1:45 PM - 3:15 PM
[SCG55-P03] Extraction of tectonic tremor waveforms using deep learning : Application to Hi-net stations in the Nankai subduction zone
Keywords:Convolutional neural network, Tectonic tremor
Recently, deep learning has been used to distinguish tectonic tremor waveforms from earthquake ones and noises. However, for the systematic detection of tremor waveforms in a wide area, deep learning has not been a typical tool. In this study, to detect tremor waveforms in the entire Nankai subduction zone, we used a CNN (Convolutional Neural Network) model with parameters trained by waveform data of Hi-net stations in the Nankai subduction zone. Here, the CNN model uses spectrograms as input data and outputs probabilities for noises, tremors, and earthquakes.
The performance of the models varies with the frequency ranges of the input spectrograms. The model using the data of a 2–30 Hz frequency range shows the best performance. This best model provides high performance of discrimination and generalization for the test data and other stations that were not included in the training data. The results of the application to teleseismic earthquakes with the lack of high frequency components supports that the best model is appropriate for tremor detection as well. A comparison of the detected tremors from the model with tremors from the existing catalogs shows that 80.1 % of the events in the catalogs are correctly detected as tremors. We expect that more tremors can be detected by changing the combination of stations. The comparison also reveals that the model has the ability the detection of tremors in high activity periods of tremors, in which the locating of tremors often fails. Further investigation and studies are needed to determine the source location of tremors detected by the CNN model.
<Acknowledgments>
We used waveform data recorded by Hi-net and the tremor catalog created by a hybrid method (Maeda and Obara, 2009), and a hybrid clustering method (Obara et al., 2010). We also used the arrival time data and the unified earthquake catalog of JMA to create our earthquake catalog.
The performance of the models varies with the frequency ranges of the input spectrograms. The model using the data of a 2–30 Hz frequency range shows the best performance. This best model provides high performance of discrimination and generalization for the test data and other stations that were not included in the training data. The results of the application to teleseismic earthquakes with the lack of high frequency components supports that the best model is appropriate for tremor detection as well. A comparison of the detected tremors from the model with tremors from the existing catalogs shows that 80.1 % of the events in the catalogs are correctly detected as tremors. We expect that more tremors can be detected by changing the combination of stations. The comparison also reveals that the model has the ability the detection of tremors in high activity periods of tremors, in which the locating of tremors often fails. Further investigation and studies are needed to determine the source location of tremors detected by the CNN model.
<Acknowledgments>
We used waveform data recorded by Hi-net and the tremor catalog created by a hybrid method (Maeda and Obara, 2009), and a hybrid clustering method (Obara et al., 2010). We also used the arrival time data and the unified earthquake catalog of JMA to create our earthquake catalog.