The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008 
1191 
where / = spectral and texture features of multi-spectral images, 
backscattering feature of radar image 
X{ = spectral and texture features of multi-spectral images, 
backscattering feature of radar image for the seed point 
X f (i,j) = spectral and texture features of multi-spectral images, 
backscattering feature of radar image for pixel (i,j) 
a, (3 = constants smaller than 1 
d x 
Figure 4. E d vs. D x with different a 
It can be seen from Figure 4 that E D is decreasing monotonely 
with the increasing of D x , when D x = 0 , that is 
X{i,j) = Xo, E d will be the maximum. 
Finally, the speed function is defined by the following formula: 
F = El x El x Et x E n 
(6) 
Figure 2. E s vs. R x with different a 
R* 
Figure 3. E R vs. R x with different f3 
It can be seen from Figure 2 and 3 that E R decreases 
monotonely with the increasing of R x . When R x > 1, that is, 
X 0 > X{i,j) , E R is going to 0 fastly. Otherwise, E R will be 
larger value. When both a and /? are smaller than 1, E R of all 
features will be closer to the medium value of the road in the 
image. 
But when applying function (4) to the spatial auto-correlation 
scale of SAR remote sensing image, we failed to get the desired 
result. It may be due to the range of the spatial auto-correlation 
scale covers only the integers from -6 to 6. Therefore, the 
following function is used in the speed function for this feature. 
D x (i,j) - X 0 - X(i,j) 1 (5) 
E D {i,j) = exp(-flx|£> x (i,7)|)J 
where S = the spectral feature of multi-spectral images 
T- the texture feature of multi-spectral images 
A = the backscattering feature of radar image 
By taking function (6) to function (2), it is easy to calculate T t ., 
the time when the surface passes the pixel (j,j) according to 
the methods proposed by Sethian and Telea (Sethian,1996; 
Telea,2004). And then, the edge of the road is obtained. 
It is conducive to enhance the road information of the image if 
all the extracted features are formed in the speed function with 
equation (5). But the non-road objects will be enhanced at the 
same time. And if the road is incontinuous in one feature, it 
can’t be made up by other features in which the road is 
continuous. Function (4) makes the road in all features to be the 
medium value and non-road objects to be bigger or smaller 
values depending on their difference to the road. If there is one 
of features extracted satisfies that the ratio between road and 
non-road is bigger, it will be smaller in function (6) while road 
will be bigger. For example, house has bigger value in the 
spectral feature extracted from the multi-spectral image, but it 
will be zero in the texture feature extracted from the same 
image because of its big and flat roof. Although grassland has 
big values in the texture feature of the multi-spectral image, it 
will be zero in its spectral feature. Road can be distinguished 
from linear water in the multi-spectral image with the spectral 
feature of it although they are similar to each other in the 
microwave radar image. Therefore, function (6) is proved to be 
a good way to enhance the road information with the 
combination of different features from both microwave radar 
and multi-spectral images. 
where X 0 = the spatial autocorrelation scale of the seed point 
X(i,j) = the spatial autocorrelation scale of pixel (i,j) 
a = constant smaller than 1 
5. EXAMPLE 
5.1 Road Extraction with Fusion of ERS-2 SAR and 
Landsat ETM + Images