Low-Cost Pavement Distress Detection and Mapping Using Smartphone Imagery and Data-Efficient DDPM Augmentation with YOLOv8

Abstract

This study presents a cost-effective, scalable, and automated framework for pavement distress detection using widely available smartphone imagery and advanced deep learning architectures. GPS-enabled smartphone imagery was collected across major urban highways, producing a spatial-visual dataset for pavement condition assessment. To address class imbalance and the limited availability of alligator crack samples, a Denoising Diffusion Probabilistic Model (DDPM) was trained using only 111 real alligator crack images to generate synthetic training data. A key finding of this study is that rotation-based augmentation (90°, 180°, and 270°) significantly improved the quality of DDPM-generated alligator crack images, despite the limited number of training samples. Comparative experiments further demonstrated that rotation-based augmentation was more effective than simply increasing the amount of training data for low-resource DDPM training, producing clearer and more realistic crack images at earlier training stages while achieving faster convergence. The augmented dataset was subsequently used to evaluate multiple YOLOv8 variants for pavement distress detection. Among the tested models, YOLOv8m with an input resolution of 640×640 pixels achieved the best overall performance. For linear crack detection, the model achieved a precision of 83.5%, recall of 85.9%, F1-score of 84.6%, and mean Average Precision (mAP) of 86.7%. For alligator crack detection, it achieved a precision of 95.8%, recall of 88.6%, F1-score of 92.1%, and mAP of 92.9%. By integrating deep-learning predictions with GPS coordinates, a geospatial mapping framework was developed to accurately localize pavement distresses and support network-level condition monitoring. The proposed approach provides an accessible and practical solution for road agencies seeking efficient and scalable pavement management practices.

Keywords: Pavement distress, DDPM, YOLOv8, Data augmentation, Geospatial mapping