Multimodal Post-Disaster Building Damage Assessment with Weakly Supervised Learning

Rapid post-disaster building damage assessment is critical for effective humanitarian response [1]. We present an edge-optimized framework that fuses optical and Synthetic Aperture Radar (SAR) satellite imagery with crowd-sourced point annotations to produce real-time damage maps. An adaptive fusion module dynamically adjusts optical-to-SAR weighting based on estimated cloud cover, preserving all-weather capability [11]. A tiered Independent Bayesian Classifier Combination (IBCC) strategy scales computational load to available hardware [7], while progressive neural architectures span devices from smartphones$(\mathbf{4 7 ~ M B})$to workstations$(\mathbf{8 9 0 ~ M B})$[14]. Experiments on the xBD and BRIGHT datasets show 78% accuracy retention under total cloud occlusion, 64% annotation throughput improvement over expert-only labeling, and robust cross-disaster generalization-establishing practical viability for time-critical disaster scenarios.