A neural-network based ionosphere model for GNSS receiver bias estimation

Abstract

Applying machine learning (ML) techniques together with fast computing machines, more sophisticated ionosphere models featuring large scale ionospheric irregularities and anomalies can be developed nowadays [Adolfs et al. 2021]. We recently developed such a fully connected neural network (NN) based total electron content (TEC) model using Global Ionospheric Maps (GIMs) covering data from previous two solar cycles. We found that the NN based TEC model can successfully reconstruct ionospheric features that are not always visible such as the Nighttime Winter Anomaly (NWA). The NWA feature is only visible in the Northern Hemisphere at the American sector and in the Southern Hemisphere at the Asian longitude sector during low solar activity, winter and local night-time conditions. Our investigation shows that the same TEC model inherits also other features such as the Mid-latitude Ionospheric Trough (MIT) and the longitudinal variation of the Equatorial Ionization Anomaly (EIA) crests. Being motivated by its performance in ionosphere reconstruction we utilized the TEC model for differential code bias (DCB) estimation for a network of ground GNSS receivers. For this, we have derived an empirical version of the NN based TEC model which is portable and can be run independently without installing libraries that required during model training (e.g., TensorFlow). We found that the receiver DCBs can be computed by the NN-based TEC model with sufficient accuracy. The obtained accuracies are comparable to those obtained by the conventional method of DCB estimation by fitting GNSS TEC data to the ionospheric basis function represented by NTCM approach or spherical harmonics [Jakowski et al. 2011]. The application of NN based TEC model for GNSS receiver bias estimation will not only simplify the operational requirements but also improve near-real-time ionosphere monitoring service.