IEEE Geoscience and Remote Sensing Magazine - September 2021 - 76

by various researchers and their major contributions. Descriptions
of the techniques are found where appropriate
throughout the article.
SUCH A FRAMEWORK MAY
BE USEFUL IN GUIDING THE
COMPOSITION OF A
PROCESSING CHAIN WHEN
A PARTICULAR LAND COVER
CLASSIFICATION PROBLEM
IS ADDRESSED.
UBS APPROACHES
In contrast to SBS and SSBS, a more intense exploitation
of UBS strategies is observed, probably because the GT is
not always available. Figure 2 depicts the importance (the
approximate density in reviewed articles) of each strategy.
A series of UBS techniques is proposed, where information
maximization is carried out with an external objective that
does not consider the class
information; the approaches
are generally characterized by
spatial entropy [37], [39], [41],
[56], [62], [66], [68], [76], [100],
[102], [111], [113], [117]-[133].
A comparative analysis of five
UBS strategies [uniform spectral
spacing, the first spectral
derivative, linear prediction
(LP), the N-finder algorithm
(NFINDR), and entropy] is
presented in [120], where LP
outperformed the NFINDR and entropy. Later, in [111], the
volume gradient-based BS (VGBS) method was proposed.
The technique considers each band as an LP from the other
bands plus an additive error term that is required for a band
to reach its actual value from the predicted one. The amount
of error in predicting a specific band gives a clue about the
band's actual information content. An interesting approach
for predicting the overall accuracy without GT information
along with redundancy is proposed in [37]; this is essentially
a constrained quadratic optimization model for BS.
Various aspects of such UBS strategies can be framed as
four subsequent steps involving 1) the preremoval of noisy
bands, 2) feature incorporation, 3) objective formulation,
and 4) selection strategy. An overview of the initial two
steps as a data selection strategy is presented in Figure 3.
The literature focuses on UBS strategies, so the discussion
of the magenta-colored steps is limited. An exploration of
other steps is presented later in this article.
HSI PREREMOVAL/PROCESSING FOR BS
Since the BS process should sort out the most distinctive
and informative bands, the bands that are known to be
uninformative can be removed a priori, including water
absorption and low-signal-to-noise-ratio (SNR) bands,
which are common in most of the BS environment. Instead
of manually performing the process, the preremoval of
poor-quality bands is suggested via the guidance of spectral
correlation between adjacent bands and within-band spatial
relationships. In [134], a histogram-based preremoval
approach was shown to be helpful for eliminating noisy
bands that do not fall within the water absorption range.
Generally, the noise level varies across bands, and the
higher the noise level, the lower the correlation with other
bands. Whereas noise whitening can be useful, the required
noise estimation is difficult to achieve. In [135],
however, it is pointed out that the net effect of noise whitening
and data whitening is similar. Hence, instead of
noise whitening, easier data whitening can be carried out
through the eigendecomposition of the data covariance
matrix. Experimental results [120] suggest that, if data
whitening is carried out prior to BS, the BS results are more
informative than those obtained from the original, nonwhitened
bands. In [136], noise reduction is performed
prior to BS, resulting in improved performance. After obtaining
the optimal bands, fusion can also be carried out
[137] for compression. However, in the present literature,
the applicability of preremoval is limited to the classification
problem.
As a whole, BS is more informative if it is followed by
two subsequent steps: 1) poor-quality band removal (a low
SNR/water absorption/automatic removal) and 2) data
whitening. Thus, instead of manual preremoval, algorithmic
noisy band preremoval followed by data whitening is
expected to generate more informative BS results. Although
preremoval can eliminate the less informative bands, the
surviving bands are still redundant and highly correlated.
Hence, further redundancy reduction should be carried out
by incorporating proper FE, objective formulation, and selection
processes, as discussed later in this article.
Data Selection
Spectral
Spatial
Hybrid
Supervised
Semisupervised
Unsupervised
Preremoval and/or Preprocess
FIGURE 3. Preparing data for BS.
76
SPECTRAL/SPATIAL/HYBRID FEATURES FOR BS
GT data provide samples of the spectral responses of interest
in a particular HSI under study; however, class-specific
spectral responses do not exhaust the available information.
There exist pixels and regions that are unlabeled in the GT
but that still contain useful information. The labeled samples
(objects) have (spatial) relations to their unlabeled, adjacent
pixels. Moreover, labeled samples contain class-specific
spatial (as opposed to spectral) relationships, making them
unique. So, depending on the feature incorporation, a BS
technique can be spectral, spatial, or hybrid, i.e., using both
spectral and spatial information (Figure 2). Indeed, various
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE SEPTEMBER 2021
Intra/Interband
Inter/Intraclass
IEEE Geoscience and Remote Sensing Magazine - September 2021

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