SCRD: A Spatiotemporal Cues-Guided Residual Diffusion Model for Precipitation Nowcasting

Precipitation nowcasting is crucial in the field of weather forecasting, and it impacts various public services ranging from rainstorm warnings to flight safety. The existing deterministic model-based methods tend to yield blurry extrapolation maps. In contrast, probabilistic generative models focus on producing realistic predictions with more details, but often have unsatisfactory forecasting accuracy. To address high forecast clarity and high-accuracy balance challenge, we propose a spatiotemporal conditional cues-guided residual diffusion (SCRD) network for precipitation nowcasting, where the spatiotemporal conditional cues-guided (STCG) module and shift window cross-interaction (SWCI) module working with residual prediction strategy extract and adaptively feed the multiscale spatiotemporal (ST) auxiliary cues to the noise generation process, enhancing the prediction accuracy of the diffusion-based model. Experiments on the real-world radar echo dataset demonstrate that the proposed SCRD significantly outperforms typical diffusion-based MCVD and ingenious deterministic models in both heavy rainfall prediction accuracy and image clarity.