The Research On Model And Algorithm Of Hyperspectral Blind Unmixing Based On Sparse Representation And Nonnegative Tensor Factorization

Hyperspectral images（HSIs）are three-dimensional image data produced by imaging technology and spectral technology.However,due to the influence of sensor’s low spatial resolution and complex ground distribution,there are a large number of mixed pixels in HSIs,which restrict the analysis and utilization of HSIs.Therefore,hyperspectral unmix-ing（HU）came into being,aiming at decomposing mixed pixels into several pure materials（i.e.,endmembers）and their corresponding fractions（i.e.,abundances）.Hyperspectral unmixing method based on nonnegative matrix factorization（NMF）can obtain endmembers and abundances at the same time,which is one of the typical blind unmixing methods.However,NMF expanded an HSI into a 2-D matrix for processing,which destroyed the spatial spectrum structure of HSI.Hyperspectral unmixing method based on nonnegative tensor factorization can represent HSI lossless,but its objective function is still nonconvex,which leads to the nonunique solution space and long solution time.To improve these problems,this thesis fully excavates the prior information of abun-dance maps.Under the framework of the matrix-vector nonnegative tensor factorization（MV-NTF）,a new unmixing model adopting a weighted nuclear norm and an L_1/2norm is proposed,which constrains the low-rank property and sparsity of each abundance map simultaneously.Instead of using low-rank matrix decomposition of MV-NTF,this model uses weighted nuclear norm to constrain the low-rank property,which avoids estimating the rank of the abundance map in advance.In addition,in order to describe the character-istics of abundance maps in detail,this model builds an adaptive update mechanism:each low-rank and sparse constraint are treated differently by using the self-adaptive parame-ters according to the actual scene although considering these characteristics of abundance maps simultaneously.Furthermore,an augmented multiplicative iterative algorithm is de-signed to solve the proposed model.Specially,the algorithm designed for the tensor model uses the equivalence relation between MV-NTF and NMF to reduce the computation of tensor operation.Finally,the proposed unmixing method is tested in the simulation ex-periments and real experiments.Experiments demonstrate that the proposed method has obviously improved both the unmixing effect and the solving speed compared with several blind unmixing methods of the same type.