Scalable variational learning for noisy-OR Bayesian networks with normalizing flows for complex cascading disaster systems

Sudden-onset disasters like earthquakes trigger multiple cascading hazards and impacts, causing human and economic losses. While remote sensing technologies enable rapid hazard assessment, current methods have two key limitations: they struggle to decouple co-located hazards and impacts, and cannot adapt to new information during disaster response. We present online-DisasterVINF, a framework addressing both challenges through Noisy-OR Bayesian networks and normalizing flows. This approach models how multiple hazards independently contribute to observed impacts while providing a more expressive alternative to log-linear relationships. The framework continuously incorporates ground truth data from post-disaster reconnaissance through a novel variational inference approach that jointly approximates posteriors by leveraging causal relationships and remote sensing techniques. Evaluation across multiple global seismic events demonstrates that online-DisasterVINF significantly improves estimation accuracy compared to existing methods, while showing substantial performance enhancement as ground truth becomes available, highlighting its utility for real-time disaster response.