GPR-based automatic volume estimation of pavement base-layer loosening via a theory–data-driven framework

Loosened base-layer structures in asphalt pavements reduce load-bearing capacity and service life, but accurate estimation of their size remains difficult. This study proposes an automatic area and volume calculation framework based on ground-penetrating radar (GPR) imaging, combining theoretical analysis with deep learning. Considering the reduced dielectric constant of loosened regions, formulas for vertical area and volume were derived and validated using irregular models generated by finite-difference time-domain (FDTD) simulation. The average errors were 4.14% for area estimation and 5.20% for volume estimation. A lightweight instance segmentation model, LSD-YOLOv11n-seg, was then developed to detect loosened regions in GPR B-scan images. Compared with YOLOv11n-seg, it improved precision, recall, AP50M, APM, and F1-score by 4.1%, 7.0%, 4.4%, 1.2%, and 5.6%, respectively, across two datasets. For automatic estimation, the minimum relative errors were 1.04% for area and 1.10% for volume, while the average errors were 4.71% and 7.28%. The method provides effective two- and three-dimensional quantification of loosened regions, supporting intelligent pavement assessment and maintenance planning.