Research On Hyperspectral Remote Sensing Image Classification Based On Multi-scale And Multi-feature Joint Sparse Representation

With the development of hyperspectral technology,the spectral and spatial resolution of hyperspectral remote sensing image has been significantly improved.It has a wide range of applications in marine exploration,environmental monitoring,military reconnaissance,geological exploration,precision agriculture and other fields.Hyperspectral remote sensing image combine space-spectrum features and have unique advantages in feature category recognition.Classification has become an important direction for the application of hyperspectral remote sensing image.In view of the fact that the joint sparse representation algorithm can not only adapt to the problem of few labeled samples of hyperspectral remote sensing image,but also better solve the problem of high dimensional hyperspectral data and difficult processing.This thesis will focus on the classification of hyperspectral remote sensing image based on the joint sparse representation algorithm.The classification performance of the joint sparse representation algorithm for hyperspectral remote sensing image is mainly explored from four aspects: spatial position relationship of pixels in the neighborhood,multi-scale,multi-feature and labeled sample size.The main research contents are as follows:(1)In the traditional joint sparse representation algorithm,when determining the category of cells to be measured,a spatial neighborhood is used instead of a single pixel and atoms of each class to calculate the residual matrix,and the category of cells to be measured is obtained according to the size of the residual value by performing a2-norm calculation on the residual matrix.However,the importance of the residual values within the residual matrix in determining the cell category to be measured is not taken into account.In this thesis,considering that the cells that are generally closer to the central cell in the spatial neighborhood are more likely to belong to the same category as the central cell,the reciprocal of the distance from each cell to the central cell in the spatial neighborhood is used as the weight,and the obtained residual matrix is processed in an inverse distance weighted manner.At the same time,in order to make full use of the scale information of the spatial neighborhood,the results of the multi-scale joint sparse representation algorithm are fused together by the voting method,and the multi-scale joint sparse representation algorithm VDJSRC(Voting Inverse Distance Weighted JSRC)based on inverse distance weighting and voting method is obtained,which is verified on the Indian Pines and Pavia University datasets.Compared with the original JSRC algorithm,the overall accuracy(OA)of classification is improved by 3.45% and 1.24%,respectively,and the improved joint sparse representation algorithm has better classification accuracy and stronger model stability.(2)Aiming at the problem that only the original spectral features are used to cause insufficient utilization of hyperspectral image features,the Gabor,DMP and EMP features of the hyperspectral image are first extracted,and the four features including the original spectral features are fused to obtain the spatial neighborhood,training dictionary and sparse coefficient matrix of the cells to be measured for multi-feature fusion,and then the residual matrix after multi-feature fusion is calculated to determine the category of the cells to be measured.At the same time,aiming at the problem of low classification accuracy due to the small number of labeled samples of hyperspectral image,this thesis expands the sample by using local spatial distance and correlation constraints to give unlabeled samples a pre-label,so as to increase the number of training samples.The resulting sample expansion group joint sparse representation algorithm SE-GJSRC(Sample Expansion Group JSRC)based on sample expansion is verified on the Indian Pines and Pavia University datasets,and the OA is improved by 3.43% and 2.41% respectively compared with the original JSRC algorithm,and the improved joint sparse representation algorithm has better classification accuracy and better model stability.(3)The multi-scale and multi-feature methods obtained by the above verification are fused to obtain SE-VGDJSRC,a multi-scale multi-feature joint sparsely represented hyperspectral image classification algorithm based on sample expansion.Validation on the Indian Pines dataset shows a 3.70% improvement in OA compared to the original JSRC algorithm.Experimental results show that the sparse representation algorithm combining multi-scale and multi-feature can better use the characteristics of "space-spectrum integration" of hyperspectral image and have better classification accuracy.This thesis has 35 figures,29 tables,and 113 references.