Classification of VHR Images Based on SVM and Multiobjective Evolutionary Optimization

This paper proposes an efficient classification system suitable for VHR images based on SVM and multiobjective optimization. The aim is to detect the optimal number of structural elements to be applied to the VHR image in addition to the optimal SVM parameters in a fully automatic way. To this end, the search process is guided by the simultaneous optimization of two different criteria. The first is the cross-validation accuracy computed on the training set, whereas the second one is related to class separability. In particular, we adopt the between and within class distance in the higher dimensional kernel space. To solve this problem, we use a multiobjective evolutionary algorithm based on decomposition (MOEA/D). MOEA/ED decomposes the multiobjective optimization problem into a number of scalar optimization sub-problems and optimizes them simultaneously. At convergence, a set of Pareto optimal solutions is obtained. For generating the final solution we propose to aggregate the Pareto optimal solutions by means of a simple majority voting (MV) rule. Experiments on a VHR image are reported and discussed.