Rapid Epicentral Distance and Magnitude Estimation via Collaborative Attention Network

Abstract

ABSTRACT Rapid and reliable estimation of epicentral distance and magnitude are pivotal for damage assessment in earthquake early warning (EEW). However, conventional epicentral distance and magnitude estimation approaches for EEW rely primarily on manually defined features to construct empirical prediction equations. Meanwhile, many existing machine learning models focus solely on single tasks (either epicentral distance or magnitude estimation alone), neglecting the correlation and impact between epicentral distance and magnitude. In this study, we design a collaborative attention network (CAN) framework for EEW, named EEWCAN, which combines global and local attention mechanisms. In addition, we explore the feasibility of EEWCAN for estimating epicentral distance and magnitude from single-station data in the Japan region. EEWCAN integrates global and local attention mechanisms to automatically capture more critical features pertinent to epicentral distance and magnitude from seismic waveforms. The EEWCAN framework is composed of two models: the EEWCAN-Dis model for epicentral distance estimation and the EEWCAN-Mag model for magnitude estimation. Meanwhile, the epicentral distance estimated by the EEWCAN-Dis model is incorporated into the EEWCAN-Mag model. We trained and tested EEWCAN using seismic events recorded by the Kyoshin network in Japan, with waveforms serving as the input to EEWCAN. The test results demonstrate that, when data from 3 s after P-wave onset are used, EEWCAN can obtain reliable estimates of epicentral distance and magnitude, outperforming the baseline models. Furthermore, by incorporating the estimated epicentral distance information, EEWCAN enhances the reliability of magnitude estimation. This result demonstrates that our proposed method can provide a certain degree of reliability for epicentral distance and magnitude estimation in EEW.