Deep learning-based pavement subsurface distress detection via ground penetrating radar data

Pavement subsurface distress endangers driving safety and road serviceability. Ground penetrating radar (GPR) can non-destructively provides high-resolution profiles of road. However, the automatic interpretation of radar signals remains challenging. This study proposed an automatic pavement subsurface distress detection method using traditional signal processing and deep learning. Firstly, a piecewise linear function for radar signal automatic gain was proposed. Wavelet transform was applied to identify road layers for function segmentation. Each signal's power spectral density was calculated to determine the gain function coefficient. A specific pseudo-color mapping method was designed to convert reflected signal for deep learning model training. A radar reflection simulation was built to pre-train the model to enhance model performance. More than 270 km of field test data were collected for model training and validation. The proposed model has shown a 72.39% accuracy in shallow cracks and 68.74% in subsurface voids.