A Study Of Sparse Representation For Hyperspectral Unmixing Algorithm

Hyperspectral remote sensing images contain both a lot of spatial information and abundant spectral information,which is a hotspot in recent years.However,in the actual hyperspectral image,due to the limitation of the spatial resolution of the sensor and the complexity of the natural objects,a single pixel usually gathers a variety of characteristic features,which are mixed according to a certain proportion to form a mixed pixel.The existence of mixed pixels hinders hyperspectral image interpretation,target recognition and classification,which requires the emergence of unmixing technology.Sparse unmixing is a popular linear spectral unmixing tool,It can be worked out in semi-supervised fashion by taking the advantage of the spectral library known in advance.In the past,most sparse regression methods were based on convex relaxation,and they attempted to obtain the global solution of a well-defined optimization problem.Recently,due to the need for low computational complexity,more and more people are beginning to focus on the greedy algorithms for sparse unmixing,among of them,subspace matching pursuit(SMP)is a preferable one to recover the optimal endmembers according to the r different columns of the original image.In this paper,we study and summarize the related techniques of unmixing,for the phenomenon that the real hyperspectral image is seriously affected by noise,on the basis of the existing sparse unmixing algorithm,some mature similarity matching algorithms are borrowed and their advantages and disadvantages are compared,finally,we propose an improved algorithm of SMP,using the Dice coefficient method instead of the inner product method as the new matching criterion.By calculating the arithmetic mean of all spectral signals,consider the information of the endmember itself,and not just the relevance information between residual and spectral library.In addition,this paper also adds a pre-partitioning strategy to avoid the problem that the algorithm is in the local optimal when the number of endmember is large,and also makes better use of the spatial information.