Event-Guided Spatiotemporal Transformer with Conditional Diffusion Refinement for High-Intensity Precipitation Nowcasting

Abstract

Accurate nowcasting of high-intensity precipitation is critical for urban flood control and short-term hydrological risk management. However, the high stochasticity of convective systems poses a significant challenge for traditional deep learning models in generating accurate predictions. Existing regression-based models, often constrained by Mean Squared Error loss, tend to produce over-smoothed results, leading to severe underestimation of heavy rainfall centers. To address this challenge, this study proposes an event-guided two-stage nowcasting framework, named EGN-Nowcast, which incorporates intensity-aware auxiliary supervision through an implicit regularization strategy. This framework integrates an event-aware spatiotemporal Transformer with a Conditional Diffusion Refiner. Specifically, an event-aware auxiliary mechanism is introduced to increase the optimization emphasis on high-intensity precipitation regions. The conditional diffusion module then refines the coarse precipitation prediction to recover fine-scale structures while preserving the spatiotemporal consistency learned by the first-stage predictor. This strategy is designed to alleviate over-smoothing and improve spatiotemporal consistency under the evaluated KNMI radar setting. Experiments based on the KNMI radar dataset from 2016 to 2025 show improvements in several pixel-level, categorical, and visual verification metrics, including a reduction in the false alarm rate for high-intensity precipitation events (>8 mm/h). These results suggest that the proposed framework can improve the representation of localized high-intensity precipitation structures and reduce false alarms under sparse high-intensity precipitation conditions on the selected KNMI radar dataset.