Automatic Classification of Pavement Distress Using 3D Ground-Penetrating Radar and Deep Convolutional Neural Network.

The realization of nondestructive detection and classification of pavement distress is of great significance for putting forward a reasonable maintenance scheme and prolonging the service life of the road. In this paper, the advanced air-coupled 3D ground-penetrating radar (3D-GPR) was used to detect Li-Ma Expressway (Jiangsu Province, China) to achieve the purpose of rapid, accurate, and nondestructive testing. Through on-site coring at the corresponding radar anomaly signal, it is verified that the 3D-GPR can identify the characteristics of pavement distress. To address the issue of a large amount of 3D-GPR images required to be processed and low efficiency of current manual classification, VGG16 and ResNet50 models were established based on deep convolutional neural network to classify four pavement distresses automatically. The results showed that the maximum accuracy, precision, and recall obtained from the VGG16 model training results were 98.73%, 97.98%, and 96.17%, respectively. In addition, The overall processing time of VGG16 to identify 2525 radar images with $200 \times 200$ pixels is 232 s. And VGG16 was selected as the automatic classification model of 3D-GPR image for pavement distress of Li-Ma Expressway. [ABSTRACT FROM AUTHOR]