Precip-UNetGAN: An Attention-Guided Multi-Scale GAN for Radar-Based Precipitation Nowcasting with Metric-Oriented Optimization

Abstract Radar-based precipitation nowcasting must preserve the location, intensity, and spatial organization of rapidly evolving precipitation while maintaining skill at operationally relevant thresholds. This study introduces Precip-UNetGAN, a 0–30 min radar nowcasting framework designed to reduce smoothing of localized high-intensity echoes and the mismatch between pixel-wise training objectives and event-based verification metrics. The generator combines a UNet encoder–decoder with multi-scale spatial attention, while adversarial learning constrains the spatial realism of predicted precipitation structures. A metric-oriented objective combines reflectivity-weighted mean squared error with a time-weighted differentiable critical success index (Soft-CSI). On the Royal Netherlands Meteorological Institute dataset, Precip-UNetGAN achieved the highest critical success index and Heidke skill score among the evaluated baselines at thresholds of 0.5, 2, 5, and 10 mm h(^{-1}), with the clearest gains for moderate-to-heavy precipitation. Experiments on the Shanghai radar dataset reproduced the same overall ranking after dataset-specific training, indicating cross-dataset reproducibility while confirming the need for local calibration. Overall, combining scale-aware feature selection with threshold-oriented optimization improves the retention of intense precipitation structures and balances detection against false alarms, providing a practical basis for further development of short-term radar nowcasting。