HAL Id: hal-00578886
https://hal.archives-ouvertes.fr/hal-00578886
Submitted on 22 Mar 2011
HAL is a multi-disciplinary open access
archive for the deposit and dissemination of sci-
entic research documents, whether they are pub-
lished or not. The documents may come from
teaching and research institutions in France or
abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est
destinée au dépôt et à la diusion de documents
scientiques de niveau recherche, publiés ou non,
émanant des établissements d’enseignement et de
recherche français ou étrangers, des laboratoires
publics ou privés.
Classication of hyperspectral images by using extended
morphological attribute proles and independent
component analysis
Mauro Dalla Mura, Alberto Villa, Jon Atli Benediktsson, Jocelyn Chanussot,
Lorenzo Bruzzone
To cite this version:
Mauro Dalla Mura, Alberto Villa, Jon Atli Benediktsson, Jocelyn Chanussot, Lorenzo Bruzzone. Clas-
sication of hyperspectral images by using extended morphological attribute proles and independent
component analysis. IEEE Geoscience and Remote Sensing Letters, IEEE - Institute of Electrical and
Electronics Engineers, 2011, 8 (3), pp.542-546. �10.1109/LGRS.2010.2091253�. �hal-00578886�
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 8, NO. 3, MAY 2011 541
Classification of Hyperspectral Images by Using
Extended Morphological Attribute Profiles
and Independent Component Analysis
Mauro Dalla Mura, Student Member, IEEE, Alberto Villa, Student Member, IEEE,
Jon Atli Benediktsson, Fellow, IEEE, Jocelyn Chanussot, Senior Member, IEEE,and
Lorenzo Bruzzone, Fellow, IEEE
Abstract—In this letter, a technique based on independent
component analysis (ICA) and extended morphological attribute
profiles (EAPs) is presented for the classification of hyperspec-
tral images. The ICA maps the data into a subspace in which
the components are as independent as possible. APs, which are
extracted by using several attributes, are applied to each image
associated with an extracted independent component, leading to a
set of extended EAPs. Two approaches are presented for including
the computed profiles in the analysis. The features extracted by the
morphological processing are then classified with an SVM.
The experiments carried out on two hyperspectral images proved
the effectiveness of the proposed technique.
Index Terms—Attribute filters, decision fusion, extended
attribute profile (EAP), independent component analysis (ICA),
mathematical morphology, remote sensing.
I. INTRODUCTION
T
HE EXPLOITATION of spatial information is very impor-
tant for the classification of high-resolution hyperspectral
images, particularly when considering urban areas, and it is
advisable to consider geometrical features in the analysis in
order to derive spatially accurate maps [1].
Manuscript received July 19, 2010; revised October 14, 2010; accepted
October 25, 2010. This work was supported in part by the European com-
munity’s Marie Curie Research Training Networks Program, Hyperspectral
Imaging Network (HYPER-I-NET), under Contract MRTN-CT-2006-035927
and in part by the Research Fund of the University of Iceland and the University
of Trento.
M. Dalla Mura is with the Department of Information Engineering and
Computer Science, University of Trento, 38100 Trento, Italy, and also with
the Faculty of Electrical and Computer Engineering, University of Iceland, 107
Reykjavik, Iceland (e-mail: dallamura@disi.unitn.it).
A. Villa is with the Grenoble Images Parole Signals Automatics Labora-
tory, Signal and Image Department, Grenoble Institute of Technology, 38402
Saint Martin d’Hères, France, and also with the Faculty of Electrical and
Computer Engineering, University of Iceland, 107 Reykjavik, Iceland (e-mail:
alberto.villa@hyperinet.eu).
J. A. Benediktsson is with the Faculty of Electrical and Computer Engineer-
ing, University of Iceland, 107 Reykjavik, Iceland (e-mail: benedikt@hi.is).
