the
spectral characteristics of the remote-sensing data set(s) that are inputs to the mapping
procedure.
Context-aware urban mapping: a set of procedures to refine or include in the mapping
results the information coming from neighbouring pixels. The driver for their choice
will be the spatial resolution of the above-mentioned remote-sensing data set(s).

International Journal of Remote Sensing

761

Pixel-based mapping
Multispectral
data set
Classification
Classifier
combination

Feature
extraction

Hyperspectral
data set

Shape-based mapping

Downloaded by [81.180.20.66] at 14:51 09 December 2014

Final
map
Context-based mapping

Figure 1.

A graphical representation of the overall framework proposed in this article.

Shape-based urban mapping: a set of techniques to include in the urban mapping

procedure some a priori available information about the elements that may be recognized in the urban environment and that belong to specific mapping classes. For this
part of the framework, the driver for a specific choice will be the class legend, i.e. the
final aim of the mapping procedure.

Of course, the options available for each of these three steps are numerous. To the aim of
explaining the framework that is proposed in this article, a small selection is proposed, as
per the following list.

Pixel-based urban mapping:

a supervised spectral classifier based on adaptive resonance theory map
(ARTMAP) neural networks;

the most generally acclaimed best classifier for hyperspectral data sets, i.e.
support vector machines (SVMs);

a hierarchical binary decision tree (HBDT) classifier based on the combination

of multiple processing chains;

a probability ensemble classifier combination.

Context-aware urban mapping:
a class-aware Markov random field (MRF) adaptive classifier, which exploits
the MRF framework to incorporate the spatial context for different classes;

a spatial pattern re-classifier based on the same ARTMAP scheme exploited for
the first step, but fed only with context information.

Shape-based urban mapping:

a shape refining approach adaptive to the class and the level of regularization
required by the objects in the scene according to the spatial sampling of the
sensor grid.

762

K. Bakos et al.

In the following, a more precise discussion about different options is presented and the
criteria useful to achieve a consistent selection of algorithms for a given mapping problem
is proposed.

Downloaded by [81.180.20.66] at 14:51 09 December 2014

2.1.

Pixel-based urban mapping

First, the pixel-based analysis can and should be performed differently whether we consider multispectral or hyperspectral data as input. Multispectral data sets may be easily
analysed by more conventional supervised classifiers (such as a maximum likelihood (ML)
classifier), but in cases where spectral and possibly spatial features (e.g. textures) are considered in addition to the original spectral information, more robust and non-parametric
approaches should be considered. One example is the neural network approach based on a
Fuzzy ARTMAP architecture in Gamba and Houshmand (2001) and successfully used for
the joint exploitation of very different data sets. This approach is supervised and has proved
to be robust enough to be useful in many situations.
When the number of features becomes extremely large and the size of the training set is
small, the SVMs are instead usually considered to be the best choice. This is indeed the case
for hyperspectral data sets, even without the addition of other spectral or spatial features.
SVMs are usually employed because of their robustness to the small size of the training
set and their generalization properties. A disadvantage is that they are usually difficult and
take a long time to train, especially for inexperienced users, and results may be even worse
without a proper parameter tuning and kernel choice.
It is true, however, that many research works published in the technical literature show
that no classifier is able to perform equally well for very different mapping problems, with
different legends, and using different (i.e. more or less abundant) training sets. For instance,
hyperspectral data are usually interpreted in different research works, even referring to the
same urban area, by means of different processing chains, all of them including a feature
extraction/reduction step followed by the actual classification. A combination of multiple
chains is in the authors opinion a solution, because it makes the whole system able to
exploit the advantages of all of the chains.
In this work, we rely on a novel methodology recently introduced in Bakos and Gamba
(2011), which is able to select among multiple processing chains the one that is most
suitable for a specific map class. More specifically, the original mapping problem is disassembled into an HBDT algorithm. Each class is mapped using a (potentially different)
processing chain, starting from the class that is the easiest to discriminate against the other
ones. Iteratively following the same scheme, the specific feature extraction/classifier pair
is selected for all the classes.
The algorithm implementation requires us to design the structure of the HBDT in an
automatic way. This is done by computing the confusion matrices on a validation data set
(different from both the training set and the ground truth) for all the considered processing
chains. Following this computation, a rank list is compiled using ranking parameter Rx for
the xth class, by combining the geometric or arithmetic mean of the users Ux and producers
Px class accuracy values:
Rx = (Ux + Px )/2 or

Ux2 + Px2 .

