POLSAR Image Classification Based on Polarimetric Decomposition and Generalized Discriminant Analysis

Abstract

In this paper, a new classification scheme of fully polarimetric SAR images is proposed. This is based on the joint use of the Freeman-Durden decomposition and generalized discriminant analysis, a new method for Feature extraction. After getting the powers of the three scattering mechanism components through Freeman-Durden decomposition, the Feature extraction algorithm is introduced to well exploit the information available in the full polarimetric coherency matrix. The experimental results show that using this exploited information as new features of Fisher classification, can provides fine performance and good compactness. Data acquired by polarimetric SAR(POLSAR) are directly related to physical properties of natural media. One of the most challenging applications of polarimetry in remote is natural media classification using fully polarimetric SAR images. The classification of POLSAR images, based on targets' backscattering properties, has been become a subject of highest interest for the past years. The classification method though dealing with the scattering vectors, implementing targets decomposition to extract classification features is must popular and effective means of classification currently. In recent years, many new statistical learning theories have been developed and gradually applied to classify natural media. The model-based Freeman-Durden decomposition (1) models the covariance matrix as three scattering mechanisms: Volume scattering, Double-bounce, Surface or single-bounce scattering. Based on Freeman-Durden decomposition, Lee (2) judged each pixel's dominant scattering mechanism, assigned them to one of the three scattering mechanism types, and then implemented the Wishart classification algorithm. In the procedure of classes' combination iterative process, it just realized initially divided of each pixel, and must be assumed that the complex polarization covariance matrix obeyed the Wishart distribution. Much more detailed polarization information polarization contained in covariance matrix was not well developed yet. Using feature extraction and feature selection can extract more effective features from target's polarimetric covariance matrix (coherence matrix) and achieve well classification performance. Generalized discriminant analysis (shorted as GDA) is the use of kernel methods to promote linear discriminant analysis proposed by Baudat(3) in 2000. This paper will introduce GDA to the POLSAR image feature extraction, and then selects an appropriate kernel function and proper parameters to extract polarization characteristics. Firstly, the Freeman-Durden decomposition is applied. As a result, the image is classified into three basic types of scattering mechanisms. Then, GDA feature extraction is introduced to deal with the polarization covariance matrix of these scattering classes respectively. Finally, a kind of classification schemes is applied to POLSAR images using these features. This approach not only establishes a relation between the medium physical properties and polarimetric transformations, but also realizes that the physical scattering mechanisms and statistical learning theory complement each other.