Classification of Hyperspectral Image Based on SVM Optimized by a New Particle Swarm Optimization

Support Vector Machine (SVM) is used to classify hyperspectral remote sensing image in this paper. Radial Basis Function (RBF), which is most widely used, is chosen as the kernel function of SVM. Selection of kernel function parameter is a pivotal factor which influences the performance of SVM. For this reason, Particle Swarm Optimization (PSO) is provided to get a better result. In order to improve the optimization efficiency of kernel function parameter, firstly larger steps of grid search method is used to find the appropriate rang of parameter. Since the PSO tends to be trapped into local optimal solutions, a weight and mutation particle swam optimization algorithm was proposed, in which the weight dynamically changes with a liner rule and the global best particle mutates per iteration to optimize the parameters of RBF-SVM. At last, a 220-bands hyperspectral remote sensing image of AVIRIS is taken as an experiment, which demonstrates that the method this paper proposed is an effective way to search the SVM parameters and is available in improving the performance of SVM classifiers.