As an emerging technology, digital twin (DT) is increasingly valued in bridge management for its potential to optimize asset operation and maintenance (O&M). However, traditional bridge management systems (BMS) and existing DT applications typically rely on standalone building information modeling (BIM) or geographic information system (GIS) platforms, with limited integration between BIM and GIS or consideration for their underlying graph structures. This study addresses these limitations by developing an integrated DT system that combines WebGIS, WebBIM, and graph algorithms within a three-layer architecture. The system design includes a common data environment (CDE) to address cross-platform compatibility, enabling real-time monitoring, drone-enabled inspection, maintenance planning, traffic diversion, and logistics optimization. Additionally, it features an adaptive data structure incorporating JSON-based bridge defect information modeling and triple-based roadmap graphs to streamline data management and decision-making. This comprehensive approach demonstrates the potential of DTs to enhance bridge O&M efficiency, safety, and decision-making. Future research will focus on further improving cross-platform interoperability to expand DT applications in infrastructure management.
Automating disaster severity assessment from social media imagery is critical yet challenging due to the scarcity of labeled data and the ambiguity of damage levels. While Large Multimodal Models (LMMs) show promise, they often struggle with the fine-grained, ordinal nature of severity estimation. To address these limitations, we propose a Logic-Guided Ordinal Distillation framework. Unlike traditional semi-supervised learning where teachers are randomly initialized, we introduce a Logic-Guided Teacher initialized with linguistic priors from CLIP's text encoder. This effectively mitigates confirmation bias, enabling the generation of reliable pseudo-labels for large-scale unlabeled "silver data''. Furthermore, we design an Ordinal-Aware Student trained with a geometric ordinal loss to enforce hierarchical severity constraints (Mild → Measured → Severe) and a weighted sampling strategy to handle long-tailed class distributions. Extensive experiments on the CrisisMMD benchmark demonstrate that our method achieves an average F1 score of 70.61%, significantly outperforming strong supervised baselines and state-of-the-art zero-shot models, particularly in identifying critical long-tail categories. Cross-dataset evaluation on MEDIC and ablations against modern semi-supervised learning (SSL) baselines (FixMatch, FlexMatch, SoftMatch, Mutual-KD) further confirm the generality of the framework.