In order to alleviate the common "fuzzy" problem in deep learning radar echo extrapolation prediction, a radar echo extrapolation prediction method (PhyDNetSGAN) with organic fusion of PhyDNet and frequency-domain matching generative adversarial network was developed, which can predict the combined radar reflectance factor in Jiangsu and its upstream region in the future 3h. By comparing PhyDNetSGAN, PhyDNet (without generative adversarial network), PhyDNetGAN and Sprog (improved optical flow method), the applicability of the new method in severe convection weather prediction was verified. The results show that: (1) Compared with Sprog, the deep learning method can better reflect the nonlinear evolution of strong echoes. (2) PhyDNetGAN and PhyDNetSGAN with the addition of generative adversarial network can obtain 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 proposed PhyDNetSGAN can not only improve the forecast precision, but also better capture the form, position and central intensity of strong echoes, so as to obtain better prediction skills and extend the effective forecast time. (4) Compared with TS, the newly proposed comprehensive score index of TS, Bias and FID can better reflect the test effect of approaching forecast which is consistent with the subjective experience of forecasters.