Research On Unmixing Technique For Hyperspectral Images

Hyperspectral remote sensing effectively make the spectral feature and geometric characters of objects together. To provide unprecedented information for human beings. One of the prominent problems in the extraction and recognition of hyperspectral remote sensing information is the mixed pixel.The method to solve mixed pixel is called mixed pixel decomposition.Hyperspectral unmixing is a process to unmix the mixed pixels of a hyperspectral image which is composed of several substances and its corresponding proportions. Non-negative matrix factorization is an algorithm which decomposes the non-negative data into the multiplication of two non-negative matrices. This paper mainly focuses on linear model of mixed pixel and image characteristics, and based on the non-negative matrix decomposition algorithm, then it has conducted research and analysis on mixed pixels solution.Firstly,The cause of the mixed pixels is analyzed. Then, the linear mixed model, the extraction of the endmember, the abundance and the evaluation index of the hyperspectral unmixing are discussed in detail. The last, the existing algorithms of the endmember extraction and mixed pixel decomposition are analyzed and compared. The advantages and disadvantages of each algorithm are compared.To solve the problem of large solution space and a mass of local minima in the traditional non-negative matrix factorization(NMF), an improved NMF with sparseness and smoothness constraints(INMFSSC) algorithm was proposed. Firstly, endmembers extracted by vertex component analysis(VCA) in hyperspectral image to initialize the endmember matrix. Then, the least squares method to extract the abundance as the initial value of the abundance matrix. The last, the traditional NMF was extended by smoothness and sparseness constraint to achieve better separation of mixed pixels. The experimental results on simulated and real hyperspectral image demonstrate that the proposed algorithm can overcome the shortcomings of traditional NMF and obtain more accurate end-members and corresponding abundance.The minimum volume constraint nonnegative matrix factorization(MVCNMF) algorithm based on minimum size constraint does not need to assume the existence of pure pixels; and when the endmembers are automatically extracted, the abundance charts corresponding to each endmember are obtained at the same time. However, this method does not take into account the sparseness of the abundance matrix. The smoothed constraint minimum volume constraint nonnegative matrix factorization(SCMVCNMF) algorithm based on sparse constraints was proposed. The smoothed L0 model sparse constraint is introduced into SCMVCNMF algorithm, and the accuracy of the algorithm is further improved.