The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008 
142 
Cloude decomposition to calculate the parameters which would 
be used in the following steps. Then the implementation flow of 
the algorithm proposed in this paper is introduced. Before the 
segmentation the edges within the image are detected initially. 
Then regions are grown between these edges. So the edge 
detection is crucial to the following steps. 
2.1 Pre-Processing of PolSAR Data 
First of all the Cloude decomposition is implemented to 
compute the parameter H (scatting entropy) and alpha (scatting 
angle) according to PolSAR data. 
Usually the PolSAR image is stored in the form of compressed 
Stokes matrixes (or Muller matrixes) M. Before implementing 
the algorithm mentioned in this paper, it is necessary to convert 
Stokes matrixes M to coherence matrixes T. Firstly, the raw 
data is decompressed into matrix M. Details of decompressing 
steps can refer to the relevant papers [5] . The computation of 
coherence matrix form M and Cloude decomposition of 
coherence matrix please browse the related articles for 
detailed information. 
2.2 Introduction of Edge Detection 
The edge-detection stage in this process as follows. 
2.2.1 Pre-Processing: Pre-processing is implemented to 
eliminate the influence of speckles. Here a 3 by 3 averaging 
filter is applied to each channel of coherence metrics 
respectively. 
2.2.2 Edge Enhancing Templates Construction: 3X7 and 
7X3 windows as edge enhancing templates are constructed 
around each pixel to enhance horizontal and vertical edge 
information. Edge enhancing templates as follows: 
-1 
-1 
-1 
0 
1 
1 
1 
-1 
-1 
-1 
0 
1 
1 
1 
-1 
-1 
-1 
0 
1 
1 
1 
Table 1. 3 by 7 Template 
-1 
-1 
-1 
-1 
-1 
-1 
-1 
-1 
-1 
0 
0 
0 
1 
1 
1 
1 
1 
1 
1 
1 
1 
Table 2. 7 by 3 Template 
2.2.3 Edge Enhancement: The edge enhancing templates 
constructed above is applied to H and alpha parameter images. 
2.2.4 Edge Detection: Then canny edge detector is 
implemented. First, we use canny edge detector to process the 
two parameter images respectively. Then the derived edges are 
added to derive the final edge detection results. 
2.3 Introduction of Region Growing 
The region-growing stage of the algorithm processed as follows. 
2.3.1 Initialization: A series of structure elements are laid 
on the image so as not to contain any detected edges [3 f 
2.3.2 Region & Region Merge: Only small regions and 
adjacent regions are merged to form larger regions according to 
distance between coherences matrices of two regions. The 
distance between coherences matrices of two regions are 
defined as below: 
^(r 1 ,7’ 2 )=-xz(i^-4|) o) 
n j 
Where is element of T\ at the position of (i, j). T\,Ti 
are the centers coherence matrices of the two regions 
respectively. 
The center coherence matrix can be calculated as below: 
T = 
(2) 
Where T i is the coherence of i-th pixel of the region. N is 
the total number of pixels of the region. 
2.3.3 Discrete Points to Region Merge: Discrete Points (at 
the position of edges) are merged into regions according to 
point-to-region distance. Each point would be merged into 
region with the shortest point-to-region distance. Here we use 
Wishart distance [8] , which often used by PolSAR classification, 
as the point-to-region distance: 
d pr (T,X m ) = ln T m +tr\T m T\ (3) 
Where tr\T m T j is the trace of T m T matrix. 
T m is the center coherence matrix of the m-th 
region. 
2.4 Algorithm Features and Improvements 
Some important features and improvements of this algorithm 
proposed in this paper are as follows: