Sparse Representation For Hyperspectral Image Classification

In recent years, sparse representation has become a focus in the fields of imageprocessing and pattern recognition. The important theory of sparse representation hasbeen applied to remote sensing images, especially in hyperspectral images (HSI).Hyperspectral image has the ability to acquire rich spectral information. Because of therich spectral information, it’s possible to address various additional applicationsrequiring very high discrimination capabilities in the spectral domain. Meanwhile, itsrich spectral bring many complex factors, such as the curse of dimensionality,uncertainty of data and problem of small samples. The sparse representation hasobviously advantages in high dimension. In addition, distribution characteristics of theobserved data are not necessarily assumed for spare representation algorithm. Moreover,under different situation, special constraints could be introduced to spare representationmodel, so the flexibility of the model is increased.In this thesis, the sparse representation was applied to the classification of HSI.The airborne OMIS-I HSI and Spaceborne Hyperion HSI classification were performedusing sparse representation method. Compared with Support Vector Machine (SVM),sparse representation was a good performance for HSI classification. The main work ofthis thesis included:(1) Sparse model which was used in HSI classification was established on the basisof sparse representation theory and method. An unknown pixel was expressed as thesparse by the dictionary that is constructed by the selection samples. The class label ofthe unknown pixel was determined by the redundancy calculation of each classreconstructed by sparse representation. Compared with the traditional SVM classifier,simulation results showed that the sparse representation yielded a favorableperformance over the SVM.(2) This thesis studied the sparse reconstruction algorithm. Sparse reconstructionmainly adopted greed pursuit algorithm, smoothing constraint model and joint sparsitymodel, including Orthogonal Mctching Pursuit (OMP), Subspace Pursuit (SP), SmoothOrthogonal Mctching Pursuit (S-OMP), Smooth Subspace Pursuit (S-SP), Simultaneous Orthogonal Mctching Pursuit (SOMP) and Simultaneous Subspace Pursuit (SSP). Theexperimental results certified that the sparse representation outperforms the classicalSVM. In the sparse representation algorithms, the classification result with OMP wasbetter than SP, but SP was much efficient. Considering the contextual information, thesmoothing constraint (S-OMP and S-SP) and the joint sparsity (SOMP and SSP) hadbetter performance than the methods based on single pixel (OMP and SP). Thesmoothing constraint and the joint sparsity had the similar classification capability, butcomplexity of the joint sparsity algorithm was lower. Taking the total classificationaccuracy of OMSI-I as an example, SVM method was86.32%, OMP method was92.51%, SP method was89.56%, S-OMP method was93.43%, S-SP method was90.82%, SOMP method was93.17%，SSP method was90.84%.(3) Sparse dictionary reconstruction was also studied and analyzed in this thesisThe Overcomplete dictionary was constructed by the samples and the K-SVD algorithm,respectively. The result of experiment showed that the sparse representation withK-SVD based overcomplete dictionary had better performance than the directly selectedsamples dictionary. Taking the total classification accuracy of OMSI-I as an example,OMP method was93.03%, SP method was91.95%, S-OMP method was94.67%, S-SPmethod was92.09%, SOMP method was96.45%, SSP method was92.31%.