Hyperspectral Image Classification Method Based On Sparse Representation

Hyperspectral image is an imaging technology which combines spectroscopy and spatial geometric.The spatial resolution of hyperspectral image is constantly improving,and the related research and application are more and more extensive,such as the application in precision agriculture,geological exploration,national defense security and environmental monitoring and other fields.Ground object classification is a hot topic in hyperspectral image research and application,the purpose of classification is to find the feature category to which each image element belongs,so as to carry out subsequent research or application.Many classification strategies and algorithms have been proposed one after another,among which joint sparse representation can get better classification effect when classifying hyperspectral images.In this thesis,we mainly consider the effects of both spatial information and training sample size on the joint sparse representation for classification.On the one hand,since the joint sparse representation needs to find the neighboring image elements of the image element to be measured,the spatial information of the joint neighboring image element is sparsely reconstructed,and the accuracy of obtaining the neighboring image elements will affect the classification results.On the other hand,the image elements to be tested are sparsely reconstructed by training samples,and if the training samples are too small,the reconstruction error of the image elements to be tested will be very large,which makes the image elements to be tested misclassified.Therefore,the main research of this thesis is derived as follows.1.In the process of joint sparse representation,the full utilization of spatial information of neighboring image elements can improve the classification accuracy of hyperspectral images.Therefore,this thesis firstly adopts the spatial preprocessing of clustering to divide the hyperspectral image dataset into different subsets,so as to obtain the spatial information of hyperspectral images.Secondly,the subsets obtained by spatial preprocessing are used to construct the joint sparse representation model,and different weights are assigned to each image element by the weight calculation formula.Finally,the classification results are determined and corrected by the known information of training samples.The method is simulated with the commonly used Pavia University and Salinas datasets,and the experimental results show that the hyperspectral image classification method can effectively improve the classification accuracy.2.To address the problem of large errors in classification of hyperspectral images based on joint sparse representation due to insufficient training samples,this thesis first processes the spectral bands of the original hyperspectral image data by hierarchical clustering,clusters the spectral bands into clusters,and selects the band with the largest kurtosis value from each spectral band cluster to reduce the dimensionality of the data.Then,the subsets of the new hyperspectral datasets are subdivided by combining the spatial information,the known information of the training samples is used to label the pixels of each subset that may be the training samples,and the discrimination is made according to the spectral similarity,so as to obtain the optimized training sample dictionary.After that,the new hyperspectral data set is divided into subsets,and the possible image elements in each subset are labeled by the known information of the training samples and discriminated according to the spectral similarity to obtain the optimized training sample dictionary.Finally,the obtained dictionary is utilized to construct a joint sparse representation model for hyperspectral image classification.The proposed classification method is simulated by the commonly used Indian Pines and Pavia University datasets,and it is demonstrated that the method has a good classification effect even with a small amount of training samples.