DALi-CGAN: domain adaptive LiDAR-based conditional generative adversarial network for earthquake damage assessment

Due to the extensive structural damage caused by earthquakes, prompt and precise building damage assessment is essential for efficient disaster response. Traditional methods often fail to perform well in new regions where labelled data are scarce, underscoring the need for domain adaptation. In this context, our research presents a domain-adaptive LiDAR-based conditional generative adversarial network (DALi-CGAN) architecture, which is designed to address the complexities of cross-domain damage assessment using 3D LiDAR point cloud data. By incorporating graph attention, the proposed DALi-CGAN architecture enhances the feature extraction process, allowing for improved classification accuracy by capturing complex spatial relationships within the data. The model is evaluated under cross-domain, limited-data conditions, demonstrating effective generalization across the Haiti and Kumamoto earthquake-affected regions. The proposed DALi-CGAN demonstrates superior generative performance by achieving the lowest Fréchet point distance (FPD) of 0.58, improving upon the best-performing baseline (FPD = 0.81) by approximately 28%. The DALi-CGAN augmented dataset achieves a classification accuracy of 90.9% and an F1-score of 0.9085, outperforming existing approaches. Additionally, cross-validation confirms the robustness and generalization capability of the DALi-CGAN, with an average test accuracy of 0.9082 ± 0.0088 and an F1-score of 0.9085 ± 0.0088. Our findings highlight DALi-CGAN’s potential to deliver reliable damage assessment across geographically distributed earthquake-affected regions.