5:15 PM - 6:45 PM
[SCG50-P06] Automated detection and hypocenter determination in tectonic tremors using convolutional neural network
Keywords:Convolutional Neural Network, Tectonic tremor, Hypocenter determination, Semblance analysis
・Introduction
Semblance analysis (Neidell and Taner, 1971), which was used in tremor analysis, often may detect noise and earthquakes as tremors. Furthermore, the semblance value of tremor waveforms with high noise levels may not exceed the threshold value and may not be detected as tremors. Recently, convolutional neural networks (CNNs) have been used for highly accurate and efficient processing of tremors in large amounts of seismic waveform data. CNN methods can classify noise, tremors, and earthquakes with high accuracy (e.g., Nakano et al., 2019). Rouet-Leduc et al. (2020) showed that deep learning interpretation methods can extract characteristics of time-frequency components of tremors. In this study, we develop an automated method for the classification of tremors, earthquakes, and noises, and for the determination of tremor source location, based on the method of Hulbert et al. (2022) for extracting characteristic time-frequency components of tremors and envelope cross-correlation approach.
・Data and Methods
In this study, we used three-component velocity waveform data obtained from the array observation network installed on the Kii Peninsula by the National Institute of Advanced Industrial Science and Technology (AIST) between April 2011 and December 2015. 1-minute spectrograms of velocity waveform data with a 2-8 Hz bandpass filter were used as input data for the CNN. The training and test data were 162,100 (April 2011 to March 2014) and 16,300 (April 2014 to December 2015), respectively. Stations with tremor probabilities greater than 0.9 were selected to create interpretation maps. Interpretation maps were averaged over stations for each event. We used the average interpretation map as a filter and reconstructed waveforms based on the average interpretation map and a spectrogram of each station for each event. Semblance analysis was performed using the reconstructed waveforms, and grids with a maximum semblance value of 0.3 or greater were defined as the location of tectonic tremors.
・Results and Discussion
The trained CNN model achieved high accuracy, with both goodness-of-fit and repeatability rates exceeding 97%. Regarding events listed in the tremor catalog (Imanishi et al., 2011) and JMA catalog from April 2011 to December 2015, about 70% of the earthquakes with epicentral distances of 200 km or greater were correctly classified. However, we recognize that the classification accuracy of tremors with epicentral distances greater than 40 km decreased, but even with epicentral distances greater than 80 km, they could be identified as tremors when the waveforms have a high S/N. A S/N ratio of at least 2 is required to classify tremors accurately. Integrated Gradients (IG) (Sundararajan et al., 2017) and Score-CAM (SC) (Wang et al., 2020) were used as CNN interpretation methods for comparison. The results showed that both methods are based on areas with large amplitude values of spectrograms. This is consistent with the results of Liu et al. (2019), who showed that the maximum amplitude and number of peaks are important in discriminating between noise and tremors in the k-neighborhood method. We conducted semblance analysis for reconstructed waveform data from July 1 to July 15, 2014. The result showed that IG based method can determine 4 times as many locations than a conventional method based on a bandpass filtering of 2–8Hz. The average SC and IG filtering can locate two times as many tremors compared to the filtering without averaging, indicating that the averaging procedure is effective in locating tremors. However, about 20% of tremors cannot be located despite being classified as tremors by CNN. In this presentation, we will discuss the characteristics of these tremors and methods to determine their locations.
Semblance analysis (Neidell and Taner, 1971), which was used in tremor analysis, often may detect noise and earthquakes as tremors. Furthermore, the semblance value of tremor waveforms with high noise levels may not exceed the threshold value and may not be detected as tremors. Recently, convolutional neural networks (CNNs) have been used for highly accurate and efficient processing of tremors in large amounts of seismic waveform data. CNN methods can classify noise, tremors, and earthquakes with high accuracy (e.g., Nakano et al., 2019). Rouet-Leduc et al. (2020) showed that deep learning interpretation methods can extract characteristics of time-frequency components of tremors. In this study, we develop an automated method for the classification of tremors, earthquakes, and noises, and for the determination of tremor source location, based on the method of Hulbert et al. (2022) for extracting characteristic time-frequency components of tremors and envelope cross-correlation approach.
・Data and Methods
In this study, we used three-component velocity waveform data obtained from the array observation network installed on the Kii Peninsula by the National Institute of Advanced Industrial Science and Technology (AIST) between April 2011 and December 2015. 1-minute spectrograms of velocity waveform data with a 2-8 Hz bandpass filter were used as input data for the CNN. The training and test data were 162,100 (April 2011 to March 2014) and 16,300 (April 2014 to December 2015), respectively. Stations with tremor probabilities greater than 0.9 were selected to create interpretation maps. Interpretation maps were averaged over stations for each event. We used the average interpretation map as a filter and reconstructed waveforms based on the average interpretation map and a spectrogram of each station for each event. Semblance analysis was performed using the reconstructed waveforms, and grids with a maximum semblance value of 0.3 or greater were defined as the location of tectonic tremors.
・Results and Discussion
The trained CNN model achieved high accuracy, with both goodness-of-fit and repeatability rates exceeding 97%. Regarding events listed in the tremor catalog (Imanishi et al., 2011) and JMA catalog from April 2011 to December 2015, about 70% of the earthquakes with epicentral distances of 200 km or greater were correctly classified. However, we recognize that the classification accuracy of tremors with epicentral distances greater than 40 km decreased, but even with epicentral distances greater than 80 km, they could be identified as tremors when the waveforms have a high S/N. A S/N ratio of at least 2 is required to classify tremors accurately. Integrated Gradients (IG) (Sundararajan et al., 2017) and Score-CAM (SC) (Wang et al., 2020) were used as CNN interpretation methods for comparison. The results showed that both methods are based on areas with large amplitude values of spectrograms. This is consistent with the results of Liu et al. (2019), who showed that the maximum amplitude and number of peaks are important in discriminating between noise and tremors in the k-neighborhood method. We conducted semblance analysis for reconstructed waveform data from July 1 to July 15, 2014. The result showed that IG based method can determine 4 times as many locations than a conventional method based on a bandpass filtering of 2–8Hz. The average SC and IG filtering can locate two times as many tremors compared to the filtering without averaging, indicating that the averaging procedure is effective in locating tremors. However, about 20% of tremors cannot be located despite being classified as tremors by CNN. In this presentation, we will discuss the characteristics of these tremors and methods to determine their locations.