講演情報
[P22-12]Leveraging a BERT-based deep learning model for earthquake detection in seafloor Distributed Acoustic Sensing data
*GERARDO MANUEL MENDO1, Hiromichi Nagao1, Shinya Katoh1, Masanao Shinohara2 (1. Research Center for Computational Earth Sciences, Earthquake Research Institute, The University of Tokyo, 2. Research Center for Geophysical Observation and Instrumentation, Earthquake Research Institute, The University of Tokyo)
Monitoring offshore earthquake activity is important for seismic hazard management due to the generation of large megathrust earthquakes near subduction zones. Several seafloor seismic observatory systems have been deployed to retrieve continuous records of seismic activity in the oceans. In recent years, Distributed Acoustic Sensing (DAS) has emerged as a promising technique to monitor earthquake activity. DAS provides high-resolution strain rate recordings associated with earthquakes using telecommunication cables, reducing the cost of installation and maintenance. Due to the vast data volumes it produces, deep learning models are suitable to process DAS data. Although there are deep learning models for inland DAS processing, seafloor-based methods to process DAS data are still scarce. For this reason, we propose a deep-learning-based model for earthquake detection using seafloor DAS data. This model integrates Bidirectional Encoder Representations from Transformers (BERT) architecture, a transformer-based encoder widely used in natural language processing tasks such as speech recognition and machine translation. To test this model, we process the DAS data obtained from Sanriku seafloor observation system, a 120-km long telecommunication cable off the northeastern section of Iwate prefecture, Japan. The training and validation datasets from the Sanriku database were obtained using SVR-DAS, a machine-learning-based method for earthquake detection. We present preliminary findings and assess the viability of using LLM-based models such as BERT to perform seismic processing, opening the possibility to identify earthquakes in low signal-to-noise (SNR) data.