mation takes the important role in the visual interpreta- 
tion. The accuracy of the visual discrimination of partic- 
ular object will increase with the pixel resolution higher. 
Road feathure is one of the particular target. Auto-rec- 
gonition of it by computer is usually implemented by line 
detecting techniques. Conventional linearment detecting 
methods are solely on the basis of the abrupt change of 
grey level. Beside roads, other linear feature with the 
similar situation will also be detected and demonstrated 
by these methods. 
It is known that roads possess two characteristics. One is 
spectral information. It is impossible to extract road us- 
ing sole spectral information because there exist other 
objects with the same spectral reflectance. Another char- 
acteristic is its linear contextual or shape information. 
The ratio of the lenth to width can be regarded as its sig- 
nificant contextual information to be used in the discrim- 
ination of the road. It is possible that the integration of 
the spectral and shape information can reach the effec- 
tive extration of the roads. 
This paper designs an approach which integrates the 
spectral information and shape information. The proce- 
dure contains two steps. First step is to classify and seg- 
ment the roads using multiband spectral data. Second 
step is to remove other noises and remain the road points 
using the ratio of its lenth to width. One algorithm to 
caltulate the lenth and width is also developed. A mass 
centring method is also developed to connect the continu- 
al road pixels by which the adjacent points near the road 
are not affected. The test shows that the result by this 
method are almost identical to those interpreted from the 
aerophotographes and those produced by field investiga- 
tion 
3. ]. Spectral Classification 
  
It is significant to consider the combination of the two 
steps to reach the best achievements. Completeness of 
road must be emphasized in the spectral classification in 
order to compute the ratio of the lenth to width. The 
process of classification should focuses on the connexion 
of the road points instead of the precision of the classifi- 
cation. While the next shape classification focuses on re- 
moving produced noises and extracting roads. Thus, the 
spectral classifier is designed as follows; 
Assume j bands are token into classifier. Select some 
samples for trainning. Let a; and s;(1 <:X 7) stand for 
average and mean square deviation respectively of the 
road grey level in ith band ( 1 xCéxC j) , (1,75 ,*** 7) 
stand for the spectral vetor of any pixel in the image. If 
|r; — 4 | Xcsifor i — 1,2,**:,7, cis a constant, then 
this pixel was discriminated to the road. According 
statistic theory , when c — 2 , 95 percent of all road pix- 
els were correctly classified. Thus a road distribution 
picture was obtained in which road shape was completely 
remained. 
3. 2. Shape Classification 
It is obvious that in the picture obtained from spectral 
classifier exist lots of other non-road objects or noises. 
Almost all these objects or noises are not linear in the 
shape. Shape information can be used to eliminate the 
noised. 
One pixel in the image possesses eight neighbor points as 
fig. 6. If we regard two opposite points as one direction, 
every pixel radiates outside on four directions. We can 
calculate the four radiated distances s; , 5» , 71, P2 on these 
four directions for each pixel (fig. 7). Let R — 
maz(si/$» , 82/81, 91/92 »p2/ 1 )- If the pixel is true road 
point, the value R on this pixel must be greater than 
those false road points because the road has large lenth 
  
DD 
  
  
  
Qy |: |W 
410 1m 
6 
  
  
  
Fig. 6 The dark square at the center shows the pixel 
in which one is interested. The pixel has 8 neighbors 
from 0 to 7 and four directions from 0 to 3. 
436 
Fi 
and 
pixe. 
It is 
brok 
diffi 
tring 
The 
Winc 
give 
strat 
this 
near 
the | 
pixe 
lack 
used 
Fig. 
proc