Analysis of False Detection Causes in Pothole Detection Models using Explainable AI

Potholes on the road can cause sudden braking, resulting in serious accidents and damage to the tire or vehicle's underbody. While deep learning-based pothole detection technologies have been developed and achieved high accuracy, significant barriers to commercialization remain: the “false positive” problem, where tire skid marks, road patches, or manholes are misidentified as potholes, and the “false negative” problem, involving the failure to detect small potholes. This study uses explainable artificial intelligence (XAI) techniques, specifically Grad-CAM and Eigen-CAM, applied to the YOLOv8s model to visually and quantitatively investigate the causes of these detection errors. In the experiment, the model inference results were classified using a public dataset of 652 images, and heat maps were generated to analyze the model's focus areas. Regarding false positives, skidmarks were the leading cause, accounting for 35.3% (47 133 false positives), whereas 87.3% (317 363 false negatives) were attributed to small potholes. This study confirmed that the deep learning model relied on low-level visual features rather than semantic properties of potholes, identifying this as the fundamental cause of various misdetection types.