Picking Induced Seismicity with Deep Learning (piSDL)

Training deep-learning picking models with several published data sets can be easily done through the Python toolbox SeisBench. Most of the data sets contain earthquakes recorded at local, regional and teleseismic distances, with only limited data in the low magnitude, close distance region. Applyi...

Full description

Saved in:
Bibliographic Details
Main Authors: Janis Heuel, Vincent Maurer, Michael Frietsch, Andreas Rietbrock
Format: Article
Language:English
Published: McGill University 2025-08-01
Series:Seismica
Online Access:https://seismica.library.mcgill.ca/article/view/1579
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1849232743798407168
author Janis Heuel
Vincent Maurer
Michael Frietsch
Andreas Rietbrock
author_facet Janis Heuel
Vincent Maurer
Michael Frietsch
Andreas Rietbrock
author_sort Janis Heuel
collection DOAJ
description Training deep-learning picking models with several published data sets can be easily done through the Python toolbox SeisBench. Most of the data sets contain earthquakes recorded at local, regional and teleseismic distances, with only limited data in the low magnitude, close distance region. Applying current published PhaseNet models to induced seismicity data leads to only a few events being detected and trained PhaseNet models are not able to outperform well-established workflows in seismology. Here we present a new seismological data set and trained PhaseNet models for picking induced seismicity with deep-learning (piSDL). PhaseNet was trained with 171,182 three component waveforms from 40,576 events. Noise samples were added in the training data set to reduce the number of false picks. In this study, we noticed that a good earthquake training data set and noise samples from the analysed area are both important to detect more seismic events with a newly trained PhaseNet model. We validated our new PhaseNet models at a geothermal site in Rittershoffen (France). The models trained with the new data set and noise samples clearly outperform PhaseNet’s original published model and traditional methods in seismology by detecting up to 62% more events compared to a seismicity catalogue published by an agency.
format Article
id doaj-art-1f3b9b80aac8489583d3da0064b0c810
institution Kabale University
issn 2816-9387
language English
publishDate 2025-08-01
publisher McGill University
record_format Article
series Seismica
spelling doaj-art-1f3b9b80aac8489583d3da0064b0c8102025-08-20T22:27:24ZengMcGill UniversitySeismica2816-93872025-08-014210.26443/seismica.v4i2.1579Picking Induced Seismicity with Deep Learning (piSDL)Janis Heuel0Vincent Maurer1Michael Frietsch2Andreas Rietbrock3Geophysical Institute, Karlsruhe Institute of Technology, Karlsruhe, GermanyÉS-Géothermie, Strasbourg, FranceGeophysical Institute, Karlsruhe Institute of Technology, Karlsruhe, GermanyGeophysical Institute, Karlsruhe Institute of Technology, Karlsruhe, Germany Training deep-learning picking models with several published data sets can be easily done through the Python toolbox SeisBench. Most of the data sets contain earthquakes recorded at local, regional and teleseismic distances, with only limited data in the low magnitude, close distance region. Applying current published PhaseNet models to induced seismicity data leads to only a few events being detected and trained PhaseNet models are not able to outperform well-established workflows in seismology. Here we present a new seismological data set and trained PhaseNet models for picking induced seismicity with deep-learning (piSDL). PhaseNet was trained with 171,182 three component waveforms from 40,576 events. Noise samples were added in the training data set to reduce the number of false picks. In this study, we noticed that a good earthquake training data set and noise samples from the analysed area are both important to detect more seismic events with a newly trained PhaseNet model. We validated our new PhaseNet models at a geothermal site in Rittershoffen (France). The models trained with the new data set and noise samples clearly outperform PhaseNet’s original published model and traditional methods in seismology by detecting up to 62% more events compared to a seismicity catalogue published by an agency. https://seismica.library.mcgill.ca/article/view/1579
spellingShingle Janis Heuel
Vincent Maurer
Michael Frietsch
Andreas Rietbrock
Picking Induced Seismicity with Deep Learning (piSDL)
Seismica
title Picking Induced Seismicity with Deep Learning (piSDL)
title_full Picking Induced Seismicity with Deep Learning (piSDL)
title_fullStr Picking Induced Seismicity with Deep Learning (piSDL)
title_full_unstemmed Picking Induced Seismicity with Deep Learning (piSDL)
title_short Picking Induced Seismicity with Deep Learning (piSDL)
title_sort picking induced seismicity with deep learning pisdl
url https://seismica.library.mcgill.ca/article/view/1579
work_keys_str_mv AT janisheuel pickinginducedseismicitywithdeeplearningpisdl
AT vincentmaurer pickinginducedseismicitywithdeeplearningpisdl
AT michaelfrietsch pickinginducedseismicitywithdeeplearningpisdl
AT andreasrietbrock pickinginducedseismicitywithdeeplearningpisdl