Pan et al. (2026) Regional flood frequency analysis using generalized additive models, random forest, and extreme gradient boosting for South-East Australia

Identification

• Journal: Environmental Earth Sciences
• Year: 2026
• Date: 2026-01-13
• Authors: Xiao Pan, Gokhan Yildirim, Ataur Rahman, Taha B.M.J. Ouarda
• DOI: 10.1007/s12665-025-12800-5

Research Groups

• School of Engineering, Design and Built Environment, Western Sydney University, Penrith, NSW, Australia
• Department of Civil Engineering, Faculty of Engineering, Aksaray University, Aksaray, Turkey
• Canada Research Chair in Statistical Hydro-Climatology, Institut national de la recherche scientifique INRS-ETE, Quebec, Canada

Short Summary

This study develops a new regional flood frequency analysis (RFFA) model using Generalized Additive Models (GAM), Random Forest (RF), and XGBoost (XG) within the Peaks Over Threshold (POT) framework for southeastern Australia. GAM is found to be superior in accuracy, with a median absolute relative error of 33%, significantly enhancing flood quantile estimation compared to RF (37%) and XG (40%).

Objective

• To develop and compare the performance of Generalized Additive Models (GAM), Random Forest (RF), and XGBoost (XG) within a Peaks Over Threshold (POT) modelling framework for regional flood frequency analysis (RFFA) in southeastern Australia.
• To determine if these advanced machine learning and non-linear regression models can outperform traditional linear regression methods in accurately and robustly estimating flood quantiles across various return periods, particularly in hydrologically complex regions.

Study Configuration

• Spatial Scale: 145 gauged catchments in southeastern Australia (55 in New South Wales, 90 in Victoria). Catchment areas range from 11 km² to 1010 km², with an average of 360 km².
• Temporal Scale: Streamflow data records span 27 to 83 years, with an average of 42 years. Flood quantiles were estimated for nine return periods, from very frequent (12 events per year, EY) to frequent (10-year average recurrence interval, ARI).

Methodology and Data

• Models used: Generalized Additive Models (GAM), Random Forest (RF), and XGBoost (XG). All models were integrated within the Peaks Over Threshold (POT) modelling framework, specifically using POT3 flood series (averaging 3 events per year) fitted with a Generalized Pareto Distribution (GPD). Model validation was performed using Leave-One-Out Cross-Validation (LOOCV).
• Data sources: Streamflow records and seven physio-meteorological catchment descriptors (catchment area, mean annual rainfall, catchment shape factor, mean annual evapotranspiration, stream density, mainstream slope, and forest fraction) were obtained from state hydrological authorities.

Main Results

• Generalized Additive Models (GAM) consistently outperformed Random Forest (RF) and XGBoost (XG) in predictive accuracy for flood quantile estimation.
• The median absolute relative error for GAM was 33%, compared to 37% for RF and 40% for XG.
• GAM exhibited the best alignment with observations, with ordinary-least-squares slope (β₁) values close to 0.65 and coefficients of determination (R²) around 0.59 across all return periods. RF and XGBoost showed slightly lower slopes (≈ 0.45–0.55) and R² values (≈ 0.51–0.52).
• GAM demonstrated the narrowest interquartile range (IQR) for relative error and absolute relative error, indicating more consistent and precise predictions.
• Spatial analysis revealed GAM's robustness, particularly in inland areas with drier climates and regions with high stream densities, where it reduced relative error margins by up to 20% compared to other models.
• RF and XG models showed a greater tendency to overestimate flood quantiles, especially in catchments with high stream densities and in specific geographical regions (e.g., upper New South Wales and along the New South Wales-Victoria border).
• Poorly performing catchments (where all models struggled) were characterized by smaller areas, higher stream densities, lower forest coverage, lower mainstream density, and lower baseflow peak factors.

Contributions

• This study provides the first comprehensive comparison of Generalized Additive Models (GAM), Random Forest (RF), and XGBoost (XG) within a Peaks Over Threshold (POT) framework across nine return periods (12EY–10ARI) for southeastern Australia.
• It demonstrates that integrating advanced machine learning methods within the POT framework significantly enhances the accuracy of flood quantile estimation, particularly with GAM.
• The research offers actionable guidance for adopting POT-RFFA in practice by highlighting GAM's systematic generalization advantage over RF and XG, especially in high stream-density terrain.
• The findings are expected to assist in upgrading the Australian Rainfall and Runoff (ARR) – the national guideline for flood estimation.
• Unlike prior POT-RFFA studies that focused on linear or regularized linear models, this work employs a unified Leave-One-Out Cross-Validation (LOOCV) protocol, common error metrics, and spatial diagnostics for non-linear learners.

Funding

No funding was received for this study.

Citation

@article{Pan2026Regional,
  author = {Pan, Xiao and Yildirim, Gokhan and Rahman, Ataur and Ouarda, Taha B.M.J.},
  title = {Regional flood frequency analysis using generalized additive models, random forest, and extreme gradient boosting for South-East Australia},
  journal = {Environmental Earth Sciences},
  year = {2026},
  doi = {10.1007/s12665-025-12800-5},
  url = {https://doi.org/10.1007/s12665-025-12800-5}
}