[S-CG55] Driving Solid Earth Science through Machine Learning

[SCG55-P07] Developments of a classification method of tectonic tremors, earthquakes, and noises and locating of tectonic tremors using deep learning

*Amane Sugii¹, Yoshihiro Hiramatsu¹, Takahiko Uchide², Kazutoshi Imanishi² (1.Kanazawa University, 2.National Institute of Advanced Industrial Science and Technology)

Keywords:tectonic tremor, Convolutional Neural Network, semblance analysis, Score-CAM

Tectonic tremors (hereafter referred to as tremors) observed in subduction zones (e.g., the Nankai Trough) around the world show characteristics different from those of ordinary earthquakes. Known tremor detection methods include envelope cross-correlation (Obara, 2002) and semblance analysis (Neidell and Taner, 1971). However, these methods have the problem of detecting earthquakes and noise as tremors. In particular, the detection accuracy of tremors decreases with waveform data with a low signal-to-noise ratio. To solve this problem, event classification using deep learning has been studied extensively in recent years. Noise, tremor, and earthquake classification methods using deep learning have been shown to identify events with high accuracy for test data selected by analysts (Nakano et al. 2019; Takahashi et al. 2021). However, it has been pointed out that the identification accuracy is reduced for tremor waveforms with low signal-to-noise ratios and that seismic waveforms with attenuated high-frequency components are misclassified as tremors with similar characteristics (Nakano et al. 2019). In addition, few studies have newly conducted source determination using waveforms classified as tremors. Therefore, in this study, we develop a discrimination model for noise, tremors, and earthquakes, including distant earthquakes, and perform a semblance analysis on waveforms classified as tremors.

Three-component velocity waveforms from the array observation network installed on the Kii Peninsula by the National Institute of Advanced Industrial Science and Technology (AIST) were used as seismic waveform data. The model used in this study discriminates between noise, tremors, and earthquakes by probability, using spectrogram images as input data. The model consists of two convolutional layers and a pooling layer, one convolutional layer and a GlobalAveragePooling layer, and three fully connected layers. When trained on a spectrum image of 20,400 waveforms with distinct noise, tremor, and earthquake characteristics, the discrimination accuracy was 98.2% compared to the test data. The training results including earthquake waveforms with attenuated high-frequency components showed that noise, tremors, and earthquakes could be discriminated more correctly. Furthermore, by using Score-CAM (Wang et al. 2019), we succeeded in visualizing the parts that the model considers important in identifying events. We found that the model identifies tremor and seismic features as well as human visualization, and we applied the model to continuous waveforms over one year from April 2013 to March 2014 and detected 8851 tremor waveforms. These included waveforms at times when no events were listed in the tremor catalog (Imanishi et al. 2011), and visual inspection confirmed that these waveforms were likely to be tremors. In this presentation, we also report the results of the semblance analysis of 8851 tremors and a new method of tremor source determination using Score-CAM and semblance analysis together.