Developing a Seismicity Catalog at Mayotte With Deep-Learning-Based Picking and Phase Association
Description:
The active seismic sequence off the coast of Mayotte island, in the Mozambique channel, that began in 2018 has drawn much attention, due to the high seismic activity which was strongly felt by the local population. The detected seismic events have been made primarily with either human analyst assistance (Saurel et al., 2022) or a combination of deep-learning-based (DL) picking and traditional phase association methods (Retailleau et al., 2022). These catalogs have detected 1000s of earthquakes and concentrated seismic activity at ~35 km depth, within ~30 km of the newly developing volcanic vent. Nonetheless, the high density of events in time apparent from visual inspection (>1000 events per day during the highest activity) indicates a large fraction may still be missed, such as for small events and those occurring close to one another in time and space. Similarly, false positives may result from using traditional associators (e.g., Earthworm) on such high pick rate sequences, since such associators were designed before the era of DL-based pickers.
In this work, we revisit the Mayotte seismic sequence and process the local ocean bottom seismometer (OBS) data using both PhaseNet picking, and the recently developed graph neural network associator, GENIE (McBrearty and Beroza, 2022), initially trained for regional scales in northern California. We compare our developed catalog using GENIE with the existing automated approach (Retailleau et al., 2022). We measure the rate of matched and distinct events obtained in either catalog, for a range of different threshold and hyper-parameter choices. On average, we find ~70% matched events. The non-overlapping seismic events appear to be the smallest events, and we look for additional systematic trends that can explain differences in the catalogs. Our findings reveal new insights into the Mayotte seismic sequence and demonstrate the generalizability of GENIE to local networks.
Session: Opportunities and Challenges for Machine Learning Applications in Seismology
Type: Oral
Date: 4/19/2023
Presentation Time: 02:15 PM (local time)
Presenting Author: Ian W. McBrearty
Student Presenter: Yes
Invited Presentation:
Authors
Ian McBrearty Presenting Author Corresponding Author imcbrear@stanford.edu Stanford University |
Lise Retailleau retailleau@ipgp.fr Institut de Physique du Globe de Paris |
Gregory Beroza beroza@stanford.edu Stanford University |
|
|
|
|
|
|
Developing a Seismicity Catalog at Mayotte With Deep-Learning-Based Picking and Phase Association
Category
Opportunities and Challenges for Machine Learning Applications in Seismology