Li et al. (2026) Multi-Machine Learning Ensemble Regionalization of Hydrological Parameters for Enhancing Flood Prediction in Ungauged Mountainous Catchments

Identification

• Journal: Hydrology and earth system sciences
• Year: 2026
• Date: 2026-01-14
• Authors: Kai Li, Linmao Guo, GenXu Wang, Jihui Gao, Xiaoyu Sun, Peng Huang, Jinlong Li, Jiapei Ma, Xinyu Zhang
• DOI: 10.5194/hess-30-205-2026

Research Groups

• State Key Laboratory of Hydraulics and Mountain River Engineering, College of Water Resources and Hydropower, Sichuan University, Chengdu, 610000, China

Short Summary

This study proposes a multi-machine learning ensemble method (GBM-KNN-ERT) to enhance Topography-Based Subsurface Storm Flow (Top-SSF) model parameter regionalization for flood prediction in ungauged mountainous catchments. Validated across 80 catchments in southwestern China, the ensemble achieved a Nash-Sutcliffe Efficiency (NSE) greater than 0.9 for 90% of catchments, demonstrating superior accuracy and robustness to climate change and donor catchment variability.

Objective

• To propose and evaluate a novel multi-machine learning ensemble regionalization method (GBM-KNN-ERT) for sensitive parameters of the Topography-Based Subsurface Storm Flow (Top-SSF) model.
• To significantly enhance flood prediction accuracy and robustness in ungauged mountainous catchments, particularly in Southwest China.
• To assess the ensemble method's stability under climate change and with varying donor catchment configurations.

Study Configuration

• Spatial Scale: 80 mountainous catchments in Southwestern China (Sichuan, Yunnan, Guangxi, Guizhou, and Chongqing provinces). Average area: 1586 km² (ranging from 109 to 6564 km²). Elevations: 63 to 6284 m.
• Temporal Scale: Hourly flow and rainfall data (2015–2018). Hourly meteorological variables (ERA5, 1940–present). Historical (1901–2021) and projected future (SSP585, 2022–2100) temperature and precipitation data.

Methodology and Data

• Models used:
  • Hydrological model: Topography-Based Subsurface Storm Flow (Top-SSF) model.
  • Machine learning models for regionalization: Gradient Boosting Machine (GBM), K-Nearest Neighbors (KNN), Extremely Randomized Trees (ERT), Decision Tree (DT), Random Forest (RF), Support Vector Machines (SVM).
  • Ensemble method: GBM-KNN-ERT (integrating GBM for Szm, td, C; KNN for lnTe, qsf, t; and ERT for Sfmax).
• Data sources:
  • Hourly flow data (2015–2018): Hydrological Bureau of the Ministry of Water Resources, China's hydrologic yearbooks.
  • Hourly rainfall data (2015–2018): Ground meteorological stations across China (http://en.weather.com.cn).
  • Meteorological variables (temperature, wind speed, dewpoint temperature, surface net solar radiation): ERA5 hourly dataset (Hersbach et al., 2023).
  • Historical (1901–2021) and projected future (SSP585, 2022–2100) temperature and precipitation data: "A Big Earth Data Platform for Three Poles" (http://poles.tpdc.ac.cn).
  • Topographic data (30 m resolution Digital Elevation Model): EARTHDATA (https://search.earthdata.nasa.gov/search).
  • Forest cover data (30 m resolution): Global Forest Cover and Forest Change Map (https://www.noda.ac.cn/).
  • Bulk density (BD) data: Soil Database of China for Land Surface Modelling (Dai et al., 2013).
  • Soil hydraulic parameters (saturated hydraulic conductivity Ks_CH, pore-connectivity parameter L): China Dataset of Soil Hydraulic Parameters Using Pedotransfer Functions for Land Surface Modeling (Shangguan et al., 2013).

Main Results

• The GBM-KNN-ERT ensemble method achieved superior flood prediction performance, with 90% of ungauged catchments exhibiting a Nash-Sutcliffe Efficiency (NSE) greater than 0.9.
• This represents a 67.44% increase in the number of high-accuracy predictions (NSE > 0.9) compared to the best single machine learning method (GBM).
• The ensemble method demonstrated improved robustness to climate change, showing a smaller increase in maximum error in runoff modulus (56.65 m³ s⁻¹ km⁻²) compared to the GBM4-KNN3 method (68.46 m³ s⁻¹ km⁻²) under projected future climate scenarios.
• It also exhibited greater stability with varying donor catchment quantities, achieving an optimal balance between predictive accuracy and computational efficiency with a relatively limited set of 20–40 high-quality donor catchments (NSE > 0.85).
• Specific machine learning models were empirically found to be optimal for different sensitive Top-SSF parameters: GBM for Szm, td, and C; KNN for lnTe, qsf, and t; and ERT for Sfmax.

Contributions

• Proposes a novel multi-machine learning ensemble regionalization method (GBM-KNN-ERT) for hydrological model parameters, specifically for the Top-SSF model, to address limitations of single ML methods.
• Demonstrates significantly enhanced flood prediction accuracy (90% of catchments with NSE > 0.9) in ungauged mountainous catchments, representing a substantial improvement over existing single machine learning approaches.
• Provides systematic evidence of the ensemble method's superior robustness and stability under simulated climate change and varying donor catchment configurations.
• Offers a reliable and practical data-driven tool for water resource management and flood disaster mitigation in data-scarce mountainous regions.
• Empirically validates the concept of model complementarity, showing that different machine learning models are optimally suited for regionalizing distinct hydrological parameters governed by different physical processes.

Funding

• Joint Funds of the National Natural Science Foundation of China (grant no. U2240226)
• National Natural Science Foundation of China (grant nos. 42330508 and 42271038)
• National Key Research and Development Program of China (grant no. 2022FY100205)

Citation

@article{Li2026MultiMachine,
  author = {Li, Kai and Guo, Linmao and Wang, GenXu and Gao, Jihui and Sun, Xiaoyu and Huang, Peng and Li, Jinlong and Ma, Jiapei and Zhang, Xinyu},
  title = {Multi-Machine Learning Ensemble Regionalization of Hydrological Parameters for Enhancing Flood Prediction in Ungauged Mountainous Catchments},
  journal = {Hydrology and earth system sciences},
  year = {2026},
  doi = {10.5194/hess-30-205-2026},
  url = {https://doi.org/10.5194/hess-30-205-2026}
}