Probabilistic assessment of warning-stage exceedance during ice-jam flooding using physics-guided river-ice simulations and interpretable stacking ensembles

Study region The Sanhuhekou Bend Reach of the Yellow River, northern China. Study focus Ice-jam flooding produces abrupt stage rises, yet probabilistic flood-risk assessment remains challenging because of nonlinear ice–hydraulic interactions and sparse observations. We develop a physics-guided framework that couples Monte Carlo river-ice simulations with interpretable stacking-ensemble surrogates for rapid assessment of breakup ice-jam stage and warning-stage exceedance risk. The process-based model RIVICE is embedded within a Monte Carlo framework to generate 10,000 physics-consistent scenarios spanning uncertainty in boundary conditions and key ice–hydraulic parameters. An interpretable stacking surrogate efficiently emulates the calibrated physics-based simulations, while SHapley Additive exPlanations (SHAP) quantify predictor contributions and guide reduced-input model distillation. New hydrological insights for the region. Jam-toe location, upstream discharge, and downstream ice thickness were identified as the dominant controls on ice-jam stage because they govern backwater propagation, hydraulic forcing, and ice-induced flow resistance. SHAP further identified model-informed threshold ranges associated with elevated flood risk. The stacking surrogate reproduced simulated stages with high fidelity (NSE = 0.95, RMSE = 0.14 m, R = 0.97), while the distilled surrogate retained strong performance (NSE = 0.92, RMSE = 0.18 m, R = 0.96) using fewer predictors. Surrogate-based ensembles reproduced the observed annual-maximum stage distribution and return-period behaviour, enabling rapid probabilistic assessment of warning-stage exceedance risk and supporting flood-risk management under data-scarce conditions.