Unmixing Pixels In Hyperspectral/multispectral Image

The reflection or emission spectrum of the grand is recorded in the form of pixel by sensor. Each pixel which is corresponding to a certain surface material is a comprehensive spectral. In remote image processing, if a pixel contains only the spectrum of a ground targets, we called this kind of pixel as pure pixel; if a pixel contains a variety of ground targets, such spectrum is called mixed pixel. The impact of the surface features such as scattering, in the junction of many types of surface targets, mixed pixel are easily produced. In remote image processing, mixed pixels brings a big problem in interpretation, and hinderes the development of remote sensing technology.Nowadays, in the research of decomposition of mixed pixels, the general research model is like this: extraction the image endmember, and get abundance of each components in the mixed pixel, then in sub-pixel level doing sub-pixel mapping with the information of abundance and endmember, after those processes, we can get a new sense of image, the classification accuracy of the new image improved a lot.The research of multi-spectral mixed pixel is few, because the image resovle is low and spectrum information is not rich enough, the solution of mixed pixel in multi-spectral is not the same as hyper-spectral. Mixed pixel solutions in multi-spectral are carried out by single image, and finally we integrates each band's result and gets the unmixing results.Hyperspectral and multispectral pixels unmixing carried out separately.In hyperspectral image, we use AMEE(automatic endmember extraction) method to extract endmembers, and then brings a new least square method with constraints to get abundance of each components in mixtures, at last we use a sub-pixel mapping method based on the credibility; While unmixing multispectral image,because there is not enough spectrum, we fist doing unmixing in a single band image, then integrates them together in stat.. In single image we use mountain clustering method to extract endmembers, and get abundance of components by gray correlation method, finally we use a modified cellular automata method with constraints in sub-pixel mapping, in this article, in remote sensing image unmixing experiments we have achieved good results, the classification accuracy raised a lot.