In: Wagner W., Székely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Voi. XXXVIII, Part 7B 
547 
AN ADABOOST-BASED ITERATED MRF MODEL WITH LINEAR TARGET PRIOR 
FOR SYNTHETIC APERTURE RADAR IMAGE CLASSIFICATION 
Xin SU a , Chu HE a ’ \ Xinping DENG a , Wen YANG a , Hong SUN a 
a School of Electronic Information, Wuhan University, Wuhan, 430079, Chian 
Commission III/3 
KEY WORDS: SAR, image classification, Linear Targets Prior, Ratio Response, MRF, AdaBoost 
ABSTRACT: 
A supervised classification method based on AdaBoost posterior probability and Markov Random Fields (MRF) model with Linear 
Targets Prior (LTP) is proposed in this paper. Firstly in contrast with most existing regions {superpixels) based models, this 
approach captures contiguous image regions called superpixels from ratio response maps of original images. Secondly, Adaboost 
classifier is employed to get likelihood probability for Markov Random Filed (MRF). Meanwhile, linear targets prior information 
(LTP) is introduced into MRF model combining with Potts prior model to engage better edges in classification results. Finally, 
iterative strategy in MRF model improves the performance of classification. Compared with traditional MRF model, the proposed 
approach has effective improvement in SAR images classification in the experiments of this paper. 
1. INTRODUCTION 
Since they can operate days and nights and under any weather 
conditions, Synthetic Aperture Radar (SAR) has been widely 
used in many fields. Furthermore, the resolution of SAR images 
has become higher and higher, which makes automatic analysis 
of SAR images rivet more people’s attention. Nevertheless, 
strong speckle noise existing in SAR images leads to difficult 
image processing. And, many articles are still published on this 
issue, such as segmentation presented in F. Galland, 2003 and 
R.F.Rocha, 2008, classification in C. Tison, 2004. 
Segmentation, classification and annotation are the fundamental 
tasks of images automatic analysis, which are called as image 
parsing (Zhouwen Tu, 2005). Recently, popular approaches for 
image parsing can be considered as a combination of three 
strategies, Pre-Segmentation, Features Extraction and Model. 
See fig. 1. 
f 
Pre-Segment 
• Pixels 
• Patch 
• Superpixel 
• Over- Segment 
»< 
extraction 
•Color 
•Texture 
• Shape 
»< 
r 
Model 
• Generating 
•Discriminative 
•Description 
N J 
C. J 
L J 
Fig 1. Framework of popular approaches for images parsing 
Recent publications present many pre-segmentation methods. 
Such as 20X20 patches are extracted in a pLSA based MRF 
classification method (Verbeek, 2007). Superpixel over 
segmentation is used in Regional Label Features based CRF 
method (Stephen, 2008). Meanshift over-segmentation method 
(Dorin, 2002) has been widely used in some classification 
articles. Multiscale segmentation based on geodesic 
morphology is used to get local regions for spatial reasoning 
(Jordi Inglada, 2009). However, there are some approaches * 
using pixels directly without pre-segmentation. In general, a 
pixel can be seen as a specific style of pre-segmentation. 
After pre-segmentation, features descriptors calculate the 
features of local regions. General features are color, texture and 
shape, such as SIFT-color (Joost van de Weijer, 2006) Gabor 
(B.S.Manjunath, 1996), LBP (T. Ojala, 2002)0, HOG (N. Dalai, 
2005) and so on. Because of imaging principle, SAR images get 
specific features. Only one kind of general features can’t 
describe SAR image sufficiently. Gray histogram and SoftLBP 
(Ahonen T, 2007) are used in this paper. 
The most popular image models can be seen as one of the three 
basic models, or combination of two or three of them. The three 
basic models are (F. Han, 2008): generating model, description 
model and discriminant model. Generating model is a model 
which infers prediction from samples such as pLSA (Verbeek, 
2007) and LDA (David M. Blei, 2003). Description model 
describes the relations of samples such as MRF (Verbeek, 2007). 
Discriminant model has discriminative functions which can get 
results from samples directly such as Adaboost (Robert E. 
Schapire, 2003). Meanwhile, there are some models combining 
two of the basic models. Specially, CRF is a unified model 
which combines discriminant model with description model 
(S.C. Zhu, 2006), and it can integrate different kinds of features 
and sorts of prior in a unified model more easily, and get better 
results by optimization. 
In common sense, land surface on one side of a certain length of 
road always belongs to the same category, and rivers, railways 
and other liner targets have similar situation. The idea in this 
paper is to introduce this linear targets prior into MRF 
description model. This paper focuses on improving the edges 
of regions in SAR images classification results. There are three 
contributions of this paper: 1) Linear targets priors are 
introduced into MRF model. The Potts model prior can infer 
* Corresponding author. Phone: +86-27-87548181; Fax: +86-27-87548181; E-mail: chuhe@whu.edu.cn