The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B3b. Beijing 2008 
with 3 pixel of road width will get max matching result with a 
narrow road in image, and the template with 25 pixel of road 
width will get max matching result with a broad road in image. 
While the width of a road in image is unknown, it can be 
obtained by multi-scale template matching method. 
The road in image may be either bright or dark strip comparing 
to the background, so two series of templates are designed: the 
first series of multi-scale templates are bright ridge-like (fige.2), 
and the second series of multi-scale templates are dark 
ridge-like (fig.3). The former match bright road more efficient, 
and the latter match dark road more efficient. Whether the road 
is bright or dark could be judged by this means. 
E snake ( v ) - “ J J^int + E image + E con W S 
LSB-Snake* 101 is an efficient model to extract liner-like features, 
it describes Snake curve using B-spline with parameters, and 
iterative to minimum energy by using the algorithm of least 
square estimation, allocate the place of node points by the 
complexity of B-spline. 
Before the extraction of road by LSB-Snake model, it need 
manual input the width and dark or bright character of the road 
to be extracted, Manual input may not be accuracy and hold 
down the extract efficiency. Besides this, the LSB-Snake model 
is not robust while the initial seed points are not dense enough. 
In our method, we obtain each road’s width and dark or bright 
attribute by self-adapt template matching and take it as initial 
value of LSB-Snake, this value is accurate and trusty, so the 
manual input is avoided. And, comparing with LSB-Snake 
model in extraction of road, the method this paper put forward 
can use not only the initial seed points, but also the new added 
seed points created by self-adapt template matching method, 
this makes the extracting result of roads more robust. 
2.2 Self-adapt Template Matching 
Self-adapt template matching method can obtain the dark or 
bright attribute and the width of a road, at the same time, it 
creates some new added seed points for LSB-Snake model. 
The character of a road in remote sensing image can be 
described by gray scale, geometry, topology, function and 
conjunction or context obligation etc. Among these characters, 
gray scale is the most important one, the gray scale of road can 
be expressed as linear feature with gray difference between the 
sides and the middle, so the ribbon-like (for ideal road) or 
ridge-like (for general road) template can be applied to match 
the road. This paper takes the initial seed points as initial place, 
to match the updated RS image by multi-scale template 
matching method. The results are: 
1) The maximum template matching point; 
2) the width of the road in image; 
3) whether the road in image is bright or dark comparing with 
the background; 
The templates this paper designed are a series of ridge-like 
templates with multi-scale in width (Fig.2), they are 
one-dimensional templates. g m axis represents template gray 
scale, y axis represents the template width, the middle part with 
even g m value represents the width of road. The difference of the 
width of template and the width of road is a constant. A series of 
templates were designed, they are differ in width of road, the 
width are respectively 3, 5, 7,..., 25(pixel)..., etc. In figure 3a, 
the width of template is 13, the width of road is 3; in figure 3b, 
the width of template is 15, the width of road is 5.The template 
Fig. 2 bright ridge-like multi-scale template 
Fig. 3 dark ridge-like multi-scale template 
After manually input several initial points, dividing the distance 
between each two points by several segments, using templates 
of different widths to match each segment with the RS images 
along the vertical direction respectively, compute the value of 
correlation coefficient, the formula is as follows. 
Pic,r) = - 
Z Z sij g'i+rj.c--—(Z Z s-./XZ Z s’«.-.,«) 
«=1 7=1 m ' n i=l y=1 i=l y = l 
EIs'«-„„<11 S,,)'I1IIs'V,.. -“<11 s',-.;»)' 
1 7=1 m n /=1 7=1 /ml 7=1 m ' n i = l j-1 
Where, m and n represent the row and column of image block 
respectively, r and c represent the searching scope, g and g’ 
represent the grayscale of template and image respectively.