Abstract: [Purposes]To address the limited availability of high-quality post-disaster optical imagery,as well as the significant modality gap and registration uncertainty between pre-disaster optical imagery and post-disaster SAR imagery,this study proposes a robust modeling method for instance-level building damage assessment to improve the reliability of building-level damage recognition in complex disaster scenarios.[Methods]Building instance priors are first constructed from pre-disaster optical imagery,while a damage evidence field is extracted from post-disaster SAR imagery.On this basis,a candidate alignment modeling and probabilistic evidence aggregation mechanism is introduced to achieve building-level evidence attribution under registration uncertainty.Combined with damage grading and uncertainty calibration,the proposed framework performs instance-level building damage assessment in a unified manner.Main experiments,ablation studies,robustness tests under registration perturbations,and cross-event generalization experiments are conducted on the BRIGHT dataset.[Findings]The results show that the proposed method achieves strong performance in instance-level segmentation accuracy,building-level damage discrimination,and prediction reliability.Specifically,it reaches an mAP of 0.66,a F1 of 0.65,and an ECE of 0.02.Compared with the variant without alignment modeling,the full model improves mAP by about 0.01 and F1 by about 0.05.[Conclusions]This study reformulates the task as a building-level evidence attribution problem under registration uncertainty,rather than a simple cross-modal feature fusion problem.The proposed framework provides a robust and interpretable technical solution for instance-level building damage assessment using pre-disaster optical and post-disaster SAR imagery.

Key words: building damage assessment, instance-level analysis, optical and SAR, registration uncertainty, evidence attribution