Fang et al. (2025) Improving the fine structure of intense rainfall forecast by a designed generative adversarial network

Identification

• Journal: Geoscientific model development
• Year: 2025
• Date: 2025-12-08
• Authors: Zuliang Fang, Qi Zhong, Haoming Chen, Xiuming Wang, Zhicha Zhang, Hongli Liang
• DOI: 10.5194/gmd-18-9723-2025

Research Groups

• China Meteorological Administration Training Center, Beijing, China
• State Key Laboratory of Severe Weather Meteorological Science and Technology, Beijing, China
• Zhejiang Meteorological Observatory, Hangzhou, China

Short Summary

This study proposes a Generative Fusion Residual Network (GFRNet), a generative adversarial network (GAN)-based framework, to integrate multi-source numerical weather prediction (NWP) forecasts and generate 3-hourly quantitative precipitation forecasts for North China up to 24 hours in advance. GFRNet significantly improves the fine structure and intensity control of intense rainfall forecasts compared to traditional NWP and deep learning baseline models.

Objective

• To improve the fine structure and accuracy of intense rainfall forecasts by developing a Generative Adversarial Network (GAN)-based framework (GFRNet) that integrates multi-source Numerical Weather Prediction (NWP) forecasts for 3-hourly quantitative precipitation forecasts up to 24 hours in advance over North China.

Study Configuration

• Spatial Scale: North China (35–44.55° N, 112–121.55° E), with a target resolution of 0.05° × 0.05° (approximately 5 km), represented by 192 × 192 grid points.
• Temporal Scale: 3-hourly accumulated precipitation forecasts, with lead times up to 24 hours. The study utilized data from the rainy seasons (July–August) of 2019–2024.

Methodology and Data

• Models used:
  • Proposed: Generative Fusion Residual Network (GFRNet), a GAN-based model comprising a U-Net-like generator (FRNet) and a discriminator.
  • Baselines: Multi-Model Similarity Ensemble Method (MSEM), FRNet (GFRNet's generator without adversarial training).
  • Input NWP models: European Centre for Medium-Range Weather Forecasts (ECMWF, approximately 9 km resolution), East China Regional Numerical Center (CMA-SH9, 9 km resolution), China Meteorological Administration (CMA-3KM, approximately 3 km resolution).
• Data sources:
  • Ground truth: CMA Multi-source merged Precipitation Analysis System (CMPAS) at 0.05° × 0.05° spatial resolution and hourly temporal resolution.
  • Input features: 3-hourly accumulated precipitation from ECMWF, CMA-SH9, and CMA-3KM; META features (elevation in meters, latitude in degrees, longitude in degrees); temporal features (forecast cycle and lead hour encoded using trigonometric functions).
  • Dataset: Rainy seasons (July–August) from 2019 to 2024, divided into training (2019–2022, with image-level sampling), validation (2021), and independent test sets (2022–2024).

Main Results

• GFRNet consistently outperforms all three NWP models (ECMWF, CMA-SH9, CMA-3KM) and baseline deep learning models (MSEM, FRNet) across light, moderate, heavy, and extreme rainfall thresholds (0.1, 10, 20, and 40 mm per 3 hours).
• Compared to CMA-3KM, GFRNet improves Threat Scores (TS) by 4% (at 0.1 mm), 28% (at 10 mm), 35% (at 20 mm), and 19% (at 40 mm).
• GFRNet improves Fractions Skill Scores (FSS) by 13% (at 10 mm), 18% (at 20 mm), and 15% (at 40 mm) compared to CMA-3KM.
• GFRNet achieves the highest Multi-Scale Structural Similarity (MS-SSIM) scores and significantly reduces Root Mean Square Error (RMSE), demonstrating robust spatial structure recovery, stable intensity control, and strong generalization ability across independent test years (2022–2024).
• The model's advantages are particularly pronounced and statistically significant (p < 0.001) for organized high-impact heavy rainfall events (Top 10% coverage subset) at 20 mm and 40 mm thresholds.
• GFRNet maintains consistently high FSS and TS across all lead times (3–24 hours) with stable Bias Scores (BIAS) ranging from 0.8 to 1.2, indicating robust intensity estimation under varying temporal conditions.

Contributions

• Proposed GFRNet, a novel GAN-based model that dynamically integrates multi-source NWP forecasts (ECMWF, CMA-SH9, CMA-3KM), significantly enhancing fine-scale precipitation structure reconstruction and mitigating the blurriness common in deep learning precipitation forecasts.
• Introduced a targeted evaluation strategy for high-impact precipitation, utilizing a stringent 40 mm per 3 hours criterion and a "Top 10% coverage" subset to explicitly assess model performance in organized, severe rainfall events.
• Conducted a comprehensive multi-year validation (2022–2024) using diverse metrics (TS, FSS, RMSE, MS-SSIM) and paired t-tests, providing robust evidence of skill improvements and clarifying the sources of these gains in precipitation forecasting.

Funding

• National Natural Science Foundation of China (grant nos. U2142214 and 42030611)
• National Key Research and Development Program of China (grant no. 2023YFC3007502)

Citation

@article{Fang2025Improving,
  author = {Fang, Zuliang and Zhong, Qi and Chen, Haoming and Wang, Xiuming and Zhang, Zhicha and Liang, Hongli},
  title = {Improving the fine structure of intense rainfall forecast by a designed generative adversarial network},
  journal = {Geoscientific model development},
  year = {2025},
  doi = {10.5194/gmd-18-9723-2025},
  url = {https://doi.org/10.5194/gmd-18-9723-2025}
}