From noise-to-signal: enhancing the sensitivity of long-term GNSS by explaining non-linear station motions

Abstract

Reducing uncertainties in long-term GNSS is crucial for providing the most stable realization of terrestrial reference frames. Besides the epoch-wise handling of errors in the GNSS processing of the raw observations, this includes the description of linear and non-linear station motions by using functional and stochastic time series models (trajectory models). While the state-of-the-art trajectory models explain common signals (such as station velocity, periodic motions and offsets), a main challenge is to explain the variety of signals that are still present in the residuals (i.e., transients and artifacts).We take the time series of almost 2000 stations from Europe and the Western U.S., and fit extended trajectory models (ETMs) using the Hector software. We show that correcting the vertical observations by geophysical loadings and common-mode errors (CME) enhances median GNSS sensitivity by a factor of two, which we demonstrate for important parameters of the ETM, for both, linear and non-linear motions. The analysis reveals a median sensitivity of 0.5 mm/year for station velocity, and 0.6 mm for annual seasonal motions. Supported by the comparison with the CME-corrected case, we highlight potential shortcomings of the non-tidal atmospheric and hydrological loading models (related to phase shifts and the presence of transients), and demonstrate biasing effects on uncertainty.Based on the post-fit residuals of the ETMs and unsupervised learning methods, yet unexplained geophysical transients can be classified and clustered, which is a significant step towards automatic data interpretation and quality control. In view of an optimal GNSS station deployment, this approach also helps to separate stable from unstable regions, for example regions of irregular subsidence.  We conclude, that this algorithmic framework cannot only be successfully used for GNSS data, but also for other geodetic time series data.