IEEE Geoscience and Remote Sensing Magazine - September 2021 - 95

Society Data Fusion Contest, and it is still offered as an
open benchmark. It was captured by the Itres hyperspectral
Compact Airborne Spectrographic Imager-1500 sensor
above the University of Houston campus. The selected HSI
is composed of 349 # 1,905 pixels, on 144 bands, with a
spatial resolution of 2.5 m. The wavelengths range from
364 to 1,046 nm, with a typical bandwidth of 10 nm. Information
about the 15 defined classes and the number
of training and test samples is given in Table 2. Similarly,
the Houston2018 data set was offered following that year's
Data Fusion Contest. It was acquired for the same area and
using the same sensor, and includes 601 # 2,384 pixels at a
ground resolution of 1 m. The 50 10-nm-wide bands span
the wavelength range between 380 and 1,050 nm. Information
about the 15 defined classes and the number of training
and test samples is in Table 3. The Indian Pines data
set was acquired by the Airborne Visible/Infrared Imaging
Spectrometer sensor across a wavelength range from 400
to 2,500 nm. It contains 220 bands, 16 classes, and images
of 145 # 145 pixels. It has a spatial resolution of 20 m per
pixel. Information about the 16 defined classes and the
number of training and test samples is in Table 4.
STATE-OF-THE-ART BS APPROACHES
Various BS algorithms, including clustering, ranking, correlation,
hybrid, and search strategies, were applied to two
hyperspectral data sets to validate the methods' performance.
Details of state-of-the-art BS techniques are offered in Table 5,
which reports the Houston2013 and Houston2018 data
sets' BS execution times on a computer with an Intel i5 processor
and 12 GB of random-access memory.
RESULTS AND DISCUSSION
The various BS approaches that were considered were validated
through their performance in 1-nearest neighbor
(1NN) and random forest (RF) classifications on spectral
and spatial-spectral features. The results of spectral classification
with the 1NN classifier for the Houston2018 data set
are presented in Table 6. In general, clustering approaches,
such as dual clustering (DC) BS contextual analysis (CA),
DSC, superpixel ML (SML)-AP, and superpixel representation
learning (SRL)-AP, perform better than ranking and
search strategies. However, the GA-PSO, Jaya optimization,
and PSO-based heuristic searching methods follow
clustering strategies. The information entropy (IE)- and
ID-based ranking methods achieve significantly lower accuracy
levels. Visually assessing the RF classifier maps for
the Houston2013 and Houston2018 data sets, in Figure 11 and
Figure 12, respectively, confirms that clustering approaches
lead to generally better results than ranking techniques.
Moving to spatial-spectral feature classification, the results
for the Houston2018 data set are presented in Table 7. As
one can note, the results of spatial-spectral classification
using the 1NN classifier are aligned with those from spectral
approaches but with increased average accuracy. Specifically,
the spatial DCBSCA approach yields an 8% positive
SEPTEMBER 2021 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE
95
TABLE 8. THE HOUSTON 2018 DATA SET RF SPATIAL-SPECTRAL CLASSIFICATION RESULTS WITH 5% TRAINING SAMPLES AND 10 SELECTED BANDS. (CONTINUED)
CLASS AP
19
20
OA
AA
RELIEFF PSO IE
0.08 0.34
56.27 41.59
0.15 1.36
66.77 68.33
0.89 0.66
74.83 75.81
0.21 0.39
60.03 63.87
0.9
1.35
l # 100 66.9 68.14
1.16 0.34
ID
1.25 0.02 1.3
LCMV-
BCM MDPP
0.12
0.79
0.01
LCMV-
BCC
0.75
1.15
FTRC-
OC-
FDPC
0.33
1.27
JAYA
OPTIMIZATION
0.86
56.9 61.35 56.66 57.72 61.86 67.77 65.35 60.86
0.16 0.78 0.21 0.84
68.85 72.55 69.89 72.2
1.41 0.98 0.68 1.26
1.38
69.36 80.68 81.32 80.18
0.8
0.83
0.64
0.18
1.02
0.74
0.55
76.29 76.65 76.67 76.85 77.27 77.38 77.54 77.87
0.13 0.64 0.39 0.45
1.06
63.67 64.56 64.38 63.46 65.12 63.63 65.85 65.48
1.41 0.33 0.72 0.56
0.79
0.48
0.4
0.91
0.1
1.37
0.65
68.8 69.29 69.33 69.53 70.07 70.18 70.35 70.87
0.56 1.34 1.28 0.6
1.36
GA-
PSO
NC
OC-
IE
NC
OC-
MVPCA SSR
0.18 1.01 0.9
SQ
LCMV-
BCM DSEBS
0.08 0.08
57.7 59.29 68.02 53.4 63.9
1.48 0.76 0.41
0.15 0.76
79.03 89.28 82.35 84.96 81.8
0.36 0.87 1.3
0.61 0.17
0.59 0.01
0.23
SC
LCMV-
BCM
0.03
1.14
SML-
AP
DCBSCA
0.65 1.25
68.51 68.41 68.41 68.94
0.88
0.12 0.99
82.46 77.45 90.22 88.91
1.25
0.45
0.6
0.33
0.28 0.21
0.67
78.15 78.34 78.58 79.01 79.14 79.41 81.49 81.99 82.31
0.37 1.18 0.11
67.9 66.35 64.79 67.65 65.37 66.12 70.99 70.88 72.05
0.26 0.04 1.43
0.65 1.44
71.15 71.43 71.78 72.32 72.5
0.89 0.5
1.28
0.66 1.36
1.14
0.05
0.01
1.02 1.06
72.85 75.62 76.27 76.7
0.8
1.07 0.27
SRL-
AP
DSC ALL
0.93 0.78 1.3
62.33 71.14 86.99
0.78 0.26 1.41
93.63 94.38 98.64
0.61 0.59 0.54
82.71 84.32 86.15
0.26 0.21
0.4
72.38 74.61 83.04
0.97 0.83 0.33
77.22 79.39 82.05
0.28 0.48
0.5
IEEE Geoscience and Remote Sensing Magazine - September 2021

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