6a, 6b, and 6c, we can find that when the number of iterations
exceeds 20,
DSC-CASSL
starts to widen the gap with other al-
gorithms. This can be attributed to the fact that: two verifica-
tion classifiers are trained on the labeled data set gradually
expanded by
AL
strategies. Two verification classifiers are
applied together with the classifier to promote pseudolabeling
accuracy and consistently accelerate
DSC-CASSL
.
In Tables 5, 8, 10, 11, and 12, T is the iteration number,
and Increase means that, compared to
CASSL
, the improve-
ments of accuracy or Kappa value achieved by
DSC-CASSL
.
We can observe that
DSC-CASSL
obtains better results by
labeling fewer samples, which means that potential infor-
mation-rich samples have more chance to be reselected. As
shown in Table 6, when the classifier is 83% for OA, the cost
is 635 by using
DSC-CASSL
algorithm, and 765 by using
CASSL
.
DSC-CASSL
saves about 130 manual labeling cost (about 17%).
To obtain the same accuracy,
MCLU
needs to manually label
825,
nEQB
needs to mark 925 samples, it’s sufficient to reflect
the effectiveness and superiority of
DSC-CASSL
.
In addition, the computation time of the algorithm also
represents the performance of algorithm. For the time cost, all
of the algorithms in our experiments are implemented using
Python Scikit-learning package. Compared with
CASSL
, we
can find that
DSC-CASSL
are more effective from Table 7. Gener-
ally speaking,
DSC-CASSL
may require more computation time
to train two check classifiers and make double verifications.
However, as a parallel programming framework,
DSC-CASSL
can exploit multiple architectures. It generates a set of paral-
lel informative samples selected form different
AL
algorithms
and assigns label to each one at each iteration. In summary,
the proposed framework can use less time to achieve better
performance.
To allow a visual inspection, when all the comparative
methods achieve the 60 iterations, the classification maps
obtained with the Indian Pines data set are displayed in Fig-
ure 7. The misclassifications between the spectrally similar
classes, such as soybean-min till, soybean-clean, and grass-
trees are obvious in Figure 7a and 7c. The classification map
of
DSC-CASSL
appears clear, which affirms the efficiency of
proposed framework.
Experiment on the KSC Data Set
From Table 8, we can observe that
DSC-C
able improvement on the KSC data set.
are labeled, compared to
CASSL
, the accu
achieved by
DSC-CASSL
are more than 2%. It means that
DSC-
CASSL
is more efficient to boost the learning accuracy when
there are certain amounts of labeled data. However, from the
25th iteration, the accuracy of
CASSL
finishes improving and
nearly collapses. This can be attributed that the verification
classifier in accord with the base classifier, and there is no
suitable unlabeled samples labeled, which leads to the algo-
rithm converges quickly and degrade the performance of the
final classifier. In summary,
DSC-CASSL
can address problems
created by
CASSL
.
DSC-CASSL
utilizes the double verification
methods to avoid classifiers always having same prediction
on a unlabeled sample. And different selection algorithms can
complement for each other and stimulate the performance of
overall framework.
As shown in Figure 8c, in the KSC data,
DSC-CASSL
needs
the least effort to label samples, when these methods achieve
the same classification performance. When OA achieves
85%,
DSC-CASSL
only needs 75 human labeling, while
CASSL
requires 105 human labeling, so the savings rate is 28%.
This case reflects the effectiveness of
DSC-CASSL
algorithm.
From Table 9, we can observe that
DSC-CASSL
need more time
to achieve the classification. The main reason is that
CASSL
doesn’t increase the accuracy from 25th iterations. However,
Table 6. The comparison of labeling cost with different
algorithms (
MCLU
,
nEQB
,
CASSL
, and
DSC-CASSL
) on the Indian
Pines dataset.
OA, %
Algorithm, CNY
MCLU nEQB
CASSL DSC-CASSL
77
345
375
355
325
79
435
485
445
385
81
575
635
555
505
83
825
925
765
635
85
1085
1245
1025
935
87
1545
1685
1485
1405
Table 7. Total training time of five compared algorithms on
the Indian Pines dataset.
Algorithm, s
RS MCLU nEQB CASSL DSC-CASSL
Time of
90 iterations
297.41 393.10 1553.34 4816.27 4360.22
Table 8. The comparison of Overall Accuracy between the
compared algorithms and
DSC-CASSL
on the
KSC
data set.
T
Algorithm, %
Increase, %
RS MCLU nEQB CASSL DSC-CASSL
10 80.78 85.35 83.12 84.47
86.61
+2.53
15 82.99 87.59 86.57 87.35
89.31
+2.24
20 84.46 89.53 88.42 88.27
90.78
+2.84
25 86.06 90.69 89.98 88.92
91.37
+2.67
30 86.69 91.27 90.83 88.99
91.91
+3.28
35 87.54 92.00 91.45 88.98
92.48
+3.93
40 88.34 92.36 92.10 89.07
92.91
+4.31
(a)
(b)
(c)
(d)
Figure 7. Comparison of the classification map of different
framework on the Indian Pines data set. (a)
MCLU
. (b)
nEQB
.
(c)
CASSL
. (d)
DSC-CASSL
.
848
November 2019
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
