2:00 PM - 2:15 PM
[STT38-02] Development of a Deep Low-Frequency Earthquake Detection Method Using Deep Learning with Multiple Traces as Inputs
Keywords:deep low-frequency earthquake, deep learning
Deep Low-Frequency Earthquakes (DLFs) occur in various regions, including the deep parts of plate boundaries in oceanic areas, deep sections of transform faults on land, the vicinity of Quaternary volcanoes, and non-volcanic zones (Hasegawa & Nakajima, 2022). In non-volcanic areas of central-northern Kinki and western Tottori Prefecture, DLFs occur locally near the Moho discontinuity and around the deeper edges of reflector in the lower crust (Doi and Nishigami, 2007; Katoh et al., 2018). Because fluids are implicated in generating DLFs, an accurate assessment of DLF activity is essential for understanding the distribution of fluids in the subsurface and fluid diffusion patterns.
This study aimed to develop a deep learning model for detecting DLFs occurring in inland non-volcanic zones. Similar waveform search methods, such as template matching, have been used to detect DLFs (Shelly et al., 2007; Kurihara and Obara, 2021). This method is limited in detecting new seismic waveforms because it relies on pre-created templates. Therefore, we created a deep learning model that can flexibly detect DLF waveforms more effectively than traditional template-based methods.
The developed model uses waveform data recorded at multiple seismic observation points as input and can learn information about the wave field. In previous studies developed a method for detecting low-frequency tremors using deep learning (Nakano et al., 2019; Rouet-Leduc et al., 2020; Takahashi et al., 2021), a waveform recorded at a single observation point is used as input. However, in this study, multiple observation points were used as inputs because we considered it more accessible to detect DLFs with small amplitude when multiple traces are input than when a single trace is input.
We detected DLFs occurring in non-volcanic areas of central-northern Kinki and western Tottori Prefecture using the developed model. These areas have been equipped with dense observation networks with station intervals of less than 5 km (Manten network: Miura et al., 2010; Iio, 2011; Iio et al., 2017) from 2009 to 2022. By analyzing data with significantly improved spatial resolution obtained from Manten network in addition to stationaly observation networks like Hi-net, the detection capability for DLFs improved.
This presentation will introduce the developed model's structure and detection performance.
This study aimed to develop a deep learning model for detecting DLFs occurring in inland non-volcanic zones. Similar waveform search methods, such as template matching, have been used to detect DLFs (Shelly et al., 2007; Kurihara and Obara, 2021). This method is limited in detecting new seismic waveforms because it relies on pre-created templates. Therefore, we created a deep learning model that can flexibly detect DLF waveforms more effectively than traditional template-based methods.
The developed model uses waveform data recorded at multiple seismic observation points as input and can learn information about the wave field. In previous studies developed a method for detecting low-frequency tremors using deep learning (Nakano et al., 2019; Rouet-Leduc et al., 2020; Takahashi et al., 2021), a waveform recorded at a single observation point is used as input. However, in this study, multiple observation points were used as inputs because we considered it more accessible to detect DLFs with small amplitude when multiple traces are input than when a single trace is input.
We detected DLFs occurring in non-volcanic areas of central-northern Kinki and western Tottori Prefecture using the developed model. These areas have been equipped with dense observation networks with station intervals of less than 5 km (Manten network: Miura et al., 2010; Iio, 2011; Iio et al., 2017) from 2009 to 2022. By analyzing data with significantly improved spatial resolution obtained from Manten network in addition to stationaly observation networks like Hi-net, the detection capability for DLFs improved.
This presentation will introduce the developed model's structure and detection performance.