Abstract:

As the primary carriers of population and economic activities， buildings are highly vulnerable to disasters， and their damage status directly affects emergency response and post-disaster reconstruction. Therefore， rapid and accurate acquisition of building damage information has become a critical requirement in disaster management. To address the challenges of geometric misalignment， background interference， and feature alignment difficulties in multi-temporal and cross-modal remote sensing imagery， we propose a Pre-Disaster Footprint-guided change-aware damage detection Network （PDF-Net）. Specifically， the framework first employs a twin pyramid vision transformer to extract multi-level features from pre-and post-disaster imagery and generates pre-disaster building masks to introduce guidance information. Subsequently， a change-aware gated attention module is designed to enhance differential representation of low-level detail features， highlighting local changes， while a grouped cross-temporal attention mechanism with overlapped windows is introduced to explicitly align high-level semantic features， thereby reinforcing the structural change representation of buildings. Finally， fine-grained damage detection is achieved through cross-level feature fusion. Experiments conducted on the xBD dataset （pre-disaster optical-post-disaster optical） and the Bright dataset （pre-disaster optical-post-disaster SAR） demonstrate that the proposed method achieves significant improvements in both intra-modal and cross-modal tasks. Specifically， B-PriorNet surpasses the current state-of-the-art methods by 0.58% in mean Intersection-over-Union （mIoU） on the xBD dataset and by 1.97% on the Bright dataset， showing stronger robustness and generalization ability in cross-modal detection scenarios. These results validate the effectiveness and practical value of the proposed framework in complex disaster environments.

Key words: remote sensing, deep learning, disaster, change detection, building damage