Hyperspectral image classification by steepest ascent relevance vector machine

Abstract

As the adjacent bands of a hyperspectral image are highly correlated,it is not optimum to classify the hyperspectral image in the high dimensional space.To solve the problem,a novel hyperspectral image classifier based on Steepest Ascent and Relevance Vector Machine(SA-RVM) was proposed in this paper.The SA was used to search an optimum feature space and to eliminate redundant features of the image firstly.Then,RVM was trained in the optimized feature subspace and used to classify the test samples.Experiments were performed for four sets of data,it is shown that the accuracies of RVM have raised more than 2.5% in the feature subspace selected by SA,which is close to those of Support Vector Machines(SVMs).For the two data sets with fewer training samples,the accuracies of RVM increase by 5.63% and by 6.2% in the subspace.In addition,benefiting from the sparse solution,the SA-RVM requires very short time in predicting the class labels of unknown samples.It concludes that the SA-RVM has higher precision and efficiency in the prediction,and it is suitable for processing the large-scale hyperspectral images.