Uncertainty‐Aware Machine Learning Models for Flash Flood Prediction

ABSTRACT Flash floods have intensified in recent years, and machine learning (ML) models are increasingly used for real‐time prediction. However, most ML‐based flood‐forecasting studies remain largely deterministic and provide limited evidence on how input uncertainty propagates to forecasts, compromising reliability for operational early‐warning systems. This study addresses this gap through four main contributions: (i) a public 5‐year high‐resolution flood‐hazard dataset for benchmarking uncertainty‐aware ML flood forecasting, (ii) an uncertainty‐analysis approach based on synthetic noise injection to evaluate input uncertainty propagation, (iii) an uncertainty‐aware ensemble providing probabilistic predictions, and (iv) three new metrics to quantify robustness and uncertainty propagation. Five ML methods (LSTM, MLP, XGBoost, LightGBM, NuSVR) were evaluated for 12 lead times with and without synthetic noise. Multi‐criteria assessment revealed no single algorithm consistently outperformed others in predictive performance, computational cost, robustness, and uncertainty quantification. All data and code are publicly available to support reproducible research.