Hyperspectral Image Classification Algorithm Based On Gaussian Process

It is very complicated to processing hyerspectral images because of its disadvantages like high dimensional non-linearity, serious pixel mixture, information redundancy and so on. So Gaussian Process Classifier(GPC) is adopted in order to avoid the Hughes phenomenon, i.e. curse of dimensionality in hyperspectral data processing. GPC is a kind of Bayesian-based randomization sorting algorithms, and is also an randomization expression of complete Bayesian. It has been successfully used in fields like pattern recognition and industrial soft sensing. On the basis of Gaussian process and hyperspectral data, this paper makes a intensive study on sorting algorithms of Gaussian process of hyperspectral images. The main work of hyperspectral images classification in this paper are given as follows:1.An introduction to fundamental principals of Gaussian process is made firstly, followed by a brief analysis on the basic theory of classification and characteristics of hyperspectral images. Then an evaluation index of the classification will be given and the Gaussian Process hyperspectral image classification model is to be obtained.2.Take the Laplace approximation method for example to illustrate the multi-target classification algorithm of the direct method. The hyperspectral classification will be carried out on the foundation of binary classification to realize an indirect multi-target classification method, i.e. two-to-two Gaussian process multi-target classification. The hyperspectral images based on binary classification is relatively more simple in algorithm implementation. Optimizing the binary classification means optimizing the multi-classification at the same time.3.Based on analyzing the different features of kernel function, several combined kernel function will be applied in hyperspetral images classification, and the result is good. The advantage of combined kernel function is that the new-built kernel function, combined by several kernel functions with different property, has the learn ability of the kernel functions with better partial performance as well as the better generalization ability of the kernel functions with better overall performance.4.In the final part, the fundamentals of Parzen window likelihood estimation are explained firstly. Then based on the combined kernel function, Parzen window-combined function of Gaussian process classification algorithm is adopted in hyperspectral classification,and the result is good.