RISE ‐net: A deep‐learning model for improving fine‐scale summer precipitation nowcasting in the Beijing–Tianjin–Hebei region

Accurate precipitation nowcasting is crucial for disaster prevention, mitigation and enhancing socioeconomic resilience. Traditional radar‐based extrapolation nowcasting methods suffer from a rapid decline in accuracy over hourly timescales, while advances in artificial intelligence in recent years offer promising pathways for improving the weather forecast accuracy. This study develops a deep‐learning model named Rapid‐refresh Integrated Seamless Ensemble network (RISE‐net), which improves 1–3‐h precipitation nowcasting with 500‐m resolution and 10‐min update frequency in summer over the Beijing–Tianjin–Hebei (BTH) region. It utilizes high spatiotemporal resolution rainfall data from the RISE system and can serve as its practical postprocessor. Built upon U 2 ‐net architecture, RISE‐net introduces the Coordinate Attention (CA) to enhance the target localization and recognition, and newly develops a novel Hierarchical Upsampling Fusion Module (HUFM) to enhance feature recovery with fine‐grained spatial details. A new dynamic fusion strategy has also been developed at the final output stage to adaptively merge deep‐learning‐based extrapolated results with RISE forecasts. Forecasting performances are assessed using Threat Score (TS), Bias Score (BS), Equitable Threat Score (ETS) and Heidke Skill Score (HSS). Results show that RISE‐net not only outperforms the traditional cross‐correlation extrapolation forecast, but also beats the deep‐learning postprocessing models of U‐net and U 2 ‐net. For strong precipitation thresholds, RISE‐net achieves improvements of up to 48.86% (TS), 52.02% (ETS) and 46.02% (HSS) at most, while BS approaches the optimal value of 1, as compared with the RISE baseline forecast. Ablation experiments demonstrate that the HUFM significantly contributes to reducing forecast errors, while CA plays a marginal role, and their joint combination further improves the rainfall capture. The dynamic fusion strategy provides important references for deep‐learning‐based postprocessing methods including integrating multiple data sources and extending the effective forecast time of extrapolation. Results of this study help strengthen the early‐warning capabilities and disaster preparedness in BTH.