Data fusion of ground-based GNSS, radio occultation and empirical model to predict the ionospheric peak electron density via artificial neural networks

Abstract

Listen

Translate article

In this work, we develop a model to predict the ionospheric peak electron density based on artificial neural network (ANN) utilizing long-term observation data from the Constellation Observing System for Meteorology, Ionosphere, and Climate (COSMIC) spanning from 2008 to 2018. New constraints such as the International Reference Ionosphere peak electron density results (IRI-NmF2) and vertical total electron content (VTEC) are considered. A preliminary regression analysis is performed via the random forest algorithm to assess the significance of the input parameters including year, month, day, local time, latitude, longitude, F10.7, Kp, IRI-NmF2 and VTEC. Results show that the root-mean-square error (RMSE) of predicted NmF2 validated by testing dataset is reduced from 1.739 × 105 to 1.417 × 105 el/cm3 when applying additional constraints. The aided ANN model performs better in the quiet time (with RMSE 9.433 × 104 el/cm3) than in disturbed time (1.784 × 105 el/cm3). Furthermore, the ANN predictions are compared with the original COSMIC data and ionosonde observation data. Separate discussions are conducted for different latitudes. For the COSMIC data in 2014, the RMSEs for the low-, middle- and high-latitude data are 2.728 × 105 el/cm3, 1.511 × 105 el/cm3, and 1.076 × 105 el/cm3, respectively; for the ionosonde data at different latitudes, the errors for ANN-fitted NmF2 are 3.314 × 105 el/cm3, 1.585 × 105 el/cm3, and 1.403 × 105 el/cm3, respectively. Under different solar activity conditions, the ANN model demonstrates superior prediction performance compared to the IRI model.