higher than the corresponding individual accuracy of
PCA
,
FCM
, and
EM-MRF
. The better performance of our proposed
method clearly indicates the benefit of the complementary
feature selection in our proposed method for urban change
detection.
Discussion and Conclusions
This paper addresses the detection of land-cover change from
the bi-temporal remote-sensing images. The proposed proce-
dure mainly uses information from a very novel group of ob-
servations: luminance and saturation. Their nature for identi-
fying different types of change occurred in urban area has been
exploited; a procedure based on combining the two features is
created by integrating automatic radiometric normalization, T-
point thresholding, Bayes Fusion and Markov Random Field.
For overall accuracy assessment, the proposed procedure is
superior over three earlier referenced unsupervised methods.
The key component for our proposed model is feature
design. We think the best feature number should be two since
more features would greatly increase computation for MRFs
modeling (e.g., three features will produce 27 initial classes
for implementing MRFs). An efficient procedure for designing
features should include the consideration of both (a) feature
independence, and (b) separability of our change of interest
from multiple changes. Since our features are derived from
HSL
color modal, their independence can be guaranteed for
the subsequent Bayes classifier fusion; both visual and quan-
titative tests in our paper have indicated their perfect comple-
mentary nature for identifying only change of interest while
keeping noisy types excluded.
Luminance feature is mainly contaminated by
low reflec-
tance
to
high reflectance
(
LR->HR)
, which mostly occurs for
the built-up area, such as the example of ‘A’ region in Plate
1a and Figure 4b. This can be explained by the fact that hu-
man activities often modify the surface of the built-up area
(such as roof or road renovation), which results in reflectance
change. Saturation feature easily results in false inclusion of
inter-class changes of vegetation (
IC
) and water quality change
(
QC
) as these two noisy changes mainly modify the color
information of the land surface and slightly affect the reflec-
tance level. Generally, saturation is less affected by local-re-
flectance changes since such changes are considered to exert
roughly equivalent influences on three bands.
The results from the experiments indicate that the pro-
posed procedure offers measureable advantages over the
earlier unsupervised change detection (Plate 2 and Figure 4).
The traditional techniques, such as
FCM
and
EM-MRF
, select
changed pixels based only on the “measureable distances”
to the center of changed and unchanged class, without any
step for feature selection. This would lead to some errors. For
example, if change is determined based on spectral bands
(such as Red, Green, and Blue band), the noisy change of
varied local illumination would exert changed magnitude for
all the feature bands. As a result, some unchanged land-cover
has a high variance of pixel values with large distances to the
center of general unchanged class in the feature space, and
thus is easy to be falsely classified as change class. Although
PCA
transform can be thought of as a method based on feature
selection (the second component is chosen), it is not efficient
since there is only one feature used for change detection.
It is noteworthy that the task of feature selection is prob-
lem-dependent, and heavily relies on the knowledge of the
application domain. The proposed method is only tested for
urban change-detection; for other applications such as forest
damage or wetland monitoring, the complementary nature
of lightness and saturation cannot be guaranteed since “real
change” and “noisy change” need to be redefined. For future
research, the exploitation of more change features and intro-
duction of supervised frameworks remains to meet a variety
of application scenes.
Acknowledgments
This research was supported by a National Science and Engi-
neering Research Council (NSERC) Discovery Grant awarded
to Chen. We would like to thank Masroor Hussain, the editor
and the two anonymous reviewers for their constructive com-
ments and suggestions for improving the quality of the paper.
References
Benedek, C., and T Szirányi, 2009. Change detection in optical aerial
images by a multilayer conditional mixed Markov model,
IEEE Trans-
actions on Geoscience and Remote Sensing
, 47(10):3416–3430.
Bruzzone, L., and D.F. Prieto, 2000. Automatic analysis of the
difference image for unsupervised change detection,
IEEE Trans-
actions on Geoscience and Remote Sensing
, 38(3):1171–1182.
Bruzzone, L., and D.F. Prieto, 2002. An adaptive semiparametric and
context-based approach to unsupervised change detection in
multitemporal remote-sensing images,
IEEE Transactions on
Image Processing
, 11(4):452–466.
Chen, D., and S, Ye, 2015. Comparison of threshold selection methods
for change detection from remotely sensed images,
Journal of
Applied Remote Sensing
, (In review).
Cheng, H.-D., X. Jiang, Y. Sun, and J. Wang, 2001. Color image
segmentation: Advances and prospects,
Pattern Recognition
,
34(12):2259–2281.
T
able
2. E
rror
M
atrixes
for
C
hange
D
etection
R
esults
from
PCA, FCM,
EM-MRF
and
the
P
roposed
M
ethod
Classification
data
Reference Data
Change
No
change
Row
total
User’s
(%)
PCA: Overall accuracy=83.8% Kappa statistic: 50.6%
Map data Change
100
41 141 70.9
No change
89 570 659 86.5
Column total
189 611 800
-
Producer’s (%)
52.9 93.3 -
-
FCM: Overall accuracy=82.0% Kappa statistic: 46.6%
Map data Change
98
43 141 69.5
No change
101 558 659 84.7
Column total
199 601 800
-
Producer’s (%)
49.2 85.6 -
-
EM-MRF: Overall accuracy=91.9% Kappa statistic: 73.1%
Map data Change
116
25 141 82.3
No change
40 619 659 93.9
Column total
156 644 800
-
Producer’s (%)
74.4 96.1 -
-
Proposed: Overall accuracy=95.1% Kappa statistic: 83.3%
Map data Change
122
19 141 86.5
No change
20 639 659 97.0
Column total
142 658 800
-
Producer’s (%)
85.9 97.1 -
-
644
August 2015
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
