Robust Building Damage Detection in Cross-Disaster Settings Using Domain Adaptation
Researchers Asmae Mouradi and Shruti Kshirsagar have published a study addressing the challenge of domain shift in automated building damage assessment from remote sensing imagery. Models trained on multi-disaster benchmarks often fail in unseen geographic regions due to distributional mismatches, undermining trust in human-machine systems for disaster management. The study explores a two-stage ensemble approach utilizing supervised domain adaptation (SDA) to classify building damage across four severity levels. By adapting the xView2 first-place method to the Ida-BD dataset, the authors systematically investigated the impact of various augmentation components. Comprehensive ablation experiments revealed that SDA is indispensable; removing it caused complete failure in damage detection. The proposed pipeline achieved its most robust performance using SDA with unsharp-enhanced RGB input, attaining a Macro-F1 score of 0.5552. These findings highlight the critical role of domain adaptation in creating trustworthy automated assessment modules for integrated disaster response systems, ensuring more reliable situational awareness for decision-makers during timely emergency operations.
Wire timeline
Robust Building Damage Detection in Cross-Disaster Settings Using Domain Adaptation
Researchers Asmae Mouradi and Shruti Kshirsagar have published a study addressing the challenge of domain shift in automated building damage assessment from remote sensing imagery. Models trained on multi-disaster benchmarks often fail in unseen geographic regions due to distributional mismatches, undermining trust in human-machine systems for disaster management. The study explores a two-stage ensemble approach utilizing supervised domain adaptation (SDA) to classify building damage across four severity levels. By adapting the xView2 first-place method to the Ida-BD dataset, the authors systematically investigated the impact of various augmentation components. Comprehensive ablation experiments revealed that SDA is indispensable; removing it caused complete failure in damage detection. The proposed pipeline achieved its most robust performance using SDA with unsharp-enhanced RGB input, attaining a Macro-F1 score of 0.5552. These findings highlight the critical role of domain adaptation in creating trustworthy automated assessment modules for integrated disaster response systems, ensuring more reliable situational awareness for decision-makers during timely emergency operations.
cs.AI updates on arXiv.org