.
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008
3.2 Information Extraction from Radarsat-1 SAR Image
Considering some of the fractals have the same fractal
dimensions but have the different textures, on the basis of
textural analysis using fractal dimensions, this research further
introduces multi-fractal theory and the second-order statistic
lacunarity as the effective supplements to fractal dimension for
textural analysis. Samples of typical ground objects were
extracted, then the multi-fractal q-D(q) curve and the lacunarity
L-C(L) curve were plotted in order to quantitatively determine
the optimum parameters for effective fractal features extraction
(Figure 2,3). From Table 2 and Table 3, we can see that when q
equals to 8 and -8, L equals to 2, there is good separability
among ground objects, so they were selected as the effective
parameters for feature extraction.
q-D(Q)
-waterbody
-bare land
village
town
-industrial premises
-reed
-cropland
-vegetation plot
-10 -9-8 -7 -6 -5 -4 -3 -2 -1 0 2 3 4 5 6 7 8 9 10
q moment
Figure 2. Multi-fractal q-D(q) curve of the typical ground
objects
Lacunarity
1.5
j—I—I 1—I—I—I—I
•build-up land
-waterbody
bare land
vegetation
2 3/ 4 5 6 7 8 9 10
Figure 3. Lacunarity L-C(L) curve based on DBC method
Based on SVM, and using the seven multi-scale and multi
texture textural features mentioned in 2.2 as input, Radarsat-1
SAR image can be classified (Figure 4).
Legend
bui Id-up land
water body
bare land
vegetation
Figure 4. Classification result of SAR image based on multi
scale and multi-texture feature fusion and SVM
(part of the test site)
In order to test the classification performance of different
textural features, based on SVM, this research designed a
classification scheme (Table 3). From Table 3 and Table 4, it
can be seen that the fused seven features can reach better
classification performance than any single or any other
combination of the features and the improvement of the
classification accuracy is significant.
Group
No.
Femme for classification
Overall
accuracy %
Kappa
coefficient
1
multi-scale and multi-texture feature (Multi-scale
GLCM+Three types of fractal features)
69.8926
0.4916
2
Multi-scale GLCM
67.2711
0.4680
3
Three types of fractal features (FD+multiFD+Lacu)
66.8600
0.4452
4
Multi-scale GLCM+FD
68.1001
0.4690
Multi-scale GLCM+multiFD
68.2594
0.4700
Multi-scale GLCM+Lacu
68.5601
0.4751
5
Multi-scale GLCM+FD+multiFD
68.5857
0.4789
Multi-scale GLCM+FD+Lacu
69.4600
0.4803
Multi-scale GLCM+multiFD+Lacu
69.4831
0.4889
6
SAR intensity image
54 4739
0.2395
SAR hackscattering coefficient image
45.9268
0.2139
Table 3. Classification accuracy analysis of SAR image based
on multi-scale and multi-texture feature fusion and SVM
Z value
Group 1
Group 2
Group 3
Group 1
Group 2
37.76787
Group 3
57.63155
19.93228
Table 4. Comparison of Z statistic
The classification performance between traditional maximum
likelihood classification and SVM was also compared in this
research. Using the same seven fused textural features as input,
the classification accuracy is compared and Z statistic is given
in Table 5. From Table 5, we can see that overall accuracy of
SVM classification is higher than that obtained using MLC by
10%, Kappa coefficient is improved, and the Z statistic is
71.6227. It can be concluded that SVM is an effective classifier
for textural features, and it can significantly improve the
classification accuracy compared with the MLC classifier.
1080