文档下载 文档下载导航
设为首页 | 加入收藏
搜索 请输入内容:  
 导航当前位置: 文档下载 > 所有分类 > 自然科学 > SVM- and MRF-Based Method for Contextual Classification of Polarimetric SAR Images
侵权投诉

SVM- and MRF-Based Method for Contextual Classification of Polarimetric SAR Images

基于SVM-MRF的极化SAR图像上下文分类 作者:巫兆聪、欧阳群东

SVM- and MRF-Based Method for Contextual Classification of Polarimetric SAR Images

Zhaocong Wu, Qundong Ouyang

School of Remote Sensing and Information Engineering

Wuhan University Wuhan, P.R. China vip_oyqd@yahoo.com.cn

Abstract—Classification of polarimetric SAR images based on the classifiers). Drawbacks of such classifiers lie in the fact that information provided by individual pixels cannot generally some pixels would be misclassified inevitably, as the scattering produce satisfactory results due to speckle. By introducing the measurements of individual pixels are often affected by spatial contextual information between adjacent pixels, a novel speckle.

classification method, using support vector machine (SVM) and markov random field (MRF), is proposed in this paper. It consists of two steps: first, a supervised pixelwise classification using SVM is applied to polarimetric SAR images; then, by means of MRF regularization, the spatial contextual information is introduced for refining the classification results obtained in the first step. The effectiveness of this method was demonstrated using a JPL/AIRSAR polarimetric SAR image. Compared with other three frequently used methods, better classification performance as well as less isolated pixels was observed, using the proposed method.

Since, in general, certain land cover types may be more likely to be placed adjacently than others [7]. Also, in many polarimetric SAR images, objects on the ground are much greater than the pixel element size, so neighboring pixels are more likely to come from the same class and form a homogeneous region. Therefore, a SVM-based contextual classifier that utilizes both spectral and spatial contextual information will be more reliable for classifying polarimetric SAR images.

Incorporating contextual information into classification process can be done in many ways. One simple method is to Keywords- image classificaiton; polarimetry; synthetic aperture

radar (SAR) use majority voting within a prescribed window. In such a

method, the central pixel will be changed to the class that occurs most frequently in the window. This is a common post-I. INTRODUCTION

processing technique after a pixelwise classifier has been

Classification of polarimetric synthetic aperture radar implemented. There are more elegant ways for modeling such (SAR) images is arguably one of the most important contextual behavior. A class of contextual model known as applications in remote sensing [1]. Many supervised and markov random fields (MRF) can be useful for modeling unsupervised classification methods have been proposed so far. context in a more precise way [8]. Earlier methods classify polarimetric SAR images mainly

Markov random fields are probabilistic models that are based on their statistical characteristics [2-3]. The performance

of such classifiers depends on the statistical model and commonly used to integrate spatial contextual information into parameters estimation. In supervised classification, the reliable image classification problems, which can be realized by parameters estimation requires a large number of training constructing a priori probability in Bayesian sense so as to samples. However, the representative samples available are accomplish the maximum a posteriori (MAP) estimate. Such a very limited in practice due to speckle existing. Hence, the MRF-MAP solution often provides more satisfactory results estimated parameters are often less accurate, and the than pixelwise classifiers. However, conventional MRF is subsequent classification performance is less accurate then oversimplified as it fixes the shape of its neighborhood, which

is predefined for a whole image. A major drawback of this need be.

fixed neighborhood lies in the fact that some fine structures

Recently, support vector machine (SVM) has become an may disappear while processing with MRF [9]. To avoid such increasing popular tool for classification in machine learning errors, adaptive neighborhood (AN) is realized in this paper tasks. Compared with the traditional statistical classifiers, it based on the local homogeneity of total power image. does not need the statistic characteristics of training samples,

To improve the classification results of polarimetric SAR and has good performance for classifying remote sensing

images even when the training samples are limited [4]. For images, a supervised contextual classification method that above reasons, much attention has been devoted to SVM in combines SVM and MRF is proposed. The method consists of

two steps: first, a SVM-based pixelwise classification is polarimetric SAR image classification [5-6]. However, most of

applied to polarimetric SAR images; then, spatial contextual these presently used methods that based on SVM classify each

pixel independently, without considering the correlations information is introduced through MRF for refining the between spatially adjacent pixels (so called pixelwise classification results obtained in the first step. Experimental

The research is supported by 863 Program (2007AA12Z143),

National Natural Science Foundation of China (40201039; 40771157).

第1页

你可能喜欢

热门文档

相关文档

站点地图 | 文档上传 | 侵权投诉 | 手机版
本站所有资源均来自互联网,本站只负责收集和整理,均不承担任何法律责任,如有侵权等其它行为请联系我们.
文档下载 Copyright 2013 doc.guandang.net All Rights Reserved.  -  email