Polarimetric Feature Mining For Terrain Classification Using Graph-Based Low-Rank And Sparse Decomposition

Polarimetric Synthetic Aperture Radar(Pol SAR)can obtain more target scattering information,which provides good data support for detailed analysis of target scattering characteristics.With the increasing amount of polarimetric SAR data,how to effectively extract features from polarimetric data and classify them is an urgent problem to be solved.Polarimetric target decomposition is a way to extract the features of polarimetric SAR data,but the features obtained by different target decomposition methods are not actually independent.Moreover,it is difficult to effectively improve the classification performance even by obtaining multi-dimensional polarization features through different polarization decomposition methods.In addition,since the polarization image contains speckle noise,some existing polarization denoising algorithms will change the original structural characteristics and polarization scattering characteristics after suppressing speckle,which is not conducive to the subsequent classification of polarized objects.Therefore,it is urgent to study the feature mining method of polarimetric SAR data to improve the identifiability and universality of features and lay the foundation for accurate ground object classification.Based on the theories and methods of super-pixel decomposition,robust principal component analysis and generalized principal component analysis,this paper studies the feature mining methods of polarimetric SAR data.The main work can be summarized as follows :1.The traditional feature mining methods such as principal component analysis,robust principal component analysis,nonnegative matrix factorization and graph nonnegative matrix factorization are analyzed and studied,and these methods are used for polarimetric feature mining and polarimetric SAR image classification.The limitation of the traditional feature mining method in polarimetric SAR feature mining is verified by the ground object classification experiment of polarimetric SAR data,and the reasons are analyzed.2.Polarimetric feature mining method using Graph-based low-rank and Sparse decomposition(PFMGLS)is proposed,which effectively improves the identification performance and generalization performance of polarization features.The details are as follows : 1)Firstly,the simple linear iterative clustering(SLIC)is used to segment the polarization data.Then,aiming at the problem of irregular superpixel size,the double cubic interpolation method is used to process the superpixel twice.2)Considering the correlation between super-pixel data,the neighbor graph containing the relationship between super-pixels is constructed.Then,the low-rank sparse graph model of polarimetric SAR data is constructed by combining robust principal component analysis.The speckle noise and outliers are eliminated by the low-rank sparse decomposition of polarimetric SAR data.At the same time,the correlation between super-pixels is used to mine the polarization characteristics with small intra-class differences and large inter-class differences.3)Aiming at the non-Gaussian characteristics of polarization data,the generalized principal component analysis(GPCA)is used to reduce the dimension of polarization high-dimensional features mined by PFMGLS.On the basis of PFMGLS high-dimensional features,the low-dimensional strong identification principal component features are extracted.The effectiveness of PFMGLS algorithm is verified by feature mining and ground object classification experiments based on GF-3 polarimetric SAR data.3.The influence of distance measurement on PFMGLS feature identification is analyzed.The Euclidean distance,Manhattan distance,Chebyshev distance and Min distance are studied,and these methods are used to construct PFMGLS nearest neighbor graph.The influence of distance metric on PFMGLS feature identification is verified by feature mining of GF-3 polarimetric SAR data and ground object classification experiments.