A self-supervised mask-enhanced dual-branch convolutional neural network for reconstructing regional TEC over China

Accurate modeling of the ionospheric total electron content (TEC) is crucial for space-weather monitoring and enhancing GNSS-based navigation and positioning. Traditional TEC models rely primarily on smooth mathematical functions, resulting in excessive smoothing and limited spatial resolution. Machine-learning methods capture complex nonlinear relationships. They have driven advances in ionospheric modeling via image-inpainting. However, these methods depend on high-precision external ionospheric products – mainly global ionospheric maps from International GNSS Service (IGS) analysis centers such as Center for Orbit Determination in Europe (CODE) – as training labels. The intrinsic accuracy and resolution limits of these products constrain model performance. We propose a self-supervised regional ionospheric modeling framework based on a self-supervised mask-enhanced dual-branch convolutional neural network (S-DBN). This approach eliminates dependence on external labels and enables high-accuracy reconstruction of regional TEC over China at 0.2° resolution. Its performance is validated using observations from 241 continuous GNSS stations of the Crustal Movement Observation Network of China (CMONOC) and surrounding regions for the year 2023. Experimental results show that the proposed model achieves a daily root-mean-square error (RMSE) ranging from 1.0 to 2.1 total electron content unit (TECU) throughout the year. The annual mean RMSE of S-DBN is 1.50 TECU, lower than those of dual-branch convolutional neural network (DB-CNN) (2.64), spherical cap harmonic analysis (SCHA) (1.82), and CODE global ionospheric maps (GIM) (4.37). Compared with these three models, S-DBN improves accuracy by 43.2%, 17.6%, and 65.7%, respectively. During quiet conditions, S-DBN achieves RMSEs of 1.17 and 1.52 TECU on May 12 and 24 August 2023. During disturbed conditions, the RMSEs are 1.81 and 1.72 TECU on November 5 and 1 December 2023. The error remains stable with minimal fluctuations, and the hourly and latitudinal RMSEs remain below 3.0 TECU on all days, significantly outperforming the other methods. These results demonstrate the effectiveness of self-supervised learning for ionospheric modeling and provide a new pathway for global and regional TEC reconstruction.