Hyperspectral image classification by genetic relevance vector machine

The adjacent bands of hyperspectral image are highly correlated. It is not optimum to classify the hyperspectral image in the high dimensional space. In addition, optimizing the parameter of classifier by the cross validation method is not a trivial task. Aiming at the two targets, the classification of the hyperspectral image with genetic relevance vector machine (GA-RVM) was proposed. GA-RVM searches the best parameter and feature space for relevance vector machine (RVM), to reduce the redundant information and simplify the parameter optimization procedure. GA-RVM was evaluated by several experiments. Nearly 50% of the bands are eliminated during the optimization, leading to a 3% increase in the overall accuracy. The improvements are obvious for the hard-to-separate classes. Two kinds of soybeans that have the most misclassifications acquire an 8% improvement in accuracy.