Research On Hyperspectral Anomaly Detection Based On Sparse Representation

As a comprehensive observation technique for detecting landmark information,hyperspectral remote sensing provides detailed information of observed scene and has advantages that other technologies cannot reach.Anomaly detection,as an important application of hyperspectral remote sensing,can be directly output as a detection result or regarded as a preprocessing method for other applications.As a result,it is widely used in military reconnaissance,mineral exploration,environmental monitoring,etc.This dissertation focuses on the sparse characteristics of hyperspectral images.Based on this,the sparse features of abnormal pixels are analyzed,and the corresponding anomaly detection methods are explored.Starting with the basic theory,the mathematical model of sparse representation is introduced,and the general sparse coefficient solving method is given.Then based on the sparse representation theory,the sparse characteristics of hyperspectral images ar e explored,and the corresponding hyperspectral image sparse representation model is provided.Furthermore,the related characteristics of the anomaly under the sparse representation model are analyzed.Finally,extending the sparsity of a single pixel to the low rank characteristics of the whole image,the low rank and sparse matrix decomposition model of hyperspectral image are proposed.Based on the above theories and models,we focus on the hyperspectral image anomaly detection methods under the sparse representation.A local sparse difference index algorithm utilizes the differences in the distribution of coefficients of the anomaly and the background.Then to address the problem that the performance of the local algorithm is limited by the window param eters,a sparse score estimation algorithm uses the utilization of each atom in the dictionary to retrieve the anomaly information and completes the exploiting of image anomaly information from another perspective.The low rank background matrix and the sparse anomaly matrix are obtained from the low rank and the sparse matrix decomposition model.The sparse matrix is used to directly obtain the anomaly information,while the low rank matrix is used in the improved algorithm and enables full use of hyperspectral imagery information.Considering other typical features of hyperspectral images,the anomaly detection algorithm is optimized from different angles.In the local sparse difference index algorithm,the compression sample matching pursuit algorithm is used instead of the orthogonal matching pursuit algorithm,and the operation efficiency is improved on the premise of guaranteeing certain detection accuracy.Then aiming at the problem of sparse score estimation algorithm with low detection accuracy,the non-negative constraint is introduced into the hyperspectral image sparse representation model,and a non-negative sparse score estimation algorithm is proposed to improve the detection accuracy of the algorithm.At last,according to the nonlinear characteristics of the hyperspectral image,the kernel function is used to improve the algorithm,so that it can obtain the ability to process nonlinear data,and the stability of the algorithm is further improved.