new threshold values. 
include more than 
through narrow 
the shrinking-and-expanding 
to isolate each 
instructions such as 
If a region was likely to 
one feature connected 
channels, 
method was employed 
independent feature. 
  
  
  
Figure 18. New 
buildings extracted 
from the SPOT image 
from the first pass. 
Figure 17. 
Segmentation result 
of the SPOT image 
from the first pass. 
The segmentation result for the second pass 
on the SPOT image is shown in Figure 20. 
Regions which were incorrectly segmented in 
the first pass now appear to have more 
building-like shapes. For example, the 
buildings indicated by the arrow were 
successfully decomposed by the shrinking- 
and-expanding technique. In addition, 
existing regions for which only the edges 
were segmented in the first pass, were 
successfully segmented with new threshold 
values. New buildings extracted in the 
Second pass were merged with those 
extracted in the first pass. In this way, 
newly extracted buildings in each pass were 
accumulated in the building file. The final 
result for the test area after four 
iterations is shown in Figure 21. 
  
  
  
  
  
Mask Figure 20. 
Segmentation result 
of the SPOT image 
from the second 
pass. 
Figure 19. 
image of focused 
areas where further 
processing is 
required. 
The accuracies of this result were 
calculated by comparing Figure 21 with the 
buildings identified by manual techniques 
on the original SPOT 10-m pixel image 
(Figure 22). The. ratio. of Correctiy 
extracted. features to manually detected 
changes (i.e., accuracy) was 82 % for the 
experiment. The omission errors were mainly 
small or dark features which might not have 
been extracted by photo-interpreters. 
574 
  
Figure 21. 
  
  
  
New 
Figure 22. 
buildings manually 
extracted from the 
SPOT image. 
Final 
result of new 
buildings extracted 
from the SPOT image. 
The same feature extraction procedure was 
applied to other areas of the SPOT image. 
In addition, the SPOT 10.m data were 
resampled to 2.5 m pixel resolution and 
feature extraction attempted with these 
"higher resolution" data (Figure 23). 
Machine feature extraction of buildings 
showed definite improvement using these 
resampled 2.5 m pixel resolution data 
(Figures 24 and 25). 
GF a Y 
   
Figure 23. SPOT image of 2.5 m pixel 
resolution for building extraction. 
  
  
  
  
Figure 25.Result 
of new building 
extraction from 
the SPOT image 
of the same area 
as Figure 23 but 
of 10 m pixel 
resolution. 
Figure 24. Result of new 
building extraction from 
the SPOT image in Figure 
23. 
  
5. CONCLUSION 
The map data used in this study provided 
the initial threshold values for image 
segmentation. The map data also 
demonstrated the capability of detecting 
large changes in existing buildings in the