The Inter national Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008
1079
2.4 Uncertainty of Change Detection Result
Uncertainty, or error, exists from remote sensing data
acquisition to each stage of data processing. Understanding the
nature and spatial distribution of uncertainty during the process
of change detection result analysis, can reduce the risk of wrong
decision based on the uncertain data. Based on the classification
results, there are several uncertainty propagation models of
change detection such as the product rule-based approach, the
certainty factor-based approach (Shi and Ehlers, 1996), and the
probability entropy approach (Van der Wei et al., 1998; DE
Bruin and Gorte, 2000). In this research, we extend the
probability vector which is usually used for maximum
likelihood classification to the object-oriented fuzzy
classification and the multi-scale and multi-texture feature
fusion classification based on SVM, then further analyze and
evaluate uncertainty of the change detection result at the scale
of pixels based on the probability entropy model, which is
briefly described as below:
Different temporal image classification can be regarded as
independent. In other words, the posterior probability vector of
a pixel at time t2 is calculated irrespective of the class or feature
vector at the previous time, tl. According to Shannon’s
information theory, entropy is calculated as
M M
M 1=1
!X n )\o&(P(C„n IX n ))-YA C J.n IX n )\o&{P{C m !X n )) ( 1}
m M
Where,
P, = P(C, n ,C jT2 /X n ,X T2 ) = P(C, n /X TX )P(C jT2 /X r2 ),
M represents the number of classes. The range of H is from 0 to
log 2 (MM) , which means uncertainty varies from absolute
certain to absolute uncertain. 3
3. EXPERIMENTAL RESULTS AND ANALYSIS
3.1 Information Extraction from SPOT5 Image
Both SPOT5 2.5-meter panchromatic band and four 10-meter
multi-spectral bands are used in the object-oriented fuzzy
classification. The first level classes are built-up land,
vegetation, water body and bare land. Where, built-up land
consists of three subclasses which are building, building
shadow and road; vegetation consists of four subclasses which
are forest, grassland, cropland (including irrigated field and
nonirrigated field), and vegetable plot (including vegetable
greenhouse); water body consists of two subclasses which are
river/canal and lake/pond. NDVI image is also produced as the
additional band for classification. During the multi-resolution
segmentation, the weight of SPOT5 panchromatic band and
multi-spectral bands are assigned to one respectively. After
several times of experiments, a network of three layers is
constructed according to features of ground objects. The setting
of parameters is given in Table 1.
Level
Scale
Color
Shape
Composition of shape
Smoothness
Compactness
Level 3
110
0.8
0.2
0.5
0.5
Level 2
90
0.9
0.1
0.4
0.6
Level 1
40
0.8
0.2
0.4
0.6
Table 1. Parameter setting for multi-resolution segmentation
After the object hierarchy has been established, the nearest
neighbour classifier and the classifier of membership function
are integrated to obtain spectral, textural, shape, positional and
contextual information of image objects. Then, by establishing
the classification rules to realize the effective separation to
ground objects in the complex scene (Figure 1).
Legend
Building
Building
shadow
Road
Water
Bareland
Forest
Grassland
(Cropland
Vegetable
plot
Figure 1. Fuzzy object-oriented classification result of the test
site (partial)
From Table 2, it can be seen that for the second level
classification, the overall accuracy is 88.53%, Kappa coefficient
is 0.861; for the first level classification, the overall accuracy
reaches to 90.19% and Kappa coefficient reaches to 0.872.
Class
Producers
accuracy
%
User's
accuracy
%
Overall
accuracy
%
Kappa
coefficient
Building
85.28
92.11
building shadow
86.03
90.27
road
79.34
81.42
forest
83.00
71.18
grassland
84 47
73.65
88.53
0.861
cropland
90.91
84 43
vegetable plot
91.18
82.36
river/canal
79.16
76 02
lake/pond
92.56
86.72
bare land
80.81
85.75
built-up land
87.90
92.73
bare land
80.81
85.75
90.19
0.872
1
vegetation
92 25
85.97
water body
90.90
84.48
Table 2. Accuracy evaluation for the fuzzy object-oriented
classification
High performance of information extraction from optical
imagery is the basis for change detection with SAR data at the
information processing level.