A NOVEL WATERSHED SEGMENTATION METHOD FOR
SAR IMAGE
Jinjing Shen
1
, Long Ma
1
, Liang Chen
1*
, He Chen
1
, Ying Yu
2
1
Beijing Key Laboratory of Embedded Real-time Information Processing Technology, Beijing
Institute of Technology, Beijing 100081, China
2
School of Information Science and Engineering, Yunnan University, Kunming 650091, China
chenl@bit.edu.cn
Keywords: Watershed segmentation, SAR image, Otsu
method, Context analysis, Statistical characteristic.
Abstract
This paper proposes a new SAR (Synthetic Aperture Ra-
dar) image watershed segmentation method based on con-
textual statistical features analysis. This method firstly
homogenizes all the sub-blocks of image by analysing
statistical features, and then extracts the feature of each
homogenized sub-block. Finally, the watershed area of
this image can be segmented by means of Otsu algorithm.
The experimental results show that our proposed method
outperforms the conventional methods in both accuracy
and time consuming.
1 Introduction
In recent years, synthetic aperture radar (SAR) images
have been widely used in military and civilian domain for
the capacities of all-weather and all-time usage. Mean-
while watershed segmentation in SAR image plays an
important role in applications such as ship detection, mari-
time surveillance, oil spills detection and so on, it has
drawn great attention of scholars both at home and abroad.
Image segmentation is the procedure to divide image into
several regions which are consistent in interior and statis-
tically separated with each other [1]. Specifically, most of
the watershed segmentation algorithms in SAR image can
be classified into three categories [2]. The first category
contains approaches based on edge. This kind of methods
divide image into different regions by extracting edge
points and then connect these points into contour accord-
ing to certain strategies. In [3], Andreas and Edzard use
edge extracted by wavelet transform to achieve watershed
segmentation. The second category belongs to algorithms
which do segmentation according to similarity of region
attributes [4]. Usually, extracting features such as texture,
gray, statistical characteristic is needed in such methods.
In [5], Dai extracts multilevel local pattern histogram fea-
tures of region, then use them to make classification by
support vector machine (SVM). In [6], Otsu proposes a
widely used thresholding segmentation method that max-
imizes between-class variance of gray. The third category
segmentation algorithms make a combination of edge and
region. In [7], Shan considers both edge and region infor-
mation into energy function and introduces an active con-
tour model to approach targets. However, the pretty accu-
rate border got in such methods is at the cost of segmenta-
tion speed. In general, scholars in remote sensing domain
have proposed many watershed segmentation algorithms
for SAR image and get a lot of meaningful breakthroughs,
but proposing an algorithm which satisfies both demands
on accuracy and efficiency remains a great challenge.
To partially solve the aforementioned problems, this paper
proposed a novel watershed segmentation algorithm for
remote sensing SAR image. The proposed algorithm first
separates whole image into overlapping H*H sub-blocks
which are all of the same size, then removes heterogene-
ous parts of these regions, finally do threshold segmenta-
tion according to the homogenized average gray. For the
proposed algorithm adopts low order statistical feature to
represent characteristics of region, its ability of real-time
processing is guaranteed.
2 Methodology
By taking both segmentation precision and computation
speed into account, the proposed watershed segmentation
scheme in this paper is showed as Figure 1. Where, sub-
block division is to divide image into same size blocks,
the size of block is dependent on the resolution of image.
Sub-block homogenization is the key link of watershed
segmentation. It eliminates the heterogeneous components
of block to attain the principle feature. At last, image
scenes are interpreted to land or water by adaptive thresh-
old calculated from Otsu method. In general, the algo-
rithm consists of two main parts: sub-block homogeniza-
tion and threshold segmentation.
2.1 Sub-block Homogenization
There are always small components that have land fea-
tures in the water region, such as ships, ship wakes, small
islands. And small shadow regions which are similar to
water may exist in the land region. Obviously, these local
Original
SAR
image
Calculation
Threshold
Sub-block
Homogenization
Divide
Sub-blocks
Threshold
Segmentation
Result
image
Figure 1. Diagram of SAR image watershed segmentation.