Discontinuity Detection in GNSS Station Coordinate Time Series Using Machine Learning

Global navigation satellite systems (GNSS) provide globally distributed station coordinate time series that can be used for a variety of applications such as the definition of a terrestrial reference frame. A reliable estimation of the coordinate time series trends gives valuable information about station movements during the measured time period. Detecting discontinuities of various origins in such time series is crucial for accurate and robust velocity estimation. At present, there is no fully automated standard method for detecting discontinuities. Instead, discontinuity-catalogues are frequently used, which provide information about when a device was changed or an earthquake occurred. However, it is known that these catalogues suffer from incompleteness. This study investigates the suitability of machine learning classification algorithms that are fully data-driven to detect discontinuities caused by earthquakes in station coordinate time series without the need for external information. For this study, Japan was selected as a testing area. Ten different machine learning algorithms have been tested. It is found that Random Forest achieves the best performance with an F1 score of 0.77, a recall of 0.78, and a precision of 0.76. Overall, 525 of 565 recorded earthquakes in the test data were correctly classified. It is further highlighted that splitting the time series into chunks of 21 days leads to the best performance. Furthermore, it is beneficial to combine the three (normalized) components of the GNSS solution into one sample, and that adding the value range as an additional feature improves the result. Thus, this work demonstrates how it is possible to use machine learning algorithms to detect discontinuities in GNSS time series.

Automated summary

This study investigates the use of machine learning algorithms to detect discontinuities caused by earthquakes in GNSS time series, and finds that Random Forest algorithm performs the best. Splitting the time series into chunks of 21 days, combining the components into one sample, and adding value range as an additional feature improve the detection results.