ViTConvKAN: a hybrid transformer–KAN modeling for post-earthquake building damage mapping using VHR imagery

Abstract

Accurate mapping of earthquake-damaged buildings is essential for rapid post-disaster assessment and urban recovery planning. Very high-resolution (VHR) satellite imagery enables large-scale geospatial damage detection; however, conventional convolutional neural networks (CNNs) often struggle to model complex spatial dependencies and nonlinear damage patterns efficiently. To address these limitations, this study proposes ViTConvKAN, a hybrid architecture that combines vision transformers (ViTs) for long-range spatial modeling with convolutional Kolmogorov–Arnold networks (ConvKANs) for nonlinear feature representation. The framework includes data preprocessing, supervised training, building-level damage classification, and quantitative evaluation. The method is validated on post-event VHR datasets from Haiti-Port-au-Prince and Iran-Bam, including cross-region transfer experiments under zero-shot and few-shot fine-tuning settings. Across both regions, ViTConvKAN achieves an average OA of 88.09% and a KC of 73.07%, outperforming several established CNN-based baselines. The results indicate that integrating transformer-based contextual modeling with nonlinear ConvKAN feature extraction improves building-level damage discrimination while maintaining scalable deployment characteristics. This hybrid approach provides a robust framework for post-earthquake damage mapping using VHR imagery.