Post-Disaster Building Damage Assessment Using Drone Imagery and Deep Learning

Accurate assessment of building damage is essential for effective disaster response and recovery. This paper introduces a UAV-based dataset of post-disaster imagery annotated according to the Federal Emergency Management Agency (FEMA) Preliminary Damage Assessment (PDA) guidelines. The dataset contains high-resolution images with pixel-level segmentation labels for four building damage categories: Destroyed, Major Damage, Minor Damage, and Affected. Using this dataset, we benchmarked a state-of-the-art YOLOv11-L instance segmentation model to automatically detect and classify building damage. The model was trained on the dataset with a standard training and validation split. Results show strong performance in identifying extreme classes such as Destroyed and Affected, while distinguishing between mid-severity categories (Major versus Minor) remains challenging. This work provides a standardized dataset and a baseline model to support scalable and reproducible research on UAV-based damage assessment, with the goal of enabling faster and more reliable decision-making in post-disaster response and recovery.