Neural disaster simulation for transferable building damage assessment

Abstract

Timely and reliable building damage assessment is essential for effective disaster response and humanitarian assistance. However, the diversity of disaster types, geographic regions, and data distributions poses significant challenges to transferring building damage assessment models to new disaster scenarios (i.e., target domains). In addition, limited availability of post-disaster training imagery in target domains further hinders progress. Recent approaches, such as single-temporal change adaptation, enable adaptation using only target pre-disaster images by constructing pseudo bitemporal damage samples at an unexplainable embedding level. While effective, these methods produce representations that are difficult for human experts to interpret, inspect for errors, or adjust categorical distributions to ensure reliable model performance. In this paper, we propose Neural Disaster Simulation (NeDS), a deep disaster generative model that synthesizes realistic post-disaster image from pre-event image and customizable disaster information (i.e., disaster types and disaster intensity). Thanks to damage data generation based solely on pre-event imagery, NeDS enables adaptation at any time, effectively bypassing the limitation of post-disaster training image availability. Furthermore, by explicitly modeling disaster effects at the image level, NeDS mitigates distribution shifts between historical training data and unseen disaster events, enhancing both model transferability and visual interpretability. Extensive experiments conducted on both global-scale and local-scale study areas demonstrate that NeDS adaptation outperforms the previous state-of-the-art, achieving a 4.3% improvement in average performance on a global dataset, as well as 3.6%, 7.9%, and 18.5% gains in damage classification performance for the 2025 Eaton fire, the 2025 Palisades fire in Los Angeles, and the 2025 Nigeria flooding, respectively.