Detection and localization of earthquakes by Deep-Learning (PhaseNet)

Study of Pyrenean seismic swarms

The procedures of detection, localization and characterization of the seismic events are based until now essentially on criteria of automatic analysis of the low level signal (detection of variations of amplitudes of the signal) and on the work of analysts visually inspecting the signals, and carrying out “with the eye” the characterization of the seismic sources (shooting of quarry, earthquake, explosion, etc). However, with the massive increase of data to be processed since the densification of the RLBP /RAP network, these procedures must be rethought and a strong interest emerges for the development of robust automatic detection methods, taking into account all the complexity of the seismological signal. Deep-Learning (DL) techniques belong to this category of algorithms.

Beyond Résif, the explosion in the number of low-cost sensors (e.g. RaspberryShake) and the deployment of dense networks also encourage the development of automatic methods.

At the Observatoire Midi-Pyrénées (OMP), we are evaluating the ability of such an automatic pointing algorithm (PhaseNet, a supervised deep-learning algorithm developed by Zhu & Beroza, 2019, at Stanford) to allow daily monitoring of seismicity on Résif stations in southwestern France. We also investigate its contribution to the study of seismicity clusters of interest for which temporary networks have been deployed since 2020. Three Pyrenean sites in particular have been instrumented and studied by OMP staff for their scientific interest: the Lacq gas field area in the Aquitaine basin, the Arette area at the Pyrenean border of Béarn, and the Campan area in the Hautes-Pyrénées.

This news is taken from an article published in the last issue of the Résif newsletter (in french only) which details the installations carried out on these three sites.

To find out more

Lettre d’information Résif n°20, décembre 2021