Tadayon et al. (2025) Enhancing long-lead rainfall forecasting in data-scarce large watersheds using multi-model fusion

Identification

• Journal: Journal of Hydrology Regional Studies
• Year: 2025
• Date: 2025-11-20
• Authors: Amirreza Tadayon, Mahta Nazari, Reza Kerachian
• DOI: 10.1016/j.ejrh.2025.102936

Research Groups

• School of Civil Engineering, College of Engineering, University of Tehran, Tehran, Iran

Short Summary

This study developed a comprehensive multi-model fusion framework to enhance long-term monthly precipitation forecasts in data-scarce large watersheds. It demonstrated that integrating bias-corrected numerical weather prediction (NWP) models using machine learning techniques significantly outperforms individual NWP models, improving forecast accuracy and reliability.

Objective

• To develop a robust framework for enhancing long-term (1–6 months lead time) monthly precipitation forecasts in data-scarce, topographically complex large watersheds by integrating bias-corrected outputs from multiple Numerical Weather Prediction (NWP) models using advanced machine learning-based data fusion techniques.

Study Configuration

• Spatial Scale: Karkheh River Basin, western Iran (approximately 51,000 square kilometers), specifically focusing on six 1 degree x 1 degree grid cells upstream of the Karkheh Dam.
• Temporal Scale: Monthly precipitation forecasts with 1–6 months lead times. Hindcast period: January 1993–December 2016. Training period: January 1993–December 2012 (240 months). Testing period: January 2013–December 2016 (48 months). Operational period case study: January–June 2022.

Methodology and Data

• Models used:
  • Bias Correction: Quantile Mapping (QM).
  • Data Fusion (Machine Learning): Multilayer Perceptron (MLP), Support Vector Regression (SVR), Bayesian Model Averaging (BMA), Random Forest (RF), Adaptive Boosting (AdaBoost), Extreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM).
  • Hyperparameter Tuning: Tree-structured Parzen Estimator (TPE) algorithm with 5-fold cross-validation.
• Data sources:
  • NWP Models: Monthly precipitation forecasts from four Copernicus Climate Change Service (C3S) models: ECMWF-SEAS5, M´et´eo-France-System8, UKMO-GloSea6, and DWD-GCFS2.1 (1993–2016, 1 degree x 1 degree spatial resolution).
  • Reference Data: ERA5-Land reanalysis dataset (monthly precipitation, 1993–2016, aggregated to 1 degree x 1 degree resolution), validated against observations from six ground-based meteorological stations (January 1980–December 2022).

Main Results

• The ERA5-Land reanalysis dataset demonstrated strong consistency with ground-based observations (mean Pearson's r > 0.9, Nash–Sutcliffe Efficiency (NSE) and Kling–Gupta Efficiency (KGE) generally > 0.6) and acceptable ability to predict extreme precipitation events (Probability of Detection (POD) = 0.68–0.92, Critical Success Index (CSI) = 0.51–0.63).
• Raw NWP forecasts exhibited limited accuracy, with ECMWF consistently performing best and DWD weakest. Forecast accuracy declined with increasing lead time.
• Quantile Mapping (QM) substantially improved forecast accuracy, reducing the Root Mean Square Error (RMSE) by an average of 8.09 millimeters and improving NSE and KGE by 0.24 and 0.17, respectively, across most models and lead times.
• Hyperparameter-optimized fusion models consistently outperformed individual bias-corrected NWP models across all lead times and grid cells. The optimal fusion model increased Pearson's r by 0.04, reduced RMSE by 3.25 millimeters, and improved NSE and KGE by 0.08 and 0.06, respectively, compared to the best bias-corrected NWP model.
• For extreme precipitation events, the fusion model showed substantial improvements, with Pearson's r increasing by 0.11–0.13, RMSE reducing by 6–9 millimeters, and POD and CSI increasing by up to 0.27 and 0.28, respectively.
• The Multilayer Perceptron (MLP) was the most frequently selected optimal fusion model (15 out of 36 grid cell–lead time combinations), followed by AdaBoost (6) and LightGBM (5).
• Ensemble forecasts from the fusion models provided reliable uncertainty coverage, with Prediction Interval Coverage Probability (PICP) values exceeding 0.7 in most cases. Continuous Ranked Probability Score (CRPS) and Average Relative Interval Width (ARIW) increased with lead time, reflecting rising uncertainty (e.g., average ARIW for 95% bands increased from 2.00 at 1-month to 2.83 at 6-month lead time).

Contributions

• This study is the first to apply NWP-based precipitation forecasting through multi-model fusion in large, topographically complex, and data-scarce watersheds.
• It addresses existing gaps in regional research by proposing a unified framework that combines multi-model fusion, advanced post-processing (Quantile Mapping), and a comparative analysis of diverse machine learning algorithms (neural networks, kernel-based methods, statistical approaches, and tree-based ensembles).
• The systematic comparison of various machine learning algorithms identifies the best-performing fusion models tailored to the basin's complex topography and climate.
• The framework demonstrates substantial improvements in precipitation forecast accuracy and reliability, particularly for extreme events, offering a robust solution for water resources management in arid and semi-arid regions with limited observational data.

Funding

Not specified in the paper.

Citation

@article{Tadayon2025Enhancing,
  author = {Tadayon, Amirreza and Nazari, Mahta and Kerachian, Reza},
  title = {Enhancing long-lead rainfall forecasting in data-scarce large watersheds using multi-model fusion},
  journal = {Journal of Hydrology Regional Studies},
  year = {2025},
  doi = {10.1016/j.ejrh.2025.102936},
  url = {https://doi.org/10.1016/j.ejrh.2025.102936}
}