Cross-scenario damaged building extraction network: Methodology, application, and efficiency using single-temporal HRRS imagery

Haifeng Wang, Wei He, Zhuohong Li, Naoto Yokoya

Abstract: The extraction of damaged buildings is of significant importance in various fields, such as disaster assessment and resource allocation. Although multi-temporal-based methods exhibit remarkable advantages in detecting damaged buildings, single-temporal extraction remains crucial in real-world emergency responses due to its immediate usability. However, single-temporal cross-scenario extraction at high-resolution remote sensing (HRRS) encounters the following challenges: (i) morphological heterogeneity of building damage which causes by the interplay of unknown disaster types with unpredictable geographic contexts, and (ii) scarcity of fine-grained annotated datasets for unseen disaster scenarios which limits the accuracy of rapid damage mapping. Confronted with these challenges, our main idea is to decompose complex features of damaged building into five attribute-features, which can be trained using historical disaster data to enable the independent learning of both building styles and damage features. Consequently, we propose a novel Correlation Feature Decomposition Network (CFDNet) along with a coarse-to-fine training strategy for the cross-scenario damaged building extraction. In detail, at the coarse training stage, the CFDNet is trained to decompose the damaged building segmentation task into the extraction of multiple attribute-features. At the fine training stage, specific attribute-features, such as building feature and damage feature, are trained using auxiliary datasets. We have evaluated CFDNet on several datasets that cover different types of disasters and have demonstrated its superiority and robustness compared with state-of-the-art methods. Finally, we also apply the proposed model for the damaged building extraction in areas historically affected by major disasters, namely, the Turkey–Syria earthquakes on 6 February 2023, Cyclone Mocha in the Bay of Bengal on 23 May 2023, and Hurricane Ian in Florida, USA in September 2022. Results from practical applications also emphasize the significant advantages of our proposed CFDNet.

External IDs:doi:10.1016/j.isprsjprs.2025.06.028