The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B6b. Beijing 2008 
is widely used to thin linear features and here we introduce the 
concept of regions. The image is split into equally sized regions; 
so, in each region, morphological thinning operator is selected 
automatically according to local road width information. The 
linking, the last step of grooming, uses comprehensive road 
features. Size, connectivity and distance between road segments 
are considered to link road segments and eliminate tiny single 
wrong road segments. Through steps mentioned above, the 
rough result of road network is groomed and final result turns 
out, which only contains connecting and continuous road 
networks and without any noise information. 
QUICKBIRD Multi-band data 
Atmospheric and geometric correction 
Figure 4. Result of road connection algorithm. The algorithm is 
based on road knowledge including continuity, shape, and 
topology of urban roads. The road features are obvious in the 
imagery. 
Median filter 
I 
K-mean clustering algorithm 
Rough classification 
‘Road connection” algorithm 
Mathematical morphology grooming 
Figure 5. Compared image overlap road thematic image with 
the original image. It clearly shows the extraction result 
Road thematic map 
T 
Accuracy assessment 
3.3 Accuracy Assessment 
Figure 2. Road information extraction flowchart. It mainly 
contains three steps: rough classification using K-mean cluster, 
road connection based on road knowledge and mathematical 
morphology grooming. 
In order to simplify the evaluation of the result, we define main 
road as those with the width larger than 10 pixels or the length 
more than 300 pixels, and the sub-road as those with the width 
less than 10 pixels and the length between 100 and 300 pixels. 
Road segment which is also defined to evaluate the result refers 
to the segment between intersections of roads of the same level. 
Accuracy is given at last in table 1. As results turn out, 86.7% 
main road segments and 65.7% sub-road segments are extracted 
correctly, while 10.0% main road segments and 8.6% sub-road 
segments are recognized partially. Only 3.33% main road 
segments are failed to be extracted, while that proportion of sub 
road segments is 15.7%. 
Main roads 
Sub-roads 
Complete 
86.7% 
65.7% 
Incomplete 
10.0% 
8.6% 
Missing 
3.33% 
25.7% 
Manual Ref. 
30 
35 
Result 
30 
37 
Wrong 
0 
2/37=5.4% 
Table 1. Road information extraction accuracy assessment