An AI system for residential building damage estimation using multimodal remote sensing data fusion

The increasing intensity of tropical storms has created an urgent need for rapid, reliable damage assessment. Although recent AI-based approaches have improved upon manual inspection, many rely on proprietary satellite imagery or ad hoc UAV datasets, limiting scalability, repeatability, and operational deployment. In this study, we develop a multimodal AI framework for post-disaster damage assessment using only publicly available remote sensing and geospatial data, including NOAA aerial imagery, Microsoft building footprints, and FEMA damage labels. A key methodological contribution is the use of spatially disjoint training and validation regions in the Hurricane Michael dataset to prevent geographic data leakage and to more realistically evaluate generalization to unseen locations, an issue that is frequently overlooked in geospatial AI studies. Using this rigorous validation strategy, we conduct a comparative evaluation of leading deep learning architectures, including Convolutional Neural Networks (ResNet and VGG) and the Vision Transformer (ViT). Model performance in classifying buildings as affected or destroyed is assessed using disaster-relevant evaluation metrics. Results demonstrate that the Vision Transformer consistently outperforms CNN-based models, likely due to its global self-attention mechanism, which more effectively captures spatial context in dense urban environments. Overall, this work demonstrates the feasibility of a scalable, public-data-based framework for rapid damage assessment, while identifying severe class imbalance as a key operational challenge, motivating the use of specialized loss weighting and data augmentation strategies in future deployments.