Abstract: Support vector machine (SVM) is one of the most widely used algorithms in parameter inversion of rough surface. However, the penalty parameters (C) and kernel function parameters (G) in SVM affects accuracy of results. If the parameters are not used properly, the model will lead to "over learning"or "less learning", which reduces greatly prediction accuracy. Several optimization algorithms for C and G of SVM are shown, such as K-fold cross validation (K-CV), genetic algorithm (GA) and particle swarm optimization (PSO). An improved PSO algorithm based on K-CV and GA (GA-CV-PSO) is proposed. The training set and test set are constructed with rough surface backscattering coefficient obtained by the moment method (MoM). Inversion precision and calculation time of different optimization algorithms are compared. It shows that GA-CV-PSO algorithm overcomes shortcomings of single optimization algorithms, with more accurate inversion precision and stronger generalization ability.

Key words: rough surface, support vector machine, improved particle swarm optimization, inversion