To address the problem that the existing spatiotemporal downscaling deep learning methods are not enough to learn the spatiotemporal characteristics of radar precipitation images, a spatiotemporal separation based 3D deep learning model is proposed. The model takes Unet3d as the core architecture. A hybrid spatio-temporal separation convolution unit is designed to enhance the extraction of local spatio-temporal features of precipitation images, and a three-dimensional Swin Transformer is used to compensate for the loss of spatio-temporal features of precipitation images caused by traditional Unet3d encoder downsampling, so as to improve the effect of spatio-temporal downscaling forecast. The model was tested and evaluated through the open data set provided by METEO FRANCE. The results show that the designed hybrid spatio-temporal separation unit has a better ability to extract local spatio-temporal features, and the spatio-temporal separation based method can improve the spatio-temporal downscaling forecasting effect. Specifically, the 3DUST model proposed in this paper increased SSIM and PSNR evaluation indexes by 5.2% and 6.7%, respectively, and reduced the number of parameters by 3.2% compared with the comparison model.