4:00 PM - 4:15 PM
[SCG60-07] Epicenter estimation of tectonic tremors in the nankai subduction zone from a single station using deep learning
Keywords:Slow earthquake, Tectonic tremor, Location estimation, Single station
Tectonic tremors, which are low-frequency seismic phenomena frequently linked to slow slip events (SSE), are considered to play a key role in unraveling the generation process of large earthquakes. However, conventional techniques, such as envelope cross-correlation, rely on a limited set of waveform features, often leading to the omission of crucial information hidden within the seismic data. Consequently, these methods suffer from reduced tremor detectability. To overcome these limitations, we have developed a convolutional neural network (CNN) classifier that differentiates between tremors, earthquakes, and noise, as well as a Vision Transformer-based epicenter locater for tremors.
In this study, we used waveform data recorded at from 129 Hi-net stations in the Nankai subduction zone, spanning the period from 2008 to 2016. The CNN classifier was trained using one-minute segments of labeled data, with data from 2016 aside for validation. The epicenter locator outputs parameters based on a normal distribution, specifically the mean, standard deviation, and correlation coefficients of relative latitude and longitude, derived from single-station velocity waveforms as input.
For stations included in the training dataset, the epicenter locator reduced the average epicenter error to approximately 5 km, a substantial improvement compared to the 28 km associated with random sampling. However, we observed a marked decrease in performance when stations not included in the training dataset were tested, suggesting limited regional generalizability. In cases involving noise and earthquake waveforms, the predicted epicenters were distributed radially outward from the station.
By integrating tremor probabilities (more than 0.9) from the CNN classifier with thresholds applied to the predicted standard deviations, we effectively filtered out the majority of non-tremor signals. When applied to continuous seismic data, our integrated approach yielded a consistent epicentral distribution, which aligns closely with an existing tremor catalog and successfully reproduced the characteristic tremor migration patterns. Furthermore, predictions from multiple stations during the same time periods consistently converged on coherent epicenter locations, indicating that combining individual station estimates may provide a reliable framework for multi-station tremor monitoring.
In this study, we used waveform data recorded at from 129 Hi-net stations in the Nankai subduction zone, spanning the period from 2008 to 2016. The CNN classifier was trained using one-minute segments of labeled data, with data from 2016 aside for validation. The epicenter locator outputs parameters based on a normal distribution, specifically the mean, standard deviation, and correlation coefficients of relative latitude and longitude, derived from single-station velocity waveforms as input.
For stations included in the training dataset, the epicenter locator reduced the average epicenter error to approximately 5 km, a substantial improvement compared to the 28 km associated with random sampling. However, we observed a marked decrease in performance when stations not included in the training dataset were tested, suggesting limited regional generalizability. In cases involving noise and earthquake waveforms, the predicted epicenters were distributed radially outward from the station.
By integrating tremor probabilities (more than 0.9) from the CNN classifier with thresholds applied to the predicted standard deviations, we effectively filtered out the majority of non-tremor signals. When applied to continuous seismic data, our integrated approach yielded a consistent epicentral distribution, which aligns closely with an existing tremor catalog and successfully reproduced the characteristic tremor migration patterns. Furthermore, predictions from multiple stations during the same time periods consistently converged on coherent epicenter locations, indicating that combining individual station estimates may provide a reliable framework for multi-station tremor monitoring.