Research On Hyperspectral Image Unmixing Technology

Hyperspectral images have the characteristics of integrating spatial and spectral information into a whole,and are the focus of hyperspectral remote sensing research.Due to the low spatial resolution of the hyperspectral sensor,it is easy to generate a large number of mixed pixels in the collected hyperspectral images,which will adversely affect the analysis of hyperspectral images.The purpose of hyperspectral unmixing is to decompose the mixed spectra into a set of reference spectra(endmembers that describe the macroscopic materials present in the scene)and their proportion in each image pixel(abundance).Although a large number of unmixing algorithms have been developed in recent years,it is still a great challenge to accurately estimate the characteristics and abundance of endmembers.The linear spectral mixing model is an effective model to describe the phenomenon of mixed pixels.Non-negative matrix factorization(NMF)has been widely used in HU because it has a similar structure compared to linear mixing model.Due to the non-convexity of HU objective function based on NMF,the solution will often fall into local extremum,which will affect the final unmixing result.In view of the fact that the hyperspectral unmixing based on traditional NMF is susceptible to the initial value and noise,a series of improvements are made in this thesis.The main contents and innovations are presented as follows:Firstly,aiming at the problem that the hyperspectral unmixing based on NMF is easy to fall into local minimum value and is greatly affected by the initial value,a linear unmixing algorithm based on NMF of sparse and orthogonal constraints is proposed.First of all,from the perspective of the traditional NMF hyperspectral linear unmixing,the physical and chemical properties of the hyperspectral data is analyzed.The sparsity of the abundance and the independence of the endmember are combined together,two methods of sparse and orthogonal non-negative matrix factorization are applied into hyperspectral unmixing.The simulation and real experimental data show that this method achieves a better unmixing result in the comparison of similar algorithms.Secondly,aiming at that traditional single-layer NMF is not effective for complex hyperspectral data processing,hyperspectral unmixing method based on gradient-optimized sparse constrained multi-negative NMF is proposed.First of all,the advantages of multi-layer NMF in processing hyperspectral data compared with traditional single-layer NMF are analyzed.Then,the multi-layer structure superior to multi-layer non-negative matrix is similar to the multi-layer structure of neural network.Therefore,combined with the gradient optimization in the neural network,the hierarchical structure of multi-layer NMF is preprocessed,and the preprocessing can effectively reduce the influence of the initial value.Finally,effectiveness of the algorithm is verified by simulation data experiments and real data experiments.