Optimization of SVM parameters using High Dimensional Model Representation and its application to hyperspectral images

Support vector machines (SVM) is one of the most important methods which has been frequently used in classification of remote sensing images. The classification performance of the SVM strictly depends on choice of convenient kernel function and its kernel parameters called model selection. In the case that the parameters are not appropriately chosen, SVM may result in relatively poor performance. Therefore, the choice of suitable kernel and its parameters is an important topic in classification problems. In this paper, we studied on the optimal selection of the radial basis kernel parameters of SVM using High Dimensional Model Representation (HDMR) which was recently proposed as an efficient tool to capture the input-output relationships in high-dimensional systems for many problems in science and engineering. The performance of the proposed approach was first analysed with some mathematical functions whose optimums are analytically known in comparison to the grid search method. Different experiments were also conducted with synthetic and hyperspectral datasets. The main advantage of the approach over the grid-search is to require relatively few number of training evaluation and hence less computational time in order to optimize the parameters. Therefore, training time required for SVM is significantly reduced.