Lyu et al. (2026) A GEFS-ensemble-driven dual-gate machine-learning system for 3-hourly quantitative precipitation forecasting over Brazil: independent dual-reference verification and calibration for flood early warning

Identification

• Journal: Journal of Hydrometeorology
• Year: 2026
• Date: 2026-10-14
• Authors: Guanru Lyu, Yue Jiang, Jiandan Zhong, Qi Wang, Linyan Luo, Yu Luo, Qunfu Huang, Tongzhe Liu
• DOI: 10.1175/jhm-d-26-0116.1

Research Groups

• Hunan Meteorological Data Center, Changsha, Hunan, China
• Hunan Sub-Bureau of Air Traffic Management, Civil Aviation Administration of China, Changsha, Hunan, China
• Chengdu University of Information Technology, Chengdu, Sichuan, China
• National Meteorological Information Centre, China Meteorological Administration, Beijing, China
• Hunan Climate Center, Changsha, Hunan, China

Short Summary

This study develops a GEFS-ensemble-driven dual-gate machine-learning system (V2.0C) for 3-hourly quantitative precipitation forecasting over Brazil, demonstrating significantly improved heavy-rain detection and overall skill compared to raw GFS, with a removable wet bias suitable for flood early warning.

Objective

• To develop and verify a machine-learning post-processing system for 3-hourly quantitative precipitation forecasting over Brazil, addressing challenges of convective, intermittent rainfall and sparse observing networks, and to calibrate it for flood early warning.

Study Configuration

• Spatial Scale: Brazil, on a 0.25° grid.
• Temporal Scale: 3-hourly forecasts. Verification periods include April–May 2026 and the austral-summer convective season (October 2025–March 2026).

Methodology and Data

• Models used: A dual-gate machine-learning system (V2.0C) coupling a LightGBM precipitation-occurrence classifier with a Tweedie amount regressor, and a dual-gate routing scheme featuring a physically guarded heavy-rain (≥5 mm) hard route. The system is driven by GEFS ensemble statistics.
• Data sources: GEFS ensemble statistics (71-dimensional feature set) as input. Verification against IMERG satellite product (areal) and the INMET automatic-station network (point). Raw GFS forecasts used for comparison.

Main Results

• The V2.0C system achieved the highest correlation against IMERG (0.324 and 0.286) and a heavy-rain Critical Success Index (CSI) (≥10 mm) of 0.063–0.071, which is 3.3–4.4 times higher than raw GFS.
• The system demonstrated stable month-to-month skill, including during the South Atlantic Convergence Zone (SACZ) peak.
• A common mask analysis localized 60–65% of the ensemble-driven correlation gain to the southern third of the domain.
• The guarded routing scheme increased heavy-rain detection by 6–10% over a no-expert control, while unguarded use halved heavy-rain CSI.
• Gauge verification confirmed the heavy-rain advantage (CSI ≥10 mm an order of magnitude above raw GFS) and identified a wet bias.
• The detection-oriented wet bias is removable a posteriori using leakage-free quantile mapping, resulting in a bias of +0.02 mm and a Mean Absolute Error (MAE) reduction of 69% without reducing detection, supporting a dual-product strategy.

Contributions

• Development of a novel GEFS-ensemble-driven dual-gate machine-learning system (V2.0C) for sub-daily quantitative precipitation forecasting in Brazil, a region challenged by convective rainfall and sparse observations.
• Introduction of a physically guarded heavy-rain hard route within a dual-gate routing scheme, enabling the safe deployment of a high-recall heavy-rain expert.
• Demonstration of significant improvements in heavy-rain detection and overall QPF skill compared to raw GFS forecasts.
• Proposal of a dual-product strategy for flood warning and hydrological applications, facilitated by the identification and a posteriori removal of a detection-oriented wet bias using leakage-free quantile mapping.
• Comprehensive independent dual-reference verification against both satellite and gauge data.

Funding

Funding information is not explicitly provided in the abstract or paper text.

Citation

@article{Lyu2026GEFSensembledriven,
  author = {Lyu, Guanru and Jiang, Yue and Zhong, Jiandan and Wang, Qi and Luo, Linyan and Luo, Yu and Huang, Qunfu and Liu, Tongzhe},
  title = {A GEFS-ensemble-driven dual-gate machine-learning system for 3-hourly quantitative precipitation forecasting over Brazil: independent dual-reference verification and calibration for flood early warning},
  journal = {Journal of Hydrometeorology},
  year = {2026},
  doi = {10.1175/jhm-d-26-0116.1},
  url = {https://doi.org/10.1175/jhm-d-26-0116.1}
}