Bayesian Approach for Identification of Multiple Events in an Early Warning System

Abstract

The 2011 Tohoku earthquake (M_w 9.0) was followed by a large number of aftershocks that resulted in 70 early warning messages in the first month after the mainshock. Of these warnings, a non‐negligible fraction (63%) were false warnings in which the largest expected seismic intensities were overestimated by at least two intensities or larger. These errors can be largely attributed to multiple concurrent aftershocks from distant origins that occur within a short period of time. Based on a Bayesian formulation that considers the possibility of having more than one event present at any given time, we propose a novel likelihood function suitable for classifying multiple concurrent earthquakes, which uses amplitude information. We use a sequential Monte Carlo heuristic whose complexity grows linearly with the number of events. We further provide a particle filter implementation and empirically verify its performance with the aftershock records after the Tohoku earthquake. The initial case studies suggest promising performance of this method in classifying multiple seismic events that occur closely in time.