construct the feature data at level-1, and the constraint of
object information after the pixel-wise change map reduced
the commission ratio. It can be seen from the tables that the
accuracy of
OBCD
is significantly higher than that of the
PWCM
because the
PWCM
is based on a single pixel, and the spectral
information in a high-spatial-resolution image is insufficient
resulting in more commission pixels. The detection results of
different heterogeneous ensemble systems are different and
the accuracy of the three heterogeneous ensemble algorithms
are not all higher than that of the individual base classifiers,
which demonstrates that the selection of the ensemble algo-
rithm has impact on the performance of the multi-classifier
ensemble system and that not all the ensemble methods can
improve the accuracy. Comparing the methods based on the
SG
hybrid ensemble system, we can observe that the overall
accuracy of
SG
-ff are higher than that of
SG
-
LS
,
SG
-
SVM
and
SG
-
MLR
due to better optimization capability of
FOA
. The commis-
sion ratio of
SG
with the parameters of
MRLR
based on
FOA
and
NNLS
is improved when we utilize the object information to
impose constraints on the pixel-wise change detection maps.
Moreover, it can be observed that the
SG
ensemble system
costs less time than other ensemble systems with higher accu-
racy. This is because in base classifiers we utilize the advan-
tages of
ELM
to classify a large part of pixels and combine the
series and parallel framework to classify other pixels.
(a) ELM
(b) KNN
(c) SVM
(d) MLR
(e) ELM-OB
(f) KNN-OB
(g) SVM-OB
(h) CVA-OB
(i) F_int
(j) D-S
(k) MV
(l) SG-LS
(m) SG-ff
(n) SG-SVM (o) SG-MLR
(p) SG-LS-OB (q) SG-ff-OB
(r) reference
Figure 7. Change detection and reference maps obtained by the different methods for the first dataset.
(a) ELM
(b) KNN
(c) SVM
(d) MLR
(e) ELM-OB
(f) KNN-OB
(g) SVM-OB
(h) CVA-OB
(i) F_int
(j) D-S
(k) MV
(l) SG-LS
(m) SG-ff
(n) SG-SVM (o) SG-MLR
(p) SG-LS-OB (q) SG-ff-OB
(r) reference
Figure 8. Change detection and reference maps obtained by the different methods for the second dataset.
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
November 2018
739
