FDB-Net: A Local-Global Collaborative Attention-Based Network for Information Extraction of Flood-Damaged Building

The sudden occurrence of flood-related natural disasters poses a severe threat to human life and property safety. Rapid and accurate extraction of building damage information is crucial for humanitarian aid and emergency response. Flooddamaged building information extraction relies on background information; however, existing algorithms for extracting flooddamaged building information predominantly focus on foreground features, resulting in incomplete damage category recognition and a high rate of false alarms. To address this issue, this paper proposes a Flood-damaged Building-Net (FDB-Net) algorithm based on Local-Global Collaborative Attention. The algorithm extracts features through foreground and background attention networks, and subsequently captures local and global features through parallel convolution and Transformer networks, thereby enhancing the network's ability to perceive foregroundbackground relationships and model long-distance dependencies. In the decoder, by modeling the relationship between dualtemporal features, the algorithm effectively integrates semantic and spatial information from multi-temporal images, combining spatial features with temporal change characteristics to alleviate false detection and missed detection issues. Through feature interaction and collaborative attention, the model increases sensitivity to flood-damaged building areas while suppressing irrelevant or redundant information. Experimental validation on the xBD-Flooded dataset demonstrates the effectiveness and advantages of FDB-Net, showing significant improvements over several existing building damage extraction methods.