Automatic Identification of Pavement Subsurface Distresses from Ground Penetrating Radar Data Using Few-Shot Data Augmentation and YOLOv11

Accurate identification of pavement distress using Ground Penetrating Radar (GPR) data is essential for non-destructive maintenance and safety assessment of roadway structures. However, the scarcity of annotated GPR datasets limits the performance of deep learning models. This study presents a method for automated pavement distress detection using few-shot data augmentation and YOLOv11 (You Only Look Once, version 11). A generative framework based on an improved ExSinGAN (Explainable Single-Image Generative Adversarial Network) is used to generate realistic GPR images from a small number of original samples. The generated data are combined with the Urban Roadway Disease Dataset (URDD) to train the YOLOv11 detector to identify subsurface distresses, such as cavities and looseness. Comparative experiments with gprMax-simulated data and field validation confirm the reliability and adaptability of the proposed approach. The results show an increase in mean average precision (mAP) from 84.69% to 89.31%, precision from 77.73% to 83.13%, and recall from 81.30% to 83.37% compared to the baseline trained on URDD alone. These findings demonstrate that few-shot generative augmentation effectively mitigates data scarcity and improves the robustness of GPR-based distress detection. The developed model was further validated on field data, confirming its generalization to real-world GPR measurements. The proposed method supports non-destructive monitoring of pavement structures. It provides practical value to civil engineers by facilitating early detection of underground defects and reducing the risk of pavement collapse from sinkholes.