Research On Hyperspectral Image Unmixing Methods

The low spatial resolution of spectral imaging techniques and the complex diversity of ground distribution make some pixels a mixture of multiple substances,which are called mixed pixels.The way to the quantitative development of remote sensing is mainly hindered by the presence of mixed pixels.To find out solutions to the mixed pixel problem,hyperspectral unmixing has become an essential technique and besides that,unsupervised hyperspectral unmixing has been drawing more concerns for the restrictions of the unknown endmember information.The non-negative matrix factorization which is known as a classical algorithm in unsupervised unmixing has been broadly utilized in hyperspectral unmixing recent years.but there are still some problems in this method such as the large solution space,the tendency to fall into local optimum solutions and so on.To settle these problems,this thesis launches the study focusing on the hyperspectral unmixing method on the basis of non-negative matrix factorization.Our major work in this thesis is summarised as follows.1.Hyperspectral unmixing which is based upon non-negative matrix factorization has the problems of large solution space and easily falling into local minima.We proposed a non-negative matrix factorization method based on abundance and endmember constraints for linear unmixing of hyperspectral images to settle the problems above.Firstly,the reweighted sparse regularisation is introduced into the non-negative matrix factorization model to promote the sparsity of the abundance results due to the sparsity of abundance matrix.Secondly,based on the similarity prior knowledge of the distribution of the same feature in neighbouring pixels,the total variance regularisation is further introduced into the non-negative matrix factorization model to improve the piecewise smoothness of the abundance matrix.And then,a potential function in a Markov random field model is brought as a constraint term in the non-negative matrix factorization model in this thesis to promote the piecewise smoothness of the endmember spectra due to smoothness of the piecewise of the hyperspectral endmember spectra.Finally,the effectiveness of the proposed method is validated using simulated data and real datasets,and the results show that the proposed method has improved the performance of unmixing compared with the existing methods.2.To solve the problems that the hyperspectral unmixing based on multilayer NMF impose a single type of constraints on pixels with different sparsity in hyperspectral images leading to the limitations of the performance of unmixing,a linear unmixing method based on data-driven constraints for multilayer nonnegative matrix factorization is proposed in Chapter three.Firstly,the unmixing performance cannot be further improved by using a single sparsity constraint because different pixels in the entire image have different sparsity,so in this thesis,on the base of the multilayer NMF model,different regularizers are imposed on each layers according to the sparsity of image pixels.L_1/2 regularizer is used for the image pixels with high sparsity to promote the sparsity of the abundance results,andL₂ regularizer is used for the low sparse image pixels to promote the uniform variation of abundance.Finally,experiments are conducted using simulated dataset and real datasets,and the results are compared with other classical algorithms.The results show that our method can improve the unmixing performance to a certain extent for data with scattered sparsity distribution.