Automatic identification and location of underground defects in urban roads via ground penetrating radar and deep learning approaches

Abstract

Underground defects in urban roads endanger driving safety and hinder road usability. These defects are primarily identified using technologies such as ground penetrating radar. The current intelligent algorithms used for identifying underground road defects rely heavily on large datasets of on-site road images. However, the automatic detection of defects remains challenging due to small datasets, limited image availability, and inconsistent on-field image quality. This paper proposes a novel approach to address these challenges through a model based on actual road conditions and forward simulations of road defect images. To improve the quality of both real and simulated field images, we apply a joint denoising method that combines wavelet transform, the K-SVD algorithm, and bilateral filtering. This denoising process enhances both real and simulated field images and expands the image dataset, transforming it into a mixed database, and strengthens the distinctive features of each defect, facilitating more accurate algorithm-based detection. In the first and second stages of the study, we conduct a comparative analysis of various deep learning-based object detection models. We then propose a deep learning model, optimized with the joint denoising model, that is best suited for practical road evaluation projects. The model was trained and validated across 100 km of high-quality field measurement data collected from various districts and counties in Beijing. Experimental results showed that the model can achieve a prediction accuracy of 82.3% for Looseness, 92.6% for Cavities, and 50.9% for Voids, with an overall Mean Average Precision of 75.3%. These results demonstrate that the method proposed in this study can enhance the detection accuracy for various subsurface defects.