Multi-Modal Remote Sensing for Post-Disaster Building Damage Detection: Lessons from Five Events Across Earthquakes and Wildfires

Rapid post-disaster building damage assessment is critical for emergency response and recovery, yet field-based inspection of large affected areas can take weeks with additional revisions afterwards. We investigate whether systematic fusion of multiple remote sensing modalities—Synthetic Aperture Radar (SAR) coherence-derived Damage Proxy Maps (DPM), ShakeMap ground motion, geologic mapping, building inventory attributes, and high-resolution optical imagery processed through pretrained deep learning—improves building-level damage classification beyond any single source. Five disaster events spanning California and Puerto Rico are analyzed in progression: the 2020 Puerto Rico M6.4, the 2014 South Napa M6.0, the 2022 Ferndale M6.4 earthquakes, and the January 2025 Eaton and Palisades wildfires in Los Angeles. Random Forest fusion of DPM and ShakeMap features achieves 0.70–0.78 balanced accuracy across earthquake events, while standalone DPM thresholding generalizes to wildfires with 0.77–0.80 balanced accuracy on 30,000+ structures, demonstrating applicability beyond seismic hazards. For an 878-building subset of the Palisades fire with cloud-free Planet SkySat imagery, fusing DPM with the xView2 deep learning model boosts balanced accuracy to 0.83. The central methodological finding is that preserving raw classification probabilities from pretrained vision models, rather than collapsing to argmax labels, provides substantially richer fusion features, accounting for a gain of 8 percentage points. Notably, the model assigns destroyed buildings the highest probability for their background class out of the five classes (background, no damage, minor, major and destroyed), a signal lost under standard inference. The modular per-building feature framework allows new modalities to be integrated without redesign, supporting operational deployment for rapid post-disaster situational awareness and risk assessment.