Machine learning-based modelling of VLBI station heights using meteorological and land surface state variables

Abstract

Very long baseline interferometry (VLBI) plays a pivotal role in geodetic and geophysical research by enabling high-precision estimation of station positions. However, these measurements are notably affected by geophysical loading effects, particularly non-tidal surface loading (NTSL), which includes atmospheric, hydrological, and oceanic components. While physics-based models are traditionally used for NTSL corrections, they can suffer from unmodeled signals and systematic residuals due to simplifying assumptions and limited representation of local environmental variability. This study explores the potential of a machine learning (ML) framework to model the non-linear variations in the vertical motion of VLBI stations. We developed and compared two distinct ML-based strategies for 41 globally distributed sites using meteorological and land surface variables from the ERA5-Land, GLDAS-2, and VMF3 datasets over the period 1990–2023. The first approach, ML_RESD, directly models residual station height variations without prior physics-based NTSL correction. The second, ML_HYBD, adopts a hybrid approach that models residuals after accounting for physics-based NTSL corrections from the International Mass Loading Service (IMLS). Station-specific XGBoost regression models were trained and evaluated using station height repeatability (SHR) derived from reprocessed VLBI data using Vienna VLBI and Satellite Software (VieVS). The ML_RESD model improved SHR at approximately 64.5% of stations, with a mean improvement of 0.9% (0.12 mm) across all the stations. Its overall performance was comparable to the physics-based models in absolute terms. The ML_HYBD model demonstrated the strongest performance, improving SHR at roughly 67.7% of stations and achieving an overall mean improvement of 2.1% (0.21 mm), surpassing all other correction methods. These findings highlight the complementary strengths of standalone and hybrid ML approaches in improving VLBI-derived station positions and demonstrate the operational potential of integrating physical and ML models for advancing the accuracy of reference frame realisations. Additionally, SHapley Additive exPlanations (SHAP) analysis identifies key contributing variables, offering station-wise insights into dominant environmental drivers that are not adequately captured by existing physics-based models.