Abstract
Global ionospheric maps (GIM) are commonly used ionospheric products in high-precision Global Navigation Satellite System (GNSS) applications. To meet the increasing demand for real-time (RT) applications, the International GNSS Service (IGS) officially started a real-time service in 2013. One of the tasks of the real-time service is the calculation of real-time GIMs. However, the accuracy of current real-time GIMs is still significantly worse than that of the final GIMs, which are the most accurate ionospheric products but have a latency of several days. The IGS RT GIMs exhibit an RMSE of around 3.5-5.5 total electron content units (TECU) compared to the final GIMs. This study focuses on improving the accuracy of existing real-time GIMs through machine learning (ML) approaches, specifically convolutional neural networks (CNN) and conditional generative adversarial networks (cGAN). We apply our method to the IGS combined real-time GIMs and to Universitat Politècnica de Catalunya (UPC) GIMs. We consider over 130'000 pairs of real-time and final GIMs. Over a 3.5-month test period, the proposed approach shows promising results with a reduction of more than 30% in mean absolute error for the real-time GIMs. Especially for regions with high VTEC values, we find a significant improvement of nearly 50%. The ML-enhanced real-time GIMs also exhibit improved positioning performance for single-frequency GNSS positioning with reductions in the 3D error up to 21 cm. Overall, our proposed method demonstrates great potential in generating more accurate and refined real-time GIMs.