Ionospheric foF2 Prediction Using an Interpretable XGBoost Model and COSMIC‑1 and 2 Radio Occultation Data

Abstract

The critical frequency foF2 of the ionosphere’s F2 layer plays a key role in shortwave communications. Its variability is influenced by factors such as solar activity and geomagnetic conditions, making accurate prediction of foF2 essential for reliable communication in navigation, aviation, and emergency scenarios. This paper presents a global foF2 prediction model based on occultation data from 2007 to 2023 and the eXtreme Gradient Boosting (XGBoost) algorithm. The model incorporates the SHAP method for interpretability analysis, which identifies the core factors driving its predictions. The results show a coefficient of determination (R2) of 0.886 and a root mean square error (RMSE) of 0.964 MHz on the test set. The model captures key ionospheric features, including the single peak at the magnetic equator, the double peak of the equatorial anomaly, the Weddell Sea anomaly, and the winter anomaly. Verification with independent datasets from the GRACE satellite and GIRO radiosonde demonstrates that the model outperforms the IRI‑2020, NPDM, and random forest (RF) models, achieving the highest R2 and reducing errors by over 50% at mid‑ and high‑latitudes. When COSMIC‑2 data were added, the model’s performance improved, with a 2.39% reduction in mean absolute error (MAE) and a 2.63% reduction in RMSE for latitudes between –42.55° and 42.91°. SHAP analysis highlighted the core driving factors of the model, including latitude, time of day, annual cumulative day, F10.7 index, and longitude. Feature ablation experiments revealed that only six features were needed to achieve core accuracy, with an R2 greater than 0.85 and a ΔR2 of –0.038.