The Inter national Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008
1081
Classifier
Class
Producer's
User's
Overall
Kappa
accuracy %
accuracy %
accuracy %
coefficient
built-up land
45.57
72.34
MLC
bare land
43.29
63.76
vegetation
64.56
52.52
59.1143
0.3816
water body
53.42
48.72
built-up land
77.15
64.39
SVM
bare land
vegetation
60.76
70.98
66.61
69.8926
0.4916
77.54
water body
50.4P
53.23
Z value 71.6227
Table 5. Comparison of Z statistic for MLC and SVM
3.3 Change Detection Result and Uncertainty Analysis
By the direct post-classification comparison, the change
detection result can be obtained (Figure 5). Where, the black
represents no change, the green represents positive difference
Figure 5. Change detection result of the test site based on the
hard-decision (partial)
values and the red represents negative difference values (the
classification code is: bare land 1, built-up land 2, vegetation 3
and water body 4). In order to effectively evaluate the change
detection accuracy, special effort sampling was done and the
number of samples was calculated according to multinomial
distribution recommended by Khorram (Siamak Khorram et al.,
1999). According to Table 6, the overall accuracy of the change
detection result is 62.7%.
Reference data
Sum in row
Classification
data
unchanged
changed
unchanged
163
37
200
changed
187
213
400
Sum in column
350
250
600
Table 6. Accuracy analysis for the hard-decision change
On the basis of the extended probability vector of the
classification result with different temporal, and according to
the probability entropy model of uncertainty propagation, the
spatial distribution of uncertainty of change detection result at
the scale of pixels can be obtained, which is shown in Figure 6.
From this figure, we can see that the blacker area represents
smaller uncertainty of the change detection result; usually, there
exists higher uncertainty on the fringe of different land
use/cover types.
Figure 6. Spatial distribution of uncertainty of the change
detection result represented by probability entropy
Interval of
entropy (A)
0-0.4
0.4-0.8
0.8-1.2
1.2-1.6
1.6-2.0
Pixel percentage
(%)
39.91
6 94
30.97
2.54
14.19
Accumulative
percentage
(%)
39.91
46.85
77.82
80.36
94.55
Interval of
entropy (A)
2.0-2.4
24-2.8
2.8-3.2
3.2-3.6
3.6-4.0
Pixel percentage
(%)
4.81
0.64
0.0
0.0
0.0
Accumulative
percentage
(%)
99.36
100.0
100.0
100.0
100.0
Table 7. Uncertainty of the change detection result
Separate the entropy value into ten intervals according to the
range of entropy, then count the number of pixels falling in
between each interval. From Table 7, it can be seen that nearly
95% pixels have the uncertainty less than 2.0. It means that
although uncertainty increases during the process of error
propagation, the change detection result by hard-decision still
has the acceptable certainty level in this research.
Employ the soft-decision change detection method described in
2.3 for the further analysis (Figure 7). We can see that by
executing the rules in turn, lots of false detected changes are
eliminated, for example, after executing the rule 1-3, the area of
detected changes decreased by 22.6%, and after executing the
rule 4-6, the area of detected changes decreased by another
28.9%.