Improving the Computerized Ionospheric Tomography Performance Through a Neural Network-Based Initial IED Prediction Model

A neural network-based initial ionospheric electron density (IED) prediction model (IED-NN) is proposed to provide high-precision initial IED for the computerized ionospheric tomography (CIT). IED-NN is based on a back propagation neural network (BPNN) and is trained with IED profiles from Constellation Observing System for the Meteorology, Ionosphere, and Climate (COSMIC) and COSMIC-2 radio occultation (RO) missions. With IED observations from eight ionosondes and an incoherent scatter radar (ISR) as references, it is validated that IED-NN has better prediction performance than International Reference Ionosphere-2020 (IRI-2020), with the improvements of 54.52% in root mean square error (RMSE) and 48.99% in mean absolute error (MAE). Based on the initial IED respectively predicted by IED-NN and IRI-2020, CIT experiments are conducted in China region and North America region at three time moments of different geomagnetic activity levels. In China region, due to the better initial IED field from IED-NN, the MAEs of CIT results are improved respectively by 38%, 41%, and 51% at the three time moments. During the geomagnetic storm, although the performance of IED-NN degrades to some extent, the improvement of IED-NN compared with IRI-2020 is still considerable. Over North American region, the CIT processes based on IED-NN perform better in reconstructing the IED in the voxels lack of global navigation satellite system (GNSS) observations and obtain the average improvements of 51.27% in RMSE and 48.45% in MAE. The better initial IED provided by IED-NN also helps to reduce the IED residuals and improve the convergence speed in the CIT iterations.