Efficient stochastic computational framework for estimating and updating earthquake fatalities, An

Abstract

Following an earthquake it is crucial to estimate the true scope of impact in terms of fatalities and damage as quickly as possible for efficient post-earthquake response and recovery. The USGS Prompt Assessment of Global Earthquakes for Response (PAGER) model is the state-of-the-art for rapid loss estimation following any major earthquake anywhere in the world. In particular, the PAGER model can provide estimates of the total loss within about 20 minutes after the event based on estimates of population, shaking intensity, and the vulnerability of the population to such shaking in the affected region. The latter is modeled as a fatality rate given as a function of shaking intensity based on past earthquake data. The current PAGER model does not explicitly account for uncertainties in the population, shaking intensity, or fatality rates independently. However, PAGER's ability to calibrate these fatality rate models using historical earthquake data entails implicitly capturing some of these uncertainties. Historically, the PAGER estimates are only updated if improved estimates of shaking are available through the USGS ShakeMap product. In addition, due to the uncertainties in the mean fatality rates, estimated population, and shaking intensities over a broad region, there is a tendency to overestimate or underestimate these losses if the estimates of shaking, population, or fatality rates are inconsistent with the ground truth. The primary goal of PAGER model is to provide the most precise and accurate estimates as soon as possible. The current interest is to upgrade the PAGER-based loss estimates through efficient spatial and temporal data-driven computational modeling, for quickly obtaining and communicating the true scope of potential loss. The main aim of this thesis is centered on developing a stochastic computational mathematical framework for efficient incorporation of ground truth data on reported earthquake fatalities to improve PAGER's overall estimate. The stochastic modeling, to obtain more accurate and precise estimates of total loss, is developed by (i) incorporating additional uncertainties when estimating the total loss; (ii) updating the initial PAGER estimates using reported losses over time. The computational framework in this work is developed using a combination of forward uncertainty propagation, and inverse Bayesian approaches for both stationary and temporal fatality data. For the latter, we incorporated techniques inspired from a traditional Kalman filtering process, which are used to update the total loss, subject to partial loss data over time provided by authoritative agencies. Using real earthquake data, we demonstrate the efficiency of the proposed computational framework and its effect in terms of improving PAGER loss alerts over time. The proposed framework allows us to propagate uncertainties over time and has the ability to ingest partial loss data over time to improve PAGER's overall forecast for the total loss.