IEEE Geoscience and Remote Sensing Magazine - September 2021 - 93

to pixels, superpixels are more informative for classification
[88], due to their inherent capability of extracting spatial
features. Indeed, because stand-alone pixels convey only
spectral information, several window-based spatial extractors
are proposed in this context. Since moving-windowbased
approaches entail resource-consuming processing,
disjointed windows are generally used to reduce complexity.
A similar approach, the ASPS [195], splits an HSI into
several subblocks; however, there is no guarantee that each
subblock will be a separate class. Hence, especially for UBS
strategies, there is no other choice than applying supervised
strategies. In contrast, superpixels can be labeled as a synthetic
class because of their homogeneous nature [99]. Intra-
and interclass variability measures can be applied, as
in SBS strategies. Although superpixels work better than
pixels in a hyperspectral classification context, the strong
assumption that leads to superpixel delineation cannot be
always true, and this remains an issue.
VALIDATING THE BS PROCESS
The performance of a BS process is mostly validated by
computing the accuracy obtained from the selected bands
via spectral classifiers, such as SVMs and k-nearest neighbors.
However, if a BS process involves spatial features, similar
spatial-spectral-enabled classifiers can provide better
validation. A similar process is adopted in [236] to classify
HSIs for BS validation purposes through spatial-spectral
features by incorporating a simple Gaussian filter:
Ge ,
1
xy,
=
2rv
2 # -
xy
v
+
2 2
22
(22)
where x and y are the distances from the origin on the
horizontal and vertical axes, respectively, and v is the
standard deviation of the Gaussian distribution. In addition
to classification accuracy, statistical evaluators for
calculating intraclass separability, spectral information,
and computational complexity can strengthen the evaluation
criteria [9].
EXPERIMENTAL ANALYSIS
To corroborate and substantiate the various approaches
discussed in the first part of this article, we designed and
carried out a series of experiments to obtain quantitative
results. We chose data sets that were considered standard
benchmarks in the Earth observation community; these
are described in the following section. Various state-of-theart
BS approaches were implemented and applied to the
test data sets as discussed in the " State-of-the-Art BS Approaches "
section. Results are reported in the " Results and
Discussion " section.
DATA SET DESCRIPTION
Benchmark data sets were selected for quantitative validation;
the choice fell to three: Houston2013, Houston2018,
and Indian Pines [119]. The Houston2013 data set was generated
for the 2013 IEEE Geoscience and Remote Sensing
SEPTEMBER 2021 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE
93
TABLE 7. THE HOUSTON2018 DATA SET 1NN SPATIAL-SPECTRAL CLASSIFICATION RESULTS WITH 5% TRAINING SAMPLES AND 10 SELECTED BANDS. (CONTINUED)
CLASS AP
0.64
19
20
OA
AA
59.46
0.25
82.83
0.98
71.28
0.76
62.82
0.27
l # 100 63.05
0.86
RELIEFF IE
0.64
49.9
1.37
78.61
1.4
71.73
0.95
64.5
1.44
63.51
0.8
PSO ID MDPP
LCMV-
BCM
0.19 0.04 0.44 0.48 0.98
LCMV-
BCC
1.44
0.81
TRC-
OC-
FDPC
1.4
1.02
JAYA
OPTIMIZATION
0.69
67.13 63.46 68.5 66.21 71.49 70.01 75.06 70.13
1.34 0.77 1.05 0.23 1.43
0.05
82.68 83.91 86.24 87.35 90.87 84.48 91.3
0.25 1.38 1.19 0.87 0.66
0.39
0.95
0.53
1.13
1.5
90.22
0.34
72.72 73.5 73.59 73.74 74.07 74.24 74.44 75.07
1.43 0.67 0.09 1.3
0.34
65.94 67.08 67.17 66.98 68.01 68.02 67.49 68.89
1.39 0.34 0.56 0.13
0.4
0.96
1.48
0.27
0.04
0.07
64.81 65.82 65.93 66.12 66.53 66.74 67.03 67.84
0.41 0.37 0.68 0.34 1.21
0.8
GA-
PSO
NC
OC-
IE
SSR
NC
OC-
MVPCA
0.36 1.15 1.14 1.11
SQ
LCMV-
BCM DSEBS
1.12
0.49
0.16
0.82
SC
LCMV-
BCM DCBSCA
1.02
0.6
0.69
70.92 78.32 60.11 57.63 77.45 80.28 78.14 77.88
1.21 1.12 0.18 0.79
0.62
89.93 91.21 94.89 92.97 90.93 92.43 87.89 96.54
0.1
1.15 1.01 1.07
0.96
0.04
0.63
0.63
0.59
0.28
1.22
75.35 76.02 76.19 76.62 76.71 77.07 78.92 80.5
0.98 0.91 0.58 0.21
1.09
69.8 68.86 69.71 72.29 70.11 71.14 73.52 75.27
1.08 0.52 0.99 0.58
0.94
1.2
1.48 0.1
1.41
0.03
0.03
1.37
1.03
1.2
68.18 69.01 69.25 69.71 69.89 70.37 72.76 74.75
0.13
1.18
SML-
AP
SRL-
AP
DSC
0.32 0.15 1.24
ALL
0.26
72.53 73.21 85.32 87.42
0.27 0.38 0.03
97.59 97.67 97.5
0.48 1.22 1.18
1.39
98.98
1.28
80.93 80.96 82.72 86.43
0.56 1.26 1.1
0.86
77.1 76.8 78.33 82.85
1.12 1.22 0.57
0.93
75.33 75.35 77.62 82.42
0.8
1.33 1.35
0.94
IEEE Geoscience and Remote Sensing Magazine - September 2021

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