Kubik Kurt 
(DImage of dark roof house (2) Result of thresheld of variance (3) Edges detected by the Sobel operator (4) Result of the 
morphological transformation (5) Combined result of (3) and(4) (6) New boundary location for dark roof (7) Region of dark roof (8) 
Region overlaid on the image 
Figure 5 Results by step 2 in Figure 3 
After calculating and thresholding the variance for each pixel derived, the area of the dark roof can normally be recognized. If 
some of the points in the dark roofs have a higher variance, some small sections may not be correctly assigned as roof pixels. 
A dilation operation can be used to fill in these small areas. Since some of the boundary pixels of dark roof areas may be 
connected with other regions, an erosion operation can be used to separate them from other regions. Dark roofs can thus 
normally be extracted . By combining the results of morphological transformation with the edges derived from the Sobel edge 
detection, a more accurate region boundary can be obtained. As shown in step 2 in Figure 3, the subsequent operations are 
similar to the ones described in step 1. After steps 1 and 2, the final house regions can be obtained. Figure 5 illustrates the 
results of processing step 2 in Figure 3. 
4.2 Single Image Processing for Delineation Tree 
4.2.1. Image segmentation 
The image is processed to recognize tree areas. Figure 6 illustrates the implementation steps. Using combined results from 
step] and step2, delineation of tree areas can be more accurate than either stepl or step 2. 
  
  
  
  
  
Sobel edge Edge Threshold Co-occurrence matrices 
detector TP] sharp » edge image ] i 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
Calculate 5 texture features 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
Region Region : 
Dilation | parameters selection - Y 
: Max-min rules classification 
Region Region Region 
it > TP] selecti 
split parameters selection Step2 
Step 1 
  
  
  
  
Vv Vv 
Final delineation 
of tree areas 
  
  
  
  
Figure 6 Image segmentation and classification for delineation tree areas 
The steps of image segmentation are illustrated in the stepl in Figure 6. The process of edge detection, morphological 
functions and analyzing region parameters are described as in 4.1. Region split is an important step in delineation tree areas, 
since trees are usually close to each other or to other objects, some regions remain connected. The “region split” function is 
used to split the connected regions and eliminate undesirable regions. Detailed can be found in (KB Vision 1996). 
This task splits the regions into new smaller homogenous regions based on peak-valley analysis of the associated pixel 
intensity histogram. For each region, a pixel intensity histogram is created. Only pixels that lie within the region are used to 
compute the histogram. The task automatically selects values and use them to threshold the pixels in the region. Contiguous 
pixels that fall within the same threshold boundaries form new regions 
4.2.2 Texture analysis and image classification using co-occurrence matrices 
Co-occurrence matrices has been described in image processing literature by a number of names including gray-tone 
spatial-dependence matrices (Haralick et al 1973 & Conners & Harlow 1980). It has been widely used in texture analysis 
and classification. The co-occurrence matrix in it core is a two-dimensional histogram of the occurrence of pairs of 
  
524 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 
  
inte 
whi 
sep 
pari 
thei 
For 
inte 
of 1 
illu: 
onl 
fou: 
inci 
cn 
In | 
mat 
Haı 
H; 
val 
onl 
cla: 
The 
ass 
Sec