J. Chanussot is with the Grenoble Images Parole Signals Automatics Lab-
oratory, Signal and Image Department, Grenoble Institute of Technology,
38402 Saint Martin d’Hères, France (e-mail: jocelyn.chanussot@gipsa-lab.
grenoble-inp.fr).
L. Bruzzone is with the Department o f Information Engineering and Com-
puter Science, University of Trento, 38100 Trento, Italy (e-mail: lorenzo.
bruzzone@ing.unitn.it).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/LGRS.2010.2091253
The spatial characteristics of the objects in an image can
be modeled with several approaches. A widely used technique
for extracting spatial f eatures is based on mathematical mor-
phology. Among all the operators belonging to this framework,
morphological connected operators [2] proved to be suitable
for extracting spatial information while preserving the geo-
metrical characteristics of the structures in the image (i.e.,
without distorting the borders). In [3], morphological profiles
(MPs), a sequence of multiscale connected operators, were
applied to high-resolution hyperspectral images by reducing
the high dimensionality of the data by principal component
analysis (PCA) and computing the profiles on the first principal
components extracted, leading to the definition of extended
MPs (EMPs). Due to the limitations of PCA, when extracting
the information sources from the high-dimensional data, it was
proposed to perform independent component analysis (ICA)
before the computation of the MPs [4].
The characterization of spatial information obtained by the
application of an MP is particularly suitable for representing
the multiscale variability of the structures in the i mage, but
it is not sufficient to model other geometrical features. To
avoid this limitation, the use of morphological attribute filters
instead of the conventional operators based on the geodesic
reconstruction was proposed [5]. The application of attribute
filters in a multilevel way leads to the definition of attribute
profiles (APs) [5], which permit to model other geometrical
characteristics rather than the size of the objects. Moreover,
APs showed interesting characteristics when extended to hy-
perspectral images [6]. In greater details, analogous to [3],
the APs were applied to the first principal components ex-
tracted from a hyperspectral image, generating an extended
AP (EAP).
In this letter, we extend the work presented in [7] by present-
ing a technique based on EAPs and ICA for the classification of
hyperspectral images. Moreover, we investigate t wo approaches
for combining the information extracted by EAPs computed
with different attributes.
This letter is organized as follows. In Section II, the
ICA is discussed, while in Section III, the concepts of mor-
phological attribute filters and EAPs are presented. The pro-
posed approaches for dealing with multiple EAPs are presented
in Section IV. The experimental results are illustrated
in Section V. Finally, conclusions are drawn in Section VI.
1545-598X/$26.00 © 2010 IEEE
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
542 IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 8, NO. 3, MAY 2011
II. ICA
Hyperspectral sensors record images with hundreds of bands
and a very high spectral resolution. The very detailed spec-
tral description provided by these kinds of images increases
the capability to distinguish between land-cover classes, thus
achieving accurate classification maps. However, the analysis of
this huge amount of data presents some methodological issues
which need to be addressed. In particular, the high dimensional-
ity of the data is a critical problem, due to the appearance of the
Hughes phenomenon: After a certain threshold, if the number of
features increases, the generalization capability of the classifier
decreases when a fixed number of training samples are used.
The threshold mainly depends on the number of samples used to
train the classifier. Because of these reasons, feature reduction
is often applied as a preprocessing step before the classification
of hyperspectral data, in order to avoid the curse of dimen-
sionality, and to reduce the computation time. Although it is
not optimal for classification, the PCA is often used f or such
a task, due to its simplicity and ease of use. The principle of
PCA is to project the data into an orthogonal space, so that the
eigenvectors correspondent to the greatest eigenvalues retain
the maximum variance of the data. Because the PCA is based
on the analysis of covariance matrix and second-order statistics,
it can neglect some important information, particularly when
few components are retained. In this letter, we propose to use
ICA for feature reduction, as an effective alternative to PCA.
