Café LabEx, le 11 mars 2020

06 mars 2020

Le prochain Café LabEx sera présenté par Alexandra Renouard dans la Salle du Conseil (5 rue René Descartes) mercredi le 11 mars à 15h00.
Sa présentation est intitulée : "Contribution of machine learning to earthquake detection in high anthropogenic context "
Abstract : 

Nowadays, the intensive seismological stations’ deployment allows for the seismological observatories to archive high data volume such as high-quality seismograms. The high data volume offers the opportunity to build more complete earthquake catalogs down to small magnitudes, and reveal more intricate details on earthquake interaction, earthquake triggering, earthquake, driving mechanisms or the seismogenic potential of faults, for example.

But, paradoxically, the expected improvements in the seismological observatories, which are the main catalog providers, are actually quite limited. If the current detection procedure is identically used to lower the detection threshold, and recover low-amplitude signal, a high level of false detections (false alarms), and misdetections (real event with false arrivals) overwhelm the event alert system. In this framework, analysts waste time to screen out a huge amount of bogus events with a higher risk of missing real events, because of a significant cry wolf effect due to false alarm fatigue, and the necessity to make rapid decisions under time constraints.

In order to correct the current detection system shortcomings, we developed a fully automatic detection procedure that produces more accurate catalogs of small events with less analyst review.

This procedure uses SeisComP3 framework, can still be employed in real-time, continue to use large amounts of data, and is easily implemented in seismological observatories.

We firstly focused on performance tuning to increase phase sensitivity over higher noise level, and to improve the pick association process. We then designed two additional SeisComP3 modules. The first module is dedicated to filtering misdetections. The second one incorporates machine learning tools to classify events and label them as either false alarms, earthquakes, or quarry blasts.