AFGDiff: a new precipitation nowcasting model framework via adversarial feedback guides diffusion

Prediction nowcasting especially heavy rainfall prediction has garnered significant attention due to its profound impact across various sectors, such as flood warning systems, daily travel planning, and agricultural irrigation. Currently, the application of deep learning in modeling and extrapolating radar echo maps has gained increasing significance. However, despite substantial research advancements, two critical challenges remain inadequately addressed: (1) the absence of multi-scale modeling for complex precipitation evolution processes and (2) the distortion of local information representing extreme weather events. In this work, we propose AFGDiff, a model framework that decomposes the prediction into deterministic and diffusion generation components, thereby achieving multi-scale comprehensive modeling of short-term rainfall. By leveraging the residual feature map from the deterministic prediction to guide the diffusion process, the output prediction map achieves high-resolution accuracy and precise local coordinates. Furthermore, a closed-loop feedback mechanism was implemented to enable the model to continuously assess its prediction performance, thereby facilitating continuous self-optimization and learning. Extensive experiments on the SEVIR dataset and CKIM dataset, demonstrate that AFGDiff achieves competitive performance, especially in enhancing short-term heavy rainfall predictions.