5:15 PM - 6:45 PM
[SCG40-P06] Extraction and analysis of deep tectonic tremor waveforms using convolutional neural network : Application to Hi-net stations in the Nankai subduction zone
Keywords:Tectonic tremor, Convolutional neural network
Tectonic tremors (hereafter referred to as tremors) are characterized as weak vibration phenomena with dominant frequencies ranging from 1 to 10 Hz and durations spanning several minutes to several hours. Their detection poses challenges due to the difficulty in discerning their phases. Consequently, applying conventional earthquake detection methods to tremors is problematic. Typically, tremor hypocenters are estimated through the envelope correlation method, which involves correlating envelope waveforms from multiple seismic stations (Obara, 2002). However, this method may erroneously identify regular earthquakes and noise as tremors, necessitating manual inspection or additional processing. To address this issue, previous studies have employed convolutional neural networks (CNN), a form of deep learning, to differentiate seismic waveforms accurately. Nakano et al. (2019) and Takahashi et al. (2021) demonstrated high accuracy in discriminating noise, tremors, and earthquake waveforms using CNNs. Nevertheless, systematic extraction of tremor waveforms over extended periods utilizing CNNs and subsequent analysis remains unexplored. Therefore, this study aims to establish a CNN-based method for tremor analysis. We conduct tremor extraction using CNNs and analyze tremor activity using the extracted data.
In this study, spectrograms derived from three-component velocity waveforms recorded by Hi-net were utilized as input data. The CNN model comprises two convolutional layers, two pooling layers, and three fully connected layers, with the output layers providing noise, tremor, and earthquake probability values, respectively. Performance evaluation on test data, comprising 4,600 noise, tremor, and earthquake events, totaling 13,800 images, demonstrated that the model achieved 98.0% accuracy. During the extraction of tremors from continuous data spanning the two-year period from 2016 to 2017 using 71 Hi-net stations, 12,681 events were extracted in the Tokai region, 6,878 events in the northern Kii, 4,866 in the central Kii, 2,045 events in the southern Kii, 5,859 events in the eastern Shikoku, 6,776 events in the central Shikoku, and 16,064 events in the western Shikoku. Approximately 47% of the events listed in the hybrid catalog (Maeda and Obara, 2009) were identified as tremors by this study. Moreover, this study identified numerous events not listed in the hybrid catalog but classified as tremors. Over the analyzed period, approximately 3.8 times as many tremors were extracted compared to those listed in the hybrid catalog. This outcome suggests that the proposed method in this study can detect tremors with high activity, which is challenging to identify using conventional methods due to low hypocenter determination accuracy. In this presentation, we additionally present a comprehensive examination of event discrimination accuracy at each station and the findings of a tremor analysis that incorporates amplitude information alongside event duration.
In this study, spectrograms derived from three-component velocity waveforms recorded by Hi-net were utilized as input data. The CNN model comprises two convolutional layers, two pooling layers, and three fully connected layers, with the output layers providing noise, tremor, and earthquake probability values, respectively. Performance evaluation on test data, comprising 4,600 noise, tremor, and earthquake events, totaling 13,800 images, demonstrated that the model achieved 98.0% accuracy. During the extraction of tremors from continuous data spanning the two-year period from 2016 to 2017 using 71 Hi-net stations, 12,681 events were extracted in the Tokai region, 6,878 events in the northern Kii, 4,866 in the central Kii, 2,045 events in the southern Kii, 5,859 events in the eastern Shikoku, 6,776 events in the central Shikoku, and 16,064 events in the western Shikoku. Approximately 47% of the events listed in the hybrid catalog (Maeda and Obara, 2009) were identified as tremors by this study. Moreover, this study identified numerous events not listed in the hybrid catalog but classified as tremors. Over the analyzed period, approximately 3.8 times as many tremors were extracted compared to those listed in the hybrid catalog. This outcome suggests that the proposed method in this study can detect tremors with high activity, which is challenging to identify using conventional methods due to low hypocenter determination accuracy. In this presentation, we additionally present a comprehensive examination of event discrimination accuracy at each station and the findings of a tremor analysis that incorporates amplitude information alongside event duration.