5:15 PM - 6:45 PM
[SCG40-P10] Automatic slow slip signal detection based on machine learning of the GNSS displacement time series of southwest Japan
Keywords:signal detection, machine learning, slow slip events, GNSS time series
Detection of slow slip events (SSE) is essential to precisely evaluate the friction properties, current strain budget, and the potential of future large earthquake of the plate interface. To detect transient signal hidden in the displacement time series by Global Navigation Satellite System (GNSS) observation, various mathematical techniques have been utilized such as template matching (e.g., Okada et al., 2022), sparse estimation (e.g., Yano and Kano, 2022), and decomposition analysis (e.g., Walwer et al., 2016). Under the recent increase of GNSS sites, the requirement of such automated and objective detection method is rising. On the other hand, unlike the analysis of seismic data, signal detection based on machine learning has still rarely been applied to GNSS time series. It is largely related to smaller number of SSEs compared to regular earthquakes, causing the difficulty of acquiring sufficient amount of training data. Although some previous studies (e.g., Xue and Freymüller, 2023) made use of synthetic training data, it is not easy to completely replicate the complex spatiotemporal variation of noise in the real data.
According to the above situation, we aim to carry out the SSE detection based on machine learning of the real GNSS observations of southwest Japan. In this presentation, we first introduce the preliminary results of single site detection. We used daily coordinate time series of 770 GEONET sites covering southwest Japan obtained by precise point positioning with ambiguity resolution analysis. Then, we generated approximately 16000 training data based on the detection catalog by Okada et al. (2022), containing 284 short-term SSEs in 23 years from 1997 to 2020. We set the length of time window as 121 epoch. Then, we applied the model architecture following the Generalized Phase Detection (Ross et al., 2018), which was originally proposed for seismic wave detection. The output will be the probability whether or not the time window includes signal. We adopted the binary cross entropy for the loss function. To examine the stability of training, we independently repeated the training 10 times with randomly changing the initial model parameters.
As a result, the training converged generally within 30–50 epochs and we obtained accuracy of 80–90%. Okada et al. (2022) classified the 284 events into “Class 1” and “Class 2”, representing the reliable events and relatively suspicious events, respectively. Our results showed 1.5–3.0 times higher ratio of false negative for Class 2 signal. While approximately 90 % of the test data were correctly judged 8–10 times among 10 times of training, rest 10 % showed unstable judgements. If we pick up the signal data with less than 7 times of correct judgements, they widely distribute in space and time, but concentrate on the cases with smaller displacement. We also calculated the true positive ratio and false positive ratio for the data of each one observation site. True positive ratio is higher mainly in the western Shikoku, where the larger events occur most frequently. Contrastingly, spatial distribution of the false positive ratio is very complicated and shows higher values also in the sites with large number of trained data.
In the presentation, we explain the details of the above results and discuss its comparison between noise characteristics around southwest Japan, to propose future improvement toward robust training and precise detection.
According to the above situation, we aim to carry out the SSE detection based on machine learning of the real GNSS observations of southwest Japan. In this presentation, we first introduce the preliminary results of single site detection. We used daily coordinate time series of 770 GEONET sites covering southwest Japan obtained by precise point positioning with ambiguity resolution analysis. Then, we generated approximately 16000 training data based on the detection catalog by Okada et al. (2022), containing 284 short-term SSEs in 23 years from 1997 to 2020. We set the length of time window as 121 epoch. Then, we applied the model architecture following the Generalized Phase Detection (Ross et al., 2018), which was originally proposed for seismic wave detection. The output will be the probability whether or not the time window includes signal. We adopted the binary cross entropy for the loss function. To examine the stability of training, we independently repeated the training 10 times with randomly changing the initial model parameters.
As a result, the training converged generally within 30–50 epochs and we obtained accuracy of 80–90%. Okada et al. (2022) classified the 284 events into “Class 1” and “Class 2”, representing the reliable events and relatively suspicious events, respectively. Our results showed 1.5–3.0 times higher ratio of false negative for Class 2 signal. While approximately 90 % of the test data were correctly judged 8–10 times among 10 times of training, rest 10 % showed unstable judgements. If we pick up the signal data with less than 7 times of correct judgements, they widely distribute in space and time, but concentrate on the cases with smaller displacement. We also calculated the true positive ratio and false positive ratio for the data of each one observation site. True positive ratio is higher mainly in the western Shikoku, where the larger events occur most frequently. Contrastingly, spatial distribution of the false positive ratio is very complicated and shows higher values also in the sites with large number of trained data.
In the presentation, we explain the details of the above results and discuss its comparison between noise characteristics around southwest Japan, to propose future improvement toward robust training and precise detection.