Large-Scale Completion of Ionospheric TEC Maps Using Machine Learning Models With Constraints Conditions
Qingfeng Li, Hanxian Fang, Chao Xiao, Die Duan, Hongtao Huang, Ganming Ren · IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing · 2025
The total electron content (TEC) is a pivotal parameter for characterizing the ionosphere. However, the collection of complete TEC data, particularly over the sea surface, poses significant challenges due to the limited receiver coverage. The advent of deep learning technology has led to the development of effective methodologies for addressing these challenges. In this paper, we propose a novel approach that integrates the latent spatial modeling capability of a variational autoencoder (VAE) with the generative adversarial network (GAN) to enhance the quality and diversity of generated samples. This integration utilizes a combination of both deep learning algorithms, achieving a more balanced output than either model alone. Concurrently, this paper introduces the TEC map of the International Reference Ionosphere (IRI) model as a constraint to construct the CGVAE-label model. This approach ensures the model's ability in large-scale data complementation, robustness and adaptability, and the ability to incorporate a priori knowledge. Furthermore, the present study employs an analytical approach to investigate the impact of model complementation during periods of geomagnetic calm and disturbance, high and low solar activity years, and across diverse mask scales. The findings demonstrate that the CGVAE-label model demonstrates superior TEC-completion capability, with the mean median SSIM and RMSE values of the generated TEC-completion maps attaining 93.80% and 2.20 TECU, respectively. Furthermore, the CGVAE-label model exhibits superiority over the CGVAE model in the domain of the peak ionospheric structure. The completion capability of the CGVAE-label model is superior during geomagnetic quiet periods compared to geomagnetic disturbed periods, and it performs better during solar minimum years than during solar maximum years. The work presented in this paper offers novel insights and ideas for the application of deep learning in a broader range of geoscience fields, particularly in addressing completion problems.