Infrastructure management along highways and railways requires inventories of critical structures like retaining walls, which traditionally rely on manual inspection and documentation. Unfortunately, data in infrastructure databases is often incomplete. This study investigates the feasibility of automating retaining wall inventories using public aerial lidar data from the Swiss Federal Office of Topography (Swisstopo) combined with deep learning. We develop a pipeline for data processing and apply the state-of-the-art Superpoint Transformer architecture with SuperCluster for panoptic segmentation. Three distinct approaches are evaluated: transfer learning from general Swisstopo lidar data, transfer learning from the DALES aerial lidar dataset, and training a specialized model from scratch. The specialized model achieves the best performance with 44% Intersection over Union (IoU) for semantic segmentation and 24% panoptic quality on test data. Our findings reveal that the primary challenges stem from data characteristics-like sparse sampling of vertical surfaces due to oblique scanning anglesrather than model architecture limitations. This work provides insights into the development of automated infrastructure inventories and identifies areas for improvement, including the need for expanded training data and robust augmentation techniques.
Gasparini et al. (Mon,) studied this question.