Multimodal, uncertainty-aware structural damage assessment for post-disaster Urban Search and Rescue (USAR) decision-making

Rapid structural damage assessment is critical for life-saving decision-making in the first hours following sudden-onset disasters, yet operational Urban Search and Rescue (USAR) teams must act under severe constraints: limited ground truth, disrupted connectivity, evolving situational awareness, and the need to justify prioritisation decisions in real time. In parallel, the remote sensing community has been a key part to providing initial information for early decisions. There is a rapidly expanding ecosystem of damage-mapping methods, including deep learning approaches and foundation models providing new opportunities. Their operational suitability for humanitarian response in terms of speed, uncertainty communication, and incremental updating still needs assessment and development for many of these methods.

We present an operationally driven evaluation and system design for post-disaster structural damage assessment using multimodal information streams. The study leverages building-level damage assessment datasets collected across multiple disasters and contexts, including the Beirut explosion (2020), Haiti earthquake (2021), Türkiye-Syria earthquake (2023), and the Myanmar-Thailand earthquake (2025). We compare and integrate methods spanning classical change detection, learning-based approaches, and multimodal fusion, with a focus on workflows that can ingest heterogeneous evidence (optical imagery, SAR products, and in-situ observations) and update outputs as new information becomes available during response.

Our proposed system is designed around the realities of humanitarian operations: generating actionable outputs at the speed required for USAR sectorisation and reconnaissance planning, while explicitly representing uncertainty to support accountable decision-making. We demonstrate how combining remote sensing modalities with sparse on-the-ground observations improves the timeliness and reliability of damage estimates. The results highlight that operational performance depends not only on predictive accuracy, but also on robustness under label scarcity, interpretability for non-specialist users, and the ability to revise assessments as the response evolves.