Inversion of Global Thermospheric Density Map Based on Satellite Measurements Using Deep Learning RFR Network

The estimation error of neutral mass density is a major source of uncertainty in calculating thermospheric drag, which is the primary non-gravitational perturbation affecting satellites in low Earth orbit. Many empirical models and physics-based models are commonly used to forecast the thermospheric density, but significant deviations might occur during extreme space weather events. The inversion approach based on observation data may provide better predictive capability. Given the limited thermospheric observational resources, this study aims to reconstruct the global thermospheric pattern using sparse data and a neural network. In this paper, we develop a global thermospheric density map inversion model based on the Recurrent Feature Reasoning (RFR) neural network (RFR-Net), which treats the inversion of global density maps from limited satellite data as an image inpainting problem. To evaluate the effectiveness of the proposed method, sparse satellite observations (covering only 0.19% of the global grid) were input to the trained model. Results indicate that RFR-Net enables the generation of global thermospheric density maps and effectively captures their temporal variations. Furthermore, we compare RFR-Net's performance with traditional models. During quiet periods, RFR-Net achieves comparable accuracy to other models, with all exhibiting low errors. Notably, during geomagnetic storms, RFR-Net demonstrates significantly higher accuracy than other models.