there must be a method for managing 
uncertainties in an inference process. 
3. APPROACHES FOR BULILDING EXTRACTION IN 
CHANGE DETECTIN 
Solutions for the above mentioned problems 
are proposed, and approaches to building 
extraction for map revision are described 
in this section. 
3.1 Region Growing Method and Threshold 
Value Selection 
In this study, the region growing method 
was employed for image segmentation. This 
method assigns the same label to the pixels 
with relatively uniform digital numbers 
(DN) in a region. In satellite images, 
buildings usually have larger DNs than the 
Surrounding ground. Hence, in classifying 
the segmented regions as building 
candidates or background, a threshold value 
was employed to the average DN of each 
Segmented region. 
As discussed above, it is difFicult to 
derive an appropriate threshold value in an 
a priori manner. In map revision, however, 
old maps are available for locating areas 
containing buildings. These areas can be 
located in the image data, and sub-image of 
the buildings extracted. The histograms of 
such sub-images will form a bi-modal 
Structure as shown in Figure 1. The DN 
value at the valley point of the histogram 
is considered as the most appropriate 
threshold value for the sub-image. The 
image in Figure 1 was segmented with the 
region growing method, and then divided 
into a building candidate and background 
using the threshold value as discussed 
above (Figure 2). The sub-images of all the 
existing buildings were examined with this 
method to derive an average threshold value 
which was assumed to be applicable to the 
entire input image. 
3.2 Selection of Descriptors 
Of the seven elements employed in human 
photo interpretation, shadow, pattern, 
Lexture, and association are not useful for 
building extraction from satellite images. 
However, the rest of the elements, tone, 
Size, and shape are all useful for building 
extraction. The first two elements were 
defined as the average DN value and the 
area of each segmented region, 
  
  
Frequency 
  
  
  
e 90 110 
Digital Number 
  
  
  
  
  
Figure 1. Small section of a SPOT image 
around an existing building and its 
histogram showing bi-modal structure. 
  
v 
memes 
an ax an ae DR an UE SE 
== mm om oan == 
| an a 
meme mimm | 1 111 
St me ss mn qu us SE 
= em 
AR am Se un Sn Sb SR an a 
== 
mm am am am sm am ne am am | 
— 
CX mn onnonk nonu onhom onn oun omn ona me 
== mm nm osm mm em ose om wm am mm oom ER 
= am om 
=R a5 an zn an wm mm am 
as nr ax Sr ae an Ss 
== an an mm SR an an SS 
= 
——X 8 
= 
mmo am Se se am as 
an an an an am oon oh am dm =m un SS an an an an 
En = 
ar un mm osm am wm am me am ae 
=m mm 
an an = 
1 
i 
mue memes ee 
€ 
a= a5 an un Sr 
  
  
xxx xxx 
xx xx 
xxx 110 
i 
Si i 
RS EE : KA iQ 
  
Figure 2. Result of image segmentation with 
the threshold value derived in Figure 1. 
respectively. Since there is no one 
parameter which can properly describe 
shape, several descriptors were employed to 
indirectly define shape. These descriptors 
include elongatedness, perimeter length, 
and diagonal length of minimum bounding 
rectangle (MBR) (Figure 3). 
In order to examine the utility of these 
descriptors, the test pattern shown in 
Figure 4 was recorded in the laboratory at 
two equivalent image resolutions, 10 m 
(Test Image A) and 2.5 m (Test Image B). 
Test Image A was then rotated from 0 to 90 
degrees in 10 degree steps and resampled to 
resolutions of 10 mn and 2.5 m. Test Image 
B, on the other hand, was also rotated from 
0 to 90 degrees, but resampled only to 2.5 
m pixel resolution. The effects of rotation 
and pixel resolution on descriptor values 
are further discussed below. 
  
Perimeter 
     
  
S S. a Ve N rai 
e < Bounding 
Diagonal Length 
OE MBR (Perimeter) 2 
47 (Area) 
  
Elongatedness = 
  
  
  
Figure 3.- Definition of descriptors for 
shape. 
The descriptor values of the features in 
these rotated images are shown in Figure 5 
through 7. The shape of the features in the 
images of 2.5 m pixel resolution (Figure 6 
and 7) can be distinguished using the 
calculated elongatedness and area. The 
graphs of elongatedness for the 10 m pixel 
resolution image, however, intersect one 
another, indicating the difficulty of 
providing proper shape information. These 
graphs clearly show that the pixel 
resolution of the input image is important 
when defining shape with descriptors.