Machine Learning-Based Probabilistic Modeling of Levee Breach Initiation Under Overtopping

A reliable assessment of levee performance under extreme hydraulic loading is crucial for effective flood risk management. However, reliable prediction of the probability of breach initiation during overtopping remains a significant challenge due to the complex interactions among hydraulic, geometric, and geotechnical factors, as well as the scarcity of high-quality field data. This study addresses this challenge by developing two complementary, data-driven models to estimate the probability of riverine levee breach initiation under overtopping conditions. Using 487 documented overtopping events from the U.S. Army Corps of Engineers’ Levee Loading and Incident Database-Overtopping subset (LLID-OT), this work applies logistic regression (OTI-LR) and random forest (OTI-RF) approaches to identify key parameters controlling breach initiation, including overtopping depth, duration, hydraulic loading time, and erosion resistance. Feature engineering and model development were guided by physical relevance, statistical significance, and validation testing. Results show that the models achieve average accuracies on unseen data of 90% and 95%, respectively, with AUROC values above 0.90 and AUPRC values above 0.98, demonstrating strong discriminatory capability. Both models identify overtopping depth and material erosion resistance as dominant predictors of breach onset. The findings highlight the potential of data-driven modeling frameworks to enhance levee reliability analysis and risk assessment.