Abstract: In order to obtain higher accurate in short-term rainfall prediction, a neural network-based prediction model is proposed. It can predict the rainfall probability in 36 minutes using Doppler radar image sequence. By comparatively analyzing the neural network-based method and the traditional optical fl ow-based method, an ensemble prediction model, which combines the advantages of both methods, is also developed. The ensemble model learns more diverse rainfall changing patterns. The methods are evaluated on a large dataset that contains real radar data spanning across multiple radar stations and for months. Experimental results show that the neural network model achieves prediction with high accuracy and that the ensemble model obtains obvious improvement in overall prediction performances.