Study On The Theory And Method Of Linear Hyperspectral Unmixing And Its Application

Hyperspectral remote sensing is one of the important frontier technologies in the field of remote sensing.Imaging spectrometer can measure the electromagnetic energy of all the materials in the instantaneous field of view scattered in hundreds or thousands of spectral channels.It has a higher spectral resolution than a multispectral camera,covering the visible,near-infrared,and shortwave infrared spectral bands,which wavelength is from 0.3 micron to 2.5 micron.At present,hyperspectral remote sensing has been widely used in many fields such as resources,disasters,global change,polar,environmental monitoring,ecology,agriculture,hydrology,biomedicine,and so on.Hyperspectral unmixing is one of the key and challenging tasks in the field of hyperspectral remote sensing,and one of the key points in the analysis of hyperspectral remote sensing images.Due to the limitation of the spatial resolution of the hyperspectral imager,a pixel in the instantaneous field of view usually contains more than one type of ground materials information,thus it forms a mixed pixel.This phenomenon exists widely in hyperspectral images.At the same time,due to many constraints such as inaccuracy of the hyperspectral unmixing model with observatioan noise,bad environmental conditions,uncertainty of endmember and large size of data,the problem of hyperspectral unmixing is a challenging ill-posed inverse problem.Therefore,it is the key to hyperspectral image analysis to develop robust,stable,feasible and accurate algorithms in order to solve the problem of hyperspectral unmixing.In this paper,we study the theory and method of linear hyperspectral unmixing and its application in ground-objects identification for the decomposition of hyperspectral mixed pixels.Firstly this paper summarizes the research background and current situation for linear hyperspectral unmixing,the content and structure of the thesis.Then the mixed model of linear hyperspectral unmixing and subspace identification are studied.The content includes linear mixed models,the processing procedure,the ideas and problems,characterization of inverse problem of linear hyperspectral unmixing,and signal subspace identification.To test reliability of the method of hyperspectral signal identification for minimum error(HySime),we further study the relationship between the given subset of eigenvalues,feature subspace and the correlation matrix,i.e.,the constrained inverse eigenvalue problem and its corresponding optimal approximate problem,and provide a sufficient and necessary condition to recover the correlation matrix by given eigenvalues and eigenvectors,as well as a presentation of the solution of the corresponding optimal approximate problem.Third the principle and method of hyperspectral unmixing based on geometric theory are studied,which including N-Finder,PPI,VCA,SISAL,AVMAX and SVMAX,and studied the application in endmember extraction of hyperspectral data.Fourthly,three kinds of hyperspectral unmixing algorithms based on nonnegative matrix factorization are studied,including complexity constrained NMF(CC-NMF)algorithm,NMF algorithm with minimum volume constraint(MVC-NMF)and NMF algorithm with both complexity and minimum volume constraints(CMVC-NMF).These algorithms are applied in ground-objects identification in city hyperspectral data.Fifth some sparse regression based algorithms in hyperspectral unmixing are studied.Some optimal problems and the alternating direction method of multipliers used in hyperspectral unmixing are concluded.Then a projected nonstationary iterative Krylov subspace regularization method for ill-posed problems is provided.The selection of regularization parameter and the convergence of the algorithm are discussed.Three different producing process of Krylov subspaces are discussed for the method.Then two kinds of algorithms for hyperspectral unmixing based on sparse regression are studied,which are sparse unmixing algorithms based on variable splitting augmented Lagrange(SUnSAL)and based on constraint SUnSAL(C-SUnSAL),as well as the application of the two algorithms in hyperspectral unmixing.The last party studies the SUnSAL-TV algorithm with total variation(TV)regularization considering the context information of hyperspectral data,the extention to a deblurring SUnSAL-TV algorithm with both the isotropic and anisotropic TV regularization,and the application in mineral identification.At the same time,we generalize the generalized Sylvester equation generated by the SUnSAL-TV algorithm to the general generalized coupled equations,and develop a conjugate gradient based iterative algorithm for solving the generalized coupled equations.Finally,some conclusions and problems to be further studied are proposed.In the present thesis,both the theory and method of linear hyperspectral unmixing are studied systematically,and the Matlab program for these algorithms is achieved.It will help to the development of hyperspectral image,and provide necessary reference of theory and method for earth observation and deep space exploration of hyperspectral remote sensing in the future.