(1)

The procedure is extremely fast, but, of course, very dependent on the training set, which
must be statistically relevant and significant for the scene. Our experiments, shown in the

Downloaded by [81.180.20.66] at 14:51 09 December 2014

International Journal of Remote Sensing

763

next section, show that with this attention, which is a basic one for trained classifiers, the
results are satisfying.
Similarly, a different methodology for multiple classifier fusion, recently introduced in
Bakos, Gamba, and Zagajewsky (2010), is also considered in this work. The approach is
based on the use of class memberships to obtain the more suitable per-class combination
of the single processing chain results. Probabilities computed using different processing
chains for each class are combined using a weight that is inversely proportional to their
rank for that class in the validation data set. The results obtained in the above-mentioned
paper and confirmed by those presented in this work show that the approach is slightly
better than the HBDT, although more computationally expensive.
Finally, it should be noted that due to the strict relationship between the feature
extraction/selection procedure and the classifier, a multiple processing chain combination
approach does not make a lot of sense for multispectral data and is recommended only
when the number of bands is sufficiently large to require the first of these two steps to
avoid the Hughes phenomenon. For multispectral data, the HBDT and class membership
approaches may have too complex results and usually provide a very small and hardly statistically significant advantage with respect to the other options discussed in the following
paragraphs.
2.2. Context-aware urban mapping
Following the graphic scheme in Figure 1, the framework proposed in this work includes
a second part, dealing with the information in the immediate neighbourhood of each pixel,
where the scale of this neighbourhood needs, of course, to be defined according to the
mapping problem.
A very powerful way to add the neighbourhood information to the pixel-based mapping
results is to exploit the MRF approach. Specifically, the MRF framework starts from the
idea that all pixels are connected in a grid, and their assignment to one class or another is
based not only on their spectral properties but also on the spectral patterns of the neighbouring pixels. A pixel and its neighbours are therefore considered to be connected in a
number of different patterns (called cliques), and the classification is performed trying to
minimize the errors in the classification, considering not only the Mahalanobis distance
between the pixel and the class representatives but also the combination of class values on
every considered clique. As a result, the final map is obtained by minimizing a weighting
(error) function based on a weighted sum of a spectral (U spectr ) and a spatial (U sp ) part,
as detailed, for instance, in Gamba et al. (2007). The minimization of this functional is
achieved by Iterated Conditional Mode (ICM), a suboptimal yet computationally efficient
algorithm, which allows the rapid reaching of a local minimum of the energy function.
The adaptive methodology added in Gamba et al. (2007) to the MRF framework for
urban areas is based on the idea that in urban area there are many sharp boundaries, where
abrupt class changes (e.g. from a roof to a road or parking lot) separate homogeneous
areas. This calls for very different scales for the cliques in both the boundary area and
the middle of the homogeneous objects. To simplify the issue, cliques can be considered
with a neighbourhood size depending on the scale of the objects in the latter cases and
totally neglected (i.e. reduced to 1-pixel patterns) in the former one. This very simple trick
improves the classification result by reducing the misclassifications in isolated pixels within
the objects, while allowing us to maintain the boundary as sharply as possible with respect
to the spatial resolution of the data.
The approach is valuable for both multispectral and hyperspectral data, but of course the
use of hyperspectral data should be coupled with an adequate spectral classifier as input to

