Abstract: To alleviate the common fuzzy problem in deep learning-based on radar echo extrapolation prediction, a new radar echo extrapolation prediction method (PhyDNetSGAN) was developed. This method organically fuses PhyDNet method and a frequency-domain matching generative adversarial network to predict the combined radar reflectivity factor for Jiangsu and its upstream region in the next 3 h. The effectiveness of PhyDNetSGAN was verified by comparing it with PhyDNet (without a generative adversarial network), PhyDNetGAN, and Sprog (an improved optical flow method) for predicting severe convection weather. The results demonstrate the following points: (1) Deep learning methods outperform Sprong method in capturing the nonlinear evolution of strong echoes. (2) PhyDNetGAN and PhyDNetSGAN, which include a generative adversarial network, produce more refined radar echo extrapolation results in line with the subjective cognition of forecasters than the other two groups of experiments and alleviate the “fuzzy” problem. (3) The newly developed PhyDNetSGAN not only improves forecast precision but also better captures the form, position, and central intensity of strong echoes, thereby improving prediction skills and effective forecast times. (4) A newly proposed comprehensive score index, combining TS (Threat Score), Bias, and FID (Fréchet Inception Distance), can better reflect the test effect of an approaching forecast, which is consistent with the subjective cognition of forecasters.