Figure 17. The Salinas hyperspectral-image classification results: (a) training set and (b) testing set. Classification maps
obtained by (c) joint collaborative representation, (d) kernel collaborative representation with Tikhonov regularization and
composite kernel, (e) support vector machine, (f) simultaneous orthogonal matching pursuit, (g) orthogonal matching pursuit,
(h) nearest regularized subspace, (i) sparse coding, (j) collaborative representation, (k) collaborative sparse coding, and (l)
collaborative sparse coding with smoothness regularization.
Table 5. Listing of 16 classes in the Salinas
hyperspectral-image data. Training and testing
sample counts are shown for each class.
Class
Samples
No. Name
Train Test
1 Brocoli_green_weeds_1
201 1808
2 Brocoli_green_weeds_2
373 3353
3 Fallow
198 1778
4 Fallow_rough_plow
140 1254
5 Fallow_smooth
268 2410
6 Stubble
396 3563
7 Celery
358 3221
8 Grapes_untrained
1128 10143
9 Soil_vinyard_develop
621 5582
10 Corn_senesced_green_weeds 328 2950
11 Lettuce_romaine_4wk
107
12 Lettuce_romaine_5wkTable 193
13 Lettuce_romaine_6wk
92
14 Lettuce_romaine_7wk
107
15 Vinyard_untrained
727 6541
16 Vinyard_vertical_trellis
181 1626
Total
5418 48711
Table 6. Classification accuracy (%) for the testing set of the Salinas image.
Boldface indicates the best-performing classifier(s) for each class.
Notes
:
SVM
=
support vector machine;
OMP
= orthogonal matching pursuit;
CRC
= collaborative
representation;
SOMP
= simultaneous orthogonal matching pursuit;
NRS
= nearest
regularized subspace;
JCR
= joint collaborative representation;
KCRT-CK
= kernel
collaborative representation with Tikhonov regularization and composite kernel;
SC
= sparse coding;
CSC
= collaborative sparse coding;
CSCSR
= collaborative
sparse coding with smoothness regularization;
OA
= overall accuracy;
AA
=
average accuracy.
Class SVM OMP CRC SOMP NRS JCR KCRT-CK SC CSC CSCSR
1
99.88
100 100 100
99.78 99.83
100
100 100 100
2
99.79
100 100 100
99.70 99.82
100
99.76 99.73
100
3
98.37
100 100
99.94 98.71 99.16
100
99.89 99.89
100
4
98.72 99.92 99.70 99.52 99.28 99.36 99.12
100
99.92 99.52
5
97.97 99.63 99.84 99.71 98.80 99.00 99.54 99.33 99.25 99.63
6
99.89
100
99.97
100
99.94 99.97 99.97
100 100 100
7
99.81
100
99.97
100
99.66 99.66 99.84 99.85 99.88
100
8
92.42 99.94 97.27 99.05 99.17 99.82 97.97 93.70 93.75
100
.86
100
99.89 99.84
100
.39 99.53 98.27 98.34
100
.86
100
99.89 99.79
100
00
100
100 100 100
13 98.18 99.87 98.85 98.67 99.76 99.51 98.67
100 100 100
14 98.44 96.78 95.77 98.75 98.65 98.75 96.05 96.46 96.57 99.27
15 87.74 98.21 90.80 91.88 91.18 96.18 98.07 89.65 94.92 99.99
16 98.58
100
99.65
100
97.91 98.28 99.94 98.95 98.89
100
OA 96.18 99.65 97.93 98.62 98.23 99.14 99.13 97.01 97.72 99.95
AA 97.84 99.64 98.72 99.20 98.69 99.22 99.29 98.48 98.80 99.99
κ
95.75 99.62 97.70 98.46 98.04 99.05 99.03 96.67 97.46 99.95
training and testing samples for each class are shown in Table
5. The training and testing sets are visually shown in Figure
17a and 17b, respectively. The classification results, which
include the classification accuracies for each class, the overall
accuracy, and the kappa coefficient, are summarized in Table
6. It can be seen that the proposed
CSCSR
method yields the
best overall accuracy, kappa coefficient, and classification ac-
curacy for most classes, where the weight parameters were set
as
λ
1
= 0.5 and
λ
2
= 0.05. The classification maps of the labeled
pixels are shown in Figure 17. As can be seen, the proposed
CSCSR
method provides the best visual classification results
compared with other methods.
The effects of the parameters on the classification results of
the Salinas image are shown in Figure 18. As can be seen, the
optimized settings of
λ
1
and
λ
2
have a certain positive effect
on the classification results, which are better than those of the
other methods for any parameter setting within an appropriate
range. When the percentage of the training data is varied as
670
September 2019
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
