with 
a to 
Ors 
zx 
at 
90 
d co 
ge 
from 
2425 
tion 
es 
  
n 
the 
e 6 
xel 
Se 
ant 
  
   
1 
  
  
  
Figure 4. Test pattern of building of 
different shape. 
In order to further clarify the 
relationship between pixel resolution and 
fidelity of descriptor value, another test 
image with 10 rectangles (Figure 8) was 
created and processed in the same way as 
above. Figures 9 through 11 show the 
results of calculated descriptor values. 
Those calculated from 10 m pixel resolution 
image show the expected values for the 5 
largest rectangles (6-10), whereas 
descriptor values derived from 2.5 m images 
are correct for rectangles 3-10 indicating 
the superiority of the higher resolution 
images. Similar results were derived for 
the other descriptors. 
A 10 m pixel resolution image is displayed 
in Figure 12. Human interpreters can easily 
define shape of rectangles as small as 
rectangle No. 3. In: order to obtain 
equivalent results by machine 
interpretation, however, descriptors must 
be derived from 2.5 m pixel resolution 
images. This indicates that building 
recognition with shape descriptors may 
require an image of four times smaller 
pixel resolution (16 times larger data 
volume) to be comparable to human 
interpreters. 
3.3 Inference with Uncertainty 
In this study, the expert system approach 
was employed since the interpretation of 
building candidates by applying human 
knowledge to their descriptor values is 
similar to the process in diagnostic expert 
systems. In implementing an expert system, 
uncertainty management and knowledge 
representation are the two most important 
factors to be specified. This study 
employed the approach of MYCIN (medical 
diagnostic expert system), i.e., the 
certainty factor model for uncertainty 
management and production rules for 
knowledge representation (Buchanan and 
Shortliffe, 1994). 
An example of simple production rules 
developed in this study for map revision is 
shown below. 
  
= 
ag 
r 
ml 
Mm 
4.1160 
2 
$ 
A 5 
À 6 
Xx 
X 8 
+2 
-e 
- 
1 
= 
1 
  
  
  
so, Large Objects 
  
  
   
   
  
  
45 A An 
<+ 40 emm m mint dre miu nmt 
M5 denuo A arin.Ez YU L2 x 
a Ce 
S25 IN ar rik 
d 20 eM. oro 
———$9 L4 
C3 15 A] 
  
0° 10 20 3% 40 50 60 70 80 90 
Rotation Angle (degree) 
Figure 5. Elongatedness plotted against the 
rotation angle for the image of 10 m pixel 
resolution. 
Large Objects 
BY neste En 
tT X 
Xe = AAT 
0 30 X= Spay SD XX 
  
    
  
  
  
Sas Io err NN ET) 
Soie te 
c3 
19 
5” jn Adele temm rea rcr en 
Lr Bll-:---- L.X emm meme me mmm 
0 bee | | i i i i | 
0 10 30 40 50 60 7 80 90 
20 0 
Rotation Angle (degree) 
Figure 6. Elongatedness plotted against the 
rotation angle for the image of 2.5 m pixel 
resolution resampled from the 10 m test 
image. 
Large Objects 
  
  
60 
SO Mage rn a 
<< 
— 40 
o «X 
a 
£ 30 Kc Xe LC te AREA 
9 
e 
of 
e 
S 
m 
  
  
0 16 20 30 40 50 60. 70. 80 90 
Rotation Angle (degree) 
Figure 7. Elongatedness plotted against the