Hyperspectral Unmixing Based On Constrained Nonnegative Matrix Factorization

Hyperspectral remote sensing detects rich spatial and spectral information of the ground,and the wavelength intervals between several hundreds of continuous spectral bands are very narrow.However,combined with high spectral resolution,the spatial resolution is low in contrast because of the limits of remote sensor and the complex diversity of objects.Mixed pixels widely exist in hyperspectral images and the high spectral resolution makes it possible to unmixing pixels.Also mixed pixels become the main obstacle in hyperspectral image applications.Hyperspectral unmixing(HU)is one of the most significant research topics in hyperspectral data processing and analysis.HU aims at decomposing the captured spectra into a set of endmembers and corresponding abundance fractions.Non-negative matrix factorization(NMF)has been widely used as a blind source separation(BSS)method.Since the decomposition model of NMF is similar to HU and the matrixes in NMF are non-negative,the NMF method has been introduced into HU.Generally,NMF is NP-hard and easily influenced by the initial values and constraints.Aiming at the problem,the paper digs out the spatial and spectral information of hyperspectral images and designs new optimization HU methods based on NMF on the basis of the own physical and chemical characteristics of hyperspectral images.Main research work and conclusions are presented as follows:(1)Summarize and analyze the theoretical foundations of HU and NMF,including LSMM,estimation the number of endmembers,endmember extraction,abundance estimation and accuracy evaluation;NMF and the current HU methods based on constrained NMF.(2)Propose a preprocessing technique with the spatial and spectral information of hyperspectral image for HU with constrained NMF.Aiming at the local minimum problem in HU methods based on constrained NMF,the spatial and spectral preprocessing technique improves the unmixing result of constrained NMF by obtaining better endmember candidates from the preprocessing within the spatial and spectral information of the neighborhood.Both spatial preprocessing(SPP)and spatial--spectral preprocessing(SPP)can improve the HU accuracy of constrained NMF,such as MVCNMF and GNMF in the synthetic and real data experiment with the well-known image of Cuprite.(3)Propose a new HU methods with the sparse piecewise smoothness constrained NMF.The method aims at the physical and chemical characteristics of hyperspectral images and take advantage of the smoothness of endmembers and sparseness of abundance fractions.SPSNMF is introduced on the base of sparse-NMF and piecewise smoothness NMF.Experiments on both synthetic and real data sets(AVIRIS Cuprite Nevada data set and Washington,DC data set)are conducted.