11:00 AM - 1:00 PM
[SCG51-P07] Development of a method for detecting tectonic tremor using deep learning
Keywords:tectonic tremor, convolutional neural network
Tectonic tremors (hereinafter referred to as tremors) occur at plate boundaries, such as the Philippine Sea Plate, which is subducting from the Nankai Trough, at a depth of about 30 to 40 km, and show characteristics different from those of ordinary earthquakes (dominant frequency of 1 to 10 Hz and vibration duration of several minutes to hours). For automatic detection of tremors, the envelope correlation method (Obara, 2002) and semblance analysis (Neidell and Taner, 1971) are used, but they suffer from false detections and, in particular, poor detection accuracy for data with a low signal-to-noise ratio.
Prior studies on classifying noise, tremors, and ordinary earthquakes (hereinafter simply “earthquakes”) using deep learning (e.g., Nakano et al., 2019; Takahashi et al., 2021) worked well on their test data. However, the trained model is not explainable. In addition, Takahashi et al. (2021) showed that training a model on data with a low signal-to-noise ratio degrades the identification accuracy. Therefore, it is necessary to screen the training data, but it is not practical to screen a large amount of data visually. In this study, we developed a deep learning model that can classify noise, tremors, and earthquakes with a small amount of data, and clarified the basis for the model's decision.
For the seismic waveform data, we used a campaign seismic array set up by the National Institute of Advanced Industrial Science and Technology (AIST) in Iitaka-cho, Matsusaka City, Mie Prefecture. The model used in this study discriminates noise, tremors, and earthquakes by probability, using spectrogram images as input data. The model consists of three convolutional layers, two pooling layers, and three fully-connected layers. To reduce the number of model parameters, a GlobalAveragePooling layer was used between the convolutional and fully-connected layers. By training this model with 680 events of each noise, tremor, and earthquake selected visually, we succeeded in the classification with 99.7 % accuracy. The accuracy of the model was improved by setting the lower limit of the spectrum and replacing lower values with the lower limit value. This made the spectrogram clear and the trained model was improved. In addition, by using ScoreCam (Wang et al. 2019), we were able to visualize which part of the data was important in identifying events, indicating that our model is suitable as a discriminator. With a different lower limit of the spectrum from that for the training data, the model worked well with Hi-net data. Moreover, from a 1-hour data from the AIST array, our model detected tremors that were not recognized in a conventional method.
Prior studies on classifying noise, tremors, and ordinary earthquakes (hereinafter simply “earthquakes”) using deep learning (e.g., Nakano et al., 2019; Takahashi et al., 2021) worked well on their test data. However, the trained model is not explainable. In addition, Takahashi et al. (2021) showed that training a model on data with a low signal-to-noise ratio degrades the identification accuracy. Therefore, it is necessary to screen the training data, but it is not practical to screen a large amount of data visually. In this study, we developed a deep learning model that can classify noise, tremors, and earthquakes with a small amount of data, and clarified the basis for the model's decision.
For the seismic waveform data, we used a campaign seismic array set up by the National Institute of Advanced Industrial Science and Technology (AIST) in Iitaka-cho, Matsusaka City, Mie Prefecture. The model used in this study discriminates noise, tremors, and earthquakes by probability, using spectrogram images as input data. The model consists of three convolutional layers, two pooling layers, and three fully-connected layers. To reduce the number of model parameters, a GlobalAveragePooling layer was used between the convolutional and fully-connected layers. By training this model with 680 events of each noise, tremor, and earthquake selected visually, we succeeded in the classification with 99.7 % accuracy. The accuracy of the model was improved by setting the lower limit of the spectrum and replacing lower values with the lower limit value. This made the spectrogram clear and the trained model was improved. In addition, by using ScoreCam (Wang et al. 2019), we were able to visualize which part of the data was important in identifying events, indicating that our model is suitable as a discriminator. With a different lower limit of the spectrum from that for the training data, the model worked well with Hi-net data. Moreover, from a 1-hour data from the AIST array, our model detected tremors that were not recognized in a conventional method.