step was then performed using the edge-constrained MSEG algorithm, and specifically the Canny edge features option was used. A set of primitive objects was obtained
and object properties were extracted (spectral and shape features). The same training
test was given to the SVM classifier and a final classification of objects was obtained.
As shown in Table 9.1, the overall accuracy of the developed method outperformed
the previous tests with an accuracy of 90.29%. This shows that edge features helped
the segmentation procedure to obtain more meaningful objects that are capable of
providing very good classification results. This procedure was repeated again with
some different parameters and similar results were produced. Of course, the difference
in accuracy is not wide, but it is a measure that compatible results are produced for
further OBIA classification steps.
9.4.2 Radar Satellite Imagery
Experimental results include the application of the developed methodology at highresolution SAR data (TerraSAR-X data set). The initial SAR image is shown in
Figure 9.3a and the output results from the edge-constrained segmentation are
compared with the ones from the mean shift and watershed (Figures 9.3b and c).
The mean-shift method did not perform well and resulted in under segmentation of the
shore line (Figure 9.3b). Even if the water area was successfully segmented into one
TABLE 9.1 Quantitative Results of Classification Accuracy for High-SpatialResolution Airborne Multispectral Data Set
Vegetation
Tile Roofs
Bright Roofs
Asphalt Like
Classification Accuracy with MSEG Only
Vegetation
15,247
0
0
2,539
Tile roofs
198
2,856
15
2,849
Bright roofs
0
1
8,362
2,064
Asphaltlike
215
34
498
35,612
Overall accuracy: 88.07%
Classification Accuracy with AML
Vegetation
15,523
0
0
2,263
Tile roofs
15
3,764
124
2,015
Bright roofs
0
0
8,389
2,038
Asphalt like
583
30
482
35,264
Overall accuracy: 89.29%
Classification Accuracy with AML and Edge Enhancement
Vegetation
15,791
34
0
1,961
Tile roofs
244
4,710
45
919
Bright roofs
0
0
8,311
2,116
Asphaltlike
141
909
475
34,834
Overall accuracy: 90.29%
Note: The proposed OBIA methodology scored better, indicating that the enhancement of MSEG with
advanced edge features along with advanced scale-space representations (AML) and the kernel classifier
(SVM) outperforms earlier approaches.
186
MULTISCALE SEGMENTATION AND CLASSIFICATION
and object properties were extracted (spectral and shape features). The same training
test was given to the SVM classifier and a final classification of objects was obtained.
As shown in Table 9.1, the overall accuracy of the developed method outperformed
the previous tests with an accuracy of 90.29%. This shows that edge features helped
the segmentation procedure to obtain more meaningful objects that are capable of
providing very good classification results. This procedure was repeated again with
some different parameters and similar results were produced. Of course, the difference
in accuracy is not wide, but it is a measure that compatible results are produced for
further OBIA classification steps.
9.4.2 Radar Satellite Imagery
Experimental results include the application of the developed methodology at highresolution SAR data (TerraSAR-X data set). The initial SAR image is shown in
Figure 9.3a and the output results from the edge-constrained segmentation are
compared with the ones from the mean shift and watershed (Figures 9.3b and c).
The mean-shift method did not perform well and resulted in under segmentation of the
shore line (Figure 9.3b). Even if the water area was successfully segmented into one
TABLE 9.1 Quantitative Results of Classification Accuracy for High-SpatialResolution Airborne Multispectral Data Set
Vegetation
Tile Roofs
Bright Roofs
Asphalt Like
Classification Accuracy with MSEG Only
Vegetation
15,247
0
0
2,539
Tile roofs
198
2,856
15
2,849
Bright roofs
0
1
8,362
2,064
Asphaltlike
215
34
498
35,612
Overall accuracy: 88.07%
Classification Accuracy with AML
Vegetation
15,523
0
0
2,263
Tile roofs
15
3,764
124
2,015
Bright roofs
0
0
8,389
2,038
Asphalt like
583
30
482
35,264
Overall accuracy: 89.29%
Classification Accuracy with AML and Edge Enhancement
Vegetation
15,791
34
0
1,961
Tile roofs
244
4,710
45
919
Bright roofs
0
0
8,311
2,116
Asphaltlike
141
909
475
34,834
Overall accuracy: 90.29%
Note: The proposed OBIA methodology scored better, indicating that the enhancement of MSEG with
advanced edge features along with advanced scale-space representations (AML) and the kernel classifier
(SVM) outperforms earlier approaches.
186
MULTISCALE SEGMENTATION AND CLASSIFICATION