An automated framework for subsurface crack condition assessment in asphalt pavements

Abstract Subsurface crack assessment remains a critical yet underaddressed component of pavement condition evaluation. Existing ground penetrating radar (GPR)-based models predominantly focus on crack detection rather than comprehensive condition characterization and are often trained on constrained datasets, limiting their generalizability. To address these gaps, this study presents an automated framework for subsurface crack condition assessment integrating deep learning-based detection with fuzzy-logic reasoning. Fifteen deep learning architectures spanning single-stage, two-stage, transformer-based, and generative paradigms were trained and evaluated on two GPR datasets: the PolyU-GPR subsurface crack dataset (PGSC), collected through field surveys in Hong Kong, and the LTPP-GPR subsurface crack dataset (LGSC), sourced from 16 U.S. states. A subsurface crack condition index (SCI) was formulated using a Mamdani fuzzy inference system, incorporating crack count and depth to jointly characterize crack extent and severity. Cross-architectural evaluation revealed that dataset quality was a more critical determinant of model performance than architectural configuration. Models trained on the LGSC dataset achieved F1 scores and mean average precision 50 values of up to 85% and 89%, respectively, whereas all architectures exhibited substantially degraded performance when trained on the PGSC dataset. The SCI was computed for 188 road sections, with the majority classified as good and only 31 rated poor. Comparative analysis confirmed that the SCI captures structural information complementary to conventional surface distress metrics, establishing it as a complementary diagnostic tool for more effective pavement maintenance decision-making.