Regression-based regional flood estimation using GAM and MARS within a peaks-over-threshold framework

This study applies Multivariate Adaptive Regression Splines (MARS) and Generalized Additive Models (GAM) within a Peaks Over Threshold (POT) framework for regional flood frequency analysis (RFFA) in southeastern Australia using data from 145 catchments. Seven physiographic and meteorological variables were used as predictors, and flood quantiles were estimated via the Generalized Pareto distribution. Model performance was assessed using Mean Absolute Error (MAE) and Mean Absolute Percentage Error (MAPE) via Leave-One-Out Cross-Validation (LOOCV). Our results indicate that GAM consistently outperforms MARS, exhibiting lower error variability, narrower residual distributions, and superior generalizability, particularly in data-sparse inland catchments. GAM’s smooth function framework effectively captures non-linear hydrological relationships, while MARS shows greater prediction biases and higher variability. Across all return periods, the GAM model achieved median relative errors generally within ±10–15% and prediction ratios concentrated around unity, indicating unbiased and stable estimates.