A deep learning approach for the prediction of ionospheric total electron content (TEC) based on combined prediction and two-step loss fine-tuning

With the increasing modern space mission complexity, synthetic aperture radar (SAR) has gained attention for its all-weather, all-day capabilities. Countries are actively developing low-frequency spaceborne SAR. However, ionospheric delay due to lower operating frequencies cannot be ignored, as it directly impacts SAR positioning accuracy. The ionosphere's total electron content (TEC) directly indicates ionospheric delay magnitude, making accurate TEC estimation essential for real-time, high-precision spaceborne SAR positioning. This study employs Convolutional Gated Recurrent Units (ConvGRU) to examine the effects of model input length and prediction interval on TEC forecasts. A combined prediction strategy for medium- to long-term global ionospheric TEC improves performance by 55.6% compared to recursive methods. Furthermore, we used a two-step loss finetuning strategy to refine the combined prediction method. Statistical indicators demonstrate that our training strategy effectively improves the accuracy of combined predictions. Moreover, this strategy can be conveniently transferred to other deep learning methods.