DSC-CASSL
still optimizes the performance, which needs more
time to assign pseudolabels and label samples.
Experiment on the Pavia University Data Set
From Tables 10–12, at the beginning of iteration,
CASSL
out-
performs
DSC-CASSL
. Then, we can observe that
DSC-CASSL
is
superior to other algorithms on the Pavia university data set
when obtains certain amounts number of labeled data.
From Tables 13–14, we can find that the computational
costs for training
CASSL
is the most expensive. Compared with
DSC-CASSL
,
CASSL
requires more computation time and labeled
samples to increase accuracy. Figure 9 shows the final clas-
sification map of four comparative methods. We can observe
that
DSC-CASSL
reduces some misclassification in the middle
of classification map. These experimental results reveal the
superiority of
DSC-CASSL
.
(a)
(b)
(c)
Figure 8.
OA
,
AA
, and Kappa results of the different
algorithms on the
KSC
data set. (a)
OA
. (b)
AA
. (c) Number of
labeling costs.
Table 9. Total training time (in seconds) of five compared
algorithms on the
KSC
data set.
Algorithm, s
RS MCLU nEQB CASSL DSC-CASSL
Time of 40
iterations
47.42 54.27 173.11 197.87
424.96
Table 10. The comparison of Overall Accuracy between the com-
pared algorithms and
DSC-CASSL
on the Pavia University data set.
T
Algorithm, %
Increase, %
RS MCLU nEQB CASSL DSC-CASSL
30 87.42 90.86 90.63 91.00
90.91
-0.09
45 89.07 91.91 91.88 92.00
92.10
+0.10
60 89.89 92.64 92.63 92.60
93.02
+0.42
75 90.30 93.17 93.11 93.08
93.63
+0.55
90 90.94 93.62 93.66 93.58
94.04
+0.46
Table 11. The comparison of Average Accuracy between the com-
pared algorithms and
DSC-CASSL
on the Pavia University data set.
T
Algorithm, %
Increase, %
RS MCLU nEQB CASSL DSC-CASSL
30 87.49 89.31 88.74 89.88
89.23
-0.65
45 88.25 90.88 89.62 90.82
90.00
-0.82
60 88.91 91.29 90.94 91.00
91.02
+0.02
75 89.19 91.51 91.44 91.41
91.62
+0.21
90 89.22 91.69 92.01 91.48
92.10
+0.62
Table 12. The comparison of Kappa between the compared
algorithms and
DSC-CASSL
on the Pavia University data set.
T
Algorithm
Increase, %
RS MCLU nEQB CASSL DSC-CASSL
30 0.8338 0.8787 0.8757 0.8810 0.8791
-0.19
45 0.8549 0.8942 0.8922 0.8936 0.8948
+0.12
0.9023 0.9020 0.9071
+0.51
0.9088 0.9080 0.9154
+0.74
90 0.8796 0.9161 0.9160 0.9148 0.9204
+0.56
Table 13. The comparison of labeling cost with different
algorithms (
MCLU
,
nEQB
,
CASSL
, and
DSC-CASSL
) on the Pavia
University data set.
OA, %
Algorithm, CNY
MCLU nEQB CASSL
DSC-CASSL
91
320
315
310
315
92
460
480
450
460
93
690
720
690
615
94
930
1010
1025
885
Table 14. Examples of computational time (in seconds ) taken
from five algorithms (
RS
,
MCLU
,
nEQB
,
CASSL
, and
DSC-CASSL
) on
the Pavia University data set.
Algorithm, s
RS MCLU nEQB CASSL DSC-CASSL
Time of
90 iterations
710.44 893.84 4550.55 32791.27 27603.61
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
November 2019
849
