Rapid and reliable recognition of seismic building damage from post-disaster aerial images is critical for both scientific assessment and governance-oriented emergency response. Conventional methods often treat damage levels as unordered categories and neglect the governance implications of misclassification, particularly the severe consequences of underestimating high-risk structures. We present SeisRank-Ord, a governance-aware ordinal learning framework that integrates ordinal alignment, risk-sensitive adjustment, triage prioritization, and stability enhancement into a unified risk functional. This design embeds governance considerations directly into the learning objective, ensuring ordinal consistency, asymmetric penalties for critical errors, and stable score distributions that support resource allocation under capacity constraints. Experiments on the Yushu and Ludian datasets show that SeisRank-Ord consistently outperforms state-of-the-art baselines in seismic damage recognition while maintaining architectural generality across multiple convolutional neural network (CNN) backbones. Beyond recognition accuracy, by coupling prediction scores with policy simulation strategies including severity-first, egalitarian, and quota plus threshold regimes, SeisRank-Ord demonstrates measurable governance benefits in terms of both efficiency and fairness. These results highlight the framework as a principled bridge between computer vision and disaster governance, advancing the methodological frontier of ordinal learning while delivering actionable insights for real-world decision-making.
Zhang et al. (Thu,) studied this question.