Real-time Prediction of Distance and PGA from P-wave features using Gradient Boosting Regressor for On-Site Earthquake Early Warning Applications

Abstract

Summary On-site Earthquake Early Warning (EEW) systems represent an important way to reduce seismic hazard. Since these systems are fast in providing an alert and reliable in the prediction of the ground motion intensity at targets, they are particularly suitable in the areas where the seismogenic zones are close to cities and infrastructures, such as Central Italy. In this work, we use Gradient Boosting Regressor (GBR) to predict Peak Ground Acceleration (PGA), and hypocentral distance (D) starting from P-wave features. We use two datasets of waveforms from two seismic sequences in Central Italy: L'Aquila sequence (2009), and the Amatrice-Norcia-Visso sequence (2016-2017), for a total of about 80,000 3-components waveforms. We compute 60 different features related to the physics of the earthquake using three different time-windows (1s, 2s, and 3s). We validate and train our models using the 2016-17 datasets (the bigger one) and we test it on the 2009 dataset. We study the performances of GBR predicting D and PGA in terms of prediction scores, finding that the models can well predict both targets even using 1s window, and that, as expected, the results improve using longer time-windows. Moreover, we perform a residual analysis on the test set finding that the PGA can be predicted without any bias, while the D prediction present a correlation with the moment magnitude. In the end, we propose a prototype for a probabilistic on-site Earthquake Early Warning (EEW) system based on the prediction of D and PGA. The proposed system is a threshold-based approach, and it releases an alert on four possible levels, from 0 (far and small event) to 3 (close and strong event). The system computes the probability related to each alert level. We test two different set of thresholds, the Felt Alert and the Damage Alert. Furthermore, we consider the lead-time (LT) of the PGA to distinguish between useful alerts (positive LT) and Missed Alerts (MA). In the end, we analyze the performance of such a system considering four possible scenarios: Successful Alert (SA), Missed Alert (MA), Overestimated Alert (OA), and Underestimated Alert (UA). We find that the system obtains SA rate about 80% at 1s, and that it decreases to about 65% due to the increase of MA. This result shows how the proposed system is already reliable at 1s, which would be a huge advantage for seismic prone regions as Central Italy, an area characterized by moderate-to-large earthquakes (Mw<7).