Scalable Variational Learning for Noisy-OR Bayesian Networks with Normalizing Flows for Complex Cascading Disaster Systems

Xuechun Li, Susu XuResearch Square 2024 · DOI

Abstract

Sudden-onset disasters such as earthquakes often induce multiple cascading hazards and impacts, causing human and economic losses. While recent advances in remote sensing technologies provide valuable data for rapid hazard assessment, current methods face two fundamental limitations. First, the inherent causal relationship and co-location of disaster-induced hazards and impacts makes the decoupling of individual hazard or impact particularly challenging. Second, and critically for disaster response, existing methods cannot adapt to new information during the critical response period - they provide only static initial estimates despite the continuous influx of ground truth data from field reconnaissance. This inability to update and refine assessments in real-time severely limits their practical utility for emergency response, where understanding of the disaster's impact typically evolves significantly over time. Herein, we present online-DisasterVINF, a framework that uniquely addresses both limitations. We leverage Noisy-OR Bayesian networks, which are particularly suited for modeling how multiple hazards independently contribute to the observed impacts, combined with normalizing flows to provide a more expressive alternative to simple log-linear relationships. The framework is also capable of improving its estimation by incorporating ground truth data as it becomes available through post-disaster reconnaissance. We develop a novel variational inference approach that jointly approximates posteriors by leveraging complex causal relationships and remote sensing techniques. We evaluate our framework on multiple seismic events from diverse countries around the globe. Our results demonstrate that online-DisasterVINF significantly enhances estimation accuracy compared to existing methods, while our analysis shows the online updating mechanism substantially improves model performance as ground truth becomes available, underlining its adaptability for real-time disaster response.

Introduction

When a major earthquake occurs, it triggers a cascade of secondary hazards—such as landslides and liquefaction—that lead to building collapses and loss of life. Emergency managers require immediate, accurate information to direct resources effectively during the critical "Golden 72 Hours." However, current satellite-based Damage Proxy Maps (DPMs) face two major hurdles: signal entanglement, where different hazards produce indistinguishable surface changes, and a lack of adaptability to real-time field reports.

Methodology

The online-DisasterVINF framework addresses these limitations by integrating causal modeling with advanced deep learning techniques.

Noisy-OR Bayesian Networks

The framework utilizes Noisy-OR Bayesian Networks (BNs) to model the disaster as a causal chain. This structure is specifically chosen for its ability to represent how multiple independent causes (e.g., ground shaking and slope steepness) contribute to a single outcome (e.g., a landslide) in a physically interpretable manner.

Normalizing Flows

To overcome the limitations of simple log-linear relationships in standard BNs, the researchers integrated Normalizing Flows (NFs). NFs use a series of invertible mappings to transform simple probability distributions into complex ones, allowing the model to learn highly non-linear dependencies. As shown in Figure 1, increasing the number of flows ($K$) significantly sharpens the posterior estimates, bringing them closer to the true disaster state.

Figure 1: Improvement in building damage estimation as the number of flows (K) in the Normalizing Flow increases.

Online Learning and Scalability

The framework employs a modified Expectation-Maximization (EM) algorithm to incorporate ground truth data as it arrives from the field. To ensure scalability, a "local pruning" strategy is used to ignore physically impossible nodes (e.g., landslides on flat terrain), allowing the model to process over 15,000 $km^2$ in under 2.5 hours.

Experimental Results

The framework was validated using data from the 2021 Haiti (M7.2), 2020 Puerto Rico (M6.4), and 2018 Hokkaido (M6.6) earthquakes.

Performance and Attribution

In the Haiti case, the model achieved an Area Under the Curve (AUC) of 0.9842 for landslides and 0.9586 for building damage, outperforming standard machine learning baselines. In Puerto Rico, it achieved an AUC of 0.9424 for liquefaction, representing an 11.76% improvement over previous methods.

Impact of Online Updating

A key finding was the model's efficiency in learning from limited ground truth. As shown in Figure 5, integrating just 10-20% of field data leads to a significant jump in performance. In the damage detection task, the F1 score rose from 0.9275 to 0.9555 as ground truth integration increased to 50%.

Figure 5: Performance metrics showing the improvement of online-DisasterVINF as ground truth data is incrementally integrated.

Conclusion

online-DisasterVINF provides a robust, scalable, and adaptive solution for disaster impact assessment. By combining the causal rigor of Bayesian networks with the expressiveness of Normalizing Flows, the framework enables more accurate hazard decoupling and real-time refinement of estimates, significantly enhancing the utility of remote sensing data for emergency response.

Summary updated

Abstract

<title>Abstract</title> Sudden-onset disasters such as earthquakes often induce multiple cascading hazards and impacts, causing human and economic losses. While recent advances in remote sensing technologies provide valuable data for rapid hazard assessment, current methods face two fundamental limitations. First, the inherent causal relationship and co-location of disaster-induced hazards and impacts makes the decoupling of individual hazard or impact particularly challenging. Second, and critically for disaster response, existing methods cannot adapt to new information during the critical response period - they provide only static initial estimates despite the continuous influx of ground truth data from field reconnaissance. This inability to update and refine assessments in real-time severely limits their practical utility for emergency response, where understanding of the disaster's impact typically evolves significantly over time. Herein, we present online-DisasterVINF, a framework that uniquely addresses both limitations. We leverage Noisy-OR Bayesian networks, which are particularly suited for modeling how multiple hazards independently contribute to the observed impacts, combined with normalizing flows to provide a more expressive alternative to simple log-linear relationships. The framework is also capable of improving its estimation by incorporating ground truth data as it becomes available through post-disaster reconnaissance. We develop a novel variational inference approach that jointly approximates posteriors by leveraging complex causal relationships and remote sensing techniques. We evaluate our framework on multiple seismic events from diverse countries around the globe. Our results demonstrate that online-DisasterVINF significantly enhances estimation accuracy compared to existing methods, while our analysis shows the online updating mechanism substantially improves model performance as ground truth becomes available, underlining its adaptability for real-time disaster response.
BibTeX
@inproceedings{li2024scalable,
  author = {Xuechun Li and Susu Xu},
  title = {{Scalable Variational Learning for Noisy-OR Bayesian Networks with Normalizing Flows for Complex Cascading Disaster Systems}},
  booktitle = {Research Square},
  year = {2024},
  doi = {10.21203/rs.3.rs-5462482/v1},
}
Figure