Regional TEC modelling over Africa using deep structured supervised neural network

Abstract

A deep structured supervised neural network (NN) model has been developed for modelling of ionospheric vertical total electron content (VTEC) using observations from over 100 Global Navigation Satellite System continuously operating reference stations (GNSS CORS) across Africa. The study covers all available data during low, moderate and declining phases of solar cycle 24, from 2009 to 2017. Optimal network parameters combination for the regional model includes a combination of spatio-temporal parameters (latitude, longitude, year, day of the year, hour), geomagnetic and solar parameters (F10.7, AE, Dst indices), hidden layer of 20 neurons and a feedforward network with Levenberg–Marquardt back-propagation algorithm. The validation and the test curves do not indicate overfitting and the performance curves of the training, validation and test data show a very similar trend. Thus, the performance of the optimal network with learning data is in sync with the data not assimilated in the learning process. The African Regional Ionospheric Tec Model (ARITM) developed in this study reproduces the known spatiotemporal features of the equatorial and low-latitudes ionosphere quite well. The ARITM performs well to a considerably high degree of precision within the Africa region by comparison with the global ionospheric maps (GIMs) and the Formosa Satellite-3 Constellation Observing System for Meteorology, Ionosphere & Climate (FORMOSAT-3/COSMIC) observations. The results suggest that the developed model can efficiently be adopted as a substitute for single station ionospheric model for any single GNSS station within Africa, avoiding the need for developing independent multiple single station ionospheric models for each of the existing and future GNSS stations within the African region.

Keywords

• Africa
• GNSS
• Neural network
• Ionospheric
• TEC modelling