Item

Abstract

When analysing GNSS displacement time series with regression approaches, a common simplifying assumption is that annual and semi-annual oscillations, which are mainly controlled by geophysical fluid loading, are perfectly repeating with no variations in phase or amplitude from year to year. In this oversimplified model, sines and cosines with periods 1 year and 0.5 years can be used to represent seasonal oscillations (Fourier terms). Here we show the modification of an existing algorithm, Greedy Automatic Signal Decomposition (GrAtSiD), to now include interannually varying seasonal oscillations. This is achieved by applying a time-varying amplitude weighting on the Fourier terms of the full trajectory model. We demonstrate the algorithm’s performance first with synthetic examples, before applying it to Nevada Geodetic Laboratory’s (NGL’s) daily PPP GNSS displacement time series from South America.In addition to isolation of interannual seasonal oscillations, we demonstrate the effect of applying common-mode filtering (that relies on an initial trajectory model to generate residuals) versus a deep-learning approach that removes higher frequency scatter from the displacement time series before a trajectory model is applied. This supervised deep learning-based method assumes no spatial dependency and is applied on a station-by-station basis. The model has been trained on thousands of freely available PPP daily displacement time series from NGL, augmented by synthetic time series for a more comprehensive training set. This work demonstrates how, by combining advanced time series analysis tools, we are able to better separate tectonic and fluid loading signals in GNSS displacement time series.