Downloaded by [81.180.20.66] at 14:51 09 December 2014

764

K. Bakos et al.

the spatial part of the above-mentioned weighting function. Moreover, the boundary areas
are extracted by means of a directional filter, which needs to be applied to a synthesized
panchromatic image, or by means of a combination of each band, properly filtered, which
makes the approach more cumbersome and so far not completely effective.
A less refined scheme for improving the spectral-based classification considering the
context may be based on a spatial post-processing of the pixel-based classification map.
To this aim, the same training set used before may be exploited to record the classification
patterns and automatically extract rules to associate patterns with classes. Once these rules
are applied to all the pixels in the map, they are re-assigned to an output class according
to the patterns around them and thus to the spatial context, irrespectively of the meaning
(i.e. the spectral characteristics) of the class. The very simple idea proposed in Gamba and
DellAcqua (2003) is to use as a pattern the vector with the percentage of pixels belonging to
each class in a given neighbourhood of each pixel. It is clear that this approach is going to
reduce the salt and pepper classification noise, but it will also blur the edges among classes
as a function of the neighbourhood width. For this reason, it is recommended to maintain
this width as very small and adapt it to the scale of the objects in the scene. The width
values must be comparable to, and preferably lower than, the mean size of the elements of
the scene, to avoid as much as possible shape changes in the already detected objects.
2.3. Shape-based urban mapping
The last step in the proposed processing chain is aimed at a refinement of the objects
extracted from the scene according to their shape and the a priori information about that
specific class: the algorithm is implemented as a set of rules comparing the shapes of
the objects extracted by means of the previous steps with a set of predefined shapes (e.g.
rectangles) or more general structures (e.g. right corners).
Specifically for buildings, the requirements to be regular may be implemented into two
different rules:
(1) if the area difference between a building shape and the best fitting rectangle is lower
than a predetermined value, than the building is approximated by that rectangle;
(2) if the threshold is exceeded, then a more refined regularization approach based on
the definition of the two major directions of the shape, followed by a reduction to
90 of the angles and the elimination of small protruding parts of the objects, is
performed (more details can be found in Gamba et al. 2007).
Similar rules may also be implemented for roads, where the usual assumption of having two parallel borders may be forced to improve the overall mapping reliability and
robustness.
3.

Experimental results

The approach discussed in this article has been tested extensively using different data
sets in different geographical areas. A first test has been performed using high-resolution
hyperspectral data from the Airborne Visible-Infra Red Imaging Spectrometer (AVIRIS)
sensor by NASA, which recorded data in 224 bands in 2008 and 2009 for the area around
the Moffet Federal Airfield, California, USA. The pixel spacing here is 20 m, which makes
the data set a rather coarse one for urban mapping. As a matter of fact, the urban areas of
Mountain View and Sunnyvale (bottom) as well as Newark and Fremont (upper right) are

International Journal of Remote Sensing

765

Bare soil
Urban area
Water
Vegetation

Downloaded by [81.180.20.66] at 14:51 09 December 2014

(a)

(b)

Figure 2. Moffet Federal Airfield (California, USA) test area: (a) false colour AVIRIS image; (b)
reference map with the corresponding colour legend.

recognizable, together with water surfaces, vegetation, and bare soil. The whole 2008 flight
line, kindly provided by the NASA Jet Propulsion Laboratory (JPL) AVIRIS team, is shown
in Figure 2(a) in a false colour image, whereas the reference image used to evaluate the
mapping results is shown in Figure 2(b).
The second test was recorded instead by the airborne Reflective Optics System Imaging
Spectrometer (ROSIS) sensor by DLR over the town of Pavia, Italy. This sensor was flown
in the framework of the HySens project, managed by Deutschen Zentrum fr Luft- und
Raumfahrt (DLR) and sponsored by the European Union within the transnational access to
major research infrastructure in July 2003. The sample of the Pavia data set analysed in this
work refers the city centre and, thanks to the fine spatial resolution (around 1.3 m), clearly
shows a typical scene for an Italian historical town: buildings, very close to one another and
with very similar roof materials; a small vegetation fraction, mainly pertaining to gardens
within palaces or to small parks; and finally, mixed stone/asphalt-paved roads. The area
is depicted using ROSIS data in Figure 3(a), whereas the ground map used to validate the
results is shown in Figure 3(b).

Shadow
Buildings
Roads
Vegetation

(a)

(b)

Figure 3. Pavia data set: (a) true colour ROSIS image of the area analysed in this work; (b) reference
map with the corresponding colour legend.

