5:15 PM - 7:15 PM
[SCG45-P31] Identifying Shallow Tectonic Tremors along the Nankai Trough: A Machine Learning-Based Analysis of N-net Data
Keywords:Tectonic tremor, Machine learning, N-net, Hyuganada
Monitoring slow earthquakes is crucial for understanding stress accumulation and release processes on subducting plate boundaries. The N-net (Aoi et al., 2023), a network of cable-type ocean-bottom seismometers deployed along the Nankai Trough, has been in operation from July 2024. This network enhances monitoring capabilities of slow earthquakes, particularly tectonic tremors, adjacent to megathrust seismogenic zone in the Nankai region. After its deployment, two large earthquakes with moment magnitudes of 7.0 and 6.7 occurred in Hyuganada region, at the western edge of the Nankai Trough, on August 8, 2024, and January 13, 2025, respectively. These events highlight the critical importance of rapidly developing a slow earthquake monitoring system using the N-net. Therefore, we developed a machine learning-based tremor monitoring system using the N-net data.
We analyzed 18 N-net station data from July 1, 2024, to January 31, 2025. A machine learning-based tremor monitoring system, previously developed for the Japan Trench (Sagae et al., 2025, preprint), was directly applied to the N-net data. The monitoring system consists of three main steps. First, Discriminator for Earthquake and Tremor (DiET) model, trained using all 150 S-net stations, classifies input spectrograms as earthquake, tremor, or noise. Second, Graph-based Associator with Signal Probability (GrASP) associates tremor-detected station groups every minute. Finally, hypocenter determination is performed using a hybrid method based on envelope cross-correlation and radiated energy ratio. In this study, we made two minor modifications to the original system: (1) we adjusted the k-nearest neighbor graph used in GrASP to k=4 to account for the number of N-net stations, and (2) we performed hypocenter determination solely using envelope cross-correlation. The rest of the original system remains unchanged.
We successfully detected 3,286 tremors during the analyzed period. Notably, the DiET model, trained on a different network, correctly predicted tremor signals in the N-net. These tremors occurred in the Hyuganada region, and the tremor distribution correspond to previous studies (Yamashita et al., 2015, 2021). Additionally, the tremors were located on the up-dip side of the aftershock distribution of the 2024 and 2025 Hyuganada earthquakes. Two active tremor episodes were identified during our analysis period. One became active after the 2024 Hyuganada earthquake, while the other became active three days before the 2025 Hyuganada earthquake. The latter episode suggests that a slow slip event preceding the 2025 Hyuganada earthquake may have occurred on the up-dip side of the earthquake. This presentation will continue the monitoring and report the results.
Acknowledgements: We used N-net velocity waveform data (https://doi.org/10.17598/nied.0029). This research was supported by JSPS KAKENHI Grant Number JP21H05205 and JP21H05203 in Grant-in-Aid for Transformative Research Areas (A) “Science of Slow-to-Fast Earthquakes”.
We analyzed 18 N-net station data from July 1, 2024, to January 31, 2025. A machine learning-based tremor monitoring system, previously developed for the Japan Trench (Sagae et al., 2025, preprint), was directly applied to the N-net data. The monitoring system consists of three main steps. First, Discriminator for Earthquake and Tremor (DiET) model, trained using all 150 S-net stations, classifies input spectrograms as earthquake, tremor, or noise. Second, Graph-based Associator with Signal Probability (GrASP) associates tremor-detected station groups every minute. Finally, hypocenter determination is performed using a hybrid method based on envelope cross-correlation and radiated energy ratio. In this study, we made two minor modifications to the original system: (1) we adjusted the k-nearest neighbor graph used in GrASP to k=4 to account for the number of N-net stations, and (2) we performed hypocenter determination solely using envelope cross-correlation. The rest of the original system remains unchanged.
We successfully detected 3,286 tremors during the analyzed period. Notably, the DiET model, trained on a different network, correctly predicted tremor signals in the N-net. These tremors occurred in the Hyuganada region, and the tremor distribution correspond to previous studies (Yamashita et al., 2015, 2021). Additionally, the tremors were located on the up-dip side of the aftershock distribution of the 2024 and 2025 Hyuganada earthquakes. Two active tremor episodes were identified during our analysis period. One became active after the 2024 Hyuganada earthquake, while the other became active three days before the 2025 Hyuganada earthquake. The latter episode suggests that a slow slip event preceding the 2025 Hyuganada earthquake may have occurred on the up-dip side of the earthquake. This presentation will continue the monitoring and report the results.
Acknowledgements: We used N-net velocity waveform data (https://doi.org/10.17598/nied.0029). This research was supported by JSPS KAKENHI Grant Number JP21H05205 and JP21H05203 in Grant-in-Aid for Transformative Research Areas (A) “Science of Slow-to-Fast Earthquakes”.