5:15 PM - 7:15 PM
[SCG60-P02] Detecting and locating tectonic tremors in the Nankai subduction zone using deep learning
Keywords:Tectonic tremor, Convolutional neural network, slow earthquake, Locating epicenter
1.Introduction
In the Nankai subduction zone, a variety of slow-slip phenomena occur. Tectonic tremors (hereinafter referred as tremor) are a useful indicator of slow-slip events. The envelope correlation method, based on the similarity of waveforms across seismic stations, is widely employed to detect tremors. However, this method is prone to misidentifying regular earthquakes or noise as tremors. To mitigate these false detections, manual inspections and/or additional post-processing steps are typically required. Furthermore, the detection performance of this method deteriorates during periods of elevated tremor activity or high noise levels.
In this study, we developed a novel approach for detecting and locating tremors using deep learning aiming to minimize the false positives and missed detections that have been persistent challenges in conventional methods.
2.Data and Method
This study used 129 Hi-net stations installed in the Nankai subduction zone by National Research Institute for Earth Science and Disaster Resilience (NIED). The analyzed period was from 2008 to September 2016.
The proposed framework utilizes two types of convolutional neural networks (CNNs): a classification CNN and a regression CNN. The classification CNN processes spectrograms as inputs and outputs the probabilities for three categories: noise, tremors, and earthquakes. The regression CNN, in turn, estimates the tremor epicenters by leveraging spatial distributions of the probability values produced by the classification CNN along with amplitude data. This method assumes that the spatial distributions of high tremor probabilities and amplitudes are uniquely associated with individual tremor locations. Consequently, the regression CNN learns these relationships as a supervised regression task.
For the classification CNN, we trained on a total of 281,046 data, 93,682 for noise, tremors, and earthquakes, respectively. For the regression CNN, the tremor location determined by Mizuno and Ide (2019) was used as the correct location, and 53,600 tremors were used for training.
3.Results and Discussion
The classification CNN achieved evaluation metrics exceeding 95%, including accuracy, recall, and precision.
To further enhance prediction accuracy and quantify uncertainty, the regression CNN incorporated ensemble predictions across 100 training iterations. The mean error distance between the predicted tremor locations and those in an existing tremor catalog for the test dataset was 6.2 km, with a median error of 3.8 km, demonstrating high accuracy even without incorporating phase information. Moreover, applying this method to previously unseen continuous data from 2012 identified tremors at a frequency seven times higher than the conventional approach. Notably, periods of increased detection coincided with intervals of heightened tremor activity, underscoring the method’s ability to substantially improve detection performance during high-activity periods where conventional methods often faltered. Furthermore, the estimated tremor locations closely reproduced the migration pattern of tremors reported in previous studies, further demonstrating the robustness and high performance of the proposed method.
In the Nankai subduction zone, a variety of slow-slip phenomena occur. Tectonic tremors (hereinafter referred as tremor) are a useful indicator of slow-slip events. The envelope correlation method, based on the similarity of waveforms across seismic stations, is widely employed to detect tremors. However, this method is prone to misidentifying regular earthquakes or noise as tremors. To mitigate these false detections, manual inspections and/or additional post-processing steps are typically required. Furthermore, the detection performance of this method deteriorates during periods of elevated tremor activity or high noise levels.
In this study, we developed a novel approach for detecting and locating tremors using deep learning aiming to minimize the false positives and missed detections that have been persistent challenges in conventional methods.
2.Data and Method
This study used 129 Hi-net stations installed in the Nankai subduction zone by National Research Institute for Earth Science and Disaster Resilience (NIED). The analyzed period was from 2008 to September 2016.
The proposed framework utilizes two types of convolutional neural networks (CNNs): a classification CNN and a regression CNN. The classification CNN processes spectrograms as inputs and outputs the probabilities for three categories: noise, tremors, and earthquakes. The regression CNN, in turn, estimates the tremor epicenters by leveraging spatial distributions of the probability values produced by the classification CNN along with amplitude data. This method assumes that the spatial distributions of high tremor probabilities and amplitudes are uniquely associated with individual tremor locations. Consequently, the regression CNN learns these relationships as a supervised regression task.
For the classification CNN, we trained on a total of 281,046 data, 93,682 for noise, tremors, and earthquakes, respectively. For the regression CNN, the tremor location determined by Mizuno and Ide (2019) was used as the correct location, and 53,600 tremors were used for training.
3.Results and Discussion
The classification CNN achieved evaluation metrics exceeding 95%, including accuracy, recall, and precision.
To further enhance prediction accuracy and quantify uncertainty, the regression CNN incorporated ensemble predictions across 100 training iterations. The mean error distance between the predicted tremor locations and those in an existing tremor catalog for the test dataset was 6.2 km, with a median error of 3.8 km, demonstrating high accuracy even without incorporating phase information. Moreover, applying this method to previously unseen continuous data from 2012 identified tremors at a frequency seven times higher than the conventional approach. Notably, periods of increased detection coincided with intervals of heightened tremor activity, underscoring the method’s ability to substantially improve detection performance during high-activity periods where conventional methods often faltered. Furthermore, the estimated tremor locations closely reproduced the migration pattern of tremors reported in previous studies, further demonstrating the robustness and high performance of the proposed method.