Abstract: The selection of multiple classes of features plays an important role in the field of pattern recognition. By analyzing the mechanism of feature selection, a performance measure of feature selection is proposed in this paper. Based on it, a dynamic programming method for feature selection is presented, by which a complex process of obtaining globally satisfactory solution to feature selection can be converted into a simple piecewise optimization. It is theoretically proved that the optimum strategy consisting of optimized decisions in each phase corresponds to the globally satisfactory solution of feature selection. The proposed approach is applied to feature selection of underwater acoustical signals.