Mitigating embankment failure in haor wetlands: an explainable machine learning framework for predicting tensile hazard vulnerability

Abstract Tensile failure in earth embankments is a common phenomenon in flash flood-prone low-lying basins. The structural integrity of these embankments is highly vulnerable to prolonged flood, saturation, and seasonal desiccation, yet no operational framework exists to predict tensile hazard-prone zones. This study addresses the gap by developing an explainable machine learning (ML) based predictive methodology that converts laboratory direct tensile strength (DTS) into a tensile hazard index classification system. For this purpose, a series of experimental tests was conducted on 390 soil samples collected from vulnerable embankments of Sunamganj, Bangladesh. To capture the diverse nature of embankment soil, nine input features, including index and strength properties, were utilized. This study incorporates seven ML methods, employing Bayesian optimization to fine-tune hyperparameters. Among those models, XGBoost demonstrated the most robust predictive accuracy (R 2 = 0.9013) and the lowest deterioration rates across all evaluation metrics (2.01% for R 2 , 11.23% for RMSE, and 2.13% for MAE). The novelty of this study lies in its predictive-to-decision translation layer when DTS values were mapped onto the proposed hazard indexing. With 82.05% exact-category and 100% within-one-category agreement, the XGBoost model ensures reliable and engineer-safe prediction. Furthermore, Shapley additive explanations based on global and local interpretation identified unconfined compressive strength, moisture content, plastic limit, and dry density as the dominant predictors. Therefore, by establishing tensile strength-based hazard classifications and linking to the large-scale tensile behavior of embankment soils, this research offers a replicable tool for mitigating embankment failure-related disasters, particularly in climate-susceptible haor regions.