1 
AV == 
  
  
  
Alan Forghani 
  
  
Title: USERS\ALI\PHD\EDGESTUD\C40.eps 
Creator: JASC, inc. 
CreationDate: 6/24/1997 
  
  
  
Figure 2a. The result of Canny filtering for different threshold values with a filter of 7 by 7 
  
Title: USERS\ALI\PHD\EDGESTUD\C40M.eps 
Creator: JASC, Inc. 
CreationDate: 6/24/1997 
  
  
  
Figure 2b. Morphological operation on detected edges from Canny filtering 
1 DISCUSSION 
Three edge detectors namely Sobel, Canny, and Deriche were tested on subsets of the aerial image. The performance of 
the Sobel filter was generally lower than the Canny and Deriche filters. Based upon this, the analysis over the entire 
image was undertaken on the Canny and Deriche edge detectors. Results showed that Canny was marginally better than 
the Deriche operator (see Table 1). In respect to thresholding values, if threshold = 10% is used, many isolated spots are 
also picked out, because the threshold is too low. However when a higher threshold (eg threshold = 20), is chosen, there 
is much less noise, but similar roads are missing too. 
Classification accuracy evaluation helped to choose the best edge detection filter, and threshold parameters. According 
to previous experimentation with the test sites, it became possible to optimize the process of linear feature detection by 
means of edge detection filter and thresholding. Using the Deriche edge detection filter (à — 2) and Canny operator 
employing three different threshold values based on the introductory experiment demonstrated that a threshold of 40% 
would the most subjective threshold to be applied for the whole image. 
Using Canny and Deriche filers, significant confusion occurred between cleared land, construction, buildings, and other 
natural line features (eg field boundaries), as evidenced by high classification commission errors and omission errors for 
(76% and 34% respectively). Refer to Table 1, as well as (Forghani, 1997a) for details. This is not surprising because 
cleared areas, man-made features and natural line features occur within the spectral make-up of road edges. Both Canny 
and Deriche filters generated very similar results. However, it is clear that the Canny edge filter has lower omission 
errors compared with the Deriche operator. 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 293