5:15 PM - 6:30 PM
[STT37-P03] A Convolutional Neural Network based classification of local earthquakes and tectonic tremors in Sanriku-oki using a single station S-net data and its generalization ability
Keywords:Slow earthquakes, S-net, CNN
Slow earthquakes characterized by a longer duration compared to regular earthquakes in the same magnitude have been widely detected in many tectonic zones (e.g., Obara and Kato, 2016). Tectonic tremors are a kind of slow earthquakes and have been often located in the adjacent areas of megathrust zones. Therefore, monitoring spatio-temporal activities of tremors provides important clues to slow fault slips.
Envelope correlation method (ECM) has been widely used to detect tremors. Although the ECM is a powerful tool for tremor detections, it principally detects signals of earthquakes as well. To classify local earthquakes, tremors and noise, Nakano et al. (2019) developed a supervised, convolutional neural network (CNN)-based method. They adopted spectrograms obtained by DONET ocean seismometers deployed in the Nankai subduction zone as input images for the CNN and attained 99.5% accuracy of classifications.
Here, we attempt to classify local earthquakes, tremors, and noise observed by S-net in northeast Japan following the approach developed by Nakano et al. (2019). However, S-net seismometers have a characteristic period of 15 Hz and thus are less sensitive to the dominant frequencies of tremors (2-10 Hz) compared to the DONET seismometers. Therefore, we modify the structures and retrain the parameters in the CNN developed by Nakano et al. (2019) to adjust the S-net data. As a first step, this study attempts to classify the known local earthquakes, tremors, and noise by the CNN using a single station to investigate how our CNN works for our final purpose of monitoring tremor activities.
We used spectrograms of three component velocity waveforms for 2 minutes in the time domain and for 2-10 Hz in the frequency domain that are obtained from 16 August 2016 to 14 August 2018, at N. S4N21, located off-Sanriku. As training data for CNN, we utilized JMA unified catalog for local earthquakes with epicentral distances smaller than 80 km, and tremor catalog (Nishikawa et al. 2019). We divided the whole data periods into two data sets; waveforms from 16 August 2016–2 December 2017 were used for training and those December 2017–14 August 2018 were used for validation.
As a result, the CNN truly classified 100, 96 and 98 % of local earthquakes, tectonic tremors and noise, respectively. The output probabilities for the true classification of earthquakes and tremors have become smaller with an increasing epicentral distance and that for the true classification of earthquakes with a smaller magnitude. This indicates that the probabilities to be judged as a true classification depend on the signal to noise ratio of the original record.
In this presentation, we will also show the classification results when applying the CNN trained at N.S4N21 to the untrained surrounding S-net stations and discuss the generalization ability of the CNN.
Envelope correlation method (ECM) has been widely used to detect tremors. Although the ECM is a powerful tool for tremor detections, it principally detects signals of earthquakes as well. To classify local earthquakes, tremors and noise, Nakano et al. (2019) developed a supervised, convolutional neural network (CNN)-based method. They adopted spectrograms obtained by DONET ocean seismometers deployed in the Nankai subduction zone as input images for the CNN and attained 99.5% accuracy of classifications.
Here, we attempt to classify local earthquakes, tremors, and noise observed by S-net in northeast Japan following the approach developed by Nakano et al. (2019). However, S-net seismometers have a characteristic period of 15 Hz and thus are less sensitive to the dominant frequencies of tremors (2-10 Hz) compared to the DONET seismometers. Therefore, we modify the structures and retrain the parameters in the CNN developed by Nakano et al. (2019) to adjust the S-net data. As a first step, this study attempts to classify the known local earthquakes, tremors, and noise by the CNN using a single station to investigate how our CNN works for our final purpose of monitoring tremor activities.
We used spectrograms of three component velocity waveforms for 2 minutes in the time domain and for 2-10 Hz in the frequency domain that are obtained from 16 August 2016 to 14 August 2018, at N. S4N21, located off-Sanriku. As training data for CNN, we utilized JMA unified catalog for local earthquakes with epicentral distances smaller than 80 km, and tremor catalog (Nishikawa et al. 2019). We divided the whole data periods into two data sets; waveforms from 16 August 2016–2 December 2017 were used for training and those December 2017–14 August 2018 were used for validation.
As a result, the CNN truly classified 100, 96 and 98 % of local earthquakes, tectonic tremors and noise, respectively. The output probabilities for the true classification of earthquakes and tremors have become smaller with an increasing epicentral distance and that for the true classification of earthquakes with a smaller magnitude. This indicates that the probabilities to be judged as a true classification depend on the signal to noise ratio of the original record.
In this presentation, we will also show the classification results when applying the CNN trained at N.S4N21 to the untrained surrounding S-net stations and discuss the generalization ability of the CNN.