A systematic approach for modeling global VTEC using machine learning

Ionosphere is the ionized layer of the Earth’s upper atmosphere. Vertical total electron content (VTEC) is a highly descriptive measure of the ionosphere. Modeling and predicting VTEC is crucial, because its disturbances are indicative of severe effects in GPS signal propagation and radio communication. We present a new neural-network-based model of VTEC parametrized with geomagnetic indices, solar wind and their time histories. The model was extensively validated with nested cross-validation to ensure that it performs well during geomagnetic storms and quiet times. We applied a number of feature selection methods, namely gradient boosting, permutation feature importance, random forests and cross-correlation. We selected the best input parameters to the model. In addition to reducing dimensionality and avoiding overfitting, the proposed approach also allows to get physical insights into the dynamics of the ionosphere.