Reproduce “Multiclass Post-Earthquake Building Assessment Integrating High-Resolution Optical and SAR Satellite Imagery, Ground Motion, and Soil Data with Transformers” on OpenResearch

About this paper

Deepank Singh, Vedhus Hoskere, Pietro Milillo

Researchers at the University of Houston and German Aerospace Center introduce QuakeMetaFormer, a transformer-based model that integrates high-resolution optical imagery with seismic, SAR, and soil metadata for multiclass post-earthquake building damage assessment. The model achieved an AUC-ROC of 0.85 for 5-class damage classification, representing a 16% improvement over prior machine learning methods and demonstrating enhanced generalization across affected regions.

Key insights

• The explicit integration of diverse metadata (seismic intensity, SAR-derived parameters, soil properties) with high-resolution optical imagery significantly boosts multiclass building damage assessment performance and model generalization to unseen geographical areas.
• Transformer-based architectures, specifically QuakeMetaFormer, combined with a modified focal loss function, can robustly handle complex multi-modal data fusion and severe class imbalance to achieve precise, nuanced damage classification.
• Seismic intensity indicators (e.g., PGA, PGV, PSA) offer the most substantial contribution from metadata to damage prediction, particularly for differentiating between various levels of damage severity, complementing visual cues from imagery.

Summary by alphaXiv.