5:15 PM - 6:45 PM
[SSS03-P02] Monitoring of shallow tremors along the Nankai Trough and identification of false detections
Keywords:shallow tremor, T-phase, machine learning
Monitoring shallow tremors along the Nankai Trough is crucial, as they are closely related to the slow slip events occurring in these shallow regions. To this end, we have developed a hybrid method for determining the location of these tremors (Tamaribuchi et al., 2022) and currently conduct near real-time monitoring with hourly updates. However, while we successfully detected tremor activities in September 2023, there were numerous false detections of T-phase waves from earthquakes near Torishima occurring in October 2023. This highlights the challenge of reducing misclassification due to distant earthquakes and T-phases in real-time monitoring. This presentation attempts to distinguish between tremors and false detections using machine learning on characteristic features from the obtained tremor catalog, following the method of Tamaribuchi et al. (2023).
The data used was the shallow tremor catalog compiled by Tamaribuchi et al. (2022) from April 2016 to September 2021. Of the 40,975 total events determined by the hybrid method, 6,529 were extracted as tremors after clustering processing, based on conditions occurring four or more events within 12 hours and 20 km of the epicenter. These clustered tremors are considered concentrated over 1-3 months with several years’ intervals. The remaining 23,127 events, excluding days with one or more clustered tremors and the preceding and following days, were defined as false detections. The training and test data were split into a 7:3 ratio by period, with events up to October 2020 as training data (20,964 events) and those from November 2020 onward as test data (8,692 events).
We extracted 15 features, including depth, location errors (latitude, longitude, depth), energy rate, maximum likelihood, maximum and median of envelope correlation coefficient, median of apparent velocity, maximum and median amplitude, ratio of maximum to minimum amplitude, number of correlation coefficient pairs, and number of stations used for amplitude. According to the classification results using the pycaret library (Ali, 2020), we achieved an accuracy of 0.942 for the training data and 0.827 for the test data. The feature importance showed a tendency to particularly prioritize values of location errors.
Next, using this model, we classified tremors from April 1, 2016, to February 7, 2024, and out of 57,668 total events, 9,574 were labeled as "tremors." This result indicates the potential effectiveness of understanding isolated tremors without the need for comparison with (fast) earthquake catalog or clustering. We plan to continue our efforts to further improve the accuracy in the future.
References
Tamaribuchi, K., M. Ogiso, A. Noda (2022), J Geophys Res, doi:10.1029/2022JB024403
Tamaribuchi, K., S. Kudo, K. Shimojo, F. Hirose (2023), Earth Planets Space, doi: 10.1186/s40623-023-01915-3
The data used was the shallow tremor catalog compiled by Tamaribuchi et al. (2022) from April 2016 to September 2021. Of the 40,975 total events determined by the hybrid method, 6,529 were extracted as tremors after clustering processing, based on conditions occurring four or more events within 12 hours and 20 km of the epicenter. These clustered tremors are considered concentrated over 1-3 months with several years’ intervals. The remaining 23,127 events, excluding days with one or more clustered tremors and the preceding and following days, were defined as false detections. The training and test data were split into a 7:3 ratio by period, with events up to October 2020 as training data (20,964 events) and those from November 2020 onward as test data (8,692 events).
We extracted 15 features, including depth, location errors (latitude, longitude, depth), energy rate, maximum likelihood, maximum and median of envelope correlation coefficient, median of apparent velocity, maximum and median amplitude, ratio of maximum to minimum amplitude, number of correlation coefficient pairs, and number of stations used for amplitude. According to the classification results using the pycaret library (Ali, 2020), we achieved an accuracy of 0.942 for the training data and 0.827 for the test data. The feature importance showed a tendency to particularly prioritize values of location errors.
Next, using this model, we classified tremors from April 1, 2016, to February 7, 2024, and out of 57,668 total events, 9,574 were labeled as "tremors." This result indicates the potential effectiveness of understanding isolated tremors without the need for comparison with (fast) earthquake catalog or clustering. We plan to continue our efforts to further improve the accuracy in the future.
References
Tamaribuchi, K., M. Ogiso, A. Noda (2022), J Geophys Res, doi:10.1029/2022JB024403
Tamaribuchi, K., S. Kudo, K. Shimojo, F. Hirose (2023), Earth Planets Space, doi: 10.1186/s40623-023-01915-3