Istanbul 2004 
14). (t1, j-1) 
y of them is a 
ien pixel (i, j) 
ifts to the next 
ed in the same 
| the last pixel 
not an isolated 
je a road pixel. 
iN 
ined with the 
. 
road segment 
operation, the 
(i, j) 1s a road 
) and none of 
search join is 
is defined as 
3— lus The 
road pixel or 
Jxel is found 
ates will be 
ne subsequent 
> iteration has 
d of the road, 
pixel. Similar 
, east, west, 
ist were also 
1ed with those 
je width of a 
; roads thicker 
uch variations 
ethod of road 
ion, width of 
There are five 
). Connected 
line structures 
line locations 
hinned results 
skeleton); the 
lines (medial 
d by thinning 
ty). There are 
   
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004 
  
many different methods to implement the thinning algorithm. The 
most common one is to successively “peel” the outermost layers 
off the objects until they are one pixel wide on the connected 
network. Usually a 3x3 window is applied to the image to decide 
if the central pixel should be removed. 
The thinning algorithm adopted in this study is a two-stage, 
single-pass parallel operation within a 3x3 moving window. The 
first stage involves removal of redundant pixels to generate a 
medial line according to the set of the templates. The templates 
selected to remove the central pixel P should have the following 
two properties: (a) removal of P should not disrupt the local 
connectivity of the 3 by 3 pattern; and (b) removal of P should not 
incorrectly deform the shape of the pattern. The purpose of this 
stage is to determine if pixel P should be removed or retained 
without destroying the local connectivity. If the removal of P 
destroys the global connectivity, then P should be restored. 
Therefore, the second stage is the set of templates to restore P to 
maintain global connectivity. In this set four additional pixels 
located at above, below, and on both sides of P are used to 
determine global connectivity. If the central pixel P has the 
coordinates (i, j), then the coordinates of the four pixels are (i+2, 
j), (i, j+2), (1-2, j); and (i, j-2). Since the iteration order is from the 
top left corner to the bottom right, the two additional pixels on the 
left of and above P do not need to be considered separately. 
Thinning will lead the input road segments to a skeleton of only 
one pixel wide (Figure 4). After the thinning algorithm was 
applied to the image that had been processed for noise removal 
and road segment joining, it was saved as a file. 
  
  
  
  
  
  
Figure 4. Appearance of three representative clusters of spatially 
contiguous pixels before and after thinning. 
4. RESULTS 
After thinning, all roads have a uniform width of one pixel. 
Compared with the original image, the thinned result (Figure 5) is 
more interpretable. It contains nearly all major streets. However, 
the output result suffers from three main problems. First, some 
road segments shorter than the specified threshold still remain in 
the result (Figure 6). They should have been removed as noises, 
but were not. This is because after the threshold of removal was 
set, the length of the longest connected pixels was calculated as 
the length of the road. After the thinning process, some pixels had 
been removed to make the road uniformly wide. This may have 
caused the length of the road segment to be shorter than the 
threshold set in noise removal. On the other hand, some road 
segments that form a longer road have not been joined together. 
This is caused by the gap of larger than 4 pixels between them. 
335 
Apparently, this allowable gap between two segments is too 
conservatively defined in light of the obtained results. This 
threshold needs to be relaxed to further improve the quality 
of the mapped road networks. 
  
C A MN wr : aem À 
% EN ^ x A gi T 4 p x se 
qq EN o € acne S 
Figure 5. Results of detected roads. This output result has 
been thinned with a threshold of 5 pixels in noise removal 
and a threshold of 4 pixels in road segment joining. 
Second, some extracted roads show up as two parallel lines or 
even as a web (Figure 6). A comparison with the original 
image reveals that these roads are wide and have multiple 
lanes (e.g, major artery roads or motorways). Their 
extraction is not satisfactorily because of the improper 
assumption that all roads are one pixel wide. Under this 
assumption an operation window of 3*3 was applied. Within 
such a window size only eight neighboring pixels can be 
taken into consideration, resulting in the double line and web 
problem. 
  
Figure 6. An enlarged portion of the extracted road network. 
Road segments shorter than the removal threshold after 
thinning have not been removed. Roads wider than one pixel 
in the input image have some branches on the edge and 
appear as a web inside. 
This problem could have been avoided using other algorithms 
and a varying operating window size. During the extraction 
the width of a road is detected first. If its width is larger than 
one pixel, then the operation widow will be set larger 
accordingly. For example, if the road has four lanes, the 
operation window is increased to 5 by 5 (2*2+1) pixels. If the 
road has six lanes, the operation window is increased further 
to 7 by 7 (2*3+1) pixels, and so on. In order to accommodate 
a varying window size, dynamic templates need to be