Research On Hyperspectral Image Unmixing Technology

Hyperspectral images contain not only spatial information of general images,but also spectral dimension information,which makes hyperspectral images have more information than general images,and also makes hyperspectral remote sensing widely used.The resolution of the spectral sensor is limited,and the collected pixels are often mixed with a variety of ground objects,which is not conducive to the analysis of ground object information.Therefore,spectral unmixing technology,which can extract deeper and more accurate information of ground objects,has become a hot research field.Although a great deal of research has been carried out on hyperspectral mixing techniques in recent years and considerable progress has been made,there is still a large room for improvement in the mixing accuracy of end elements and abundance.In this thesis,two algorithms are proposed to solve some problems of the hyperspectral unmixing.First,a scattering-term constrained nonnegative matrix factorization algorithm is proposed for that,the traditional constrained nonnegative matrix factorization method which has less consideration for the physical properties of unmixing.Different from most constrained nonnegative matrix factorization algorithms which base constraints on the mathematical properties of the data,the proposed algorithm considers that the scattering of suspended solids and cement objects in the atmosphere has a significant influence on the spectral signals receiver by the imaging spectrometer.The proposed algorithm will make the Mie Scattering of the atmosphere caused by the neighborhood contribution as an interference,through the Scattering phase function as the constraint condition of constrained non-negative matrix factorization in mathematical expression,and in the abundance of target pixels and Mie Scattering of neighborhood interference constraints,and in order to achieve the objective function to make the interference effect which is caused by Mie Scattering and noise reduce effectively,and the effectiveness of the proposed method is verified by experiment.Secondly,for the problem that the non-negative matrix factorization method is easy to fall into the local optimal solution,due to the non-convexity and noise of the non-negative matrix factorization model,a reweighted sparse constraint and orthogonal constraint non-negative matrix factorization algorithm is proposed.The reweighted sparse non-negative matrix factorization algorithm is a sparse enhancing algorithm,which fully reflects the sparsity of feature abundance in the factorization of hyperspectral images,but also makes it easy to confuse who have similar features in the spectrum.On the basis of sparse non-negative matrix factorization with reweighting,the proposed algorithm introduces orthogonal non-negative matrix factorization to enhance the independence the spectrum of endmember,and further optimizes the reweighting sparse algorithm to achieve better unmixing effect.Experiments also prove the superior performance of the proposed algorithm.