Highlights • Machine learning models to predict daily ZTD value. • Increasing the data quality of ZTD products in the reference stations. • Using of support vector regression to predict the tropospheric delay. • Performance comparison of IGS-ZTD product, VMF1 and RBF-SVR models.

Abstract Modelling of tropospheric delay has a crucial place in the Global Navigation Satellite System (GNSS) as well as atmospheric and space research. Until now, many different modelling put forward and are still being developed to predict tropospheric delay. Developments in Machine Learning (ML) provide alternative approaches to the predictions of Zenith Tropospheric Delay (ZTD) in GNSS observations and allow an increase in the efficiency of current models. This study focusses on Support Vector Regression (SVR) modelling for predicting ZTDs over selected NYAL (North Europe), BAIE (North America), GOPE (Central Europe) and NKLG (Central Africa) stations in different regions globally. The datasets for the SVR are meteorological data, station coordinates (φ, λ and h) and the site-wise ZTDs obtained from the VMF1 product for the period 2019–2020. SVR model predictions were realized by using Linear, Polynomial and Radial Basis Function (RBF). Predictive results of SVR models were compared through various performance metrics such as coefficient of determination (R²), Root Mean Squared Error (RMSE), etc. The results from the NYAL station show a good level of prediction capability of the RBF-SVR model with average RMSE and R² of 17.5 mm and 0.859. This model also presents good predictions at BAIE and GOPE stations with average RMSEs of 20.1 mm and 20.3 mm, and R² of 0.810 and 0.805 respectively. The station with the lowest model success is NKLG with 24.8 mm average RMSE and 0.698 R². According to these results, it was obvious that the RBF-SVR model achieved more success in mid-high latitudes and the height differences at the stations do not affect the model. In addition, the RBF-SVR model has obtained close and realistic results that are compatible with the IGS-ZTD product. These conclusions indicated that the ML model is usable as a means of improving the data missing in the current ZTD products and predicting daily tropospheric delay.