Item

Abstract

Accurate seismic hazard assessment requires high quality and complete seismic catalogs. Anthropogenic events such as quarry blasts, rock bursts, explosions can introduce significant errors in statistical analyses, especially in areas with low background seismicity rates such as Mainland France. The BCSF-Renass, a national service of observation in seismology, characterizes the seismicity of the Mainland France, and makes available to the scientific community its seismic catalog. Until 2012, only natural seismicity was localized but since 2012, other types of events are localized and manually discriminated, in particular quarry blasts, marine explosions (demining operation), induced seismicity. These last years have seen a huge increase of new seismological stations installed in the Resif-EPOS framework. The main consequence is the location by the BCSF-Renass of many low magnitude events, most are quarry blasts (around half in 2022). The work of the analysts is considerably impacted, firstly by a huge increase of events to be processed and secondly by the difficulty of determining the events' origins with the issue of catalog contamination. This is why the BCSF-Renass, in early 2022, has developed a Deep Convolutional Neural Network (DCNN) tool to automatically discriminate earthquakes from quarry blasts. The learning step was done on the 3 components seismogram spectra on the year 2021 events. The performances are more than 99% for precision and recall, and take well into account the different regions and geological contexts of France. Moreover, this discrimination also worked well for marine explosions, and allowed us to highlight landslides or rock falls.