Ionospheric Data Fusion With GNSS, GNSS‐RO and Satellite Altimetry Based on Machine Learning

Abstract Global Ionospheric Maps (GIMs) are essential in Global Navigation Satellite System (GNSS) applications, particularly for correcting ionospheric delays in single‐frequency positioning. GIMs are mainly derived from dual‐frequency GNSS observations, but the uneven distribution of ground stations—sparse over oceans—reduces accuracy in these regions. Alternative space‐geodetic techniques, such as satellite altimetry and GNSS radio occultation (GNSS‐RO), provide ionospheric information over oceans but differ from GNSS‐derived vertical total electron content (VTEC) in terms of orbital altitudes and observation geometries. Moreover, the sparsity of satellite altimetry and GNSS‐RO data on a daily scale poses additional challenges. We present a framework that integrates GNSS, Jason‐3 altimetry, and COSMIC‐2 GNSS‐RO observations into GIMs using machine learning (ML). First, satellite altimetry and GNSS‐RO VTEC were calibrated to GNSS VTEC. To mitigate daily data sparsity, we built daily background models with XGBoost, each based on 80 day time spans of satellite altimetry and GNSS‐RO observations capturing medium‐term ionospheric characteristics. These background models generated virtual VTEC samples to supplement regions lacking GNSS data for training our Neural Network (NN)‐based GIMs. Tests over three regions (Hawaii, Atlantic and Subantarctic) in 2024 show mean absolute error (MAE) reductions of 37%, 10%, and 33% compared to GIMs without local GNSS data. Evaluation with single point positioning indicates improved 3D accuracy, and external validation with Sentinel‐6A altimetry demonstrates a 0.5 TECU reduction in residual standard deviation when including VTEC from the background model.