Assessment of Deep Learning Models for Pavement Distress Detection in High-Resolution UAV Imagery
نویسندگان
1 Civil Engineering Department, Faculty of Engineering, Al-Azhar University, Cairo, Egypt.
2 Civil Engineering Department, Faculty of Engineering, Al-Azhar University, Cairo, Egypt
3 Civil Engineering Department, Faculty of Engineering, Al-Azhar University, Cairo, Egypt.
4 Civil Engineering Department, Faculty of Engineering, Al-Azhar University, Cairo, Egypt.
doi
10.22075/jrce.2025.2415چکیده
Pavement distress detection is a critical task for ensuring road safety and maintaining transportation infrastructure, particularly in environments with limited resources. This study conducts a comprehensive evaluation of ten YOLO object detection modelscomprising five YOLOv5 and five YOLOv8 variants (n, s, m, l, x) for identifying seven pavement defect classes using 500 high-resolution UAV images. The dataset, manually annotated and split into 70% training and 30% validation sets, was used to train all models under uniform hyperparameter settings. The performance was assessed using standard metrics: mAP@0.5:0.95, mAP@0.5, Recall, Precision and F1-score. Results showed that YOLOv8 consistently outperformed YOLOv5 across all dataset sizes, with YOLOv8l reaching the highest mAP@0.5:0.95 of 0.381, and YOLOv8m providing the best balance between accuracy (0.344), training time (67 minutes), and robustness. Statistical validation using Two-Way ANOVA confirmed significant performance differences (F = 13.81, p = 0.0006, Cohen’s d = 0.73). Further analysis using Repeated Measures ANOVA and Bonferroni-corrected t-tests reinforced YOLOv8l's superiority over other variants. Despite the limited dataset size the findings demonstrate YOLOv8’s effectiveness and reliability in low-resource conditions without the need for image augmentation or preprocessing. The main innovation lies in providing a statistically validated benchmark for UAV-based pavement monitoring under small dataset conditions, highlighting YOLOv8’s superior efficiency and generalization without augmentation or preprocessing. This study provides a reproducible benchmark for real-time UAV-based pavement distress detection, offering insights for deploying lightweight deep learning systems in resource-constrained settings.
