Electromagnetic wave echo characteristic atlas and automatic recognition of internal distresses in asphalt pavement combined with 3D-GPR and deep learning

Concealed distresses in asphalt pavement structures, such as internal cracks and debonding (including loosening and voids), are commonly detected using ground penetrating radar (GPR). However, recognition accuracy is often compromised by similar echo characteristics of various distresses and structural objects (pipelines, rebars and patches). To address these challenges, the optimal acquisition parameter combination for three-dimensional GPR (3D-GPR) inspection of expressway pavements is identified, a reference atlas of 3D-GPR echo characteristics for typical internal distresses is constructed, and a deep learning-based approach is introduced to automatically recognise internal distresses from pavement GPR images. A dataset comprising 2,983 internal crack instances and 1,274 debonding instances is established using both real and simulated GPR images, with various data augmentation techniques applied. Four YOLO-series model families, YOLOv5u, YOLOv8, YOLO11 and YOLO26, each with five sub-variants (n, s, m, l, x), totalling 20 models, are trained, validated and tested on this dataset. Among all evaluated models, YOLOv8s achieves the best overall performance, with a recognition accuracy (mAP@0.5) of 98.7% and a processing speed of 111.1 frames per second, outperforming the other YOLO variants. These results demonstrate the effectiveness of the proposed methodology in improving both the accuracy and efficiency of internal distress detection in pavements.