Post-disaster building damage assessment based on gated adaptive multi-scale spatial-frequency fusion network

• Propose a two-phase Transformer framework for localization and classification. • Design a Multi-Scale Wavelet Fusion (MWF) module for feature enhancement. • Introduce Adaptive Attention (AA) to focus on severity-related regions. • Develop a Gated Multi-scale Feed-Forward Network (GMFFN) for refined extraction. • Validate transferability on the independent Ida-BD benchmark. Accurate building damage assessment is crucial for post-disaster response, yet existing methods struggle to capture complex spatial relationships and contextual features needed for distinguishing damage levels. To address this, we propose the Gated Adaptive Multi-scale Spatial-frequency Fusion Network (GAMSF), a two-phase framework for building localization and damage classification. GAMSF integrates three key innovations: (1) Adaptive Attention (AA) to dynamically prioritize critical regions, (2) Gated Multi-scale Feed-Forward Network (GMFFN) to enhance robustness by emphasizing prominent damage features, and (3) Multi-Scale Wavelet Fusion (MWF) to extract fine-grained structural details using wavelet transforms. Rigorous evaluations on the datasets, including xBD and xFBD, demonstrates that GAMSF achieves the state-of-the-art performance, with a 1.7% improvement in F1-score, a 2.1% gain in Kappa, and a 3.7% increase in minor damage identification accuracy compared to existing approaches. Furthermore, transferability experiments on the high-resolution Ida-BD dataset validate GAMSF’s superior generalization capabilities, outperforming four advanced models. These results highlight the practical value of GAMSF in enhancing disaster management, emergency response, and resource allocation strategies.