766

K. Bakos et al.

In both cases, the processing chain computational load depends heavily on the different
options. The most straightforward chain, applying ARTMAP classification and spatial reprocessing as well as the final shape refinement may last 5 minutes on the Pavia data set
using a PP4 processor running at 2 GHz with 1 GB of RAM, with an even distribution of
the time among the steps. Instead, on the same data set, the MRF-based approach and the
HBDT classifier may require 1015 minutes each, and the probability ensemble classifier
even more, around 2030 minutes.

Downloaded by [81.180.20.66] at 14:51 09 December 2014

3.1.

Urban mapping for the Moffett Airfield test site

The coarse resolution of the AVIRIS data does not allow us to discriminate among different land uses, but only simple land-cover classes. With such a simplified class legend, and
almost no attention to the details (as shown in the reference map used), one may expect
that pixel-based mapping followed by a simple spatial re-processing is the best solution to
achieve a reasonable result. Moreover, since the mapping product is relatively simple, and
the urban area is significantly wide in the scene, a complex fusion procedure is likely to
be useless and not effective. As a matter of fact, even the visual comparison between the
results shown in Figures 4(a) and (b), depicting the results of SVM classification as opposed
to HBDT combination of SVM applied after principal component analysis (PCA) and minimum noise fraction (MNF) transforms and feature reduction (arbitrarily to 15 bands),
reveals that no advantage comes from using the more complex combination procedure.
A quantitative evaluation confirms that the 83.7% overall accuracy of the first result is not
statistically different from the 83.4% accuracy of the second one. Spatial re-processing of
the first map provides instead a map Figure 4(c) with an overall accuracy of 85.8%. For this
spatial re-processing, a conveniently wide neighbourhood was considered (5 5 pixels),
corresponding to 100 100 m2 .
3.2. Urban mapping for the Pavia test site
For the Pavia test site, the fine spatial resolution of the data and the detailed mapping legend
to be matched allow evaluation of all of the steps in the proposed framework. Specifically,

(a)

(b)

(c)

Figure 4. Classification maps for the Moffet Airfield test area: spectral classification results using
(a) SVM; (b) the BHDT approach; (c) MRF after SVM.

International Journal of Remote Sensing

Downloaded by [81.180.20.66] at 14:51 09 December 2014

(a)

(b)

767

(c)

Figure 5. Classification maps for the Pavia test area: spectral classification results using (a) Fuzzy
ARTMAP; (b) the BHDT approach; (c) the class membership approach for the fusion of multiple
chains.

the hyperspectral data sets require a more sophisticated processing chain to extract satisfying results using a pixel-based approach. Therefore, Fuzzy ARTMAP, SVM, and ML
classifiers were applied to the original data set, or to the subset of features extracted using
feature selection (FS), PCA, MNF, and independent component analysis (ICA). A total of
12 processing chains were considered to build the HBDT and combine their abilities. The
maps resulting from a single processing chain (Fuzzy ARTMAP) as well as the HBDT and
probability ensemble are shown in Figure 5, and the confusion matrices for the first and last
maps are reported in Table 1. As for the context-aware mapping, a comparison of the MRF
and spatial re-processing results is available from Figures 6(a) and (b), while the injection
of a priori information about buildings leads to the map in Figure 6(c). Table 1 includes
the confusion matrices for Figures 6(a) and (c) for comparison with previous results.
The results in Figures 5 and 6 and Table 1 show that there is a clear advantage in
using the complete chain with respect to the original spectral-based classification, which
is the common way to use hyperspectral data. In the case of the spectral classifier, the
overall accuracy is lower than exploiting the spatial information using MRFs and shape
regularization. It is true, however, although this is not visible in the results shown in this
article, that the spatial resolution of the data set is an important parameter. As shown by
the previous test set and in Gamba et al. (2010), using data with coarser spatial resolution,
the advantage coming with context information is still true, but basic re-processing is more
appropriate than the more complex MRF framework.
4. Conclusions
This article shows a general framework for the analysis of multispectral and hyperspectral
data sets with enough spatial and spectral resolution to be useful in urban environments.
The availability of these data sets opens new possibilities for urbanization monitoring, urban area mapping, and the analysis of spatial relationship among environmental
indicators, such as land-use classes, temperature, and air pollution.
According to the results shown in the previous section, it is possible to define some
guidelines for the exploitation of multispectral and hyperspectral data in urban mapping.

