the robustness of the
FOA
. Therefore, we set
Z=20
for the
two datasets in the following experiments.
As previously mentioned, the
SG
ensemble system is
constructed to obtain the pixel-wise change detection
result, and then the change result is constrained by segmenta-
tion information to produce the final change map. The
OBCD
were used for comparison with the proposed method and the
objects information was used to constrain the change informa-
tion, so objects information was obtained with the method
previously described. According to the distribution of the two
regions in the experimental area, the segmentation size was
set to 20, the shape size was set to 0.4, and the compactness
size was 0.6 for the first dataset and 50-0.4-0.6 for the second
dataset. The threshold r was experimentally set to 0.15 for
the two datasets. The overlay charts of true color images and
boundaries of the segmented regions for the two datasets are
shown in Figure 6.
The accuracy of change detection for the two datasets are
shown in Tables 1 and 2, respectively.
SG
-
LS
is the
SG
ensem-
ble system with the parameters of
MRLR
obtained by the use
of
NNLS
.
SG
-ff is the
SG
ensemble system with the parameters
of
MRLR
obtained by the use of
FOA
.
SG
-
SVM
and
SG
-
MLR
are
the
SG
ensemble system with the
SVM
and
MLR
as the meta-
classifier, respectively.
SG
-
LS
-
OB
is the
SG
-
LS
algorithm refined
by object information, similarly,
SG
-ff-
OB
is the refined
SG
-ff.
All these methods, i.e.,
MV
,
D-S
, F_int, are heterogeneous
ensemble algorithms and consist of the same base classifiers.
As can be seen from the above results, the accuracy of
SG
-ff-
OB
is the highest among the methods, with the overall accuracy
of 0.9556and 0.9762, lowest commission ratio of 0.1922
and 0.1195, because the predicted results of base classifiers
and the spatial neighborhood information were utilized to
(a)
(b)
Figure 6. Overlay charts of true color images and boundaries
of the segmented regions.
Table 1. Accuracy of the different methods for the first dataset.
Method
OA
Commission
ratio
Omission
ratio Time(s)
PWCM
ELM 0.9351
0.3983
0.0309
5.08
KNN 0.9292
0.4423
0.0239 1957.15
SVM 0.9332
0.4204
0.0245 11.27
MLR 0.9317
0.4095
0.0363 0.305
OBCD
CVA-OB 0.9425
0.3540
0.0283
8.32
ELM-OB 0.9470
0.2339
0.0415
5.21
KNN-OB 0.9479
0.2030
0.0436 1818.92
SVM-OB 0.9533
0.1771
0.0388 10.29
HEAM
MV 0.9332
0.4195
0.0252 1957.94
D-S 0.9389
0.3441
0.0388 1966.74
F_int
0.9328
0.4022
0.0361 1979.52
SG
SG-LS 0.9403
0.3644
0.0300 24.92
SG-ff
0.9507
0.2462
0.0346 24.95
SG-SVM 0.9422
0.3576
0.0277 25.31
SG-MLR 0.9431
0.3461
0.0291 24.53
SG-LS-OB 0.9547
0.2244
0.0319 27.62
SG-ff-OB 0.9556
0.1922
0.0344 27.65
Table 2. Accuracy of the different methods for the second dataset.
Method
OA
Commission
ratio
Omission
ratio Time(s)
PWCM
ELM 0.9581
0.3153
0.0175
5.04
KNN 0.9522
0.3884
0.0068 1925.32
SVM 0.9614
0.3232
0.0090 25.58
MLR 0.9224
0.5271
0.0229
0.39
OBCD
CVA-OB 0.9611
0.2749
0.0203 13.80
ELM-OB 0.9628
0.2213
0.0250
8.11
KNN-OB 0.9660
0.1665
0.0261 2144.15
SVM-OB 0.9652
0.2371
0.0195 14.09
HEAM
MV 0.9610
0.3271
0.0088 1925.42
D-S 0.9657
0.2446
0.0177 1933.75
F_int
0.9517
0.3743
0.0146 1945.61
SG
SG-LS 0.9625
0.3108
0.0101 25.77
SG-ff
0.9720
0.1852
0.0162 25.84
SG-SVM 0.9621
0.3205
0.0084 26.16
SG-MLR 0.9650
0.2950
0.0094 25.37
SG-LS-OB 0.9748
0.1799
0.0131 28.67
SG-ff-OB 0.9762
0.1195
0.0174 28.84
738
November 2018
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
