Cross-disaster generalization of deep learning models for building damage assessment: a systematic evaluation using the Generalization Retention Ratio

Deep learning has significantly advanced automated building damage assessment (BDA) from post-disaster remote sensing imagery. However, most existing studies evaluate models within a single disaster type, leaving their ability to generalize across different disasters largely unexplored. This study presents a systematic investigation of cross-disaster generalization by training multiple modern object detection models exclusively on hurricane imagery from the xBD dataset and evaluating them, without retraining or fine-tuning, on unseen earthquake, flood, and wildfire datasets. Model performance was assessed using F1-score, mean Average Precision (mAP@0.5 and mAP@0.5:0.95), class-wise F1-score, and a normalized Generalization Retention Ratio (GRR) that quantifies the proportion of source-domain performance retained after transfer to an unseen disaster domain. Experimental results reveal a consistent ordering in detection performance across the evaluated target domains, with flood imagery retaining the most performance and wildfire the least. Nevertheless, the proposed BDA-Net consistently achieves the highest detection accuracy and retains the largest proportion of its baseline performance across all evaluated disasters. Flood imagery exhibits the highest cross-disaster transferability, whereas wildfire imagery presents the greatest challenges to model generalization. These findings demonstrate the importance of evaluating building damage assessment models beyond their training domain. A cross-disaster ablation study confirms that attention-guided feature refinement and semi-supervised learning each contribute to retention, and that their combination yields the strongest cross-disaster robustness. The proposed evaluation framework and GRR provide a practical basis for future research on robust and transferable deep learning models for post-disaster damage assessment.