Hyperspectral Imagery Unmixing Theory Based On Nonnegative Matrix Factorization

Remote sensing technology is a new earth observation technology.With the development of imaging spectrometer and remote sensing technology,it obtains the surface of the earth image and contains rich space,radiation and spectral triple information,hyperspectral remote sensing imagery processing technology has been applied in more and more areas.However,the spatial resolution of remote sensing image is not enough and each pixel contains many kinds of materials information.This case cause the problem of mixed pixels widely exist.This would seriously hinder the practical application of hyperspectral imagery.Hyperspectral unmixing is one of the most useful tasks in hyperspectral image processing,it makes the mixed pixels into a collection of constituent materials and their corresponding proportions.In recent years,there are many news ideas and methods have been proposed in the field of hyperspectral imagery processing.Linear mixing model has been widely employed,due to its simple and physical meaning clearly.Unmixing algorithms assume that the image pixels are linear combinations of a given number of endmembers with corresponding fractions referred to as abundances.But the existing methods are presented to estimate proportions of all endmembers in the hyperspectral data.However,we need to estimate proportions of class-of-interest rather than all constituent materials.If we only estimate class-of-interest abundance information,then other materials will be regarded as interference existing in the data.We do in-depth research in estimate the class-of-interest abundance information on the basis of predecessors' research.The main contents are as follows:Firstly,it gives a view of a linear mixture model of hyperspectral imagery,describes the steps of unmixing,and introduces some methods which be widely used,including endmember estimated algorithm,endmember extraction algorithm and abundance estimation algorithm.Then,focused on non-negative matrix factorization,and have an introduction to algorithm model and iteration rules of non-negative matrix factorization.At the same time,this paper introduces the nonnegative matrix factorization in the application of hyperspectral image processing and several kinds of typical nonnegative matrix algorithm with constraint items.A new method is proposed that class-of-interest oriented constrained non-negative matrix factorization algorithm.Endmembers and abundances are divided into two parts,class-of-interest and non-class-of-interest.Assuming endmembers of class-of-interest constitutes the vertex of convex geometry,and use the distance is measured and summed up from every endmember to their centroid as constraint item,accomplishing unmixing for class-of-interest while the existence of non-class-of-interest.We study the unmixing ability in different signal to noise ratio and number of pixels by experiment,and verify the effectiveness of the algorithm.Finally,Studied the differences of spectral distance between within-class and between-class.A spectral distance constrained non-negative matrix factorization method is proposed.Using spectral similarity and separability between different endmembers,and adopting different ways of constraints for within-class and between-class.We study the effectiveness of the algorithm,and research unmixing ability in different signal to noise ratio and number of pixels by experiment,and verify.