One day ahead prediction of global TEC using Pix2pixhd

In this paper, we make a 1-day forecasting of global total electron content (TEC) using pix2pixhd that is a novel deep learning method based on GAN. For training, we use the International GNSS Service (IGS) TEC maps from 2003 to 2018, covering the entire solar cycle with a training set (2003-2013), a validation set (2018), and a test set (2014-2017). Then we evaluate our model using Root Mean Square Error (RMSE) and Correlation Coefficient (CC) and compare our model with IRI-2016. As a result, our model shows a good performance for TEC forecast with one day in advance. From the average difference between output maps of our model and target ones (+1 day IGS TEC maps) as well as between IRI TEC maps (+1 day) and target ones and corresponding standard deviations, our model behaves better than IRI at a global scale, especially at low latitudes. Under different conditions of geomagnetic and solar activities, the prediction effect of our model is always better than IRI. Our model behaves well during geomagnetic quiet conditions, but is affected during geomagnetic storms showing stronger deviations. Besides, our model behaves better during solar maximum than minimum days. Our work demonstrates a new possibility for the application of deep learning on a broader field of geosciences, particularly for problems of prediction. (C) 2022 Published by Elsevier B.V. on behalf of COSPAR.

Automated summary

This paper presents a 1-day forecasting of global total electron content (TEC) using the novel deep learning method pix2pixhd based on GAN. The model outperforms the traditional model IRI-2016 in predicting TEC at a global scale and particularly in low latitude regions.