From manual to UAV-based inspection: Efficient detection of levee seepage hazards driven by thermal infrared image and deep learning

Abstract

Levee failures are caused mainly by river water seepage erosion inside the levee, manifesting as slope leakage and foundation piping phenomena. To address the urgent need for levee safety monitoring during flood seasons and extreme rainfall events, unmanned aerial vehicles (UAVs) equipped with thermal infrared imagers can quickly detect leakage and piping hazards based on temperature differences between damaged areas and their surroundings. In this study, we collected 5995 UAV thermal infrared images of leakage and piping on levees in four floodplains under various weather, time, and surface coverage conditions to evaluate the applicability of combining thermal infrared imaging with a deep learning model in complex natural environments. We categorized levee hazards into water piping, ground piping, and slope leakage, and developed a Mask R-CNN segmentation model. The results revealed that thermal infrared levee inspection was affected by vegetation occlusion and subtle temperature differences caused by continuous rainstorms and water body depth. The Mask R-CNN demonstrated strong generalizability to hazards with significant variability in shape, size, and temperature difference, making it suitable for detecting evolving and expanding damaged area. The mean average precision, recall, and precision of the Mask R-CNN were 0.977, 0.982, and 0.897, respectively, and the detection time was 0.015 s per image. Moreover, Eigenvector-based class activation mapping (Eigen-CAM) was used to visualize the decision basis and failure modes of the Mask R-CNN to improve interpretability. Seepage damage is a progressive process, so timely hazard identification can provide valuable time for levee repair.