141
A NOVEL EDGE-DETECTION
BASED SEGMENTATION ALGORITHM FOR POLARIMETRIC SAR IMAGES
Jie Yang 3, *, Ran Yang 3 , Shigao Li b , S. Shoujing Yin 3 , Qianqing Qin 3
a State Key Laboratory of Information Engineering in Surveying, Mapping and Remote Sensing. Wuhan University,
No. 129 Luo Yu Road Wuhan City 430079 China (yangjie@lmars.whu.edu.cn, yangran322@163.com,
shoujingy@163.com, qqqin@lmars.whu.edu.cn. )
b National Engineering Research Center For Multimedia Software, Wuhan University, No. 129 Luo Yu Road Wuhan
City 430079 China (sg51@163.com)
Commission VII, WG VII/2
KEY WORDS: Pattern Recognition, Segmentation, Image understanding, Feature extraction, Land Cover
ABSTRACT:
Aiming at overcoming the disadvantages of the algorithm proposed by White, R. G. (RGW) and making it applicable for PolSAR
data classification, a novel edge-detection algorithm based on segmentation for Polarimetric SAR images is proposed in this paper.
As one of the famous algorithms based on objected-oriented segmentation, RGW has been used extensively in image segmentation.
However, it also has some disadvantages such as a large number of small regions in the segmentation result. In this paper, after a
series of pre-processing including edge enhancement and edge detection, the initial segmented small regions are merged according
to the region distance defined by distance of coherence matrix of PolSAR data, and discrete points (at the position of edges) are
merged into regions according to Wishart distance. We applied the new algorithm to NASA/JPL AIRSAR L band data of Flevoland,
Netherlands. Compared with RGW, experiment results demonstrate that the region of Flevoland resulting from our method get
superior segment results. In addition, a better classification result is derived from the segmentation result. What’s more, the method
has the advantage of edge holding while implementing classification.
1. INTRODUCTION
1.1 General Instructions
Most recent classification works have concentrated on
point-based classification of SAR data. These methods are
mainly based on spectral characteristic of pixels. But there may
be some problems with those methods when applied to high
resolution images, e.g. there are discrete points existing in the
results and edges between regions are not clear. Also these
methods are only lower level understanding of images, since
consideration of single pixel characteristic in isolation neglects
the rich and complex information of land covers. In such cases,
misclassification often happens; also the result may contain a
significant amount of salt-pepper noise. Thus, object-oriented
classifications are proposed. Image segmentation is the
precondition and foundation of object-oriented classification of
high resolution remotely sensed images and the quality of
image segmentation greatly influences/affects the accuracy of
the following processing.
But efficient segmentation algorithms for high resolution
remotely sensed images are rarely seen, especially for PolSAR
data. RGW algorithm presented by White [l) [2] aiming at
processing common remote sensing images is an edge-detection
based segmentation algorithm which exploits the cartoon model
and includes adaptive filter size. But, the RGW segmentation
results may often contain a large number of small regions, in
which some segments result from the choice of threshold, and
some ones are probably anomalies arising from flaws in the
algorithm [3] [4] .
In order to overcome the disadvantages of RGW method
mentioned above, and make it applicable for PolSAR data
classification, one novel segmentation algorithm based on edge-
detection is proposed in this paper.
The algorithm is applied to NASA/JPL AIRSAR L band
data of Flevoland, Netherlands to test the validity.
Classification results based on the segmentation result using our
algorithm demonstrate that the region of Flevoland is well
classified and the edges are well preserved.
1.2 Paper Structure
This paper is organized as follows: Section II presents
methodology consisting of (a) pre-processing of PolSAR data,
(b) introduction of edge-detection method used in the algorithm
proposed in this paper, (c) introduction of region-growing in the
algorithm, (d) features and improvements of the algorithm
proposed in this paper. Section III, presents the experiment the
proposed method using AIRSAR data, and gives classification
results derived from the experiment. This part also gives brief
evaluation of the algorithm and the experiment results. Section
IV is the conclusion.
2. METHODOLOGY
First of all a pre-processing should be done to the PolSAR data,
including decompression, coherence matrix computation and
* Corresponding author: Jie Yang.