Rapid disaster damage assessment using deep adversarial sliced Wasserstein domain adaptation

This study investigates the use of unsupervised domain adaptation (UDA) techniques for rapid damage assessment during disasters. Traditional disaster damage assessment models are typically trained on historical disaster data, but their performance often declines when applied to new disaster events because of differences in data distributions, known as domain shift. To address this challenge, the researchers evaluate advanced UDA methods that enable models to adapt to emerging disasters without requiring labeled data from the new event. Using real-world images collected from Twitter during four major disasters, the study compares several state-of-the-art domain adaptation approaches, including DANN and AWDAN, and explores how deeper neural network architectures can improve feature extraction and transferability across disaster scenarios. 

The results show that domain adaptation significantly improves the performance of disaster damage assessment models when applied to previously unseen disaster events. The proposed approach, which combines a deeper backbone architecture with the AWDAN framework, consistently outperformed existing methods such as DANN, CORAL, MMD, and CDAN across all source–target disaster combinations. The enhanced model achieved statistically significant improvements, reaching up to 11.4% higher F1-score and 8.9% higher accuracy compared with models trained only on source-domain data. These findings demonstrate that advanced UDA techniques can effectively mitigate domain shift and provide more reliable rapid damage assessments during real-world disaster response operations.