ICA consists of finding a linear decomposition of the observed
data into statistically independent components (ICs). Given an
observation model x = As, where x is the vector of the ob-
served signals, A is a matrix of scalars corresponding to the
mixing coefficient, and s is the vector of the source signals, the
ICA finds a separating matrix W such that y = Wx = WA s ,
where y is a vector of ICs.
Independence is a much stronger assumption than decorre-
lation, which can be obtained with PCA or factor analysis.
In ICA, the concept of independence can be summarized as
follows: Each component should not provide any information
about higher (than second) order statistics of the other com-
ponents. However, there are several methods for estimating
ICA. In this letter, we have used the algorithm JADE (Joint
Approximate Diagonalization of Eigenmatrices), due to good
results shown when used for feature reduction in hyperspectral
remote sensing data [8]. Due to space constraints, we refer t he
reader interested in more details about the general framework
of ICA to [9].
III. EAPs
EAPs are based on the concept of the AP, which is a
generalization of the widely used MPs [6]. Analogous to the
definition of the EMPs, EAPs are generated by concatenat-
ing many APs. Each AP is computed on one of the c fea-
tures (FRs) extracted by a feature reduction transformation
(e.g., PCA) from a multivariate image (e.g., the hyperspectral
image) [5]
EAP = {AP (FR
1
),AP(FR
2
),...,AP(FR
c
)} . (1)
The AP is an extension of the MP, obtained by processing
a scalar grayscale image f , according to a criterion T , with n
morphological attribute thickening (φ
T
) and n attribute thin-
ning (γ
T
) operators, instead of the conventional morphological
filters by reconstruction
AP (f )=
φ
T
n
(f),φ
T
n−1
,...,φ
T
1
(f),f, γ
T
1
(f),...,γ
T
n−1
(f),γ
T
n
(f)
.
(2)
Attribute filters are connected operators which operate on
the connected components (i.e., regions of isointensity spatially
connected pixels) that compose an image, according to a given
criterion [2]. The criterion associated to the transformation is
evaluated on each connected component of the image. In gen-
eral, the criterion compares the value of an arbitrary attribute
attr (e.g., area, volume, standard deviation, etc.) measured on
the component C against a given reference value λ (which is
the filter parameter), e.g., T (C)=attr(C) >λ. If the criterion
is verified, then the regions are kept unaffected; otherwise, they
are set to the gray level of a darker or brighter surrounding
region, according to if the transformation performed is
extensive (i.e., thickening) or antiextensive (i.e., thinning),
respectively. When the criterion considered in the analysis is
increasing (i.e., if it is verified for a connected component, then
it will also be verified by all the regions brighter or darker,
according to the transformation, including the component), the
attribute thinning and thickening operators are actually opening
and closing transformations. Nonincreasing criteria do not have
a unique definition when considering grayscale images. In fact,
different effects can be obtained by the operators with a nonin-
creasing criterion according to the filtering rule selected [10].
Attribute filters can be efficiently computed by taking ad-
vantage of the representation of the input image as a rooted
hierarchical tree of the connected components of the image. The
tree is obtained by the Max-tree algorithm [10]. The approach
based on this data representation is particularly useful when
computing an AP, since the image is converted to the tree only
once (this is the most demanding stage of the filtering) and
processed several times with the different criteria. Examples of
EAPs computed with different attributes are reported in (Fig. 1).
IV. A
PPROACHES TO DEAL WITH MULTIPLE EAPs
The choice of the most suitable attribute and range of thresh-
olding values (λ’s) for extracting the information on the geospa-
tial objects is certainly a complex task, particularly when apiori
information on the scene is not available. A possible approach
attempting to overcome this issue relies on the computation of
EAPs with different kinds of attributes. However, this leads to
the problem of properly exploiting, in the analysis, the different
information extracted by the computed EAPs.
