Machine Learning Model for Predicting Global Ionospheric TEC Maps Based on Constraint Conditions

The prediction and modeling of ionospheric total electron content (TEC) have consistently been a focal point for researchers, as it holds significant implications for satellite positioning, navigation, telemetry, control, and radio wave propagation. In this context, we propose a machine learning prediction model [predictive GAN variational autoencoder-label (PGVAE-label)] using a labeled graph of image segmentation as a constraint to predict the global ionospheric TEC. We use IGS TEC maps from 2003 to 2018 as training, test, and validation sets, respectively. Subsequently, we conducted comparative experiments using the unlabeled machine learning prediction model (PGVAE) and the one-day and two-day forecast maps published by the Center for Orbit Determination in Europe (CODE). In addition, the article analyzes the effect of predictions during the periods of geomagnetic quiet and disturbance, high solar activity years, and low solar activity years. The results show that the PGVAE-label model has superior TEC prediction capability, producing TEC prediction maps with the lowest average root-mean-square error values of 1.79, 1.80, and 1.83 TECU, and that the PGVAE-label model is also superior to the PGVAE and CODE models in the region of the peak ionospheric structure. The predictive ability of the PGVAE-label model is better in geomagnetically quiet periods than in geomagnetically disturbed periods, and better in solar low years than in solar high years. The work in this article provides new ideas and thoughts on the application of deep learning to the broader field of Earth sciences, particularly in the problem of prediction.