Hyperspectral Unmixing Algorithm Combining Spatial Information And Spectral Variability

Hyperspectral Unmixing(HU)is an indispensable step in the process of Hyperspectral Image(HI)processing.It is one of the important steps before Hyperspectral Classification and Hyperspectral Target Detection.However,due to the complex structure of HI in natural scenes,the spectrum exists variability and other issues,accurate HU is still a big problem in the process of HI processing.Considering the spatial information relationship between pixels in the HI and applying it to the HU algorithm model,improving the unmixing performance has become the mainstream direction of HU.This article focuses on the pixel spatial relationship and endmember spectrum variability of the existing HU algorithms to conduct in-depth research,mainly including the following two aspects:(1)In order to overcome the shortcoming that the spatial structure of HI in natural scenes is difficult to accurately represent,a HU algorithm guided by Multi-Scale Spectral Similarity is proposed.The algorithm divides the HI into an approximate domain HI with a spatial structure,and performs approximate domain sparse unmixing according to the similarity between adjacent superpixels.Finally,the approximate domain unmixing result is converted to the original domain,and combined with the actual pixel spectrum for pixel-by-pixel unmixing of the original domain.In addition,the algorithm also uses the-norm to constrain the abundance matrix in the unmixed model to further strengthen the sparsity of the abundance matrix.We use synthetic data sets and real HI data to simulate experiments,and experimental results show that the proposed algorithm has higher accuracy and lower unmixing algorithm complexity.(2)Aiming at the problem of endmember spectrum variability in HI in natural scenes,a new HU algorithm based on Multi-Scale Spectral Spatial Weighting combined with extended linear mixing model is proposed.The algorithm model first uses superpixel segmentation to segment the HI into superpixel images in the approximate domain,and then uses the spatial information of the spectrum in the approximate domain to introduce spectral spatial weighting for sparse unmixing of the approximate domain,and finally performs the original domain HI estimation of the abundance of endmembers.In addition,the spatial relationship weighting factor is introduced to strengthen the correlation between the superpixels of the abundance map or the correlation between pixels.We conduct experiments on simulated data sets and real data sets respectively,and the experimental results prove that the algorithm proposed in this paper can more effectively unmix the spectrum variability,and the algorithm complexity is lower.