ViVi Ja 
88 
sen. 
) AND 
ıned 
For 
and 
  
  
i | I 
n 9 | 
ok | | 
ri 
c5 / | | 
scm 
^1, em 
I 
5 4 4 
Oz | | 
(OI —_-_ —_ m 
ES | | 
UE | | 
E] a E | 
small medium ! large 
low 1 
Area 
Figure 13. Two-dimensional descriptor space 
spanned by area and elongatedness. 
Building candidates with insufficient 
probability need to be processed again from 
the original image with more suitable 
parameter values since further processing 
of the segmented image would add little 
useful information. Once building 
candidates are found and analyzed in the 
image in the first processing pass, 
however, their location and appropriate 
parameter values can be given for each of 
the candidates in the following processing 
passes (focusing mechanism). Each 
processing pass outputs newly identified 
buildings which are accumulated in a 
"building file" as shown in Figure 14. 
Several processing passes are required to 
record all the detectable buildings in the 
original image in the file. 
4. EXPERIMENT WITH SPOT IMAGE 
The method described in the previous 
section was implemented with the 
programming language C in a VAX Station 
3500 installed in the Center for Remote 
Sensing and Mapping Science of the 
University of Georgia, and applied to-a 
SPOT image. The rules were recorded in a 
text file separately from the expert system 
inference routine. 
4.1 Data Used in the Experiments 
A SPOT panchromatic image recorded on May 
4, 1986 covering Atlanta, Georgia was used 
as the source of the original image for the 
experiments. A test area was selected from 
the USGS 1:24,000 topographic map, 
"Chamblee, Ga." After image rectification, 
the test area was cut out from the original 
image and resampled to an image of 300 by 
400 pixels with 5im pixel resolution as 
shown in Figure 15. 
The black and purple separates of the 
topographic map were also digitized with a 
linear array CCD camera and rectified for 
the same area and pixel resolution as 
Figure 15. Noise in the resultant map 
Separate image was eliminated with image 
processing techniques (Figure 16). 
573 
  
Initial 
Parameter 
  
PI 
y 
Segmentation 
     
  
  
  
  
Region Descriptor 
Extraction 
    
  
  
   
   
  
  
  
New 
Parameter 
Values 
     
  
  
Building 
File 
   
  
Figure 14. Flow of new building extraction 
process. 
  
  
Figure 15. SPOT 
image of test area. separate image of 
test area. 
Figure 16. Map 
4.2 Result of Experiments 
The test area has a number of existing 
large industrial buildings and factories. 
The segmentation result from the first pass 
is shown in Figure 17. Four categories of 
segmented regions may be noted: 1) 
correctly segmented regions; 2) regions 
which contain more than one feature 
(multiple feature regions); 3) bright 
background (mostly bare ground); and 4) 
regions for which only the edges were 
segmented. Multiple feature regions were 
caused by the bright background surrounding 
buildings. Also edges tend to form where 
the intensity gradient of an object edge is 
not uniform relative to the background. 
The segmented images were then processed 
with the expert system to extract building- 
like regions using shape, size and tone 
descriptor values of each region. Figure 18 
shows the regions which were confirmed as 
new buildings (from the SPOT image) in the 
first pass. The remaining uncertain regions 
were then expanded as shown in Figure 19 to 
make a mask image for the next pass. In the 
second and later passes, the input SPOT 
image was processed only within these 
masked areas with specific processing