Radar Echo Nowcasting Based on a Physics-Constrained Generative Adversarial Network Fusing GNSS-RO and GNSS-PWV

This study introduces a deep learning framework integrating physical constraints and multi-source data to improve 0–2 h radar echo nowcasting. The proposed Generative Adversarial Network (GAN) employs a dual-branch architecture to model motion and growth-decay fields separately, incorporating dynamical constraints and GNSS-derived precipitable water vapor (PWV). Evaluated on severe convection cases in northern China in 2025, the physical constraints GAN significantly enhances forecast spatial consistency and reduces intensity attenuation compared to conventional models. Quantitative results show a 28.5% improvement in Critical Success Index for strong echoes (≥ 30 dBZ) at 42–120 min lead times, with a nearly 50% reduction in missing alarms and stable false alarm rates. The performance gain is primarily attributed to the physics-based loss functions, while PWV inclusion offers additional, though modest, improvement. The framework demonstrates effective extension of useful forecast skill and supports operational nowcasting applications.