OASIS: Occlusion Aware Semantic Inference and Sensing for Fused UAV-Satellite Driven Disaster Damage Assessment

Post-disaster damage assessment demands rapid, reliable situational awareness under severe occlusion, heterogeneous viewpoints, and limited sensing opportunities. While wide-area satellite imagery provides early regional context, it lacks the resolution and perspective required to assess structural damage at street and facade levels. Conversely, unmanned aerial vehicles offer detailed local observations but are constrained by flight time, visibility, and the need for intelligent viewpoint selection. Existing approaches typically process these sensing modalities in isolation, rely on static perception pipelines, or employ heuristic coverage strategies that do not reason about occlusion, uncertainty, or the value of future observations. This paper introduces OASIS, an occlusion-aware semantic inference and sensing framework for disaster damage assessment that tightly integrates historical wide-area imagery with real-time unmanned aerial vehicle (UAV) observations. OASIS leverages a vision-language model to perform multi-scale semantic reasoning, fusing satellite priors and UAV imagery into continuous damage estimates while explicitly modeling visibility and occlusion. A probabilistic belief state captures spatially varying uncertainty across structures, and an information-theoretic planner selects UAV trajectories that maximize expected uncertainty reduction under motion and energy constraints. By embedding occlusion awareness and semantic sensitivity directly into the observation and planning process, OASIS actively resolves ambiguous regions rather than passively collecting data. Experimental results demonstrate that OASIS significantly improves information efficiency, occlusion resolution, and damage severity estimation compared to existing baselines, highlighting the importance of tightly coupled semantic reasoning and active sensing for time-critical post-disaster assessment.