A simple strategy is the stacked vector approach (SVA),
which combines the EAPs by concatenating them in a single
vector of features [also called extended multi-AP (EMAP) [6];
see Fig. 2(a)]. However, even if complementary information
can be extracted by considering different attributes, a great
redundancy is present in the features extracted. Thus, it is advis-
able that a classification algorithm with excellent penalization
capability is used for classifying the features in order to handle
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
DALLA MURA et al.: CLASSIFICATION OF HYPERSPECTRAL IMAGES 543
Fig. 1. Examples of EAPs computed on the first two PCs of a portion of the
image Fig. 3(a). Each row shows an EAP built by different attributes. Attributes
starting from the first row are area, length of the diagonal of the bounding
box, moment of inertia, and standard deviation. Each EAP is composed by the
concatenation of two APs computed on PC
1
and PC
2
. Each AP is composed
of three levels: A thickening image φ
T
, the original PC, and a thinning image
γ
T
. All the thickening and thinning transformations were computed with the
following attribute values λ’s. Area: 5000; length of the diagonal: 100; moment
of inertia: 0.5; and standard deviation: 50.
Fig. 2. Proposed approaches for dealing with multiple EAPs. (a) SVA and
(b) FA.
the increased dimensionality which can lead to the Hughes
phenomenon.
Another approach is the fusion approach (FA) that is based
on the separate classification of each EAP and on the f usion of
the results obtained by the independent classifiers in order to
generate the final decision map [see Fig. 2(b)]. In comparison
Fig. 3. ROSIS Pavia data sets: True color representation and test set for
(a), (b) University and (c), (d) Center.
to the SVA, the FA keeps low the dimensionality of the data
and increases the robustness of the results, particularly if the
different EAPs generate complementary errors.
In this letter, an SVM classifier is considered with the one-
against-one multiclass strategy. The fusion rule considered
when combining the results of the single classifiers relies on
the sum of the votes of the classifiers applied to the four EAPs,
assigning each pixel to a class, according to the majority voting
scheme. Obviously, other decision criteria can be applied.
V. E
XPERIMENTAL ANALYSIS
The experimental analysis was carried out on two hyper-
spectral images acquired over the city of Pavia (Italy) by the
ROSIS-03 (Reflective Optics Systems Imaging Spectrometer)
hyperspectral sensor. The two images have a geometrical res-
olution of 1.3 m. The first one shows the university campus
(610 × 340 pixels), while the second one was acquired on
the city center (1096 × 489 pixels). In the following, we will
refer to the two data sets as “University” and “Center,” respec-
tively. The original data are composed of 115 spectral bands,
ranging from 0.43 to 0.86 μm with a band of 4 nm. However,
noisy bands were previously discarded, leading to 103 and
102 channels for the two images, respectively. Nine thematic
land-cover classes were identified in the university campus:
Trees, Asphalt, Bitumen, Gravel, Metal sheets, Shadows, Self-
blocking Bricks, Meadows, and Bare soil. For this data set, a
total of 3921 and 42 776 pixels were available as training and
test sets, respectively. In the center area, the thematic classes
found were Water, Tree, Meadow, Self-blocking Bricks, Soil,
Asphalt, Bitumen, Tile, and Shadow. The training and test sets
for this data set were composed of 5536 and 103 539 samples,
respectively. The true color representation of the images and the
test sets taken as reference are shown in Fig. 3.
In the analysis carried out, all the samples of the training
set were used f or the University data set, while for the Center
data sets, only 50 samples (randomly chosen from the full
training set for each class) were considered. All the experiments
conducted on the latter data set were run ten times with a set of
different training samples each time.
