Research On Spectral Unmixing Of Hyperspectral Image Acounting For Endmember Variability

Spectral unmixing is a key issue needed to be addressed in many hyperspectral image applications.The traditional spectral unmixing methods assume that each type of substance has only one corresponding endmember and its endmember set is fixed.Due to the influence of the complex diversity of substance signaturesand imaging conditions,the phenomena of synonym spectrum and identical spectrum commonly exist in hyperspectral image.leading to antonym spectrum and synonym spectrum,the unmixing accuracy of pixels with a fixed set of endmember is limited.Therefore,it is of great significance to research the spectral unmixing method acounting for endmember variability,which helps to improve the application effect of hyperspectral images.This dissertation conducts research on the extraction of endmember bundle with endmember variability and multiple endmember spectral mixture analysis,of which the main contents are as follows:(1)Based on the fact that the existing endmember bundle extraction algorithm using both spatial and spectral information does not make full consideration of the removal of redundant endmembers,which leads to the increase of spectral unmixing error and the high algorithm complexity,an endmember bundle extraction algorithm based on pure pixel index and super-pixel segmentation is proposed.First,the initial candidate endmembers are extracted by PPI.In each superpixel,one candidate endmember is reserved,and its homogeneity index is calculated by its neighborhood superpixels.The reserved endmembers are filtered according to the homogeneity index.Then candidate endmembers are clustered to formulate the endmember bundle,and the redundant endmembers in the same endmember bundle are further removed.The results of experiments on both synthetic data and real world data show that the proposed method can effectively extract variability endmembers and reduce the complexity of subsequent spectral unmixing.(2)The endmember bundle extraction based on super-pixel segmentation and pure pixel index can not be effectively used for urban hyperspectral data with multiple vegetations and dense vegetation mixed with other features.In order to solve these problems,an endmember bundle extraction algorithm based on vegetation index combined with pure pixel index and super-pixel segmentation is proposed.According to the vegetation index,the endmember set extracted by PPI and super-pixel analysis is divided into three types:vegetation endmember,mixed endmember containing vegetation and other endmembers,pure vegetation endmembers are divided into two categories using their maximum spectral values.Other endmembers are screened using the homogeneity index,the final endmember bundle of various types of features are obtained by clustering.The effectiveness of the algorithm is verified by a series of experiments.(3)Multiple endmembers spectral mixture analysis is an effective method for the variability endmember problem.In order to reduce the time complexity of spectral mixing analysis and improve the accuracy in the same time,a multiple endmember spectral mixture analysis algorithm based on corse-to-fine scheme is proposed.Based on the extended endmember set for each pixel,the proposed algorithm firstly make fully-constrained spectral mixing coarse analysis to determine the initial set of end-members containing all land cover materials.On this basis,the algorithm further conducts fine spectral mixture analysis,iterative spectral mixture analysis to build end-member subsets and the optimal end-member set is finally determined according to the variation of reconstruction error.The experimental results show that compared with the iterative spectral mixture analysis method and the hierarchical multi-endmember spectral mixture analysis algorithm,the proposed algorithm reduces the error of inversion abundance and improves computational efficiency greatly.