768

K. Bakos et al.

Table 1. Confusion matrices for the maps in (a) Figure 5(a), (b) Figure 5(c), (c) Figure 6(b), and
(d) Figure 6(c).
(a)
Buildings
Vegetation
Shadow
Roads

Buildings

Vegetation

Shadow

Roads

38, 868
977
1116
397

3
2609
16
10

1565
2106
14, 046
268

4125
666
616
3411

Overall accuracy
Buildings

Downloaded by [81.180.20.66] at 14:51 09 December 2014

(b)
Buildings
Vegetation
Shadow
Roads

4139
303
1000
360

83.2%

Vegetation

Shadow

Roads

26
2721
16
19

1876
2097
14, 122
303

1340
1237
656
3404

Overall accuracy
(c)
Buildings
Vegetation
Shadow
Roads

87.0%

Buildings

Vegetation

Shadow

Roads

41, 979
1092
1567
288

3
2654
13
8

849
2096
13, 919
143

1694
516
267
3138

Overall accuracy
(d)
Buildings
Vegetation
Shadow
Roads

87.8%

Buildings

Vegetation

Shadow

Roads

41, 487
7
718
379

28
5862
265
89

1770
365
14, 920
251

1276
124
521
3367

Overall accuracy

91.8%

There are essentially three parameters to be considered: the spatial resolution, the spectral
resolution, and the scales of the objects/land-use/land-cover classes to be obtained in the
final mapping product.

Spectral resolution is important to determine the complexity of the spectral classification step. The finer this resolution and thus the higher the number of bands,
the more important is the Hughes phenomenon and the relevance of the training set
size. Accordingly, the complexity of the classifiers is larger and the procedure usually
involves fusion at different levels, either pixel, or information, or decision.
Spatial resolution is also very important to select the best option for the contextaware refinement. In our experience, lower spatial resolutions tend to favour simpler
approaches (e.g. spatial re-classification) versus more complex and adaptive ones.
When details are important, however, the larger flexibility of the adaptive approaches
comes into play.
The scale of the objects in the scene is another crucial factor, in the sense that a scene
dominated by a class with objects of a specific size will also drive the choice of the

International Journal of Remote Sensing

Downloaded by [81.180.20.66] at 14:51 09 December 2014

(a)

(b)

769

(c)

Figure 6. Classification maps for the Pavia test area: joint mapping results using (a) MRF after
Fuzzy ARTMAP; (b) spatial re-processing after Fuzzy ARTMAP; (c) MRF and geometrical rules
after Fuzzy ARTMAP.

context-aware algorithm parameters (such as the size of the neighbourhood in the

MRF framework) or the threshold for the geometrical rules (like those used in the
building shape refinement).

Acknowledgements
The authors gratefully acknowledge the financial support provided by the Hyperspectral Imaging
Network (HYPER-I-NET) Marie Curie Research and Training Network, as well as the numerous
discussions on this topic with Fabio DellAcqua and Prashanth Marpu. This work was partially funded
by the HYPER-I-NET FP6 RTN.

