A final image was created by combining the classified 
images of Fig. 7 and Fig. 8. This was done by selecting 
the same gray level class, for example the rock, in both 
Fig. 7 and Fig. 8, and then filtering out the other two 
classes. Edge detection was performed on the final image 
using the Roberts edge detection algorithm (Pratt, 1978). 
Fig. 9 is the image of detected edges. 
  
       
  
C Uva ^ 
fi se SC 50] 
| ADT. Pate f 0| 
S S ec c | 
o Boo on 9 
RMT 
co 
  
7 X ; (55. : a ^ 2 | 
Fig. 9. Detected edges after applying Roberts edge 
detection algorithm to the final image. 
  
A recursive line following scheme is implemented by 
using a chain code (Ballard and Brown, 1982) to detect 
the short, unwanted edges. The binary line image is 
scanned line by line for points from which to start a 
chain. Every continuous line is chain coded and the 
length for each line is stored. A global constraint on 
length is set to identify the ice sheet boundary and 
remove the unwanted small edges. 
The detection of ice sheet boundaries using SPOT 
imagery and the digitized aerial photograph was done 
using image tone information. First, an anisotropic 
diffusion method was applied to both images to enhance 
edges. A dynamic thresholding algorithm was then 
applied to both edge enhanced images. Fig. 10 and Fig. 
11 are classified images segmented into three gray level 
classes after application of the dynamic thresholding 
algorithm to the edge enhanced SPOT image and the 
digitized aerial photograph, respectively. The final 
images were created by selecting the  class(es) 
corresponding to the rock area and filtering out the other 
class(es) from both Fig. 10 and Fig. 11. Edge detection is 
performed on the final images using the Roberts edge 
detection algorithm. Recursive edge following is done to 
detect the boundaries. 
630 
) Ez WM i > . : s le 
Fig. 10. Classified image segmented into three gray level 
classes after applying dynamic thresholding algorithm to 
the edge enhanced SPOT image. 
  
BE ri y 
Fig. 11. Classified image segmented into three gray level 
classes after applying the dynamic thresholding algorithm 
to the edge enhanced aerial photograph. 
4. RESULTS 
Fig. 12 is the final result of the detected ice sheet 
boundary from ERS-1 SAR imagery. The constraint used 
for the length of edge was 1900 pixels (size of image I5 
750 x 750 pixels). 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B7. Vienna 1996