AI Uses Engineering Knowledge to Predict Disaster Damage from Partial Imagery

Researchers at Seoul National University, in collaboration with Iowa State University and the U.S. Air Force Research Laboratory, have developed a novel Scientific AI framework capable of rapidly predicting city-scale structural damage from incomplete satellite imagery. Published in Scientific Reports, the study introduces a data-driven methodology that bypasses traditional machine learning training pipelines while maintaining high accuracy under real-world disaster conditions. The framework integrates structural engineering principles with statistical data imputation to analyze pre- and post-disaster satellite imagery. Instead of relying solely on algorithmic pattern recognition, the system incorporates physical engineering knowledge to interpret structural changes. A core component is image entropy, which quantifies environmental disorder and structural disruption before and after an event. To address the common challenge of obscured imagery caused by clouds, smoke, or precipitation, the team employed a statistical reconstruction technique. Rather than discarding missing data or applying simple averages, the algorithm leverages spatial correlations to accurately infer obscured regions, enabling continuous damage assessment even when over half the imagery is blocked. Validation tests demonstrated the framework robustness during a major typhoon event. The system successfully mapped urban structural damage with an error rate below five percent, despite more than fifty percent of the satellite data being obscured. The methodology also proved effective in a wildfire damage assessment, indicating broad applicability across various natural and human-induced disasters. A defining advantage of this approach is its immediate deployability. Unlike conventional deep learning models that require extensive datasets and costly retraining for each new disaster, this framework generates actionable insights directly from available imagery and open-source geographic data. The ability to rapidly produce accurate, city-wide damage assessments holds significant implications for emergency response and infrastructure management. First responders and disaster management agencies can use the output to prioritize rescue missions, allocate recovery resources, and identify structures requiring urgent safety inspections. Long-term applications extend to urban resilience planning, aging infrastructure monitoring, and performance evaluation of industrial assets. Lead researcher Professor In Ho Cho emphasized that merging domain-specific engineering knowledge with Scientific AI enables complex structural analyses to be conducted rapidly and at minimal computational cost. He noted that the system reliability with degraded imagery opens pathways for broader adoption in defense, aerospace maintenance, and automated industrial inspection.