The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B3b. Beijing 2008 
588 
following algorithm which can be classified as morphological 
operators is suggested. 
For each road pixel, neighboring road pixels are counted in a 
3*3 window. If the number of road pixels is equal or less than 
one, the center pixel is omitted from the road raster map, 
otherwise it is kept unchanged. The main problem with this 
algorithm is that, in each run one pixel from the ends of main 
road skeleton is also removed. Figure 7 depicts the obtained 
result after successive iterations when all spur elements are 
removed. 
Maximum Likelihood Method 
RCC: 55.91 
BCC: 70.12 
RMSE: 0.3349 
Kappa Coefficient: 63.10 
Overall Acc.: 91.20 
* * 
■ Ï 
. * M 
I* ; i 
w * 
I 
' *. 1 * 
/} ‘ *r 
1. 1 
* * f 
iT' 
lr.-^ 
1 * f 
I * 
1 ■« ' 
Hidden Neurons: 10 
Iteration: 5000 
RCC: 73.29 
BCC: 88.85 
RMSE: 0.2240 
Kappa Coefficient: 68.34 
Overall Acc.: 94.63 
Hidden Neurons: 15 
Iteration: 1500 
RCC: 75.51 
BCC: 95.57 
RMSE: 0.2014 
Kappa Coefficient: 72.18 
Overall Acc.: 95.15 
Figure 4: Left image: Maximum-Likelihood result, Middle image: Simple BNN with 10 neurons in hidden layer from Tablel, Right image: Improved 
BNN with 15 neurons in hidden layer from Table 2