Objectives: This study developed and field-validated a camera-based framework for detecting potholes on unpaved roads using a YOLO11 object detection model, with a pavement condition index (PCI) survey as the engineering ground truth. Methods/Analysis: A YOLO11s detector was trained on 250 smartphone images from unpaved roads separate from the study sections, using only realistic augmentations. Its confidence threshold (0.21) was fixed on a 50-image validation subset, and the model was then tested without retraining on an independent field set of 236 images, one per PCI-inventoried pothole, from four unpaved road sections in Irbid, Jordan. Detections were matched to the ground truth at an Intersection over Union (IoU) ≥ 0.50 and compared with a YOLOv8s baseline. Findings: The model detected 194 of 236 potholes, with 26 false positives and 42 missed detections, giving 88.2% precision (95% confidence interval, CI: 83.2–91.8%), 82.2% recall (95% CI: 76.8–86.6%), and an 85.1% F1-score (95% CI: 81.4–88.5%), outperforming YOLOv8s (79.9%). Recall rose from 48.5% for low-severity potholes to 92.3% and 97.8% for medium- and high-severity potholes. Excluding the missed low-severity potholes raised section PCI by 0.1–2.2 points. Novelty/Improvement: The study provides an independently field-tested, low-cost smartphone approach that links object detection to PCI ground truth and quantifies how pothole severity limits camera-based detection on unpaved roads.
Copyrights © 2026