Abstract
Wildland fires' destruction of the built environment has increased sharply in the past decade. Remote sensing of wildfire destruction offers great potential for rapid damage assessment at regional to global scales, but the available approaches suffer from high uncertainty and low generalizability. Here, we introduce the Building Damage Assessment (BDA) with Kolmogorov-Arnold Networks (KAN) model, a novel deep learning architecture integrating attention mechanisms with KAN networks using pre-post normalized Digital Surface Model (nDSM) differencing. BDA-KAN achieved 97.20% and 96.47% accuracy over the 2025 Eaton Fire and Palisades Fire in California—encompassing 11,726 and 17,965 structures, respectively—with precision exceeding 97%, outperforming state-of-the-art architectures, while requiring >30 times less computational power than deep learning models like PiT and RVT. The mixture-of-experts gating mechanism in BDA-KAN dynamically routes building footprints to specialized sub-networks optimized for distinct damage manifestations, while learnable activation functions in KAN capture complex nonlinear relationships between vertical displacement and structural integrity. Cross-fire generalization experiments demonstrated robust transferability, with >90% accuracy when trained on one fire and evaluated over another fire, confirming the model's generalizability in assessing damage signatures across geographic contexts. SHAP interpretability analysis confirmed physically meaningful decision-making, with attribution concentrated over collapsed roofs, despite the noise introduced by domain shifts. BDA-KAN offers a transferable, efficient deep learning framework that can accurately assess post-fire structural damage where nDSM is available.
