← Back to Sessions

Operational Testing of an Efficient Bayesian Framework for Updating Pager Fatality Estimates

Session: Advances in Real-Time Geophysical Network Operations and Data Analytics [Poster]

Type: Poster

Date: 4/23/2021

Presentation Time: 03:45 PM Pacific

Description: 

Loss estimates and alert levels produced by the U.S. Geological Survey’s Prompt Assessment of Global Earthquakes for Response (PAGER) system have not traditionally considered event-specific reported data on casualties. However, as discussed in recent research (Noh et al., 2020), we have derived an algorithmic method for updating PAGER fatality estimates based on evolving, uncertain, initial fatality reports. That work, funded by the U.S. Agency for International Development, involves recursive Bayesian updating based upon the loss projection model and uncertainties from loss reporting and the loss model. After establishing a credible framework, we hindcasted fatalities for several past earthquakes where archived media reports showed increasing fatalities for a day or two following those events and found robust results, even when initial estimates were substantially different than the actual casualties. While that work gave us the proper tools, we had insufficient experience in an operational setting. We have since been evaluating this framework in a real-time, offline, capacity and herein report on results for two significant events in 2020: M6.4 Petrinja, Croatia (December 29) and M7.0 Western Greece (October 30). In both cases the median fatality estimates converge quickly to the actual losses and uncertainties were rapidly reduced when compared to our independent loss estimates. From a reporting perspective, the modified medians and lower uncertainties manifest as much tighter PAGER fatality histograms. Significant challenges identified in our testing include: (1) updating framework enhancements (i.e., the versatility of loss model projections and better use of training data), (2) recognizing and accessing reputable sources of loss information, (3) communicating updates and determining which warrant alert renotifications, and (4) modifying associated economic losses, since unlike fatalities, losses are not available in real-time.

Presenting Author: Davis T. Engler

Student Presenter: No

Authors