The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008
2.4.1 Applicability for PolSAR Data: Here, the algorithm
is applicable for PolSAR data segmentation and classification.
The edge-detection-based segmentation algorithms are seldom
used for PolSAR data, while PolSAR data can provide more
information than traditional SAR data. It is considered that the
parameters, H and alpha, derived from the target decomposition
of the PolSAR data, have better discrimination for objects []. So
the information of those two channels is more credible for
feature extraction than that of other channels, e.g. VV, HH, HV
etc. The experiment results in this paper give supports for this
point.
2.4.2 Edge Enhancement Template: Here 3X7 and 7X3
windows are proposed as edge enhancing templates to enhance
horizontal and vertical edge information.
2.4.3 Canny Edge Detector: Canny edge detector is
performed after a series of pre-processing procedures, such as
edge enhancement.
2.4.4 Region Merging: Large regions won’t be merged into
others in the following region merging steps while small ones
can be merged into any other one.
Large regions won’t merge with each other in order to avoid
the combination of heterogeneous regions across the undetected
edges which really exist., In condition that two large regions are
not merged into each other even though they satisfy the
criterion for merging, this won’t effect the following processing,
such as classification, enhancement and so on . As we know,
the large regions contain enough pixels to represent the
statistical characteristic of the region, which can avoid the
influence of the speckles.
Small regions can merge with each other or into large ones,
where two cases are considered. If the pixel numbers in the
small region is less than a threshold, it will be directly merged
into the nearest region. Otherwise, the similarity between this
region with the nearest one is considered, one threshold is set to
decide whether it involves into merging. The minimum size is
set to be 3 X 3 to ensure the segmentation accuracy of boundary
regions.
2.4.5 Point-to-region merge: Point-to-region merging is to
realize the segmentation of the whole image. Some regions
smaller than 3X3 and boundary pixels are left after region
merging, which are called discrete points. In order to realize the
thorough segmentation of the whole image, those discrete
points should be merged into corresponding regions. In this step,
Wishart distance of point-to-region or point-to-class is used as
the criterion for merging or discrimination.
3. EXPERIMENTS AND RESULTS
Experiment Data Introduction
as state farms. Figure 1 shows intensity image of VV
polarization synthesization of the region, Flevoland.
Figure 1. Intensity image of VV of the region, Flevoland
Experiments
Segmentation of the experiment data is performed using the
algorithm proposed in this paper. Figure 2 shows the
segmentation result, in which each region is given the mean
value of H. In order to verify the validity of the segmentation
result, a simple classification using the segment results [X] is
performed and the result is compared with that got using point
based classification method. The results are shown in Figure 3
and Figure 4.
From the analysis of results we can conclude four points bellow:
First: There are no discrete points and seldom small regions in
the segmentation results.
Second: Boundaries are very clear, which can be seen from the
segmentation results.
Third: The algorithm, at the step of edge detection, has high
efficiency and the time complexity is lower than that of RGW.