Rapid, Scalable, and Joint Assessment of Seismic Multi-Hazards and Impacts: From Satellite Images to Causality-Informed Deep Bayesian Networks

Johns Hopkins University, Baltimore MD

Investigators

Abstract

A seismic event often involves multiple hazards (e.g., ground shaking, landslide, and liquefaction) and impacts (e.g., building and infrastructure damage). Understanding the locations and extents of such hazards and impacts in high resolution immediately after an event is critical for facilitating real-time responses, such as timely evacuation, search and rescue, and effective allocations of limited resources. Researchers have been investigating using satellite imageries to extract hazard and impact information for wide affected areas; however, the co-occurrence and co-location of hazards and impacts result in mixed signals in satellite imagery, making it very challenging to directly categorize and estimate each hazard and associated impacts. This Disaster Resilience Research Grants (DRRG) project aims to develop a novel system to provide rapid, scalable, and joint assessments of cascading seismic hazards and impacts by leveraging the causal dependencies among them. It will enhance the accuracy, resolution, and timeliness of existing rapid disaster information systems by integrating satellite images with existing geospatial hazard models from The US Geological Survey and building fragility functions from the Federal Emergency Management Agency’s HAZUS tool. The revealed regional causal mechanisms aims to enable improved seismic risk analysis as well as the study of other natural disasters involving cascading impacts. This will improve overall community resilience to future natural disasters. This project will develop a causality-informed variational Bayesian network modeling framework to adaptively provide regional-scale seismic multi-hazard and impact occurrence estimates in near-real-time, by fusing information from satellite images with prior geophysical knowledge and building fragility functions through a deep causal Bayesian network. First, a novel paradigm will be established to model complex and implicit causal dependencies among cascading seismic multi-hazards and impacts as a flow-based causal Bayesian network to integrate information from prior hazards and impact models with mixed-signal satellite imagery. Further, an online variational Bayesian inference framework will be developed to jointly infer and update, in a scalable and efficient manner, the estimations of seismic multi-hazards and impacts, with or without partially observed ground truth. Third, local geospatial hazards model and building fragility functions will be updated through a novel uncertainty-aware prior model updating scheme using the event-specific patterns learned from the causal Bayesian network. The quantitative causal mechanisms of cascading seismic hazards and building damage, revealed by the causal Bayesian network, will be characterized in multiple earthquake events, to render an in-depth understanding of event-specific seismic hazards and damage patterns for improving regional resilience to future disaster events. The framework will be demonstrated on seven moderate-to-large global earthquake events. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.