3:00 PM - 3:15 PM
[SSS05-06] Automatic detection of short-term slow slip events using GNSS data: An l1-trend filtering approach
Short-term slow slip events (SSEs) are fault slip events that occur at a slower rate than regular earthquakes and are observed by geodetic instruments such as the Global Navigation Satellite System (GNSS) and tiltmeters. The durations of short-term SSEs generally ranges from a few days to a few weeks (Obara and Kato, 2016), and SSEs occur beneath subduction zones (e.g., Sekine et al., 2010, Nishimura et al., 2013). Since SSEs have close connections to the other seismic activities such as megathrust earthquakes and low-frequency earthquakes (Kano et al., 2019), detection of SSEs is important. Nishimura et al. (2013) proposed a short-term SSE detection method based on model comparison in a moving window using the Akaike information criterion (AIC) and detected short-term SSEs in southwest Japan. This method has high detection accuracy when hyperparameters such as the length of a moving window are given properly, yet it is difficult to determine these parameters properly.
In this presentation, we propose a detection method from multiple GNSS stations based on l1 trend filtering. We apply the proposed method to GNSS data in western Shikoku, and present the results of inversion analysis using Markov Chain Monte Carlo (MCMC) method for estimating the fault models of the detected events. The proposed method considers short-term SSEs as the change points in GNSS displacement data, and quantifies the uncertainty of the change points using multiple observation points. The hyperparameters in the proposed method can be appropriately determined by using Mallows' Cp.
The proposed method conducts (a) trend filtering, (b) testing using neighboring observation stations, and (c) integrating the test results in a stepwise manner. We first fit a piecewise linear function to the observed time series by l1 trend filtering (Kim et al., 2009). The length of the interval before and after the node is automatically determined by using the nodes of the l1 trend filtering. We finally integrate the p-values of the tests for detection uncertainty at multiple observation stations. l1 trend filtering is a sparse estimation method, also called higher order total variation regularization, that accurately estimates piecewise polynomial function hidden in the input. In the subsequent tests, we use the segmentation points obtained by l1 trend filtering.
To validate the accuracy of the proposed method, we use GNSS data for western Shikoku during April 1, 2004 to March 31, 2009, and synthetic data generated from this data; GNSS common mode errors are removed by subtracting the average time series of three stations in the Goto Islands (Kano et al, 2019), and the two-dimensional displacement data are projected in the direction of subduction of the Philippine Sea Plate relative to the Amur Sea Plate (N45W). The comparison using synthetic data shows that the missed event rate and noise false positive rate of the spatial average of AIC vary with the window width, while the proposed method performs as well as the spatial average of AIC with an appropriate window width. We report the results of using the proposed method on real data and the fault model estimation using MCMC.
In this presentation, we propose a detection method from multiple GNSS stations based on l1 trend filtering. We apply the proposed method to GNSS data in western Shikoku, and present the results of inversion analysis using Markov Chain Monte Carlo (MCMC) method for estimating the fault models of the detected events. The proposed method considers short-term SSEs as the change points in GNSS displacement data, and quantifies the uncertainty of the change points using multiple observation points. The hyperparameters in the proposed method can be appropriately determined by using Mallows' Cp.
The proposed method conducts (a) trend filtering, (b) testing using neighboring observation stations, and (c) integrating the test results in a stepwise manner. We first fit a piecewise linear function to the observed time series by l1 trend filtering (Kim et al., 2009). The length of the interval before and after the node is automatically determined by using the nodes of the l1 trend filtering. We finally integrate the p-values of the tests for detection uncertainty at multiple observation stations. l1 trend filtering is a sparse estimation method, also called higher order total variation regularization, that accurately estimates piecewise polynomial function hidden in the input. In the subsequent tests, we use the segmentation points obtained by l1 trend filtering.
To validate the accuracy of the proposed method, we use GNSS data for western Shikoku during April 1, 2004 to March 31, 2009, and synthetic data generated from this data; GNSS common mode errors are removed by subtracting the average time series of three stations in the Goto Islands (Kano et al, 2019), and the two-dimensional displacement data are projected in the direction of subduction of the Philippine Sea Plate relative to the Amur Sea Plate (N45W). The comparison using synthetic data shows that the missed event rate and noise false positive rate of the spatial average of AIC vary with the window width, while the proposed method performs as well as the spatial average of AIC with an appropriate window width. We report the results of using the proposed method on real data and the fault model estimation using MCMC.