| of 
rts, 
ion 
tal 
1ga 
4); 
ved 
the 
ine 
be 
age 
the 
gh, 
for 
ng. 
are 
96; 
res 
ach 
his 
use 
can 
ove 
ed. 
ers. 
wi 
the 
lon, 
Ing, 
are 
ther 
‘ve 
lian 
| is 
dge 
  
Alan Forghani 
3, Edge detectors (Sobel, Canny and Deriche) and thresholding. The Sobel can be applied in both vertical and 
horizontal directions. The Canny has different mask sizes (eg 3 by 3, 9 by 9, 21 by 21). Using a Deriche filter, 
the user has three options for the size of the filter or local maxima ( a ). 
4. Thresholding can be done interactively. Different threshold values can be tested to gain satisfactory results. 
5. Mathematical morphologic transformations. The ILFDP allows the user to apply mathematical operations for 
the binary data (Mathworks, 1995) using five basic functions, namely dilation, skeletonization, bridge, fill, and 
close. These operations help increase accuracy of linear feature detection. 
3.2 Data Analysis 
32.1 Pre-Processing 
A subset of a rural area (180 row by 160 column), a subset of a built-up area (185 row by 235 column), were selected 
from an aerial photograph for image segmentation analysis (Figure 1). Primary processing of the data was undertaken 
including standard image histogram stretching and median filtering techniques. 
If image smoothing and noise removal is not employed on the image, problems may occur in edge detection. The effects 
of noise on the responses of different operators can be suppressed by smoothing the image with a low pass filter before 
applying an edge operator. Median filtering with a 3x3 kernel was used to reduce “salt and pepper” type noise, and 
remove outlying pixel values. Thus it minimised the influence of noise, and produced a smoother image. Too large a 
filter eg bigger than 7x7 may cause fuzzy images which are not appropriate for edge detection. This especially occurs 
where edge density is high and complex such as in urban areas. 
3.2.2 Edge Detection and Thresholding 
The Sobel, Canny, and Deriche operators at different threshold values were compared. The Sobel edge masks as a 
conventional technique which estimates two components of the intensity gradient vector. These masks were used to 
detect the horizontal and vertical edges within the image. The Sobel edge finding method in the ILFDP uses a default 
estimator that consistently produces good results. 
As noted, the optimal threshold for Canny's operator was reported by other researchers to be 70 to 80 percent using 
hysteresis thresholding (Guestari, 1996). However this may be too high for some imagery including that used in this 
research. The hysteresis thresholding of Canny tries to keep the edge pixels as continuous as possible. The above 
threshold (70% to 80%) may lead to unconnected edges and may remove useful information. It was necessary to 
experiment with different threshold values for a homogeneous test area (eg the rural site), and a heterogeneous area (eg 
the built-up site). Refer to Table 1 for quantitative results. 
The Deriche filter has not been widely experimented with for RS imagery. The results of the Deriche approach have 
been applied to the test area with a range of thresholds. Inthe ILFDP program, the user has three options for the size of 
the filter (à ), and it was found that & = 2 had the best performance over images here. The Deriche operator was 
applied to variety of images with different values of the parameter a controlling the width of the impulse response. It 
was found that too many edge points were detected, particularly in regions where the contrast was low (Forghani, 
1997a). Lower threshold values present too many pixels which are not true edges. Higher threshold values cause 
breaking of connected chains of edge pixels into chains, which is not highly desirable. It was found that lower 
thresholds produced too many edges, particularly in the southern part of the image where the contrast is fairly low. For 
example, the best threshold value for the rural site was 20-40 percent, and for the urban site was 30-50 percent using a 
Deriche filter. Threshold values of 35, 40, 45 percent were applied to evaluate the edge detection result for the whole 
image (refer to Table 1). Consequently a 40% threshold obtained reasonable results to carry out morphological 
transformation in order to generate a final classified map. 
Using Canny and Deriche edge detectors requires higher threshold values for built-up areas in comparison to rural 
areas. To satisfy this objective, three thresholds T,, T, and T, were used. It was found that by increasing the 
threshold values, the roads begin to disconnect, while significant aggregation of distinct features (eg adjacent houses) 
remains. A low threshold causes better road connectivity but blurs houses into roads. More conservative thresholds 
produce neater road segments, but leave gaps, eliminate small roads, and still some building and driveways are attached. 
It was found that the Canny edge detector (with a filter size of 7) and Deriche operator (with a filter size ofa = 2) using 
à threshold value of 4096 yielded more road boundaries, and tended to trace closed contours around houses. 
After empirical testing, it was found that thresholding of between 20 to 30 percent over a rural site, and 40-50 over a 
built-up area produces edges which most closely correspond to real road boundaries. Details of the analysis are 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 291