Real-Time ZTD correction grid based on augmented GNSS network for navigation services

Tropospheric refraction is one of the major error sources in satellite-based positioning. The delay of radio signals caused by the troposphere ranges from 2m at the zenith to 20m at low elevation angles, depending on pressure, temperature and humidity along the path of the signal transmission. If the delay is not properly modeled, positioning accuracy can degrade significantly. Empirical tropospheric models, with or without meteorological observations, are used to correct these delays but they are limited in accuracy and spatial resolution resulting in up to a few decimeters error in positioning solutions. The present availability of ground-based GNSS networks and the state of the art of GNSS processing techniques enable precise estimation of Zenith Tropospheric Delays (ZTD) with different latency ranging from real time to post-processing.
We present a method for computing ZTD residual fields interpolating, through Ordinary Kriging, the residuals between GNSS-derived and model-computed ZTD at continuously operating GNSS stations. GNSS ZTD estimates, obtained in real time and in PPP mode, are augmented by a multi-prediction model based on a Graph Neural Network model trained using one year of Near Real Time ZTD observations and a model using a polynomial plus harmonic interpolation. A combination strategy is defined to merge GNSS ZTD estimates at sites with the predicted values, where predicted ZTD values act as hole fillers for stations missing from the GNSS network at the current epoch. The residual ZTD field, obtained from PPP/prediction model and ZTD empirical model, is modelled as a random process and for each epoch a variogram is estimated and fitted to characterize the spatial correlation of the process. At a known user location, ZTD value is obtained as the sum of site interpolated ZTD residual and modeled-ZTD value. The algorithm is validated with respect to GNSS ZTD estimates provided by an external provider at a selection of sites not included in the network used to fed the computation. Details about validation and possible improvements will be provided.