Item

Abstract

Seasonal oscillations in GNSS time series are a major source of noise for the interpretation of tectonic signals. While some success in isolating the seasonal signals has been demonstrated with Kalman filters and matrix factorization, seasonal signals are still generally difficult to remove, especially for time series with interannually varying amplitudes. We develop a deep learning model with the aim of predicting seasonal oscillations in GNSS time series from Earth system models describing geophysical fluid loading on the Earth’s surface (developed at ESM-GFZ) and that consist of hydrological loading, and non-tidal atmospheric and oceanic loading. For our algorithm, we use globally distributed daily PPP GNSS displacement time series from Nevada Geodetic Laboratory (NGL) and isolate the seasonal displacement signals as our learning targets. We pair each target sample with a sequence of loading grids within a set time window around the sample time and spatially constrain the grids to a small area around the respective station. We test different architectures: a 3D U-Net (also called V-Net) and a time distributed LSTM-2D-ConvNet. Finally, we evaluate the performance of our model on a hold-out data set not used during training to assess the effectiveness of our deep learning method for removing annual and semi-annual oscillations from GNSS displacement time series. We also demonstrate the effect of removing seasonals on the identification of tectonic transients, using time series from active fault zones.