From both the two hyperspectral images, four components
extracted by PCA and ICA were considered. The first four PCs
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
544 IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 8, NO. 3, MAY 2011
TAB L E I
ROSIS U
NIVERSITY DATA SET.CLASSIFICATION ACCURACIES OBTAINED BY CLASSIFYING THE HYPERSPECTRAL IMAGE (SPECT.),
THE FOUR COMPONENTS EXTRACTED (4 COMP.), EAC H SINGLE EAP, AND THE DATA WITH THE PROPOSED APPROACHES:SVAAND FA
TAB L E II
ROSIS C
ENTER DATA SET.CLASSIFICATION ACCURACIES OBTAINED BY CLASSIFYING THE HYPERSPECTRAL IMAGE (SPECT.),
THE FOUR COMPONENTS EXTRACTED (4 COMP.), EAC H SINGLE EAP, AND THE DATA WITH THE PROPOSED APPROACHES:
SVA
AND FA. THE RESULTS REPORTED ARE THE AVERAGE OF THE ACCURACIES OBTAINED IN 10 TRIALS
WITH 50 TRAINING SAMPLES PER CLASS RANDOMLY CHOSEN FOR EAC H TRIAL
contain more than 99% of the total variance of the data for
both the data sets. The components were rescaled to the range
[0, 1000] and converted to integer in order to be processed by
the attribute filters. Four EAPs were computed by considering
four different attributes on the components extracted by PCA
and ICA: 1) a, area of the regions (λ
a
= [100 500 1000 5000]);
2) d, length of the diagonal of the box bounding the region
(λ
d
= [10 25 50 100]);3)i, first moment invariant of Hu,
moment of inertia [11] (λ
i
=[0.20.30.40.5]); and 4) s,
standard deviation of the gray-level values of the pixels in the
regions (λ
s
= [20 30 40 50]). The area and the length of the
diagonal of the bounding box extract information on the scale
of the objects. The moment of inertia and the standard deviation
are not dependent on the size dimension, but they are related
to the geometry of the objects and the homogeneity of the
intensity values of the pixels, respectively. Each EAP is 36-
dimensional, i.e., it is composed of four APs with nine levels
computed on each component extracted. In the sequel, the
notation EAP
attr
denotes the EAP built with the attr attribute.
The classification maps are obtained by analyzing the features
extracted by the extended profiles with an SVM classifier with a
radial basis function kernel. The model selection in the training
phase of the classifier was based on a gradient descent method,
which proved to be computationally less demanding than the
exhaustive investigation of the parameters on a grid approach,
giving comparable results [12].
The thematic accuracies of the obtained maps (which are pre-
sented in Tables I and II) were assessed by computing the over-
all accuracy (OA), the average accuracy (AA), and the Kappa
coefficient (κ) on the available reference data. The statistical
significance of the classification maps obtained by PCA and
ICA and the same morphological processing was evaluated with
the McNemar’s test. All the results were statistically significant.
The obtained results are reported in Table I. It is clear, as
in most of the cases, by including the features extracted by
the EAPs in the analysis resulted in higher accuracies (up to
almost 17% of OA) than those obtained by considering only the
spectral features. The ICA proved to extract more informative
components from the data, leading to better results than those
generated by the PCA in all the experiments. When considering
the contribution of the single EAPs, the EAPs built with area
and the moment of inertia attributes performed the best with
the PCA and ICA, respectively. This proves how it can be
difficult to select apriorithe most suitable attribute on the
data. In these experiments, considering all the EAPs together,
in the SVA architecture, with the ICA gives excellent results
in terms of classification accuracies. As far as we know, these
accuracies are higher than all those reported in the literature for
this data set without postprocessing [1], [13]. In contrast, the
SVA approach led to low accuracies for the PCA. This can be
due to the high variation in terms of accuracy showed by the
single EAPs (more than 20% of OA), which affects the overall
performances of this approach. The FA is performing well in
average and has a robust behavior since, in all the experiments,
the accuracies obtained, when compared to those of the single
EAPs, are slightly lower than the best case (less than 2% of
OA) and better than all the others. The improved accuracies
obtained by the proposed technique are also confirmed by the
higher precision shown in the map obtained when considering
the ICA and all t he EAPs together [see Fig. 4(c)].
Table II reports the thematic accuracies obtained on the
Center data set (the correspondent classification maps are not
