Research On Sparse Representation For Hyperspectral Image Classification

As an emerging remote sensing technology,Hyperspectral remote sensing has been widely used in environmental monitoring,military reconnaissance,urban planning and many other related fields.Compared with the traditional multispectral image,Hyperspectral image contains a wealth of spectral information and spatial information for the observed target,which can provide more detailed feature information for more accurate analysis.However,Hyperspectral image has the characteristics of high characteristic dimension,strong spectral correlation and high redundancy of the data,which also brings new challenges to the follow-up of hyperspectral image processing.As the basis operation of process remote sensing date,hyperspectral image classification has always been an important part of hyperspectral image processing.It is an important research direction of the hyperspectral image classification,which can make full use of the image information to improve the classification accuracy for meeting the increasing practical application demands.In recent years,With the improvement of the theory of compressive sensing,hyperspectral image classification based on sparse representation had been widely applied.Both handling the challenges of the characteristics of hyperspectral image data and reasonablely using of the abundant spectral and spatial information are main researched in this thesis.The main research content is as follows:1.The hyperspectral image classification algorithm based on weighted-fusing kernel sparse and collaborative representations is proposed.The hyperspectral image data was projected onto high dimensional feature space by kernel function which takes the advantages in the nonlinear data processing.It is advantageous for the classification of hyperspectral image based on Representation-based classifiers in the next step.To improve the representation ability of the atoms in Representation-based Classifiers,The kernel weighted-fusing representation coefficient,which is obtained by fusing kernel sparse representation coefficient and kernel collaborative representation coefficient,is used to classify for the test sample.Experiments on two real hyperspectral image indicate that the proposed algorithm can effectively improve the classification accuracy of hyperspectral images.2.A joint fusion representation hyperspectral image classification algorithm based on region-growing is proposed.Region-growing algorithm under the description of spectral angle cosine measure is used to extract the adaptive region sample information for the test sample,which can effectively avoid the limitations of fixed neighborhood window in joint sparse representation model.At the same time,the thought of fusion representation classifier is extended to the joint sparse representation model.The Weighted–fusing representation matrix,which is obtained by fusing two joint individual representation matrix that can consider the sparsity and collaborative structure of the base atoms simultaneously,is used to classify for the test sample.The experimental results show that the proposed method,compared with the joint sparse representation model,effectively improves the classification accuracy of hyperspectral image.