BP Neural Network Classification Of Full Polarization SAR Data Based On Feature Decomposition

Full polarimetric synthetic aperture radar image classification has become a research hot spot. Compared with the common optical remote sensing image and single polarimetric SAR data, full polarimetric SAR data is more valuable, due to its data characteristics which can provide more bases to deep levels image analysis. Thus it has more potential in target detection, identification and geometric parameter extraction. Based on the imaging theory and data features of polarimetric SAR, the classification methods of full polarimetric SAR data are discussed, and a improved methods aimed to achieving a better classification performance and accuracy are presented in this thesis.This paper first introduces the imaging mechanism for polarimetric SAR image data and the main polarization scattering mechanisms. Then studies the polarization characteristics of polarimetric SAR data decomposition theory, which provides an important foundation for subsequent land cover classification. On this basis, the BP neural network classification algorithm which usually used in optical remote sensing images and SAR raw data is studied. Finally, full polarimetric SAR data classification is processed based on the polarization characteristics decomposition and the BP neural network classifier integration.Taking RADARSAT-2 full polarimetric SAR image data of Beijing area as an example, polarization information is extracted using several typical characteristics of decomposition methods. Then BP neural network classifications are processed separately with each decomposition result. Finally, the classification results of this paper are compared with the traditional BP neural network, the SVM supervised classification and the K-means unsupervised classification methods. The result shows that the BP neural network classification method base on Freeman decomposition has greatly improved the land cover classification accuracy of full polarimetric SAR data.