IEEE Geoscience and Remote Sensing Magazine - September 2021 - 102

have little experience in the field. All possible BS approaches
are shown graphically. A detailed description of the proposed
framework is provided in the following.
In Figure 14, bands are represented by circles. Let us consider
that a total of n bands is available in an HSI. From
the total available bands, a desired number of bands m is
obtained through the selection process. Heuristic search
algorithms can be directly applied to the HSI to obtain
an optimal set of bands; as an alternative, clustering and
ranking strategies may be performed, followed by heuristic
search algorithms applied to clustered and ranked bands
(represented by dotted lines). To the best of our knowledge,
such a BS strategy has not been applied in the literature. The
clustered and ranked bands may also undergo search- and
threshold-based operations to select the optimal ones. In
search-based strategies, a selection strategy or a discrimination
strategy may be incorporated to choose optimal bands.
The BS/discrimination may be single (a single BD or SBS) or
multiple (several BDs, multiple BS, band subset discrimination,
or BSS). This is also true for the thresholding approach
(threshold-based discrimination or threshold-based selection).
Search-based approaches may be applied directly
to the original HSI without the need to rank and cluster
bands. Still, the bands obtained using such a technique are
not necessarily the optimal ones, and such a method results
in a very large computational burden. However, in any of
these BS approaches, the objective formulation is an obvious
requirement. The objective function can be based on
distance (minimize/maximize), MI, theoretical measures,
and external evaluators (wrappers).
CONCLUSIONS
A unified BS framework was proposed in this work. More
than 225 research articles focused on BS were reviewed,
and a generalized architecture was conceived, refined, and
presented. Many existing BS approaches consider only spectral
information; however, spectral and spatial information
should be taken into consideration for optimal performance.
Very few spatial-spectral BS techniques are available
in the literature, and this topic can be a rich field for further
investigation and the production of novel scientific results.
When spatial features are incorporated into a BS process, the
desired outcomes can be obtained with very few bands. The
results obtained from the SML-AP and SRL-AP approaches
are convincing in terms of classification accuracy. This
strongly suggests the incorporation of superpixels, which
inherently exploit spatial features, into BS strategies. To evaluate
the performance of bands selected using spatial-spectral
features, it is advisable to use spatial classifiers. Rather
than heuristic approaches, clustering and ranking can be
modeled as wrapper techniques for accurate BS.
In certain applications whose main goal is classification, it
is necessary to find an optimal set of bands where class-specific
representations are distinct. Once BS is performed, objectspecific
analysis can be carried out offline, and the selected
bands can be used for further classification in real time; hence,
102
no further BS is needed. Such requirements should be satisfied
through supervised and semisupervised approaches (incorporated
with wrappers) whose computational complexity (time)
will not be a major constraint. For unsupervised selection,
however, information-theoretic distance measures are key to
extracting the most informative and significant bands. These
approaches, after computational time and complexity validation,
could be considered for real-time and near-real-time BS
applications, such as in-camera BS. Although deep learning
(DL)-based hyperspectral classification approaches [167] are
widely available, their applicability to BS is limited. They can
be supervised (and most of them are) [197], [208], [209] or
unsupervised [108], based on the architecture of the deep network.
DL-based BS techniques may be explored further as a
novel approach to unsupervised FE for BS.
ACKNOWLEDGMENTS
The authors are thankful to the anonymous reviewers for
their valuable feedback, which greatly helped to improve
the manuscript. The authors would like to thank the International
Institute of Information Technology, Bhubaneswar,
for making available its rich library and online
resources, and all the researchers whose work helped us
make a detailed survey of BS approaches.
AUTHOR INFORMATION
Ram Narayan Patro (ram_patro@rediffmail.com) is with
the Department of Electronics and Communication Engineering,
International Institute of Information Technology,
Bhubaneswar, Bhubaneswar, 751003, India. He is a Student
Member of IEEE.
Subhashree Subudhi (ssubudhiiit@gmail.com) is with
the Department of Electronics and Communication Engineering,
International Institute of Information Technology,
Bhubaneswar, Bhubaneswar, 751003, India. He is a Student
Member of IEEE.
Pradyut Kumar Biswal (pradyut@iiit-bh.ac.in) is with
the Department of Electronics and Communication Engineering,
International Institute of Information Technology,
Bhubaneswar, Bhubaneswar, 751003, India. He is a Senior
Member of IEEE.
Fabio Dell'Acqua (fabio.dellacqua@unipv.it) is with the
Department of Electrical, Computer, and Biomedical Engineering,
University of Pavia, Pavia, 27100, Italy. He is a
Senior Member of IEEE.
REFERENCES
[1] R. B. Gomez, D. Ghoshal, and A. Jayanna, " Quantum computation
in the analysis of hyperspectral data, " Quantum Inform. Comput.
II, vol. 5436, pp. 324-332, Aug. 2004. doi: 10.1117/12.543360.
[2] A. Agarwal, T. El-Ghazawi, H. El-Askary, and J. Le-Moigne, " Efficient
hierarchical-PCA dimension reduction for hyperspectral
imagery, " in Proc. IEEE Int. Symp. Signal Process. Inf. Technol.,
2007, pp. 353-356. doi: 10.1109/ISSPIT.2007.4458191.
[3] A. Datta, S. Ghosh, and A. Ghosh, " PCA, Kernel PCA and dimensionality
reduction in hyperspectral images, " in Advances
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE SEPTEMBER 2021
IEEE Geoscience and Remote Sensing Magazine - September 2021

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - September 2021
