Hyperspectral image classification based on compsite kernels support vector machine

To improve the utilization of spatial information when classifying hyperspectral images,this paper proposes a composite kernel SVM algorithm combining spatial and spectral information.First,the hyperspectral image was classified into a map using conventional SVM.The spatial-contextual features were then extracted based on the classified map,and combined with spectral information to construct a composite kernel SVM for classification.The spatial-contextual features were extracted again and the composite kernel SVM classified the image iteratively.The process was repeated 10 times and a proper one was chosen as the last outcome.The results show that the method increases the overall accuracy by around 10%,compared with conventional SVM.In addition,the method also demands much less training samples than usual SVM.