Effects of Different Feature Representations on Global Ionospheric TEC Model Based on Spatiotemporal Convolutions

Abstract

AbstractThe Total Electron Content (TEC) is a commonly used parameter for characterizing the morphology of the ionosphere. In this paper, we propose a novel feature representation of the time, called the Day and Night Map (DNM), for improving the global ionospheric TEC model. The specific effects of different feature variables (e.g., TEC, DNM, solar and geomagnetic indices) on forecasting global TEC maps are also analyzed. We apply factorized 3D convolutions, a new form of spatiotemporal convolutions, to construct the global TEC model. The global ionospheric TEC maps with a 2-hour temporal resolution are provided by the Center for Orbit Determination in Europe (CODE). The data sets that are resampled by a data segmentation strategy are used to train and evaluate the model. Results show that the DNM has a significant improvement in forecasting global TEC maps compared to the model with a single TEC feature, especially in the middle and low latitudes where large-scale ionospheric TEC anomalies occur frequently. Solar and geomagnetic indices facilitate the prediction capability of the global TEC model, but more indices are not always better. We demonstrate that matching appropriate feature variables with the structural characteristic of the deep learning algorithm is significantly important.