Deep learning approach for combining global ionospheric maps for GNSS positioning

Abstract

One of the most widely used sources for information regarding the state of the ionosphere is the global ionospheric maps provided by the IGS service. These maps are created through a weighted average of solutions from various centers, including CODE, ESA, JPL and UPC. As technology has advanced, the application of artificial intelligence in ionospheric research has become more prevalent, motivating us to apply this approach to improve the process of combining ionospheric maps. The objective of our research is to use deep learning in the form of recurrent neural networks to generate global ionospheric maps. The model is complemented by the inclusion of positional and temporal parameters as well as solar and geomagnetic activity indices. In the study, the total electron content (TEC) was extracted from Jason altimetry measurements that served as the reference data for the model. The Jason TECs contain electron content up to an orbit height of approximately 1336 km. Therefore the missing data above the Jason orbit was modelled using several ionospheric/plasmaspheric models. One of the key objective of this study was to identify the optimal fitting model for mapping electron content above the Jason orbit. The solution that demonstrate the most significant impact on the learning process and providing the best results was the Neustrelitz Electron Density Model (NEDM). To validate the Gdańsk University of Technology model (GUT), we conducted a comparative analysis of single-frequency positioning using maps from GUT and IGS. Our solution demostrated an improvement in positioning for over 70% out the 300+ stations studied on average for each studied day during calm or disturbed ionospheric conditions. For three-dimensional positioning errors, we obtained improvements ranging from 5 to 15% relative to IGS results.