Global data-driven predictions of seasonal non-tectonic signals in vertical GNSS displacement time series from non-tidal surface loading data

Daily displacement time series from Global Navigation Satellite Systems (GNSS) are frequently used to study deformations of the Earth’s surface due to a wide range of different geophysical processes. The recorded deformations result from tectonic activity or non-tectonic processes like volcanism, groundwater fluctuations and atmospheric loading. In addition, local disturbances of the antenna (e.g., snow cover, thermoelastic effects of the monumentation) and artifacts from GNSS processing (e.g., draconitic signals) are sometimes prominently included in coordinate time series. We use a Temporal Convolution Network (TCN) to predict non-tectonic vertical GNSS displacements on a global scale from physics-based non-tidal loading products. We train our model on a global dataset with more than 11,000 GNSS stations from the Nevada Geodetic Laboratory, active from January 2002 until June 2024, and evaluate the performance against independent estimations. Across the hold-out dataset, our TCN derives non-tidal loading GNSS signatures that when compared to the non-tectonic GNSS signal results in a global average reduction in RMSE of 4.7 % with respect to the numerical non-tidal loading models. This approach presents an initial step towards a data-driven complement to physics-based numerical loading models, improving the isolation of non-tectonic signals in GNSS time series and validation of numerical non-tidal loading models.