The increasing frequency of extreme climate events poses significant risks to slope infrastructure, while traditional inspection methods are often inefficient and unsafe. Although unmanned aerial vehicles (UAVs) combined with structure-from-motion (SfM) provide high-fidelity 3D models, they lack the semantic understanding necessary for automated damage assessment. This study addresses this gap by developing and validating an anomaly-enhanced digital twin (AEDT) framework. The proposed system integrates multiview UAV imagery, SfM-based 3D reconstruction, and a convolutional neural network (CNN) for automated anomaly classification. This information is then fused into an interactive, geographic information system (GIS)-compatible DT platform for lifecycle management. A case study on a soil and water conservation (SWC) structure in central Taiwan was conducted for verification. The deep learning module achieved a macroaverage F1-score of 0.81, demonstrating balanced performance across erosion, spalling, siltation, and collapse classes. This was validated on a held-out test set derived from a total of 2000 annotated images spanning four anomaly types with three severity levels. Furthermore, the AEDT-based workflow reduced onsite inspection time by ~63% compared to conventional manual methods. The resulting AEDT model provides a dynamic, semantically enriched 3D representation of the infrastructure, linking geometric data with damage attributes and historical maintenance records. This research demonstrates a feasible and scalable solution for intelligent infrastructure monitoring, offering a robust tool for enhancing climate resilience and enabling proactive asset management.
Loading....