Automated Identification of Internal Defects in Asphalt Pavements Based on Ground Penetrating Radar B-Scan

Abstract Ground-penetrating radar (GPR), as an effective and rapid nondestructive testing technique, has been widely applied in pavement distress detection. However, large-scale GPR data still rely heavily on manual interpretation, which is inefficient and prone to significant variability. To address this issue, a B-scan-based GPR method for detecting subsurface distresses in asphalt pavements is proposed in this study. First, the radar signature characteristics of subsurface distresses are determined through field coring, and a distress dataset is constructed. Second, a suitable detection model is selected and optimized to enhance performance. Finally, the model performance is validated using field project data.The results show that when subsurface distresses occur, the in-phase axis of the radargram exhibits high-amplitude, smooth, and discontinuous features, which can be effectively used for distress identification. Among ten YOLOv8 and YOLOv5 models compared, the YOLOv5s model achieves a good balance between detection accuracy and training speed. By replacing the backbone network with MobileNetV3, the model parameters are reduced by 48.8%, and the training time is shortened by 43.1%. Based on cross-validation with manually labeled test sections, the improved YOLOv5s-MobileNetV3 model achieves an accuracy of 80.0% for structural looseness and 87.1% for interlayer defects, demonstrating its effectiveness in practical applications.