A Hybrid ERT-IRT and Deep-Learning Approach for Early Seepage Detection in Earthen Dams

Abstract

An integrated framework combining time-lapse electrical resistivity tomography (ERT), infrared thermography (IRT), and deep learning methods is proposed for early-stage seepage detection in earthen dams. Laboratory-scale earthen dam models are instrumented with soil moisture sensors (SMS), infrared imaging, and ERT arrays to provide a comprehensive assessment of seepage processes. Upon triggering SMS at different locations inside the earthen dams, thermal images are captured at the downstream face, where IRT observations reveal near-surface thermal anomalies before the onset of visible downstream seepage, highlighting its effectiveness as an early-stage indicator. Subsequently, time-lapse ERT measurements are recorded along the length of the embankment, capturing the progressive reduction in resistivity associated with advancing wetting fronts. To automate seepage detection, a custom-labeled data set of 1,224 paired infrared and optical images (75% laboratory, 25% field) obtained from various lab-scale models and field studies is used to develop dual-input deep learning models. While standard convolutional neural networks exhibited moderate performance, transfer learning architectures significantly improved detection accuracy. Among them, EfficientNetB0 achieved the highest accuracy rate (94%) with well-balanced F1-scores for both classes (0.93 for no seepage and 0.95 for seepage), outperforming the other methods. Overall, the hybrid IRT-ERT framework augmented with data-driven analysis offers a robust and scalable solution for real-time seepage monitoring and early detection, with promising applicability for unmanned aerial vehicle (UAV)-based dam surveillance and proactive safety assessment.