Orthomosaic and Deep Learning-Based Framework for Longitudinal and Transverse Cracking Detection and Quantification in Airfield Asphalt Concrete Pavements

The integrity of airfield pavements is critical to airport safety and operational efficiency. Asphalt concrete (AC) pavements are prone to longitudinal and transverse cracking (L&T) due to thermal loading, traffic, and environmental effects; however, current inspection practices rely on manual surveys that are labor-intensive, subjective, and limited to geospatial sampling. Moreover, most prior studies rely on single-site datasets and image-level metrics, with limited validation against pavement inspection records. To address these limitations, this study proposes an integrated framework that combines small uncrewed aerial system (sUAS) imaging, deep learning (DL), and geographic information system (GIS) processing for automated identification and quantification of L&T cracking. sUAS-based RGB imagery from Custer Airport (TTF), Michigan, was used to train and evaluate three segmentation models: U-Net, SegFormer-B1, and YOLOv11m. SegFormer-B1 achieved the lowest mean absolute percentage error (11.50%) in crack-length estimation, while YOLOv11m attained the highest Dice score (80.48%) and mIOU (72.64%). The models were further evaluated on an independent dataset from Forest City Municipal Airport (FXY), Iowa, where YOLOv11m demonstrated the strongest agreement with reported cracking, yielding normalized errors between 0.48 and 1.72 ft per 100 sq. ft for five sample areas. This analysis evaluated the robustness of the model to varying surface textures and imaging sensors. Overall, the framework bridges pixel-level segmentation and engineering-scale crack evaluation, enabling scalable and efficient assessment of pavement condition.