IEEE Geoscience and Remote Sensing Magazine - June 2013 - 21

the image, while Fig. 11(b) shows nine reference classes of
interest, which comprise urban features, as well as soil and
vegetation features. Finally, Fig. 11(c) shows a fixed training
set available for the scene which comprises 3921 training
samples (42776 samples are available for testing).
Table 2 illustrates the classification results obtained by
different supervised classifiers for the ROSIS University of
Pavia scene in Fig. 11(a), using the same training data in
Fig. 11(c) to train the classifiers and a mutually exclusive
set of labeled samples in Fig. 11(b) to test the classifiers.
As shown by Table 2, the SVM classifier obtained comparatively superior performance in terms of the overall classification accuracy (OA), average classification accuracy (AV)
and kappa statistic [88] when compared with discriminant
classifiers such as LDA, QDA or LogDA. In this experiment,
the SVM was also slightly superior to the multinomial
logistic regression (MLR) classifier [89], which has been
recently explored in hyperspectral imaging as a technique
able to model the posterior class distributions in a Bayesian
framework, thus supplying (in addition to the boundaries
between the classes) a degree of plausibility for such classes
[90]. A subspace-based version of this classifier, called MLRsub [91], is also included in the comparison given in Table 2.
The idea of applying subspace projection methods relies on
the basic assumption that the samples within each class can
approximately lie in a lower dimensional subspace. However, in the experiments reported in [91] for the MLRsub it
was observed that spatial information needs to be included
in this (and other classifiers) in order to improve classification performance. In the following subsection, we summarize some techniques for spatial-spectral classification.
c. spAtiAl-spectRAl clAssiFicAtion
Several efforts have been performed in the literature in
order to integrate spatial-contextual information in spectral-based classifiers for hyperspectral data [76]. It is now
commonly accepted that using the spatial and the spectral
information simultaneously provides significant advantages in terms of improving the performance of classification techniques. Some of these approaches include spatial
information prior to the classification, during the feature
extraction stage. Mathematical morphology [92] has been
particularly successful for this purpose. Morphology is a
widely used approach for modeling the spatial characteristics of the objects in remotely sensed images. Advanced
morphological techniques such as morphological profiles (MPs) [93] have been successfully used for feature

TABLe 2. ACCuRACIeS OBTAIneD BY DIFFeRenT SuPeRVISeD
CLASSIFIeRS FOR THe ROSIS unIVeRSITY OF PAVIA SCene.
MeTRIC

LDA

QDA

LogDA

SVM

MLR

MLRsub

77.95

78.41

80.99

80.11

67.08

73.67

78.73

79.82

88.28

87.80

77.20

kappa

0.606

0.770

0.720

0.761

0.750

0.703

extraction prior to classification of hyperspectral data by
extracting the first few principal components of the data
using the PCA [3], and then building so-called extended
morphological profiles (EMPs) on the first few components
to extract relevant features for classification [94].
As shown by Table 3, the combination of EMP for feature
extraction followed by SVM for classification (EMP/SVM)
provides good classification results for the ROSIS University of Pavia scene. Recently, morphological attribute
profiles (APs) [95] were introduced as an advanced mechanism to obtain a detailed multilevel characterization of a
hyperspectral image created by the sequential application
of morphological attribute filters that can be used (prior
to classification) to model different kinds of the structural
information. According to the type of the attributes considered in the morphological attribute transformation, different parametric features can be modeled. The use of different
attributes leads to the concept of extended multi-attribute
profiles (EMAPs) which have been also used successfully
for hyperspectral image classification purposes [96].
Another strategy in the literature has been to exploit
simultaneously the spatial and the spectral information.
For instance, in order to incorporate the spatial context
into kernel-based classifiers, a pixel entity can be redefined
simultaneously both in the spectral domain (using its spectral content) and also in the spatial domain, by applying
some feature extraction to its surrounding area which yields
spatial (contextual) features, e.g., the mean or standard
deviation per spectral band. These separated entities lead
to two different kernel matrices, which can be easily computed. At this point, one can sum spectral and textural dedicated kernel matrices and introduce the cross-information
between textural and spectral features in the formulation.
This simple methodology yields a full family of new kernel methods for hyperspectral data classification, defined
in [97] and implemented using the SVM classifier thus providing a composite kernel-based SVM (SVM-CK) illustrated
in Table 3 (using the summation kernel). Recently, composite kernels have been generalized in [98], using the MLR

TABLe 3. OVeRALL ACCuRACIeS OBTAIneD BY DIFFeRenT SuPeRVISeD
SPATIAL-SPeCTRAL CLASSIFIeRS FOR THe ROSIS unIVeRSITY OF PAVIA SCene.
MeTRIC

eCHO

LDA-MLL

QDA-MLL

LOGDA-MLL

SVM-CK

eMP/SVM

SVM-W

SVM-RHSeG

MLR-MLL

MLRsub MLL

MLR-GCK

87.58

80.27

89.48

87.04

87.18

85.22

85.42

93.85

85.57

94.10

98.09

92.16

78.05

91.91

83.32

90.47

90.76

91.31

97.07

92.54

93.45

97.76

kappa

0.839

0.739

0.864

0.872

0.871

0.808

0.813

0.918

0.818

0.922

0.974

june 2013

ieee Geoscience and remote sensinG maGazine

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - June 2013