PSO Parameters Optimization Based Support Vector Machines for Hyperspectral Classification

This study investigates a new approach in hyperspectral image classification. The new method of hyperspectral classification in this paper is a new SVM algorithm based particle swarm optimization (PSO-SVM). First, three classification methods are used to classified a benchmark hyperspectral images: SVM, ML (maximum-likelihood) and K-nn (K-nearest neighbor), the performances of SVMs are compared with two other traditional classifiers (maximum-likelihood classifier and the K-nearest neighbor classifier). The study indicates that the classification accuracy of SVM algorithm is better than ML and K-nn algorithms. The over accuracy of the SVM using RBF kernel is above 90% and the over accuracy of the traditional methods is below 81%. The kernel parameters setting for SVM in a training process impacts on the classification accuracy, to select accurate parameters of the RBF kernel function, we present a SVM algorithm based particle swarm optimization (PSO-SVM) to improve the classification accuracy compared original SVM classification, the experiment indicates our proposed PSO-SVM approach can improve the classification accuracy.