Forecasting of global ionospheric TEC using a deep learning approach

There is a great interest in total electron content (TEC) forecasting. In this study, pix2pixhd, a novel deep learning method based on generative adversarial network, is applied to forecast the global TEC. Different from previous research, similar to semantic segmentation, we introduce the label map that divides the TEC in a map into the ionospheric peak structure (TEC >= 50 TECU) and the remaining area, with two masks. The historical time series of the global International Global Navigation Satellite Systems Service (IGS) TEC maps and the corresponding label maps are used as input data to develop our model. The input of label maps helps the model extract the features of the TEC peak structures and improves the prediction effect of our model at low latitudes. Our model is evaluated using root mean square error and correlation coefficient and compared with the model without labels, 1-day and 2-day Center for Orbit Determination in Europe (CODE) prediction models, and IRI-2016. Results demonstrate that our model performs well in TEC forecasting (the root mean square errors between generated and IGS TEC maps for 1-day, 2-day, and 3-day TEC predictions are 2.57, 3.32, and 4.19 TECU, respectively) and always behaves better than the other three models. The performances of our model for 1-day, 2-day, and 3-day TEC predictions worsen in turn. Moreover, our model generates the ionospheric peak structures well and behaves better in low latitudes than in mid- and high latitudes. Besides, our model performs slightly better during storm times than quiet times as well as during high-solar-activity years than low-solar-activity years. Our work demonstrates a new possibility of applying deep learning methods to a broader field of geosciences, particularly for predictions.