An advanced tool for a refined earthquake early warning system: methodology, alert decision scheme and application performance

Having a real-time, reliable image of the seismic source is crucial for realistic peak ground motion predictions, which are necessary for an efficient planning of emergency operations. Operational platforms for network-based Earthquake Early warning ground on the assumption of a point-like earthquake source model and of 1D ground motion prediction equations to estimate the earthquake impact.

We propose a new method which allows issuing an alert based on the real-time mapping of the Potential Damage Zone (PDZ), e.g. the area where the ground shaking is expected to exceed the damaging level, with no assumption on rupture extent and spatial variability of ground motion. The system processes the vertical component, velocity/acceleration data streams at each station. The characteristic P-wave period (τc) and peak amplitude (Pd, Pv and Pa) (in displacement, velocity and acceleration, respectively) are measured on expanding time windows. These evolutionary measurements around the source allow predicting geometry and extent of PDZ, and the lower shaking regions at larger distances from the source. The P-peak motion at sites where no record exists is inferred by a physics-based spatial interpolation, using the real-time event location and attenuations relationships for amplitude vs. magnitude and distance. This method naturally accounts for the effects of rupture extent (e.g. source directivity) and for spatial variability of ground motion due to crustal wave propagation and site amplification.

To provide robust alert notifications to the target, we conceived a probabilistic decision module in which the exceedance probability of a pre-determined shaking level accounts for the uncertainties of the empirical prediction equations. We implemented the idea of a variable ground shaking threshold and of a user-customizable exceedance probability level, quantifying the uncertainties related to the empirical equations. Our decision module may operate in different configurations, each of them corresponding to an increasing level of complexity.

We tested the method with playbacks of the Mw 6.5 Norcia (Italy) event and of the Mw 9.0 Tohoku-Oki (Japan) event and evaluated the performance on isolated targets (e.g. schools/hospitals) and on distributed targets (as virtual railway lines).