An ensemble deep learning and geospatial AI pipeline for automated roadway centerline and shoulder delineation

State Departments of Transportation (DOTs) frequently encounter significant data gaps in roadway characteristics inventory, particularly in budget-constrained counties, where vital infrastructure details such as the presence of shoulders are often uncataloged. This challenge is compounded by the common issue of geometrically inaccurate roadway centerline information, even in statewide open data repositories. To address these critical deficiencies, this study introduces a hybrid Deep Learning and geoprocessing framework for automated roadway centerline and shoulder extraction from aerial imagery. The proposed methodology employs an ensemble strategy that combines the outputs of two deep learning models trained in distinct geomorphological contexts (urban and rural) to maximize roadway corridor coverage. A separate deep learning model is then used to detect lane markings, enabling the extraction and topological linking of roadway centerlines within the detected corridors. The framework was evaluated through case studies in Walton and Union Counties of Florida. The ensemble approach improved roadway corridor coverage by approximately 2–3% compared to individual model outputs. Using this enhanced corridor representation, the framework generated roadway centerlines with completeness ranging from 86.5% to 88.9% and geometric accuracy exceeding 93% Intersection over Union (IoU) relative to ground truth data. In addition, the system successfully delineated roadway shoulders, achieving F1-scores between 77.1% and 93.7% across the two counties. Designed for implementation within the ArcGIS environment, this framework is inherently user-friendly and replicable, offering a scalable solution to fill critical roadway inventory data gaps beyond Walton and Union Counties and serving as a blueprint for other regions facing similar challenges.