IEEE Geoscience and Remote Sensing Magazine - December 2017 - 51

kernels) [121] in the SVM classification process. However,
classification using composite kernels and SVMs demands
convex combination of kernels and a time-consuming optimization process. Therefore, the approach has been modified to deal with convex combinations of kernels through
generalized composite kernels (GCK) [122], the results of
which are shown in Table 4, and multiple-kernel learning
[123]. In [124], a classification framework was introduced
that combines multiple features with the linear and nonlinear class boundaries present in the data without requiring
any regularization parameters to control the weights of the
considered features (the results are shown as MFL in Table 4).
SEMISUPERVISED AND ACTIVE LEARNING
As previously discussed, the number of training samples
is usually limited because the collection of such samples
is either expensive or time consuming. In such situations,
the limited number of training samples available may not
be representative of the statistical distribution of the data,
which can downgrade the quality of the classification map
obtained by supervised classifiers. To partially address this
issue, active learning, which aims to find the most informative training set, has gained popularity in the hyperspectral community.
Active learning starts an iterative process with a small and
suboptimal initial training set and then selects a few additional samples from a large quantity of unlabeled samples.
Active learning considers the result of the current model,
ranking the unlabeled samples according to a criterion that
allows selection of the most informative samples to improve
the model, thus minimizing the number of training samples
while preserving discrimination capabilities as much as possible [125]. For a complete survey on the use of active learning for remote-sensing image analysis, see [126] and [127].
Active learning and semisupervised learning share a
similar conceptual background as both types of learning
try to address the issue of limited labeled samples. In this
manner, both approaches start with a small set of labeled
samples and a large set of unlabeled data. Active learning
usually requires a labor-intensive labeling process, while
semisupervised learning, although avoiding manual labeling by assigning pseudolabels to unlabeled data, may introduce incorrect pseudolabels and consequently downgrade
classification performance [128]. Although active learning
and semisupervised learning follow different work flows,
they both aim to make the most of unlabeled data while
reducing manual labeling efforts [125]. Therefore, it is common to use both of these strategies to make the most of
these two paradigms for HSI classification. In [128], active
learning and semisupervised learning were collaboratively
integrated to form an approach called collaborative active
and semisupervised learning that improves pseudolabeling
accuracy and thus facilitates semisupervised learning. This
method was based on spectral information. In [125], active
learning and hierarchical segmentation were combined for
spectral-spatial classification of HSIs.
DECEMBER 2017

IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE

SPARSE REPRESENTATION CLASSIFICATION
SR classification (SRC)-based approaches with dictionarybased generative models [129], [130] have received considerable attention in the hyperspectral community. In this
context, an input signal is represented by a sparse linear combination of samples (atoms) from a dictionary [129], where
the training data are generally used as the dictionary. The
main advantages of such approaches are that SRC avoids the
heavy training procedure usually conducted by a supervised
classifier and that the classification is performed directly
on the dictionary. ClassificaTHE RICH SPECTRAL
tion can be improved by inRESOLUTION AVAILABLE
corporating contextual inforIN HYPERSPECTRAL DATA
mation from the neighboring
pixels into the classifier. This
CUBES CAN BE USED TO
can be performed indirectly
UNMIX HYPERSPECTRAL
by exploiting the spatial corPIXELS.
relation through a structured
sparsity imposed earlier in the
optimization process. If an adequate number of training samples is available, discriminative as well as compact class dictionaries can also be developed to improve classification performance [131].
CHALLENGES IN CLASSIFICATION
The main challenges for HSI classification are not particularly related to methodology. They are, rather, related to
the lack of appropriate benchmark data sets and the corresponding training and test samples. As can be seen in
Figure 3, most published contributions in the hyperspectral community are dedicated to HSI classification. The approaches are often capable of producing very accurate classification maps on the widely used Indian Pines and Pavia
data sets, which makes real comparison of the approaches
almost impossible. In other words, the existing data sets
have already been saturated in terms of classification accuracies. Therefore, our community is in urgent need of
more complex data sets to share (e.g., highly nonlinear
data sets with greater area coverage that are composed of
many classes). In addition, a standard set of training and
test samples should be defined for each particular data set,
to make the proposed approaches fully comparable with
each other.
SPECTRAL UNMIXING
Spectral unmixing has been an alluring exploitation goal
since the early days of HSI processing [132]. Mixed pixels
are common in remotely sensed HSIs because of the imaging spectrometer's insufficient spatial resolution or due to
intimate mixing effects. However, the rich spectral resolution available in hyperspectral data cubes can be used to
unmix hyperspectral pixels. In fact, mixed pixels can also
be obtained with high-spatial-resolution data because of
intimate mixtures. This means that increasing the spatial
resolution often does not solve the problem.
51

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - December 2017