Classification Of Hyperspectral Images Based On Gaussian Process Classifiers

Hyperspectral Remote Sensing(HRS) technology appears the earliest in the1880s. It is the first time to achieve combination of the spectral information with location for targets. Compared with multi-spectral remote sensing images, hyperspectral images contain much more information, making the more reasonable and effective analysis of this kind of spectral image possible. Whereas, the processing of hyperspectral image is complex to manage due to their character:few samples are applied with ground truth, high dimensional feature space, and the presence of noise. Gaussian Process Classifier (GPC) is a classification method based on Bayesian framework. It has been applied to human identification, hyperspectral images classification and many other fields. Based on the GPC theory, some new methods for hyperspectral image classification are proposed in this paper aiming for solving the above difficulties. The paper is organized as follows:First, the set of samples with ground truth is relatively small. To solve this, a co-training based semi-supervised method on the basis of Gaussian Process Classifier, namely Co-Gaussian Process Classifier (Co-GPC), is proposed. The proposed method is applied to two real hyperspectral image classifications. The classification performance has notable improvement compared to the original GPC.Second, high-dimensional feature space will cause "curse of dimensionality" for learning. A semi-supervised feature selection algorithm based on spectral analysis proposed in this paper. A semi-supervised similarity matrix is introduced into the spectral analysis feature selection to construct the semi-supervised feature selection algorithm. The proposed semi-supervised feature selection method was applied to real hyperspectral images to perform band selection. The time consuming for classification reduced and satisfactory classification accuracy can be obtained as well.Third, GPC is a classifier based on Bayesian framework. Compared to support vector machine (SVM), it has the advantage of automatic relevance determination, GPC can achieve good performance in hyperspectral classification, but the complexity is high due to the high number of spectral bands. In this paper, we propose a combination of automatic relevance determination and band selection using Particle Swarm Optimization (PSO). Then the hyperparameters can be obtained as well as selected bands. This method can reduce the dimensionality of the feature. It cut down the computation complexity without much loose on accuracy when used in real hyperspectral image classification.This work was supported by the National Natural Science Foundation of China (No.60803097), the National Science Basic Research Plan in Shaanxi Province of China (No.2011JQ8020), and the Fundamental Research Funds for the Central Universities (No. K50511020011).