Research On Algorithm Of Mixed Pixel Decomposition For Hyperspectral Image Linear Model

In the continuous electromagnetic spectrum from ultraviolet to short-wave infrared,three-dimensional image information for tens or hundreds of bands has been captured by the hyperspectral imaging spectrometer.Hyperspectral image includes a spectral dimension and two plane dimensions in length and width.Hyperspectral image is widely used in civil and military domains because it contains abundant spectral information,which can be used to detect ground targets and identify ground objects.Hyperspectral image can be formed in narrow bands,as a result it has high spectral resolution.However,in view of the low spatial resolution and complex distribution of ground objects,hyperspectral image contains a large number of mixed pixels.In the present spatial resolution,the containing of a large number of mixed pixels affects the image classification and target detection,in addition,it limits the quantitative development of hyperspectral.Therefore,how to decompose mixed pixels effectively is key to improve the application of hyperspectral image.Mixed pixel decomposition is the process of calculating the types(Endmembers)and the type proportion(Abundance)of ground objects.At present,mixed pixel decomposition algorithm of linear model has been widely used.The aim of this paper is to improve the accuracy of endmember extraction and abundance estimation.The main work are as follows:1.Morphological operator has larger statistical deviation in output,which affects the accuracy of endmember extraction.To solve this problem,an endmember extraction algorithm based on generalized morphology was proposed.In this algorithm,reference pixel was introduced,and the regularized modified energy function was taken to achieve the distance measure.The endmemberswere extracted by calculating the generalized open-close operator with two structural units.The experimental results showed that the algorithm can automatically extract endmembers with high accuracy.2.In view of the apparent Markov property of endmember error in the sequence endmember extraction algorithm for hyperspectral image,which degrades the endmember extraction accuracy,we proposed an endmember extraction algorithm with three endmembers as a group based on Gram-Schmidt orthogonalization.Groups of endmembers were extracted by repeatedly searching for the largest triangle in feature space,so as to reduce the Markov property of the endmember error,improving the accuracy of endmember extraction.The performances of the algorithm were verified by experiments.3.A problem of high computational complexity is involved in most abundance estimation algorithms,in which determinant operations and matrix inversion operations are needed to be performed.Herein,an abundance estimation algorithm based on orthogonal basis was proposed.Abundance estimation was obtained by calculating the ratio of the projection of the vector to be decomposed on the eigenvector and the orthogonal basis.The algorithm has less computational complexity,which only involves the inner product operation of the vectors.The effectiveness of the algorithm was verified by experiments.4.The objective function of classical nonnegative matrix factorization(NMF)is non-convexity,which affects the obtain of optimal solutions.To solve this problem,NMF unmixing algorithm based on the endmember constraints was proposed.According to the independent characteristics of the endmembers,two constraints of spectral correlation and spectral difference were introduced.The iterative operation was performed by using the projection gradient algorithm,and the endmember and abundance estimation were obtained simultaneously.The algorithm possesses better unmixing performance verified by the experiments.