1:45 PM - 3:15 PM
[SCG55-P02] Development of a machine learning model to detect short-term slow slip events from tilt records
Keywords:Slow earthquakes, Subduction zones, Convolutional Neural Network, Crustal deformation, Strain accumulation process
Short-term slow slip events (SSEs) are one of slow earthquakes in which slow fault slip lasts for days to weeks. These occur mainly on plate interfaces in subduction zones and are observed by geodetic instruments, such as tilt meter, strain meter, and GNSS. These are considered to play an important role in the strain accumulation process of megathrust earthquakes. However, there may be undetected events because short-term SSEs are detected by human inspection in most cases. A more complete SSE catalog would be useful for quantitative discussion on the strain budget around the megathrust rupture zones. In this study, we develop a machine learning model to detect short-term SSEs from tilt records.
The developed machine learning model takes a time series of one component of tilt records at a station as input and outputs three probabilities at each time epoch: (1) start of an SSE ; (2) end of an SSE; and (3) other than these. This model is a fully convolutional neural network based on PhaseNet (Zhu and Beroza, 2019) which picks P- and S-wave arrivals from seismograms and U-Net (Ronneberger et al., 2015) which performs image segmentation for biomedical images. We used a synthetic dataset for training that consists of a ramp function and a random noise simulating a tilt deformation due to an SSE. The start and end times and the magnitude of the ramp were given randomly. We also used a noise-only dataset without a ramp because a typical recurrence interval of short-term SSEs is several months and an SSE signal is not contained in all of the segments of observed tilt time series. After the training phase, we evaluated the performance of the trained model with a test dataset that is a synthetic one similar to the training dataset. As a result, the model detects 76% of SSE start time epochs (the start time of a ramp) and 72% of SSE end time epochs (the end time of a ramp). The model judges almost all noise-only data correctly. The accuracy decreases as decreasing the signal-to-noise ratio (S/N) of the test data, indicating that it becomes difficult to detect a small SSE signal that is buried in larger noise.
We applied this trained model to the high-sensitivity accelerometer (tiltmeter) time-series records installed at National Research Institute for Earth Science and Disaster Resilience (NIED) Hi-net stations located in western Shikoku, Japan to confirm the applicability to real data. We used 10 years (2003/01/01 to 2012/12/31) of tilt time series data. Compared the detected SSEs by the model with the short-term SSE catalog of Hirose and Kimura (2020), we successfully detected 20 of the start or end time of SSEs out of 32 events during the studied period. Furthermore, we obtained 99 possible SSE detections without corresponding reported events.
The obtained performance (63% =20/32) is reasonably good, but there are many possible improvements to gain a better performance, for example, multiple components used, better treatments for abnormal and missing observations contained in the real data, synthetic data with more similar noise characteristics to real data, and so on. The improved model with better detection than human inspection would be useful to better understand the strain accumulation process in subduction zones quantitatively.
The developed machine learning model takes a time series of one component of tilt records at a station as input and outputs three probabilities at each time epoch: (1) start of an SSE ; (2) end of an SSE; and (3) other than these. This model is a fully convolutional neural network based on PhaseNet (Zhu and Beroza, 2019) which picks P- and S-wave arrivals from seismograms and U-Net (Ronneberger et al., 2015) which performs image segmentation for biomedical images. We used a synthetic dataset for training that consists of a ramp function and a random noise simulating a tilt deformation due to an SSE. The start and end times and the magnitude of the ramp were given randomly. We also used a noise-only dataset without a ramp because a typical recurrence interval of short-term SSEs is several months and an SSE signal is not contained in all of the segments of observed tilt time series. After the training phase, we evaluated the performance of the trained model with a test dataset that is a synthetic one similar to the training dataset. As a result, the model detects 76% of SSE start time epochs (the start time of a ramp) and 72% of SSE end time epochs (the end time of a ramp). The model judges almost all noise-only data correctly. The accuracy decreases as decreasing the signal-to-noise ratio (S/N) of the test data, indicating that it becomes difficult to detect a small SSE signal that is buried in larger noise.
We applied this trained model to the high-sensitivity accelerometer (tiltmeter) time-series records installed at National Research Institute for Earth Science and Disaster Resilience (NIED) Hi-net stations located in western Shikoku, Japan to confirm the applicability to real data. We used 10 years (2003/01/01 to 2012/12/31) of tilt time series data. Compared the detected SSEs by the model with the short-term SSE catalog of Hirose and Kimura (2020), we successfully detected 20 of the start or end time of SSEs out of 32 events during the studied period. Furthermore, we obtained 99 possible SSE detections without corresponding reported events.
The obtained performance (63% =20/32) is reasonably good, but there are many possible improvements to gain a better performance, for example, multiple components used, better treatments for abnormal and missing observations contained in the real data, synthetic data with more similar noise characteristics to real data, and so on. The improved model with better detection than human inspection would be useful to better understand the strain accumulation process in subduction zones quantitatively.