International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004 
  
3.2 Joining road segments 
There are many methods by which the retained road segments can 
be joined together, such as Hough transform (global), optimal 
search (local), direction filter (local), overlaying with GIS layers, 
and so on. Because of the characteristics of the input image, the 
directional cone search method was used in this study. The joining 
procedure consists of several steps. The first step is to identify the 
end of a road segment through iteration under the assumption that 
both the end pixels of a segment are the ends of a road. All pixels 
in the image are searched for to determine whether they are the 
ends of a road. If a pixel is a road pixel, then a search is carried 
out further to determine if there are pixels to its left and its right. 
If so, then this pixel is not the end of a road segment. If pixels are 
found to lie next to the pixel under consideration, either to its left 
or to its right, but not to both of them, then the pixel under 
consideration is confirmed as the end of a road. A search is then 
carried out to determine whether there is another road end in its 
vicinity. If so, and if both clusters are oriented in the same 
direction and the distance between them is smaller than the 
specified threshold, then they are regarded as different segments 
of the same road, and will be joined together to form a longer road. 
  
  
  
  
  
  
  
  
  
  
  
  
  
N 
S 
Le 
N » 
A # Iw. 
Ci. D eel A 
= 
| St} 
; A 
(i i3 
E a Ka 
"n "i A se 
a, / 2 di 
qe ut. NA ride Sat 
/ ^. / ^. / ™ / 
ZN f ENS LAN PEN 
/ EN / CN / EN. A NN 
y X E 3 | X N j } 4 x i \ NS 
Á. (Xy P VN # (IX 
/} | 
Ed 
FEN 
Figure 2. Structure of the recursive search. 
This operation to determine whether pixel (i, j) is the middle of a 
road or its end is executed like this. If pixel (i, j) is a road pixel, 
then at least one pixel around it must be a road pixel. The pixel 
above it (i-1,j) will be evaluated first. If it is also a road pixel, then 
pixels (i, j* 1), (i+1, j+1) and (i+1,j) will be evaluated, as well. If 
one of them is a road pixel, then pixel (i, j) is considered as the 
middle of a road. No further processing is undertaken. The 
program moves on to the next pixel. If none of pixels (i, j*1), 
(i+1, j+1) and (i+1,j) is a road pixel, then pixel (i, j) must be the 
end of a road. The road segment joining operation is performed on 
it. If pixel (i-1,j) is not a road pixel, the program will move on to 
334 
pixel (i-1,j+1). If it is a road pixel, pixels (i+1,j), (i+1, j-1) 
and (j, j-1) will be evaluated to determine if any of them is a 
road pixel. If one of them is a road pixel, then pixel (i, j) 
forms the middle of a road and the iteration shifts to the next 
pixel. Otherwise, pixel (i, j+1) will be processed in the same 
manner as before. This process is repeated until the last pixel 
inside the operating window. Because (i, j) is not an isolated 
pixel, one of its eight neighboring pixels must be a road pixel. 
  
Figure 3. The downward road segment to be joined with the 
end of a road segment linking (i, j). 
If the pixel is found to be the end of a road, the road segment 
linking operation is performed on it. In this operation, the 
cone along the direction is first defined. If pixel (i, j) is a road 
pixel and pixel 1 is also a road pixel (Figure 3) and none of 
pixels A, B and C is a road pixel, then the search join is 
downwards. In Figure 3 the directional cone is defined as 
DE, F, GH, 1.3, K, LM, N, OÖ, P, O0, R; S—> ....... The 
join search is terminated by either finding the road pixel or 
meeting the designated threshold. If a road pixel is found 
during the search process, then its coordinates will be 
returned as ((i, j) and the pixel found) for the subsequent 
connection. If no road pixel is found and the iteration has 
reached the threshold, then this pixel is the end of the road, 
and the join operation will move on to the next pixel. Similar 
road segments in the directions of north, east, west, 
southwest, northwest, northeast, and southeast were also 
taken into account in the join operation. 
After linking, dangling road segments were joined with those 
in their immediate vicinity. 
3.3 Thinning 
Noises in remote sensing images can make the width of a 
road vary in different parts. They can also make roads thicker 
than the unit width in remote sensing images. Such variations 
can be detected by humans in the manual method of road 
extraction. In the automatic method of extraction, width of 
road networks can be unified through thinning. There are five 
requirements for road thinning (Choy, 1994). Connected 
image regions should be thinned to connected line structures 
(connectivity preservation); approximate end-line locations 
should be maintained (no excessive erosion); thinned results 
should be minimally 8-connectd (unit — width skeleton); the 
thinned results should approximate the medial lines (medial 
line approximation); and extraneous spurs caused by thinning 
should be minimized (boundary noise immunity). There are 
  
    
    
   
   
   
    
   
    
  
   
    
   
      
  
  
  
  
  
   
   
    
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
   
    
   
   
   
  
     
     
    
    
   
  
  
Interna 
many d 
most cc 
off the 
network 
if the ce 
The th 
single-[ 
first ste 
medial 
selectec 
two pr 
connect 
incorrec 
stage is 
without 
destroy 
Therefc 
maintai 
located 
determi 
coordin 
i), (i J+ 
top left 
left of a 
Thinnir 
one pp 
applied 
and roa 
Figure 
After t 
Compa 
more ir 
the out 
road se 
the resi 
but wei 
set, the 
the lens 
been re 
caused 
thresho 
segmen 
This is