Researches On Hyperspectral Image Unmixing Technology Based On Non-negative Matrix Factorization

Hyperspectral image analysis has become one of the most influential and fastest developing technologies in remote sensing field.However,because the existence of mixed pixels seriously affects the application of hyperspectral data,in order to solve this problem,hyperspectral unmixing is proposed.The purpose of hyperspectral unmixing is to decompose the mixed pixels into a set of spectra and their corresponding ratios.In recent years,hyperspectral unmixing based on non-negative matrix factorization has received a lot of research,but how to effectively use the spatial and spectral information of hyperspectral is still a research difficulty,and how to obtain the hidden layer information in the process of unmixing is also a big challenge.This thesis mainly studies the above problems.The specific research contents are as follows:First,a non-negative matrix factorization algorithm based on adaptive local neighborhood weighted constraint is proposed to solve the problem of structural fixation when the hyperspectral spatial and spectral information is not utilized enough,especially when determining the local neighborhood.According to the data characteristics of abundance,the algorithm can adaptively determine the local neighborhood of a given pixel,and the weight of the algorithm makes full use of the spatial and spectral information of the given pixel and the neighborhood pixel,which improves the performance of hyperspectral unmixing.In this thesis,the gradient descent method is used to derive the multiplication iteration rules.In order to verify the effectiveness of the proposed algorithm,the Japser Ridge dataset and Urban dataset are used for experiments,and compared with other classical methods,the results show that the proposed method has better unmixing effect.Secondly,in view of the fact that many methods based on non-negative matrix factorization are single-layer structures and ignore the hidden layer information,a deep non-negative matrix factorization algorithm based on adaptive local neighborhood weighting constraint is proposed.Through the concept of deep learning,the algorithm extends the single-layer non-negative matrix factorization to the deep non-negative matrix factorization.Combined with the content of the previous part,the spatial and spectral information of hyperspectral is effectively utilized to further improve the accuracy of unmixing.In order to verify the effectiveness of the proposed algorithm,two real data sets are used to conduct experiments,and the results show that the proposed algorithm has a good unmixing effect.