Lastly, we validate our model by applying it to detect slow-slip events in Cascadia with a real data set. Here we apply RNNs to a classification problem where the final output represents a category (whether a data point contains transient signals or not are the two categories in our study). This shows the feasibility of applying RNNs to regression problems in the geodetic field where continuous values at the next moment are predicted based on previous data. Yamaga and Mitsui ( 2019) introduced a machine learning approach, based on the RNN, to forecast postseismic deformation and constrain its characteristics. In this study, we propose a new method for transient detection in geodetic time series in a sequential manner based on a Recurrent Neural Network (RNN). With the concept of a sliding window, only the data within the time window (size of the CNN layer) are considered instead of the entire time series, which we think can be particularly useful to solve the questions. All these methods used models with a Convolutional Neural Network (CNN) layer being the first later to scan the time series. More recently, machine learning methods have been used as feature extraction tools for data mining in many other geophysical fields, such as phase picking in seismic waves (Ross et al., 2018 Zhu & Beroza, 2018), detecting volcanic deformation in satellite images (Anantrasirichai et al., 2018), and transient detection in sea floor pressure data (He et al., 2020). This makes us think about whether we can find other metrics or features like the RSI value that are useful for transient signal detection. This single feature, the RSI value, was demonstrated to be efficient for detecting transient signals in GPS time series. They used the RSI and a probabilistic approach to calculate the probability that each data point was part of a transient (termed the transient probability). To obtain more flexibility, Crowell et al. ( 2016) proposed a single-station detection method based on the Relative Strength Index (RSI), an indicator of the divergence from the typical trend of the time series. Many of the above methods generally have better performance with a network of GPS stations, exploiting the spatial coherence across multiple stations. Examples include, but are not limited to, methods built upon some kind of Kalman filter (Bekaert et al., 2016 Ji et al., 2017), decomposition analysis (Dong et al., 2006 Walwer et al., 2016), template matching (Okada et al., 2022 Riel et al., 2014 Rousset et al., 2017) or least squares with some kind of constraints (Nishimura et al., 2013 Yano & Kano, 2022). Such large data sets provide new research opportunities for geophysical problems, but pose challenges in how to efficiently extract signals of interest, such as transient deformations due to various geophysical processes.ĭuring the past two decades, many methods have been proposed to detect transient deformation in GPS time series. For example, the Nevada Geodetic Laboratory currently acquires geodetic GPS data from more than 17,000 stations and processes daily solution of more than 10,000 stations every week (Blewitt et al., 2018). There has been a tremendous expansion of automated processing and open data from the Global Positioning System (GPS). In general, our ML model detects more stations likely to be associated with nearby slow-slip events than the RSI model, especially if there are data gaps in the timeseries. As a benchmark, we compared our results in detail with those based on the Relative Strength Index (RSI). Based on our detection results, the spatial extent, duration, and migration of the major events are consistent with previous studies. The specific model is first trained and validated using synthetic data, and then used to detect slow-slip events on real Cascadia data. As a case study, we apply our method to detect slow-slip events in Cascadia between 20. Unlike most previous studies using a sliding window technique, our model uses a single data point of the entire time series as sequential input and directly outputs the transient probability for all points in the time series. Inspired by recent studies using various machine learning methods on different types of time series data (e.g., seismic, sea floor pressure), this study proposes a simple machine learning method, based on the recurrent neural network approach, for transient deformation detection in Global Positioning System time series.
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |