Autonomous Spectral Unmixing For Hyperspectral Remote Sensing Imagery

Hyperspectral remote sensing technology collects information of martials from across the electromagnetic spectrum and has the advantages of both spectroscopy and optical imaging,which allows for the more accurate identification of materials present in scene.As one of the most important breakthroughs in the field of remote sensing,hyperspectral remote sensing has been applied in the fields of geology,agriculture,environment and military.However,due to the insufficient spatial resolution of sensors,mixed pixels that contain multiple land cover types widely exist in hyperspectral images(HSIs),which seriously hinders the development of quantitative applications of hyperspectral remote sensing.Hyperspectral unmixing(HU)is the main means to solve the problem of mixed pixels.HU aims at decomposing mixed pixels into a collection of pure martials called endmembers,and their proportions,called abundances.There are three fundamental scientific problems to be addressed in HU: the determination of the number of endmembers,the extraction of endmember spectra and the inversion of abundances.In traditional HU chain,the existing algorithms for solving the above three scientific problems are independent and lack of interaction,which leads to the inability of the existing researches to achieve autonomous HU.In general,there are still four urgent problems to be solved: 1)different methods and models in the HU chain are independent of each other,and there is no integrated prototype system for the autonomous HU task;2)the determination of the number of endmembers is usually independent of the following endmembers and abundances estimation,and thus the determined number of endmembers is not suitable for HU;3)the spatial priors of endmember is important of HU but is not as fully utilized as the spectral priors of endmembers.4)the existing mixing model only considers the linear or non-linear mixing effects within a single pixel,but does not consider the non-linear mixing caused by the cross radiation between adjacent pixels.This paper aims at addressing the above common problems in autonomous HU,and the main research contents and innovations are summarized as follows:(1)A saliency-based autonomous endmember detection method(SAED)is proposed to simultaneously estimating the number and spectral signatures of endmembers,which aims at solving the problem that the estimated number of endmembers is independent of the subsequent HU process.In SAED,the endmembers are proved to be salient in abundance anomaly subspace,and then can be detected by using saliency detection methods.The number of endmembers is simultaneously determined in SAED by judging whether there is a salient target in scene.Since the endmembers number,the spectra and the abundance all tie in together,SAED can provide more accurate estimation of both the number and spectra of endmembers.(2)A linear mixing model-based blind HU algorithm,termed spatial group sparsity regularized nonnegative matrix factorization(SGSNMF)is proposed to exploit the spatial priors and sparse patterns of HSIs.Specifically,a spatial group sparsity constraint is introduced into NMF to simultaneously force the sparsity and the spatial correlation of abundances,which can effectively improve the robustness of blind HU algorithm while reducing the regularization parameters.(3)A nonlinear mixing model based blind HU algorithm,termed joint deconvolution and blind HU(DBHU)is proposed to address the adjacency effect during hyperspectral data collections.The influences of adjacency effect on mixed pixels are analyzed,and a bilinear mixing model considering the adjacency effect is proposed to eliminate the impact of the cross-radiation between adjacent pixels.Based on the proposed bilinear mixing model,the DBHU algorithm can effectively reduce the adjacency effects during HU and improve the sharpness and accuracy of abundance estimations.(4)An integrated prototype system for autonomous HU is developed based on the proposed SAED,SGSNMF and DBHU algorithms.The performances and applicability of this prototype system is analyzed,and the implementation of each module is described in details.In summary,this paper aims at establishing a fully autonomous HU framework for hyperspectral remote sensing images,and realizing the interaction between the estimation of the numbers,the spectra and the abundances of endmembers.The research is of great significance to enhance the automation of HU.