A Geospatially and Physically Guided Method for Correcting Multi‐Source Ionospheric Occultation Data Quality

Abstract To address the inconsistency in data quality between the FY‐3 satellite series and the COSMIC‐2 satellite, this study proposes a feature‐enhanced dual BP neural network framework that integrates adaptive weighting and physical constraint mechanisms. The framework constructs a multi‐source spatial feature set and adopts a two‐stage residual learning strategy, coupled with a joint weighting scheme based on altitude and solar zenith angle. Additionally, residual‐based dynamic optimization and physical boundary constraints enable multi‐level, high‐precision correction of FY‐3 electron density data. To unify cross‐platform data quality and improve the predictive accuracy of multi‐source ionospheric models, the proposed method reduces the global RMSE of FY‐3 electron density from 4.14 × 10 5 el/cm 3 to 2.18 × 10 5 el/cm 3 and the MAE from 1.94 × 10 5 el/cm 3 to 1.31 × 10 5 el/cm 3 while increasing the correlation coefficient from 0.78 to 0.9463. Latitude‐based analysis shows the highest accuracy in low‐latitude regions ( R = 0.97, RMSE = 1.64 × 10 5 el/cm 3 ), followed by mid‐latitudes. The correction is particularly effective in the critical ionospheric layer between 200 and 400 km. Seasonal validation indicates optimal performance in spring ( R = 0.9585, RMSE = 1.83 × 10 5 el/cm 3 ). Furthermore, during geomagnetic storms, the model maintains a high correlation of 0.93, demonstrating strong robustness to disturbances. Overall, the proposed correction framework significantly improves the accuracy of FY‐3 electron density data, providing essential support for three‐dimensional ionospheric modeling and multi‐source data assimilation.