Yue et al. (2026) Improving Daily Precipitation Estimates through Machine Learning-Based Downscaling, Precipitation Event Classification, and Categorical Merging

Identification

• Journal: Water Resources Management
• Year: 2026
• Date: 2026-02-27
• Authors: Zhenzhen Yue, Lihua Xiong, Chenguang Xiang
• DOI: 10.1007/s11269-026-04492-8

Research Groups

• PowerChina Huadong Engineering Corporation Limited, Hangzhou, Zhejiang, China
• State Key Laboratory of Water Resources and Hydropower Engineering Science, Wuhan University, Wuhan, China
• PowerChina Kunming Engineering Corporation Limited, Kunming, China

Short Summary

This study proposes a three-step machine learning framework for multi-source precipitation merging, integrating downscaling, precipitation event classification, and categorical merging. The framework developed a high-resolution (1 km, daily) merged precipitation dataset (MSMP) for the Pearl River Basin, demonstrating significantly improved accuracy, especially for heavy and extreme precipitation, compared to existing products.

Objective

• To overcome limitations of existing multi-source precipitation merging (MSP) methods that assume fixed relationships and often degrade performance during heavy and extreme rainfall events, by developing a novel three-step merging framework that integrates downscaling, precipitation event classification, and categorical merging using machine learning to capture nonlinear precipitation behavior and produce a high-resolution, accurate merged precipitation dataset.

Study Configuration

• Spatial Scale: Pearl River Basin, South China; 1 km spatial resolution.
• Temporal Scale: 1981–2020; Daily temporal resolution.

Methodology and Data

• Models used:
  • Random Forest (RF) for precipitation fusion (downscaling, classification, merging).
  • Ordinary Kriging (OK) for interpolation and resampling.
  • Spatial Random Forest (SRF) model for downscaling with spatial autocorrelation.
  • Comparative analysis also included XGBoost and GBDT.
• Data sources:
  • Gauge precipitation: Daily precipitation from 48 gauges provided by the China Meteorological Administration (CMA) for 1981–2020.
  • Multi-source precipitation products:
    • APHRODITE (V1101): 0.25° × 0.25°, 1981–2000.
    • GSMaP_Gauge (Version 06 Final Run): 0.1° × 0.1°, 2001–2020.
    • IMERG-F (Version 06): 0.1° × 0.1°, 2001–2020.
    • CHIRPS: 0.05° × 0.05°, 1981–2020.
    • ERA5-Land reanalysis: 0.1° × 0.1°, 1981–2020.
  • Environmental variables:
    • Total Column Water Vapor (TCWV) from ERA5 reanalysis (0.25° × 0.25°, 1981–2020).
    • Geospatial variables (elevation, longitude, latitude, slope, aspect) derived from Shuttle Radar Topography Mission (SRTM) DEM (90 m × 90 m).

Main Results

• The Multi-Source Merging Precipitation dataset (MSMP) consistently outperformed original precipitation products and conventional MSP methods in both statistical and categorical metrics.
• MSMP achieved a correlation coefficient (CC) of 0.88, representing a 10%–60% improvement over original products.
• MSMP showed an RMSE of 5.21 mm, corresponding to a 25%–59% reduction, and a Kling-Gupta Efficiency (KGE) of 0.88, reflecting a 20%–60% improvement.
• MSMP demonstrated superior performance for heavy and extreme precipitation events, achieving the lowest False Alarm Rate (FAR) and highest Critical Success Index (CSI).
• The multi-class classification strategy consistently outperformed binary classification across all tested machine learning models, with the Random Forest-based multi-class scheme showing the best overall performance.
• Variable importance analysis confirmed that gauge precipitation data and precipitation classification variables were the most influential inputs in both classification and regression models.
• MSMP exhibited improved spatial accuracy and more stable and reliable detection performance across all seasons and subregional scales within the Pearl River Basin.

Contributions

• Proposes a novel three-step machine learning-based framework (downscaling, precipitation event classification, categorical merging) that explicitly accounts for nonlinear relationships between precipitation intensity and environmental variables.
• Develops a high-resolution (1 km, daily) merged precipitation dataset (MSMP) for the Pearl River Basin (1981–2020) with significantly enhanced accuracy, particularly for heavy and extreme precipitation events.
• Demonstrates the superior performance of multi-class precipitation event classification over binary classification in improving fusion accuracy and reliability.
• Provides more reliable precipitation inputs for hydrological modeling, water allocation planning, reservoir operation, and flood risk management, especially in extreme hydrological scenarios.

Funding

• Postdoctoral project of POWERCHINA Huadong Engineering Corporation Limited (KY2024-NGH-02–06)
• National Natural Science Foundation of China (NSFC Grants U2240201)
• Yunnan International Joint R&D Center for Basin-scale Water-Energy-Ecology Regulation (Grant No. 202503AP140045)
• Science and Technology Project of Power China Kunming Engineering Corporation Limited (KD-ZDYF2024-085)

Citation

@article{Yue2026Improving,
  author = {Yue, Zhenzhen and Xiong, Lihua and Xiang, Chenguang},
  title = {Improving Daily Precipitation Estimates through Machine Learning-Based Downscaling, Precipitation Event Classification, and Categorical Merging},
  journal = {Water Resources Management},
  year = {2026},
  doi = {10.1007/s11269-026-04492-8},
  url = {https://doi.org/10.1007/s11269-026-04492-8}
}