Research on post-disaster building damage detection method based on multitemporal remote sensing images

Disasters impose significant losses on human society, among which building damage constitutes a crucial factor. Leveraging the strengths of deep learning to promptly detect buildings after disasters with varying damage levels enables efficient disaster assessment. Current methodologies predominantly rely on pixel-level change detection approaches for post-disaster building damage assessment. However, high-resolution pixel-based analytical paradigms exhibit inherent limitations, particularly their susceptibility to pseudo-changes caused by variations in imaging angles, seasonal changes, and environmental interference. Furthermore, pixel-level building damage assessment faces the challenge of intra-instance damage heterogeneity, where pixels within a single building structure may belong to different damage categories. To address the technical constraints of pixel-level building damage identification paradigms in disaster loss assessment, this study proposes an object-level building damage detection framework. This work shifts the analytical perspective from pixel-wise comparison to instance detection, effectively reduces false alarms induced by environmental variations. The proposed network architecture operates through two stages: a building localization stage and a damage detection stage. The building locating stage performs structural identification from pre-disaster images, establishing location references for subsequent analysis. The damage detection stage then conducts comprehensive analysis by processing both pre-disaster and post-disaster image pairs, enabling classification of structural damage levels through temporal comparison. The proposed architecture integrates a feature fusion network enhanced with a frequency-spatial attention mechanism. This mechanism synergistically combines discrete cosine transform-based frequency analysis with spatial relationship modeling, significantly improving the extraction of critical building arrangement features while suppressing irrelevant background noise. To overcome the persistent challenge of small building target recognition, we implement an improved composite loss function that strategically balances localization accuracy and shape similarity. Experimental validation conducted on our dataset demonstrates that the proposed method achieves 65% mean Average Precision (mAP). The work provides an efficient technical solution for post-disaster building damage assessment.