Extending near-fault seismic catalog from single-station waveforms using deep learning

Josipa Majstorovic, Sophie Giffard-Roisin, Piero Poli

Abstract:

Several studies have already shown that machine learning (ML) algorithms have tremendous potential when it comes to their applications to the standard seismological procedures, such as earthquake detection and localization, phase picking, ground motion prediction, signal denoising, etc. Their superiority lies in their capability to rapidly scan voluminous data and to learn the features of the earthquakes while being robust to highly noisy time series.

We here make use of ML algorithms to obtain a more complete near-fault seismic catalog and thus better understand the long-term (decades) evolution of seismicity before large earthquakes occurrence. We developed an end-to-end two-stage pipeline using 1D convolutional neural networks (CNN) able to detect, localize and characterize earthquakes from single-station three-component waveforms. From an intensive and rigorous hyper-parameter selection, we are presenting here the insights of what makes the difference in a deep learning algorithm. Moreover, our algorithm is robust and does not need any pre-processing of the seismograms (e.g. filtering) or any prior knowledge of the region.

We applied our pipeline on 29 years of data (1990 to 2019) recorded at the AQU station, located near the city of L’Aquila (Italy) in the Abruzzo region (Central Italy). Furthermore, we focus on the data right before the devastating L’Aquila earthquake (6 April 2009 01:32 UTC, Mw6.3). Before this event, only sparse stations were available which limits the magnitude completeness of standard catalogs. We used the available pre-existing catalogs as training and validation data, and we further tested our developed methods on continuous data.

For the goal of detecting local seismicity, our results show that we are able to detect the earthquake events among random noise waveforms with 99.9% accuracy (first stage). We are then able to determine the events that are close (less than 10km from the AQU station) from rest with a 88% accuracy as well as identifying the low magnitude events (between Mw0 and Mw2) with a 58% accuracy (second stage).