Research On The Technologies Related To Class Information Of Hyperspectral Imagery

Hyperspectral remote sensing has been become the front line of remote sensing devolopment, and plays an important role in many fields of military, civil and other applications. Higher spectral resolution, more of channels, narrower bandwidth and larger amount of data of hyperspectral image not only bring great research values for human being, but also bring larger challenges for processing them. In this paper, by studying hyperspectral image data, correlative technologys of class information are developed.First,the problems of hyperspectral imagery processing are introduced, and classification and endmember extraction, which are the important aspects of hyperspectral image processing, are analyzed briefly. Disadvantages of both traditionary supervised classification methods with class information and endmemer extraction methods without class information are pointed out to show the necessity of the study in this paper.For supervised classification methods with help of class information, two kinds of processing methods are proposed to protect classes of interest in process of least square support vector machine based classification:to reduce sample classes and to change weighting coefficients. In the reduce sample classes method, by deleting classes of not interest and reserve classes of interest, the classification effect of classes of interest is improved largely. In the change weighting coefficients method, by changing emphasis of every trained sample in training process, samples of classes of interest are emphasized, and so the classification effect is improved.For endmember extraction methods without class information, distance based endmember extraction methods are introduced. Because SGA method and N-FINDR method have obvious excellences, these methods are studyed mainly. At first, SGA method is improved with distance measure from pixel point to hyperplane instead of volume evaluation. And then, N-FINDR method is improved based on support vector machines. In the new method, endmembers are extracted by distance measure of support vector machines, and dimensional reduction processing is not needed. Experiments show that the speed of the proposed algorithms is greatly increased, and the computational burdun is decreased greatly, based on the proposed algorithms have the same selection results as original algorithms.At the same time, some pretreatment technologies of data processing of hyperspectral imagery have been proposed, as for pre-sorting of pixels, detection and deletion of outliers. They also improve the speed of original endmember extraction algorithms.