Researches On Methods Of Unmixing And Endmember Extracting For Hyperspectral Remote Sensing Image

With the improvement of the resolution of hyperspectral sensors, hyperspectral remote sensing image provide massive surface information with spectral resolution in the nanometer level, how to quickly remove the redundant and extract information has become the important issue of hyperspectral remote sensing research. The instantaneous field of view and hyperspectral imaging theory limit the spatial resolution of the hyperspectral image, which makes a single pixel to represent different real objects. Mixed pixel decomposition for multiple homogeneous pixels(endmembers) as a function of expression in linear or nonlinear models need to extract endmember first,then calculate each endmembers’ contribution ratio(abundance estimation) to mixed pixel, and then we get the abundance map that allow us continues hyperspectral image classification analysis. The three keys of hyperspectral remote sensing image processing are: the determination of the number of endmembers, the research of endmember extraction method, mixed pixel unmixing. How to extract endmember effectively and quickly is the key of hyperspectral image unmixing technology research.This paper firstly introduces the background knowledge of hyperspectral image processing, summarizes recent advances in endmember extraction algorithm, analysis the hyperspectral image processing technologies and requirements. The principle and thought of several classic endmember extraction algorithm is introduced, and a brief analysis of the algorithm is proposed. Then studies and analysis the principle and defects of N-FINDR and OSP endmember extraction algorithm, summarizes the N-FINDR improvement thoughts, proposed and implemented a new N-FINDR endmember extraction method combined with spatial information, finally using two types of data, including the artificial data and actual hyperspectral image, to prove the effectiveness of improved method, and strictly quantitative analysis and compared with prime evaluation standard. The experimental results show that the spatial-information-based endmember extraction method can effectively improve the accuracy of endmember extraction and achieve better results. Unfortunately, limited to geometric model assumptions, finding the maximum volume simplex vertex is not efficient for all kinds of complicated data. Hyperspectral images that lead to confusion of the endmembers spectral easily couldn’t extract all the pure endmember very well.