Deep Learning–Based Automated Cover Depth Estimation for Concrete Bridge Decks Using Ground-Penetrating Radar

Ensuring adequate concrete cover depth is crucial for the durability of reinforced concrete bridge decks. Ground-penetrating radar (GPR) has emerged as a prominent nondestructive testing method for evaluating reinforcement layout and concrete cover depth. Although GPR data collection is convenient, the data analysis and interpretation are intricate and demanding. This paper enhances the efficiency of GPR data analysis by introducing an automated method for cover depth estimation. The method incorporates deep learning (DL)–based rebar recognition, novel migration-based electromagnetic wave velocity estimation, and systematic travel-time calibration. To minimize human effort, a semi-automatic annotation tool was implemented to generate a training dataset containing 586,116 labeled rebar instances. The annotated data were used to train a YOLOv8 rebar detection model. Consequently, the maximum amplitude within the bounding boxes detected by the DL models served as a metric for migration-based wave velocity estimation. Simultaneously, travel time was calibrated by identifying the time-zero reference and correcting for the effects of antenna geometry. The developed module achieved 0.99 precision and 0.97 recall for rebar detection and a mean absolute error of 3% for cover depth estimation. The advanced method enables convenient evaluation of the adequacy of concrete cover depth. This allows agencies to implement maintenance activities to address nonconformities and ensure the structural integrity and durability of concrete bridge decks.