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: