A new workflow for revising the seismicity catalog for mainland France, covering the period 2010-2018

In the continuation of the work carried out over the period 1962-2009 as part of the SI-Hex project, work is on going to revise the seismic catalog of mainland France from 2010 to 2018. This time period is characterized by both an upgrade of short-period stations with broadband stations and a major deployment of new broadband stations as part of the Résif-Epos research infrastructure (now called Epos-France), significantly increasing the amount of detected and processed events.

This catalog will benefit from our advances in the use of new artificial intelligence tools, such as PhaseNet, a deep learning automatic picking method, as well as in the development of a deep learning method for discrimination between earthquakes, quarry blasts and explosions.

This catalog will be built from those of the national observation service BCSF-Renass, CEA/LDG and regional seismological observatories (Isterre, OCA, OMP). The earthquake picks from these catalogs will be supplemented by those automatically obtained by deep learning on all the waveforms from the Epos-France (formerly Résif-Epos) stations daily used by BCSF-Renass (as part of its mission to monitor seismicity in mainland France) including stations from neighboring countries (GB, LU, BE, DE, CH, IT, ES), as well as those from temporary network stations (AlpArray, CifAlps2). 

The process workflow includes several steps. The first one consists in a clustering of picks close in time to reduce the amount of picks to process; duplicated picks are removed and priority is given to the manual ones. The second step is the association of seismic phases to create events, by combining the HDBSCAN algorithm - to merge picks close in time and space - with the PyOcto one - to discard picks that did not follow typical travel-time curves. The third step consists in event location using NonLinLoc algorithm with several regional models chosen based on the prior location obtained from PyOcto. At the last step, a moment magnitude Mw is computed (when possible) from waveform spectral fitting using a modified version of SourceSpec. To compute robust magnitudes in particular for low magnitude events, we include magnitude station corrections computed from statistics on magnitude differences between event and stations. Finally, events information (ie. origins, magnitudes) coming from the various catalogs are integrated into the multi-origin catalog according to the QuakeML standard, with the preferred location being the new one computed on the third step.

This catalog currently under revision will represent an update of seismicity in France over the period 2010-2018. Preliminary results show that it will incorporate a significantly increased number of low-magnitude events, detected thanks to the inclusion of picks from artificial intelligence tools. Event labeling is consolidated using our deep learning discrimination algorithm, and a Mw magnitude is calculated for each event using waveforms.