ronics and 
23905), the 
ndamental 
    
  
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B7, 2012 
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia 
AUTOMATIC EXTRACTION OF WATER IN HIGH-RESOLUTION SAR IMAGES 
BASED ON MULTI-SCALE LEVEL SET METHOD AND OTSU ALGORITHM 
HG. Sui* C. Xu^* 
^ State Key Laboratory for Information Engineering in Surveying, Mapping and Remote Sensing, Wuhan University, 
Wuhan, 430079, PR China, xc992002@foxmail.com 
Commission VII, WG VII/7 
KEY WORDS: Synthetic aperture radar (SAR), Multi-scale level set, Water extraction, OTSU algorithm, Segmentation 
ABSTRACT: 
Water extraction has an important significance in flood disaster management and environmental monitoring. Compared to optical 
sensor, Synthetic aperture radar (SAR), which has the properties of high resolution and all-weather acquisition, has been used for 
water extraction in this paper. Due to the presence of coherent speckles, which can be modeled as strong, multiplicative noise, water 
extraction in SAR image is very difficult. In order to extract water from SAR images automatically, accurately and quickly, a novel 
water extraction algorithm combine multi-scale level set method with OTSU algorithm is proposed in this paper. Firstly, we 
introduced multi-scale framework into level set method. Multi-scale framework is a method considering both global information and 
local information of the image. The overall structural information of the image can be maintained at coarse scales and detailed 
information can be kept at fine scales. Therefore, coarser scale extraction results can be used as a prior guide for the finer scale, so 
that not only are the statistical properties of the signal-resolution image considered, but also statistical variations of multiple 
resolutions are exploited. Moreover, computational complexity is reduced since much of the work can be accomplished at coarse 
resolutions, where there are significantly fewer pixels to process. Secondly, based on the multi-scale level set framework, the 
segmentation result of OTSU algorithm is used to represent the initial segmentation curve. Finally, in order to eliminate the influence 
of buildings shadow and road, post-processing is considered in this paper. The experiments with real SAR images demonstrate the 
effectiveness of the new method. 
1. INTRODUCTION 
Synthetic Aperture Radar (SAR) has the capability of large-area 
coverage, cloud penetration and all-weather acquisition, and it 
can usually obtain massive information within a short time. 
Thus, they are more suitable than optical sensors to reliably and 
timely map inundated areas in flood situations, which usually 
occur under overcast sky conditions. The high resolution and 
the increased observation frequency of the new class of SAR 
sensors offer enormous potential in the domain of flood 
mapping. However, the improved spatial resolution of the SAR 
data results in a large variety of very small-scaled image objects, 
which makes image processing and analysis even more 
challenging. 
Segmentation is the main tool to extract water bodies from SAR 
images. However, there exist some difficulties: On the one hand, 
due to the presence of coherent speckles, which can be modeled 
as strong, multiplicative noise, segmentation of SAR images is 
generally acknowledged as a difficult problem; see Lee (1989) 
and Oliver and Quegan (1998). On the other hand, confused 
objects may influence the result of segmentation, such as 
buildings shadow and road. 
The level set method was first introduced by Osher and Sethian 
(1988), and since then, much effort has been directed towards 
image segmentation (for example: Mumford and Shah 1989, 
Zhao et al. 1996, Horritt 1999, Germain and Refregier 2001, Li 
et al. 2005, Law et al. 2008). Compared with some other SAR 
image segmentation methods (for example: Cook et al. 1994, 
  
* Corresponding author. E-mail address: xc992002@ foxmail.com 
Fjortoft et al. 1998, Xu et al. 2003), the level set method has 
the advantages of being robust in locating the boundary of an 
object, and of being able to handle topological changes in the 
curves during their evolution. 
The most general model in the level set method is the Chan- 
Vese (C-V) model (Chan and Vese 2001), which is based on 
curve evolution techniques employing the Mumford-Shah 
functional (Mumford and Shah 1989) for segmentation, and 
level sets. This model can detect objects whose boundaries are 
not necessarily defined by a gradient and can be adapted more 
easily to topological changes. Although traditional level set 
methods based on the C-V model have obtained encouraging 
results, prior information (speckle noise) of SAR images is 
commonly ignored. It is widely recognized that the gamma 
distribution is the most general model employed to represent a 
SAR image, thus, many authors employ the gamma statistical 
model instead of the C-V model to define the energy functional. 
For example, Martin et al. (2004) analyzed the level set 
implementation of region snakes based on the maximum 
likelihood method for different noise models, and obtained 
improved segmentation results. Ayed et al. (2005) investigated 
SAR image segmentation into a given but arbitrary number of 
gamma homogeneous regions via active contours and level sets. 
Silveira and Heleno (2009) adopted a mixture of lognormal 
densities for SAR image segmentation between water and land, 
and results demonstrated the good performance of their 
proposed method.