Research On Sub-pixel Mapping Technique For Hyperspectral Remote Sensing Imagery

Hyperspectral remote sensing is one of the frontier technologies of remote sensing technology development,which has rich spectral information.However,the higher the spectral resolution,the less the spatial information.Because of the complex terrain information,the hyperspectral image has a large number of mixed pixels.Each mixed pixel includes several objects.If using the traditional hard classification methods which sentences a pixel to a certain class,a lot of information will be lost.The spectral unmixing technique decomposes the mixed pixels to acquire proportion of each object in a mixed pixel,but the spatial distribution of these objects is not obtained.Sub-pixel mapping is a technique for estimating the spatial distribution of various objects in a mixed pixel.As a result,the various targets are displayed on the sub-pixel scale and the spatial resolution of the hyperspectral image is improved.In this paper,the sub-pixel mapping method for hyperspectral image is studied.The concrete work is as follows:1.The linear mixed model of hyperspectral image is studied and the method of estimating the endmember of the vertex component analysis and the method of abundance calculation of the constrained least squares algorithm are introduced.The basic principles of sub-pixel mapping is studied,and two classical sub-pixel mapping methods are introduced in detail: sub-pixel/pixel spatial attraction model and pixel swapping algorithm.Both methods are applied to the mixed pixel unmixing.The abundance accuracy of the obtained pixel has some requirements.If the abundance is not accurate enough,there will be a large number of isolated pixels in the sub-pixel mapping result image.2.Due to the basic sub-pixel mapping method affected by the precision of abundance,the sub-pixel mapping method based on Markov random field is introduced in detail in this paper,and the formula for hyperspectral image is deduced.The sub-pixel mapping method based on Markov random field also considers spectral constraints and spatial constraints,and the spatial restraint part can smooth out the abundance error caused by spectral unmixing.However,this method is often affected by the initial priori information parameters(mean and variance),so this paper combines the K-means clustering idea,proposed a Markov random field sub-pixel mapping method based on K-means clustering.The priori information is optimized in the iterative process to further improve the accuracy of sub-pixel mapping.3.In order to eliminate the error of the abundance,this paper introduces a K-P-Means endmember clustering algorithm,which first presents the idea of "purified" pixel,optimize the abundance and endmember through a two-step iteration to estimate abundance more accurately.At the same time,the sub-pixel mapping model with abundance information and spatial information is adopted and simulated annealing algorithm is used for optimization.This sub-pixel mapping method is not limited by abundance completely.A method for sub-pixel mapping based on the K-P-Means algorithm is proposed,where the abundance and sub-pixel mapping results are optimized in the iterative process to reach the optimized and convergent result.