Spatiotemporal prediction of total electron content using CONVLSTM, Patch- CONVLSTM and 3D-U-Net models

Abstract. Given the importance of predicting the total ionospheric electron content (TEC), many studies have attempted to predict its spatiotemporal nature. In this study, a patch-based convolutional neural network with long short-term memory (CONVLSTM) (with patch sizes of 5 and 15), simple CONVLSTM, and 3D-U-Net models were used to predict the spatiotemporal nature of the next day's TEC data (next 12 samples). The proposed models use the spatiotemporal nature of the previous day's TEC data (previous 12 samples) along with temporal data such as AP, KP, DST, SN, and F10.7 to predict the next day's TEC data. The results showed that the 3D-U-Net model and then the model with patch size 5 had a higher generalization ability than the classical CONVLSTM architectures, while reducing the RMSE and MAE. The execution time of the program in the 3D-U-Net model has been significantly reduced compared to other models and it has also been able to better extract the microstructural features of TEC maps.