59 
Beijing 2008 
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B6b. Beijing 2008 
i re-sampling 
referring to 
id features to 
lie image by 
two kinds of 
ordance with 
; the feature 
tilize median 
tests, 3X3 
.1 vehicles on 
is method is 
: of applying 
classes. The 
into which a 
the selection 
atic K-mean 
)ly into two 
dvantages of 
is higher in 
id costs less 
dws on roads 
rmation, the 
. speckles or 
eliminated 
ws on roads, 
1 connection 
frame work, 
roads are 
each pixel, 
ir to those of 
!) geometric 
ally smooth 
stance, road 
iognized as 
d above, for 
in a straight 
lidate pixels 
: road) to all 
; we give in 
Sed as road 
xels to road 
id the result 
i of number 
m (ij) in 9 
/hose value 
N-> number of pixels in DL in image I whose value equal 
1 
N1 -> number of pixels in DL in image I 
IfNl/N ^ Rthen 
RI (DL) = 0 
End if 
IfNl/N 2? Rthen 
RI (DL) = 1 
End if 
End for 
End for 
End for 
Compared with other methods, this method is more efficient, 
and makes full use of road knowledge. K-mean clustering 
classifier done previously quickens the search speed without the 
need to compare spectral values of pixels with those of the 
standard. Meanwhile, the algorithm can deal well with 
discontinuous roads which are occluded by shadows and other 
geo-types. After that, a rude result image is produced on which 
roads are extracted as sets of straight segments. 
2.4 Result Grooming 
The rude result image derived from the last procedure is 
groomed using mathematical morphology in this stage. The 
grooming stage relies on four basic steps: connecting, 
smoothing, thinning and linking. 
The connecting joins discrete road segments using 
morphological dilation. The smoothing, which combines 
morphological opening and closing operator, reduces the 
roughness of road edges significantly. The thinning erodes the 
road segment into one-pixel width. To achieve the goal, the 
thinning process is improved by introducing regions 
corresponding to more local information. The image is split into 
equally sized regions and in each region, morphological 
thinning operators are selected automatically according to local 
road width information. The linking, the last step of grooming 
stage, concentrates on correct connection of one-pixel wide 
road segments and final elimination of non-road information 
from the image. Geometrical features such as size, connectivity 
and distance between road segments are considered to achieve 
the purpose. Single or too short segments would be eliminated 
from the image. 
After this final step, we acquire the result image which contains 
road network information extracted from original remotely 
sensed imageries. 
3. IMPLEMENTATION OF THE AUTOMATIC ROAD 
EXTRACTION APPROACH 
In this section, we take Beijing City as a study case to 
implement the proposed approach. The data we choose is 
QUICKBIRD multi-band image. 
3.1 The Data of Study Area 
We take QUICKBIRD image for example, and the image was 
graphed in Nov. 2002. There are four multi spectral band data 
and a panchromatic data. The resolution of 4-multi spectral 
bands is 2.44meters, which is Blue band (450-520nm), Green 
band (520-600nm), Red band (630-690nm) and NIR band (760- 
900nm). The resolution of panchromatic band is 0.61meters, but 
there is only one band and the image is monochrome lack in 
spectrum information. So we choose 4 multi-spectral bands as a 
study image to extract road. 
The study area, with the size of 1000 X 500 pixels in the 
imagery, covers an area of nearly 3 km 2 in Beijing city. The 
image contains large volumes of detailed information, including 
roads, buildings, vehicles, trees, shadows, zebra crossings and 
other geo-types. And the image is seriously affected by 
shadows as other areas in the original image. 
Firstly, we do some pretreatment before classification. 
Geometrical correction, atmosphere correction and rational 
correction are done. Then we adopt the median filtering as 
mentioned above and can get the image as Figure 1 after 
correction. 
Figure 1. The image of study area. It is shown using red, green 
and blue bands, which are real colors. And the image has been 
corrected primarily. 
3.2 Automatic Road Extraction 
Then we use the proposed approach to extract roads from the 
image. The method in this paper is demonstrated in figure 2. 
After atmosphere and geometric correction, we classify the 
whole image through K-mean clustering into two classes. 
Assigning the number of classification as two, spectral 
properties of roadsides and shadows on roads are similar in this 
wide spectral range; so, darker objects like roads and shadows 
are classified into the same class while other lighter objects are 
classified into the other. So on the acquired binary image, roads 
and shadows on roads are classified into the same class thus 
ensures the continuity of roads. The rough classification result 
can be found in figure 3. 
Next, the road connection algorithm we invented in this paper is 
applied to the binary image to extract the road skeletons 
roughly. Because of the algorithm based on road knowledge 
including continuity, shape, topology of urban roads, we can get 
a satisfactory result of road network connection. The connected 
results are quite good. As shown in figure 4, we can see clearly 
that main roads are generally extracted while some of the sub 
roads are extracted too. 
After that, we adopt morphological algorithms to process 
discrete road segments, smooth road edges and erode roads to 
one-pixel width. Morphological dilation is firstly used and 
discrete road segments are connected together while the roads 
are also widened. Then morphological opening and closing 
operators are used to smooth the road edges; as a result, 
roughness of road edges is reduced significantly, which is 
beneficial for thinning in the next step. Morphological erosion