lon — 
tion, 
te the 
is de- 
d lin- 
10 ef- 
is de- 
ne in 
1 gen- 
racted 
to re- 
s that 
> road 
notely 
toring 
n. 
on the 
scien- 
have 
n the 
se ap- 
e dis- 
rieties 
icted. 
' edge 
They 
  
endi ea 
  
  
  
Fig. 1 An ideal step edge Fig. 2 A roof line 
ne as 
Fig. 3 A rectangle line 
luu X ee 
  
  
  
Fig. 4 A real line 
have no ability to integrate multi-band information. It is 
impossible for these methods to detect and extract the 
road when road has the similiar gray level with its back- 
ground on single band image. 
In this paper, one contextual method simulating the vi- 
sual interpretation is developed to extract the road fea- 
tures. It uses the shape information that is long lenth 
and narrow width to detect and extract road features. It 
also can integrate the multi-band information. It was 
found that roads extracted by this method are almost i- 
dentical to those produced by field investigation. 
2. STUDY SITE 
The TM data were obtained on August 25, 1987 from 
the Chinese Satellite Ground Station (CSGS). CSGS 
provided both the false color composited photo and com- 
puter compatible tape(CCT). The study area was select- 
ed to the Yueyang City, Huan Province (see fig. 5). 
The downtown area, in the northern part of Hunan 
Province, has about 300,000 population. Great Dongt- 
ing Lake is adjacent at its west and the Yangzhi River at 
its north. Beijing to Guangzhou railway passes the whole 
area. It is developed rapidly in the recent years and now 
is the one most important base for industry, transporta- 
tion, aquatic production and agriculture. 
Yueyang city expands median-small urban area. The 
435 
  
  
  
  
  
G A 
’ V vu uf * 
T Q Huer ç p 
0 
* 
: East 
; Do Gt tn 
; Lape Jj 
"e 
29° wikow/ 
Xiang yl | 
em] 
77 km 
LA Az 
J 
River Railway High way 
Fig. 5 Study site of Yueyang city 
complexity of road distribution increases difficulty of the 
extraction of road system from TM data. 
3. ANALYSIS AND ALGORITHM 
For more than a decade, efforts to extract information 
from multispectral remote sensing image data have 
proved increasinly successful. To a large extent, these 
efforts have focused on the application of pattern recog- 
nition techniques to the multispectral measurements 
made on individul ground resolution elements, i. e. , 
scenes have been classified pixel-by-pixel based on the 
measurement vectors associated with the individual pix- 
el. However, sole spectral information is not adequate 
to recognize the road features. There are many applica- 
tions for which the classes of interest can be better char- 
acterized if the spatial information in the remote sensing 
data is utilized in addition to the spectral information. 
Characteristic spatial features include, for example, tex- 
ture and context. Contextual information generally indi- 
cates the structural relationships between pixels. 
One way to approach contextual information is to utilize 
the shape information. One particular target is the object 
with regular shape projecting on the image. Shape infor-