Enhancing Road Maintenance Efficiency through Comparative Evaluation of Pothole Detection Models

This study evaluates the effectiveness of various pothole detection models, comparing YOLOv8m (default), YOLOv8m with parameter tuning, and YOLOv11m. The aim is to enhance road maintenance efficiency by determining which model performs best under different scenarios. The dataset, comprising 2,371 images of potholes, was divided into training, validation, and test sets (70%, 20%, and 10%, respectively). Each model's performance was assessed using metrics such as precision, recall, mean Average Precision (mAP), accuracy, and Fl-Score. The default YOLOv8m model achieved the highest precision (71%) and accuracy (44%), outperforming the adjusted YOLOv8m, which scored (59.2%) precision and (42.7%) accuracy. Meanwhile, YOLOvllm displayed a slight improvement in mAP50-95 (0.253) compared to YOLOv8m's (0.252). Scenario 1 (default YOLOv8m) demonstrated superior detection accuracy, indicating its suitability for pothole detection. Despite YOLOvllm's enhanced architecture, it only marginally improved recall rates, suggesting that default configurations may be optimal. This research concludes that default YOLOv8m is the most reliable for real-time pothole detection, balancing accuracy and efficiency, which is critical for road safety and infrastructure maintenance efforts. Further refinements, such as dataset expansion and advanced data augmentation, are recommended to improve detection accuracy across diverse environments.