Figure 7 and Figure 8 show the change maps and reference
maps for the two datasets. It can be seen from the results of
the change maps that the changed regions in the first dataset
mainly includes the changes of vegetation and bare land to
roads and buildings. Compared with the reference change
map in Figure 7(r), the change detection results of
SG
-ff-
OB
shown in Figure 7(q) is more consistent with the reference
change map. After utilizing the segmented object information
constraint, some of the salt-and-pepper noise in the change
detection result based on the
SG
hybrid ensemble system is
suppressed. As shown in Figure 7(a)- to (d),(i) to (k), it can
be observed that there are many false-alarm pixels in the left
and lower part of the change maps obtained by
PWCM
and
HEAM
, whereas the pixels belonging to false detection class
are significantly less in Figures 7(p) and (q) obtained by the
proposed methods. That is because the areas in left part are
mainly dense residential buildings where the frequency of
gray scale change is large. However, the spectral informa-
tion of high-spatial-resolution image is insufficient and the
PWCM
is based on pixel units, thus generating more noise and
commission pixels. From the results of Figures 7 (a), (b) and
(c), we can see that there are less false-alarm pixels in the left
part but some missed pixels in the middle part of the image
obtained by
ELM
, which is complementary with the other two
methods. As can be seen from the Figures 7(e), (f), (g), and
(h), some changed regions of the maps obtained by OBCDs are
mis-identified as unchanged regions in the change detection
maps due to the dependence of the segmentation scale of the
OBCD
. The
SG
methods yield less false-alarm pixels in change
detection map, due to the consideration of spatial information
when constructing the input data at level-1. The changed re-
gions in the second dataset are mainly including the increase
of buildings and roads and the decrease of vegetation and
bare land. Similarly, the change maps in Figures 8(p) and (q)
obtained by the
SG
ensemble system methods,
SG
-
LS
-
OB
and
SG
-ff-
OB
, are more consistent with the reference map in Figure
8 (r). As shown in Figures 8 (a), (b), (c) and (d), the results of
the
PWCM
and Figures 8(h), (i) and (j), the results of the three
HEAM
, contain salt-and-pepper noise in all the maps, whereas
the pixels belonging to false detection class are significantly
less in the map Figure 8(q). Meanwhile, as compared with the
reference map in Figure 8(r), the salt and pepper noise can be
effectively suppressed in Figures 8(e), (f), (g), and (h) obtained
by the
OBCD
, but due to the influence of the segmentation er-
ror, some changed areas are identified as unchanged regions
in the lower part of the change maps, whereas the regions can
be detected accurately by the proposed methods.
In order to show the generalization capability of the pro-
posed method, an additional dataset collected by the Quick-
Bird satellite is also used. The images were collected by four
multispectral bands and a panchromatic band and the fusion
resolution is 0.6 meters. The multi-temporal dataset contains
800 × 700 pixels. The true color images are shown in Figure 9.
For this dataset, we used the same parameter of as in the
ZY-3
datasets. The spectral and texture features were utilized
as the input datasets. The segmentation size was set as 40-
0.5-0.5. The accuracy of different methods for the additional
dataset is shown in Table 3. As can be seen, the accuracy of
SG
-ff-
OB
is also the highest among all the methods. We can see
the same results as in the
ZY-3
datasets, which further demon-
strated the generalization of the proposed method.
Conclusions
A simple multi-classifier hybrid ensemble system for pixel-
wise change detection and an object-based approach to
change detection using constrained analysis were proposed
in this paper. The
SG
hybrid ensemble system uses
ELM
,
KNN
,
and
SVM
as the level-0 base classifiers, and
MRLR
as the level-1
meta-classifier by using the spectral, statistical texture and
structure texture (morphological profile) features as input.
In addition to the prediction results of base classifiers, the
weighted average of eight neighboring pixels were also used
as the input of meta-classifier. The
FOA
, one of the recently
developed global optimization algorithms, was adopted to
estimate the model parameters in
MRLR
. In order to utilize
the advantages of high-resolution remote sensing images and
decrease the direct influence of segmentation, constrained
analysis on segmented objects was implemented to integrate
segmentation map and pixel-wise change map for final result.
The experimental results demonstrate that the proposed
method performs better with lower computational cost.
Acknowledgments
This research is supported in part by Natural Science Founda-
tion of China (No. 41471356), the Xuzhou Scientific Funds
(KC16SS092), and Priority Academic Program Development
of Jiangsu Higher Education Institutions.
References
Azzouzi, S.A., A. Vidal, and H. A. Bentounes, 2015. Enhancement
of the double flexible pace search threshold determination for
change vector analysis,
ISPRS - International Archives of the
Photogrammetry, Remote Sensing and Spatial Information
Sciences
, XL-7/W3(7):599-603.
Benedek, C., M. Shadaydeh, Z. Kato, T. Szirányi, and J. Zerubia, 2015.
Multilayer Markov Random Field models for change detection in
optical remote sensing images,
ISPRS Journal of Photogrammetry
& Remote Sensing
, 107:22-37.
Bruzzone, L., and D.F. Prieto, 2000. A minimum-cost thresholding
technique for unsupervised change detection,
International
Journal of Remote Sensing
, 21(18):3539-3544.
Chang, N.B., M. Han, W. Yao, L.C. Chen and S.G. Xu, 2010. Change
detection of land use and land cover in an urban region with
SPOT-5 images and partial Lanczos extreme learning machine,
Journal of Applied Remote Sensing
, 4(1):2816-2832
Chellasamy, M., T.P.A. Ferré, M. Humlekrog Greve, R. Larsen, and
U. Chinnasamy, 2014, An ensemble classification approach for
improved land use/cover change detection,
ISPRS - International
Archives of the Photogrammetry, Remote Sensing and Spatial
Information Sciences
, XL-8(1):695-701.
Chen, J., P. Gong, C. He, R. Pu, and P. Shi, 2003. Land-use/land-
cover change detection using improved change-vector analysis,
Photogrammetric Engineering & Remote Sensing
, 69(4):369-379.
Du, P., J. Xia, W. Zhang, K. Tan, Y. Liu, and S. Liu, 2012. Multiple
classifier system for remote sensing image classification: A
review,
Sensors
, 12(4): 4764-4792.
Table 3. Accuracy of the different methods for the additional
dataset.
Method
OA
Commission
ratio
Omission
ratio
PWCM
ELM 0.8745
0.3638
0.0405
KNN 0.8348
0.4510
0.0269
SVM 0.8582
0.4022
0.0387
OBCD
ELM-OB 0.8742
0.3713
0.0316
KNN-OB 0.8121
0.4868
0.0261
SVM-OB 0.8679
0.3833
0.0349
HEAM
MV
0.8606
0.4004
0.0328
D-S
0.8976
0.2752
0.0563
F_int
0.8326
0.4518
0.0699
SG-ff
0.8829
0.3447
0.0379
SG-ff-OB 0.9219
0.2087
0.0445
740
November 2018
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
