Research On The Hyperspectral Image Unmixing Algorithm Based On Nonnegative Matrix Factorization

In hyperspectral remote sensing images,due to the limitations of sensor platforms and the complexity of the real distribution of ground objects,there are mixed pixels composed of a variety of ground objects,resulting in the inability to improve the accuracy of recognition and classification of ground objects based on hyperspectral image data.Therefore,hyperspectral image unmixing technology has become an important research content of quantitative remote sensing.As a blind source decomposition algorithm,there is no need to assume that pure pixels are present in the hyperspectral image,It has become a classical method in unmixing technology,but the non-negative matrix unmixing model is easy to fall into local minimum,which is greatly affected by noise,the unmixing accuracy is low when there are "foreign matter with the same spectrum" end elements in the mixed pixels,and the matrix decomposition is not thorough enough,the decomposition efficiency is low,and so on.Therefore,this paper improves and optimizes the linear unmixing model of non-negative matrices.The main contents and results are as follows:1.In order to solve the problem that the non-negative matrix unmixing model is easy to fall into the local minimum,is greatly affected by noise and has low unmixing accuracy in the case of "foreign matter in the same spectrum" terminal element in the mixed pixel,a non-negative matrix unmixing algorithm(DLGNMF)based on the minimum distance and sparse graph regularization constraints is proposed.The algorithm combines spatial information with spectral information.At first,that distance minimization constraint of the simplex is introduce into the objective function as the constraint condition of the end-member,Secondly,L1/2 sparsity constraint and graph regularization constraint are taken as the constraints of abundance,which ensure the global sparsity and local similarity of the end-member distribution in hyperspectral images,and make the end-member distribution more close to the real distribution of ground objects.However,after testing,it is found that DLGNMF algorithm is inefficient and occupies too much computer memory,which cannot be applied to hyperspectral images with large image size.2.Aiming at the problems of incomplete decomposition of non-negative matrix unmixing model,low decomposition efficiency and the inapplicability of DLGNMF algorithm to large hyperspectral images,a multi-layer non-negative matrix unmixing algorithm(DLNMF-M)based on minimum distance and approximate sparse constraints is proposed.The algorithm retains the simplex distance minimization as an end-member constraint,Secondly,the quadratic parabolic function is added to the objective function as a sparse constraint of abundance,Finally,the idea of multi-layer iterative decomposition is added to the non-negative matrix unmixing model as an external constraint condition,so that the mixed pixel matrix is decomposed in a multi-level and more thorough way.3.In order to test the accuracy and performance of the proposed algorithms,the simulation data and the real hyperspectral data are used to test the two algorithms,and compared with VCA-FCLS algorithm,MVC-NMF algorithm and GLNMF algorithm.The results show that:1)both of the two algorithms enhance the anti-noise ability of the algorithm and solve the problem that the non-negative matrix unmixing model is greatly affected by noise;2)The DLGNMF algorithm can extract the different end-members with similar spectral curves accurately,which can improve the problem of low resolution when the end-members with the same spectrum of foreign matter exist in the mixed pixels.3)DLNMF-M algorithm can solve the problems of incomplete decomposition and low decomposition efficiency of non-negative matrix solution model,and can also be successfully applied to large-scale hyperspectral images with high resolution.4)By summarizing the experimental results,it can be found that DLGNMF algorithm is more suitable for hyperspectral images with small image size,small number of end-members and "foreign objects with the same spectrum",while DLNMF-M algorithm is more suitable for hyperspectral images with large image size,large number of end-members and complexity,and has more practical value.