Generating realistic synthetic GNSS station position time series to evaluate automatic discontinuity detectors

The adjustment of trajectory models to GNSS station position time series is an essential step in the establishment of terrestrial reference frames, but also in a wide range of geophysical studies investigating glacial isostatic adjustment, tectonics, or coastal sea-level change. This trajectory modelling step is complicated by the presence of occasional discontinuities in the time series, including outliers, mean offsets (jumps), and changes in velocity. With the increasing number and longevity of GNSS stations, traditional manual trajectory modelling of position time series by an experienced operator becomes increasingly time-consuming and even impossible for large datasets. For this reason, automatic discontinuity detection approaches are increasingly appealing for many geodetic and geophysical applications.

A central concern with automatic modelling approaches is their reliability. While past studies suggest that automatic approaches are less reliable than human experts, this situation may change with advances in artificial intelligence. One limitation to monitoring progress in automatic trajectory modelling is the absence of a standardized benchmarking approach for discontinuity-detection algorithms. In practice, each research group publishes performance measures based on different data sets of either human-labeled or synthetic time series. Unfortunately, data sets of human-labeled time series are limited in size and may be incomplete because humans are unlikely to detect the smallest discontinuities. Synthetic time series are not necessarily reliable either, as they may lack realism with respect to length, gaps, frequency and amplitude of discontinuities, and noise properties. To address these benchmarking issues, we will present a generator of realistic synthetic GNSS station position time series. Building upon the characterisation of real data sets, this generator will contribute to the development of a standardized benchmarking approach for automatic discontinuity detection algorithms. In the long run, this generator may also be employed to train new machine learning algorithms.