References
Bakos, K., and P. Gamba. 2011. Hierarchical Hybrid Decision Tree Fusion of Multiple Hyperspectral
Data Processing Chains. IEEE Transaction on Geoscience and Remote Sensing 49: 38894.
Bakos, K., P. Gamba, and B. Zagajewsky. 2010. Combining Classifiers for Robust Hyperspectral
Mapping Using Hierarchical Trees and Class Memberships. In Proceedings of IGARSS10,
Honolulu, Hawaii, 14025. Piscataway, NJ: IEEE.
Ehlers, M. 2009. Future EO Sensors of Relevance Integrated Perspective for Global Urban
Monitoring. In Global Mapping of Human Settlement: Experiences, Datasets, and Prospects,
edited by P. Gamba, and M. Herold. New York: Taylor and Francis.
Friedland, C. J., M. L. Levitan, and B. J. Adams. 2008. Suitability of Remote Sensing PerBuilding Damage, Assessment of Residential Buildings Subjected to Hurricane Storm Surge.
In Proceedings of the 6th International Workshop on Remote Sensing for Disaster Applications,
Pavia, Italy, unformatted CD-ROM.
Gamba, P., and F. DellAcqua. 2003. Improved Multiband Urban Classification Using a Neuro-Fuzzy
Classifier. International Journal of Remote Sensing 24: 82734.
Gamba, P., F. DellAcqua, K. L. Bakos, and G. Lisini. 2010. Information Extraction from
Hyperspectral Data in Urban Areas Using Contextual Information and Classifier Fusion. In
Proceedings of the ESA Hyperspectral Workshop, Frascati, Italy, unformatted CD-ROM.
Gamba, P., F. DellAcqua, and A. Ferrari. 2003. Exploiting Spectral and Spatial Information for
Classifying Hyperspectral Data in Urban Areas. In Proceedings of IGARSS03, Toulouse, France,
vol. I, 4646. Piscataway, NJ: IEEE.

Downloaded by [81.180.20.66] at 14:51 09 December 2014

770

K. Bakos et al.

Gamba, P., F. DellAcqua, G. Lisini, and G. Trianni. 2007. Improved VHR Urban Area Mapping
Exploiting Object Boundaries. IEEE Transaction on Geoscience and Remote Sensing 45:
267682.
Gamba, P., and B. Houshmand. 2001. An Efficient Neural Classification Chain for Optical and SAR
Urban Images. International Journal of Remote Sensing 22: 153553.
Gautama, S., W. Goeman, J. DHaeyer, and W. Philips. 2004. Robust Detection of Road Junctions
in VHR Images Using Improved Ridge Detection. In Proceedings of XX ISPRS Congress,
International Archives of Photogrammetry and Remote Sensing and Spatial Information Sciences
XXXV, no. 83: 1682750.
Herold, M. and D. A. Roberts. 2005. Spectral Characteristics of Asphalt Road Aging and
Deterioration: Implications for Remote Sensing Applications. Applied Optics 44: 432734.
Kux, H. J. H., and E. H. G. Arajo. 2008. Object-Based Image Analysis Using QuickBird Satellite
Images and GIS Data, Case Study Belo Horizonte (Brazil). In Object-Based Image Analysis
Spatial Concepts for Knowledge-Driven Remote Sensing Applications, edited by T. Blaschke,
S. Lang, and G. J. Hay, 57188. Berlin, Heidelberg: Springer.
Marino, C. M., C. Panigada, and L. Busetto. 2001. Airborne Hyperspectral Remote Sensing
Applications in Urban Areas: Asbestos Concrete Sheeting Identification and Mapping. In
Proceedings of 2001 IEEE/ISPRS Joint Workshop on Remote Sensing and Data Fusion over
Urban Areas, Rome, 21216. Piscataway, NJ: IEEE.
Nichol, J. E. 1996. High Resolution Surface Temperature Patterns Related to Urban Morphology in
a Tropical City: A Satellite-Based Study. Journal of Applied Meteorology 35: 13546.
Nichol, J. E, M. S. Wong, C. Fung, and K. K. M. Leung. 2006. Assessment of Urban Environmental
Quality in a Subtropical City Using Multispectral Satellite Images. Environment and Planning
B: Planning and Design 33: 3958.
Roessner, S., K. Segl, U. Heiden, and H. Kaufmann. 2001. Automated Differentiation of Urban
Surfaces Based on Airborne Hyperspectral Imagery. IEEE Transaction on Geoscience and
Remote Sensing 39: 152532.
Schiefer, S., P. Hostert, and A. Damm. 2006. Correcting Brightness Gradients in Hyperspectral Data
from Urban Areas. Remote Sensing of Environment 101: 2537.
Yamazaki, F., Y. Yano, and M. Matsuoka. 2005. Visual Damage Interpretation of Buildings in Bam
City Using QuickBird Images Following the 2003 Bam, Iran, Earthquake. Earthquake Spectra
21, no. S1: S32936.