Automated Spatial Mapping and Localised 3D Dimensional Extraction of Pavement Distress Using Deep Learning and GCP-Free UAV Photogrammetry

Abstract

Traditional pavement monitoring is labour-intensive, and while automated 2D deep learning approaches improve detection speed, they lack the critical depth information needed for structural severity assessment. Conversely, high-fidelity 3D photogrammetry typically demands labour-intensive Ground Control Points (GCPs). To bridge this gap, this study proposes a lightweight, GCP-free framework integrating UAV-RTK photogrammetry and the YOLOv11m algorithm to automatically detect and geometrically quantify 3D pavement distress. Using the six primary distress categories (ASTM D6433), the trained model was deployed directly into a geographic information system (GIS) environment via the QGIS Deepness plugin. To mitigate the inherent absolute elevation shifts in GCP-free mapping and global road gradients, a novel localised 3D depth extraction algorithm was introduced. This approach dynamically calculates the relative difference between a defect's localised Z-minimum and its surrounding reference plane on the Digital Surface Model (DSM). Validation in two study areas demonstrated overall detection accuracies of 81.74% and 88.79%. Planimetrically, the model achieved a Root Mean Square Error (RMSE) of 0.13 m for potholes and 0.14 m for alligator cracking. Crucially, the localised algorithm extracted vertical distress dimensions with centimetre-level precision, yielding exceptional depth RMSEs of 0.021 m (potholes) and 0.019 m (rutting). These findings demonstrate that integrating GCP-free spatial data with localised AI extraction yields a highly scalable, ready-to-use solution for modern Pavement Management Systems (PMS), while highlighting the need for instance segmentation to address bounding-box limitations for elongated deformations.