Polarimetric SAR Image Classification Based On Eigenvector Analysis

Polarimetric Synthetic Aperture Rader(POLSAR) is a Radar Imaging System with multiple parameters and multiple channels. It gets target polarimetric information by measuring the full polarimetric scattering echo of every resolution unit on the ground. Compared with the traditional imaging radar, POLSAR can detect the scattering feature of target objects under combinations of multipal electromagnetic waves of Polarimetric Transceiver. Therefore, it can provide more information for target interpretation. Image Classification is a major problem in POLSAR data application. This paper mainly studies the polarimetric SAR image classification methods based on the similarity measure of coherent matrix eigenvectors. We choose the eigenvectors as the classification features and propose two classification methods for polarimetric SAR image using the Support Vector Machine(SVM) algorithm or the spectral clustering algorithm. The main contents are as follows.1. We analyse the numerical characteristics of the three eigenvectors through the eigenvalue decomposition of the coherency matrix of polarization SAR data. We choose the real part and the imaginary part of the eigenvector and realize SAR image classification with the support of SVM classifier based on the eigenvectors’ good separability. The method is simple, feasible and has good classification result.2. Then we propose a similarity measure suitable for the eigenvectors of the coherency matrix through the mathematical analysis of the eigenvectors of the coherency matrix, based on the physical mechanism of polarization SAR radar imaging. In this method, the similarity of two pixels is calculated in the unitary space. Experiments show that the similarity measure is correct and available. Then we further propose a POLSAR image classification algorithm based on measuring the spectral clustering by eigenvector similarity metric. This method achieves good classification result and reduces the complexity of the classification process.This work was supported in part by the National Natural Science Foundation of China under grants 61003198, 61472306 and the Fundamental Research Funds for the Central Universities JDYB140508.