Improving the fine structure of intense rainfall forecast by a designed adversarial generation network

Abstract. Accurate short-term precipitation forecasting is critical for socio-economic activities. However, due to inherent deficiencies of numerical weather models, the accuracy of precipitation forecasts remains significantly inadequate. In recent years, deep learning has been employed to enhance precipitation forecasts, yet these forecasts frequently appear blurry and fail to meet the precision required for operational applications. In this paper, we propose a Generative Adversarial Fusion Network (GFRNet) designed to provide quantitative forecasts of 3-hour accumulated precipitation over the next 24 hours in North China, based on the outputs of multiple numerical weather models. Evaluation results indicate that GFRNet outperforms numerical models across all precipitation intensities. Specifically, GFRNet's threat scores (TS) improved by 4 %, 28 %, 35 %, and 19 % at thresholds of 0.1 mm, 10 mm, 20 mm, and 40 mm, respectively, compared to the highest spatial resolution regional numerical model of the China Meteorological Administration (CMA-3KM). Additionally, GFRNet's Fraction Skill Scores (FSS) at thresholds of 10 mm, 20 mm, and 40 mm show improvements of 13 %, 18 %, and 15 % respectively, over those of CMA-3KM. These enhancements are consistent across most spatial regions and forecast lead times. Furthermore, GFRNet outperforms all models in terms of Root Mean Square Error (RMSE) and Multi-Scale Structural Similarity Index (MS-SSIM). Compared to the deep learning-based precipitation model FRNet, which lacks a generative strategy and tends to produce blurry forecasts with over-prediction, GFRNet more accurately captures the fine structure and evolutionary patterns of precipitation, demonstrating significant operational value.