assifi- 
on re- 
3, the 
some 
nd for 
of the 
(7)? 
le, } If 
then 
rding 
] pix- 
ution 
letely 
ectral 
oises. 
n the 
te the 
nts as 
tion , 
e can 
these 
> road 
than 
lenth 
  
se x. T 
a? 
oz : 
SAL ru? 
  
  
  
Fig. 7  Radiated distances on four directions 
10.0.0000 1000000 
0100010 01009010 
0600010 0,01 0.0.1.0 
90091000 0001100 
0-000 100 00:00 1:0.0 
0:0:0 0-0 1.0 0000010 
0.0:0:0,0. 0 1 0:0:0:0:0:0:1 
Fig. 8 The broken point on Fig. 9 Connetion result after 
the road mass centring 
and lttle width. If given a value T , when R = T , the 
pixel can be discriminated to the true road. 
It is often that on real line in the image exists random 
broken points. (see fig. 8). It causes much trouble and 
difficulty for radiated distance calculating. A mass cen- 
tring algorithm is designed to fill these broken points. 
The average coordinates were calculated within a given 
window. The point on the average coordinates then is 
given the same value of the line point. Fig. 9 demon- 
strates the result by this algorithm. The characteristic of 
this algorithm is that the values on the adjacent points 
near the line will not be affected. Widow sizes determine 
the lack pixel number being connected. For example, 3 
X 3 window can connect broken points with one lack 
pixel, 5X5 window can connect broken points with two 
lack pixels. In the test, 3X3 and 5X5 windows were 
used gradually. 
4. RESULT AND DISCUSSION 
Fig. 10 is the road distribution picture produced by the 
procedure. In fig. 10, the urban road system within the 
437 
  
Fig. 10 
frame is the urban road system and outside is the rail- 
Road distribution picture. In the white 
way. Railway was combined with one city road when 
passed in to the city. 
white frame and the railway outside are extracted respec- 
tively from the TM data. The railway was combined 
with one main city road when pass in to the city and 
went out on the south. TM band 1,3,4,5 were selected 
to spectrally classify the railway and TM band 5 was 
used to segment the urban roads because these bands pos- 
sess particular characteristic for roads classification. Re- 
sult picture is also implemented thinning processing. 
The comparison with the aerophotograph and field inves- 
tigation show that this map demonstrated correctly the 
railway and indicated almost the whole urban road sys- 
tem. Only two small roads in the western city were par- 
tially or not demonstrated. This shows that the contextu- 
al method simulating the visual interpretation is correct 
and the algorithm designed is effective. 
In the process of evaluating these results one can observe 
an obvious phenominon that the whole road was isolated 
to several fregments. This was caused by two kinds of 
errors. A) errors due to the spectral classification. The 
path that road passes possesses complex surroundings. 
The reflectance of some road elements may be affected 
and changed by surroundings. B) errors due to contex- 
tual method. The greater curvature of road may cause 
the ratio of the lenth to width lower.