Research On Synchronous Decomposition Of Mixed Pixels In Hyperspectral Imagery Based On The NMF

The optical sensor can only be imaged after the radiance of the entrance pupil reaches a certain value,and the radiance of the entrance pupil is determined by the size of the pixels and the number of bands.As the spectral resolution increases,the corresponding spatial resolution decreases.Hyperspectral images often have hundreds of bands,and the corresponding spatial resolution is low,which is also the fundamental reason for the existence of a large number of mixed pixels in hyperspectral images.The existence of mixed pixels greatly limits the quantitative application of hyperspectral images,so it is particularly important to efficiently and accurately unmix the mixed pixels.As a blind source signal separation method,the synchronous decomposition algorithm based on non-negative matrix factorization has become a research hotspot in mixed pixel decomposition,which matrix factorization model is similar to the linear unmixing model,and which non-negative characteristics perfectly fit the characteristics of non-negative values in the process of hyperspectral image unmixing.The hyperspectral unmixing system based on NMF is composed of three parts:the initialization?the objective function and the iterative rules of the algorithm itself.Most scholars focus on the research of hyperspectral image unmixing algorithms,and think less about initialization problems.This paper studies from the perspective of endmember initialization,and proposes a hyperspectral image segmentation method,based on this,a new initialization method is proposed.At the same time,the applicability of the hyperspectral unmixing system based on the existing NMF was studied.The main work of this study is as follows:(1)Research on segmentation algorithm based on hyperspectral imageFor the problem of local endmember subset optimization,from the perspective of block calculation of hyperspectral images,and applying the idea of multispectral clustering to hyperspectral images,a hyperspectral image clustering segmentation method based on spatial spectrum fusion(HICS)is proposed based on spatial spectrum fusion,in order to obtain homogeneous image blocks serving for unmixing system based on NMF.The experiments show:The segmentation method can remove the redundant endmember features in the local image blocks,and select the optimal endmember subset of each homogeneous image block.When segmenting shape parameters?=7 and segmenting weight parameters _sw=0.2,the segmentation result is optimal.(2)Research on unmixing initialization of hyperspectral images based on NMFRelevant research on the problem of unmixing initialization of hyperspectral images,and analyze the advantages and limitations of the initial method of unmixing hyperspectral images based on the existing NMF.Based on the homogeneous image block and the idea of iterative update by image block as the core,a hyperspectral image block initialization method based on Spatial Spectrum Fusion(IISSF)is proposed.The experiment show:The IISSF method uses block initialization as the core and can provide initialization results superior to other initialization methods.(3)Research on the applicability of hyperspectral unmixing system based NMFAiming at the three core parts of the hyperspectral unmixing system based NMF:the initialization,the objective function,and the iterative rules,the applicability research is conducted.The following conclusions were reached through experiments:1)The block initialization method is better than the non-block initialization method.2)Blocked NMF algorithm is better than non-blocked NMF algorithm.3)Constrained NMF algorithm is better than unconstrained NMF algorithm,and the priority of blocking is higher than that of constraint.4)Compared with the multiplicative iterative rule,the gradient descent iterative rule is more limited in the block NMF algorithm and is not suitable for the random initialization process.