Evaluation of F10.7, Sunspot Number and Photon Flux Data for Ionosphere TEC Modeling and Prediction Using Machine Learning Techniques

Abstract

Considering the growing volumes and varieties of ionosphere data, it is expected that automation of analytical model building using modern technologies could lead to more accurate results. In this work, machine learning techniques are applied to ionospheric modeling and prediction using sun activity data. We propose Total Electron Content (TEC) spectral analysis, using discrete cosine transform (DCT) to evaluate the relation to the solar features F10.7, sunspot number and photon flux data. The ionosphere modeling procedure presented is based on the assessment of a six-year period (2014–2019) of data. Different multi-dimension regression models were considered in experiments, where each geographic location was independently evaluated using its DCT frequency components. The features correlation analysis has shown that 5-year data seem more adequate for training, while learning curves revealed overfitting for polynomial regression from the 4th to 7th degrees. A qualitative evaluation using reconstructed TEC maps indicated that the 3rd degree polynomial regression also seems inadequate. For the remaining models, it can be noted that there is seasonal variation in root-mean-square error (RMSE) clearly related to the equinox (lower error) and solstice (higher error) periods, which points to possible seasonal adjustment in modeling. Elastic Net regularization was also used to reduce global RMSE values down to 2.80 TECU for linear regression.