Hyperspectral Unmixing Based On Nonnegative Tensor Factorization

Hyperspectral remote sensing images contain abundant spectral-spatial information,improving the earth observation ability of traditional remote sensing technology.Therefore,hyperspectral remote sensing images have been widely used in geological survey,military reconnaissance,and other fields.However,due to the limited spatial resolution of spectrometer,terrain complexity,and multiple scattering,there are a large number of mixed pixels in the image data,which seriously restricts the application of hyperspectral remote sensing images.Consequently,hyperspectral unmixing has been an important strategy to solve the problem caused by mixed pixels,which decomposes the spectrum of a mixed pixel into a set of component spectra（named endmembers）and their corresponding proportions（named abundances）.In the current research about spectral unmixing,nonnegative tensor factorization（NTF）-based unmixing methods have the advantages of preserving image structure information,explicit physical meaning,and obtaining endmembers and abundances at the same time.Therefore,this thesis analyzes a tensor decomposition model named MV-NTF and proposes two MV-NTF-based spectral unmixing algorithms.The main works are summarized as follows:（1）A new sparsity-constrained coupled nonnegative matrix-tensor factorization for hyperspectral unmixing was proposed.MV-NTF can avoid the structure information loss caused by the hyperspectral data unfolding in nonnegative matrix factorization（NMF）.However,MV-NTF ignores local spatial information due to directly dealing with data as a whole and the forceful rank constraint in low-rank tensor decomposition.Based on analyzing the characteristics of NMF and MV-NTF,a new spectral unmixing approach of coupled decomposition is proposed in this thesis.From the perspective of multi-view,MV-NTF and NMF are coupled with each other by artfully sharing endmembers and abundances to retrain the intrinsic structure information of HSI data and exploit more detailed spatial information.Due to the representations for abundances in different dimensional decompositions are distinct,abundance sharing is achieved indirectly by introducing an auxiliary constraint.Besides,a sparsity constraint is imposed to further improve the unmixing accuracy.Moreover,the coupled decomposition can be seen as an implicit constraint to further reduce the solution space and enhance stability from a mathematical perspective.A series of experimental results illustrate the superiority of our proposed method in preserving the structure information of hyperspectral remote images and the spatial details of abundance maps.（2）A new double weighted sparse nonnegative tensor factorization for hyperspectral unmixing was proposed.For MVNTF,the forceful rank constraint and the way of directly dealing with data tensor lead to the loss of some local details.In order to solve this problem,a double weighted sparse regularization term is introduced into MV-NTF model.To make up for the defect of traditional weighted sparse regularization in fusing spectral information and spatial information,in our proposed method,the first weight focuses on penalizing the nonzero coefficients in abundances,resulting in a sparser solution.Meanwhile,the second weight is designed by exploiting local spatial information,which can further enhance the sparsity of abundance maps and preserve more details simultaneously.The spectral-spatial double weights strategy is firstly applied in NTF-based unmixing method,wherein the weighting factors are reasonably redesigned to adapt to the higher-dimensional factorization.Different from other regularizations of spatial constraint,the double weights avoid the introduction of additional parameters and reduce the complexity of unmixing model.Experimental results demonstrate that the proposed method can enhance the abundance sparsity and exploit the underutilized local spatial information to characterize more detailed textures.