Research On Hyperspectral Unmixing Techniques

With the continuous development of hyperspectral remote sensing technology,hyperspectral images have remarkable characteristics: high spectral resolution,Atlas one,and widely used in various fields.However,the main obstacle to further development of remote sensing technology to quantitative direction is the widespread existence of mixed pixel.In order to break through the barriers and objects,the spatial resolution of remote sensing image with low complexity and diversity of the effects of various types of objects often contained within a single pixel independent to get the real attribute information of mixed pixels in sub-pixel level accuracy,improve the image classification accuracy.In hyperspectral image,one of the key problems is how to decompose mixed pixels effectively,which has attracted wide attention and has been deeply studied.This paper first describes the related technology research and application,and describes the present situation of the mixed solution of hyperspectral unmixing problems,such as the unmixing effect is not ideal,the objective function of the slow convergence speed of the algorithm,image classification is not accurate,time-consuming etc.In view of the above problems,based on the NMF algorithm,3 hybrid pixel decomposition algorithms are proposed:(1)a mixed pixel decomposition algorithm based on graph regularization and sparse constraint semi supervised NMF.The algorithm adds Laplasse map regularization constraint and partial sample category information,and applies sparse constraints to the abundance matrix.Finally,it is fused into the same objective function,which can improve the effect of unmixing.(2)INMF hyperspectral unmixing based on graph regularization and sparse constraint.The algorithm combines sparse nonnegative matrix factorization and incremental learning,which not only reduces average running time but also improves image classification accuracy.(3)dual graph-regularized constrained NMF.mixed pixel unmixing.The algorithm not only takes account of the geometry structure of hyperspectral data and manifold,but also applies the known label category information to the nonnegative matrix factorization,which greatly accelerates the convergence speed of the objective function and improves the effect further.