Research On Hyperspectral Image Classification Method Based On Sparse Representation

In recent years,hyperspectral images with rich spectral information and high-resolution spatial information have become the focus of research in the field of remote sensing.Among them,theories and techniques related to hyperspectral image classification have developed rapidly,highly improving utilization rate of spectral information and spatial information.Therefore,the true ground information corresponding to the hyperspectral image can be fully mined.At present,the classification technology of hyperspectral images is widely used in agriculture,geological environment detection,military investigation,marine exploration and other fields.This thesis analyzes the traditional and existing sparse representation theory-related classification algorithms for the deficiencies of hyperspectral image-rich spectral information and spatial information utilization and the sparse representation theory-related algorithms,a hyperspectral image classification algorithm based on joint sparse representation of kernel function and hyperspectral image classification algorithm based on dictionary learning and sparse spectrum joint sparse representation is proposed.The main research contents of this thesis are:1.A joint sparse representation hyperspectral image classification algorithm based on kernel function is proposed.This algorithm is aimming at the problem of insufficient utilization of spatial information of hyperspectral imagery in the joint sparse representation classification algorithm,especially for the recognition of different terrain objects in the boundary region,and proposes several improved joint kernel functions to calculate each detected pixel and neighbor.The size of the domain pixel weight is selected by setting the weight threshold to select the optimal neighborhood window for each detected pixel,Finally,the detected pixel and its neighboring cells are classified by a joint sparse representation model.Experimental simulation shows that the joint sparse representation classification algorithm based on kernel function can effectively improve the classification accuracy of hyperspectral images.2.A hyperspectral image classification algorithm based on dictionary learning and sparse spectral joint sparse representation is proposed.This algorithm improves the existing sparse representation algorithms such as insufficient utilization of residuals and selection of training dictionaries.Firstly,the kernel functions are used to calculate the weights of the pixels to be measured and the dictionary atom,and each dictionary atom is assigned with the corresponding weight to the dictionary atom to form a second-level learning dictionary,and the obtained second-level dictionary is used as the matching tracking training dictionary.Secondly,aiming at the insufficient utilization of the sparse reconstruction residuals in the existing algorithms,the boosting idea is proposed to be integrated into the matching pursuit process,that is,the residuals are added to the features of the enhanced band of the tested samples to perform multi-iteration optimization,in order to make full use of the spatial information of hyperspectral images,and make the band characteristic of the sample to be tested has good stability in the matching tracking process.The kernel function is used to calculate the weight of the sample to be measured and the neighborhood pixel to select the most similar pixel and perform joint matching tracking.Simulations are performed on three commonly used hyperspectral image datasets.The results show that the proposed algorithm can effectively improve the classification accuracy.