[S21P] PM-P

[S21P-07] Development of deep-learning methods for seismic event detection in seafloor DAS data: creation of training/validation datasets

*GERARDO MANUEL MENDO¹, HIROMICHI NAGAO¹, SHINYA KATOH¹, MASANAO SHINOHARA² (1. Research Center for Computational Earth Science, Earthquake Research Institute, The University of Tokyo, 2. Research Center for Geophysical Observation and Instrumentation, Earthquake Research Institute, The University of Tokyo)

Distributed Acoustic Sensing (DAS) is one of the promising technologies revolutionizing seismology for two reasons: 1) ground motion can be recovered from DAS recordings obtained fiber optic cables, and 2) the optical fiber cable can be considered as a multichannel array, allowing the recording of the wavefield radiated from earthquakes with very high resolution. Challenges such as manageability of large volumes of data and variable signal-to-noise ratio makes difficult to use conventional seismological processing techniques to detect and identify seismic traces from DAS traces. In order to address these issues, state-of-the-art research in DAS systems aims to develop AI-based algorithms to process DAS recordings. Deep-learning-based methods have been successful worldwide due to the availability of large earthquake catalogs created by human analysts through the years. However, there is still room for improvement. In this work, we present the results from applying the template matching algorithm to identify and retrieve seismic events from DAS traces. The aim to create training and validation datasets to use them in deep-learning methods for training and test stages. As starting templates, we used three types of signals: a 20-s time window that comprises a seismic event with both P and S phases and two 5-s time windows that comprise either a P or an S phase. We also use templates from earthquake signals where only S phase is distinguishable. Template matching was applied using both waveforms and signal envelopes. We search throughout the data from the seafloor cable located in Sanriku, Japan. Two main differences are observable: 1) correlation coefficients ranges from 0.2 – 0.4 and 0.85 – 0.95 using waveform and envelope correlation respectively, and 2) envelope correlation detects events farther from the cable. For events with source locations close to the cable, we successfully identify events with both waveform and envelope. We present the main results of this work, highlighting that this method can be useful to obtain good datasets to use in deep learning models.