An automatic target recognition (ATR) system based on rough set-support vector machine (RS-SVM) for SAR targets is proposed in this paper. The system combines the strong feature selection ability of rough set (RS) with the excellent classification ability of SVM together. The wavelet invariant moments firstly are extracted, then selected by using forward greedy numeral attribute reduction algorithm (FGNARA) as the optimal feature subset to indicate targets and fed to SVM for target recognition. Experiments with neural network (NN) and SVM on both original and selected feature set demonstrate the selection of optimal feature subset is meaningful and RS-SVM is efficient in ATR of SAR.