537 
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B3b. Beijing 2008 
In fig 4, the AB, C> D and E are initial seed points, the bold 
line above represents a road in RS image, the broken lines 
represent the direction of template matching. 
Choose the maximum correlation result as the only matching 
result of the segment. If the only matching result is larger than a 
given threshold, it’s an efficient matching result. Take several 
maximum efficient matching results as adopted matching 
results. 
At the matching place of adopted matching results add seed 
points automatically, these seed points accompany with the 
initial seed points, composed the seed points of LSB-Snake 
model. 
After processing above, on the one hand, we compute the sum 
of the number of corresponding templates of the efficient 
matching results, take the template of maximum sum as the 
width of the road; on the other hand, compute the frequency of 
“bright ridge” and “dark ridge” template responding to the 
adapted matching result respectively, educe the road’s attribute 
of dark or bright by the higher frequency. 
3. EXPERIMENTS 
A program is developed to realize the method above using 
Visual C++ language |ll] , a series of experiments is presented to 
extract roads from IKONOS and QuickBird image, to compare 
the result of road extraction by our method and LSB-Snake 
model. 
Fig. 5 uses IKONOS image, in this experiment, there almost 
doesn’t interfered land object around the road. Input several a 
few seed points as fig. 5a, the round points are inputted 
manually, and the extraction result of LSB-Snake model is 
presented, the extraction result of auto-initial-value LSB-Snake 
model as fig. 5b, the rectangle points are added by self-adapt 
template matching method. 
a b 
From this experiment we can see that, if the initial seed points 
are not dense enough, the result of road extraction by 
LSB-Snake model is incorrect, while the auto-initial-valued 
LSB-Snake model can extract road correctly, it’s more robust 
and automatic than LSB-Snake model. 
Fig. 6 uses QuickBird image, in this experiment, there exist 
heavy shade or shelter around of the road by trees and houses, 
fig. 6a isthe result of road extraction by LSB-Snake model, fig. 
6b is the result of road extraction by auto-initial-valued 
LSB-Snake model. 
b 
Fig. 6 Comparing of Road Extraction Results 
From this experiment we can see that, if the shade or shelter is 
heavy, the extraction result would be incorrect as fig 6a 
indicates, while the auto-initial-valued LSB-Snake model can 
overcome the shelter and extracts road correctly, more powerful 
in anti-jamming than LSB-Snake model. 
4. CONCLUSIONS 
Based on the research of LSB-Snake model, this paper put 
forward an auto-initial-value LSB-Snake model using self-adapt 
template matching method to extract road feature in high 
resolution remote sensing image semi-automatically. IKONOS 
and QuickBird image are taken for experiments, manually input 
the same amount of initial seed points, comparing the extraction 
result of the LSB-Snake model and auto-initial-valued 
LSB-Snake model. Experiments indicate: 
The auto-initial-valued LSB-Snake model can add seed points 
automatically by self-adapt template matching. Given the same 
amount of initial seed points, our method is more robust than 
LSB-Snake model. 
Needn’t manual input of the parameter, such as the width of 
road, dark or bright, the auto-initial-valued LSB-Snake model is 
more automatic than LSB-Snake model. 
The auto-initial-valued LSB-Snake model can overcome the 
shade or shelter of land objects such as building and trees, and 
more powerful in anti-jamming than LSB-Snake model. 
Fig. 5 Comparing of Road Extraction Results 
While, because the condition of the road in High resolution 
remote sensing image is complex, our method of road extraction