Baida et al. (2026) Calibrated uncertainty for continental design flood estimation using conformal probabilistic machine learning

Identification

• Journal: Journal of Hydrology
• Year: 2026
• Date: 2026-06-14
• Authors: Maelaynayn El Baida, Mimoun Chourak, Farid Boushaba
• DOI: 10.1016/j.jhydrol.2026.135864

Research Groups

• Laboratory of Modeling and Scientific Computation (LMCS), National School of Applied Sciences of Oujda, Mohamed 1st University, Morocco
• Industrial and Seismic Engineering Research Team, Mechanics and Applied Mathematics Department, National School of Applied Sciences of Oujda, Mohammed 1st University, Morocco
• African Disaster Mitigation Research Center (ADMIR), NRIAG, Cairo, Egypt

Short Summary

This study evaluates the performance of four uncertainty-aware machine learning models for predicting 100-year design flood discharge in ungauged African watersheds. The results show that a conformalized quantile regression extreme gradient boosting model outperforms traditional methods and provides robust uncertainty quantification.

Objective

• Investigate the accuracy and reliability of uncertainty-aware machine learning models for predicting 100-year return period flood discharge in data-scarce regions, specifically ungauged African watersheds.

Study Configuration

• Spatial Scale: Continental scale (African watersheds)
• Temporal Scale: Long-term (historical climate and hydrological data)

Methodology and Data

• Models used: Natural Gradient Boosting (NGBoost), NGBoost with conformal calibration, Quantile Regression extreme gradient boosting (QR XGBoost), Conformalized Quantile Regression Extreme Gradient Boosting (CQR XGBoost)
• Data sources: Gauged African stations (821) for model training and evaluation

Main Results

• The CQR XGBoost model emerged as the optimal predictor, combining accuracy (R2 = 0.885) with robust uncertainty quantification (85.7% coverage).
• The results show that the ML models significantly outperform traditional empirical power-law fit (R2 = 0.62).

Contributions

• This study provides an uncertainty-aware calibrated machine learning framework for continental design flood prediction.
• The work delivers practitioners a promising predictive framework and region-specific insights for infrastructure planning and engineering implementation in African unregulated or lightly regulated watersheds.

Funding

• No specific funding information is provided in the paper.

Citation

@article{Baida2026Calibrated,
  author = {Baida, Maelaynayn El and Chourak, Mimoun and Boushaba, Farid},
  title = {Calibrated uncertainty for continental design flood estimation using conformal probabilistic machine learning},
  journal = {Journal of Hydrology},
  year = {2026},
  doi = {10.1016/j.jhydrol.2026.135864},
  url = {https://doi.org/10.1016/j.jhydrol.2026.135864}
}