RSG-GAN: A GAN-Based Precipitation Nowcasting Model Integrating Radar QPE, GOES-16 SWD, and GNSS ZTDs

Accurate precipitation nowcasting with high spatiotemporal resolution is essential for various applications, including meteorological services, ecological conservation and atmospheric research. The current nowcasting models, which are primarily based on single radar echo data, exhibit limitations in accurately capturing the complex and fast-evolving nature of precipitation patterns. Consequently, there is an urgent need to incorporate supplementary data sources that offer high spatiotemporal resolution, and the capability for all-weather, all-day monitoring. In this study, we propose an enhanced precipitation nowcasting model, named RSG-GAN (Radar-Satellite-GNSS Generative Adversarial Network), based on the Generative Adversarial Network (GAN). It effectively combines the strengths of radar quantitative precipitation estimation (QPE), Geostationary Operational Environmental Satellite-16 (GOES-16) split window difference (SWD), and Global Navigation Satellite System (GNSS) Zenith Total Delays (ZTDs) to improve nowcasting performance. The American west coast (36° N to 48° N, 118° W to 124° W) is considered as the experimental area. The RSG-GAN model is compared with the traditional optical flow method as well as two deep learning models of utilizing solely radar data (Radar-only model) and integrating radar and satellite data (Rad-sat model). Results of the cases studies exhibit that compared to the optical flow model, the deep learning models demonstrate enhanced ability in capturing rainfall intensity variations, spatial shifts, and achieving outstanding performance in both image quality and precipitation nowcasting metrics, with the RSG-GAN model showing the most notable improvements. Statistical analysis across 189 precipitation periods reveals that the RSG-GAN model achieves the lowest average Mean Absolute Error (MAE) of 0.34 mm/h and Root Mean Square Error (RMSE) of 0.61 mm/h over a 120-minute lead time, with reductions of 36.3% and 41.6%, respectively, compared to the optical flow method. Additionally, at intermediate and higher rainfall intensity thresholds, the RSG-GAN model consistently outperforms other methods, with significant improvements in Critical Success Index (CSI) and Fractions Skill Score (FSS), while maintaining robust nowcasting performance even when other models struggle to predict precipitation. Compared with three deep learning-based methods (CM-STJointNet, MM-RNN and MM-STMixGAN), the RSG-GAN model consistently shows superior performance in both prediction accuracy and event detection. Furthermore, transfer learning experiments on the publicly dataset Storm EVent ImageRy (SEVIR) also demonstrate the remarkable generalization capability of RSG-GAN model.