Integrating electrical resistivity tomography–derived metrics with ensemble machine learning to predict the depth-to-water in earthen levees

For earthen levees, the dominant failure mechanism is seepage-induced internal erosion. Conventional levee monitoring uses piezometers, which require drilling invasive boreholes yet provide only sparse spatial coverage. In this study, an ensemble machine-learning framework was evaluated for estimating profile-scale depth-to-water (DTW) from three-dimensional electrical resistivity tomography (3D ERT)-derived seepage metrics during a controlled levee-flood experiment along the Körös River, Hungary. Five machine-learning algorithms were compared using leave-one-out cross-validation (LOO-CV) for profile-wise evaluation, complemented by repeated fivefold cross-validation to assess robustness. Feature selection retained three ERT-derived seepage metrics for predicting DTW: leakage distance, time to initial detection, and leakage persistence index. Among the five algorithms, the gradient boosting machine achieved the lowest LOO-CV prediction error, with R 2 = 0.983 and a mean absolute percentage error (MAPE) of 1.47%. Augmenting the ERT-derived feature set with the sparsely sampled borehole-derived parameters available in this experiment did not improve predictive performance (MAPE = 1.92%). Shapley additive explanations showed near-equal predictive contributions from all three metrics and revealed consistent model-specific associations across the monitored seepage gradient. These findings indicate that ERT-derived seepage metrics can support profile-scale DTW estimation in fine-grained levees and may reduce dependence on dense piezometric monitoring networks.