Study On Blind Signal Separation In Hyperspectral Images

Hyperspectral sensors can collect images in many narrow adjacent spectral bands,providing unprecedented rich information for hyperspectral remote sensing images.However,the spectral values recorded by the sensors are usually the composite spectra of various ground objects,which cannot directly match the spectra of pure ground objects in the spectral library.Spectral separation is used to distinguish and locate the various types of substances present in the scene.When this process is done without the help of a material spectral reference library,the problem is called blind signal separation.The existing research on blind signal separation of hyperspectral remote sensing images mainly has the following problems that need to be solved urgently: the performance of the original ICA method is reduced or even fails to deal with the outliers of ICA data;The robustness of fast fixed-point algorithm;The traditional nonlinear unmixing algorithm has poor convergence and instability.The traditional preprocessing of hyperspectral data by stripping is not effective.Lack of a framework for full automatic spectral decomposition.Based on the research,the following results were obtained:(1)The decomposition of mixed pixels in hyperspectral images is discussed by using the statistical analysis method.The problems of independent estimation of the number of end elements and subsequent spectral decomposition are solved.Using a rapidly fixed-point algorithm for linear and nonlinear mixed simulated hyperspectral images,real hyperspectral image mixed solution,by setting the rapid fixed point(serial)orthogonalization algorithm steps and rapidly fixed point symmetry(parallel)orthogonalization algorithm,different nonlinear function,the number of components extracted from the yuan,simulated white gaussian noise and impulse noise from three aspects,analyzes the parameters on the robustness of mixed results.(2)The robustness problem of the fast fixed-point algorithm is solved.The traditional fixed point algorithms for real signals are compared by using the nonpolynomial function given by Hyvarinen.A robust non-polynomial iterative function is constructed and applied to the decomposition of mixed pixels of hyperspectral images by using the negative entropy approximation and piecewise estimation method proposed by predecessors.The results show that the polynomial can still maintain its effective separation performance.(3)The multilayer non-negative matrix factorization algorithm is studied as a spectral decomposition method.The parameters of the multilevel non-negative matrix factorization algorithm are evaluated to determine their influence on the unmixing results.Verify the robustness of the algorithm in the linear mixing model of hyperspectral image unmixing.The multi-layer non-negative matrix factorization algorithm is introduced to the unmixing problem of the hyperspectral image of a nonlinear mixed model.By analyzing the error of the unmixing results,it is shown that the algorithm can achieve better results although its convergence speed is slow.(4)Hyperspectral image restoration based on frame regularization low-rank nonnegative matrix factorization.The low-rank nonnegative matrix factorization is used to describe the hyperspectral image in low-rank subspace,and the algorithm can converge to the minimized subset in the coordinate system.The validity of the model is proved quantitatively and qualitatively in the data experiments of the HJ-1A hyperspectral experimental area.(5)Many unmixing algorithms are typically evaluated using synthesized physical data.In this experiment,Luding County of Dadu River Basin was taken as the area for result analysis and accuracy verification.In the field environment,the spectral database of spectral data of ground objects in the test area is completed.To HJ-1A hyperspectral image using a variety of experimental spot algorithm is the spectral information extraction,and then mixed with solution after the pure spectra of compared with spectral field acquisition,the results show that the ICA algorithm solution mixed end dollars spectral information closer to the end in the study area and using the spectral library,basin for the plateau transition zone can be fast matching unknown ground object to provide the reference.To sum up,starting from spectral unmixing of hyperspectral remote sensing images,this paper establishes a framework for fast matching unknown ground endmembers,realizes automatic assessment of endmembers spectrum,and improves the level of ground object recognition of remote sensing images.