Plasma Density Estimation from Ionograms and Geophysical Parameters with Deep Learning

Abstract. Accurate estimates of the ionospheric electron density are essential for various space-weather applications but are especially challenging at high latitudes due to strong spatial and temporal variability driven by auroral precipitation and complex ionospheric convection. This study presents an assimilative empirical model designed to improve regional electron-density estimates in Northern Scandinavia. The model uses ionogram images, the local magnetic field, the auroral electrojet, the ring current and solar-activity indices as inputs. These inputs are fused by a multimodal neural network and trained with incoherent-scatter-radar (ISR) observations of electron density profiles as the target. The model remains functional with only a subset of input, leading to a modest accuracy degradation. Comparative analysis demonstrates that our neural-network–based assimilative model outperforms the ARTIST 4.5 ionogram scaler and the state-of-the-art E-CHAIM model, especially during auroral activity. Overall, our model achieves an R² score of 0.74 on an independent test dataset, whereas ARTIST 4.5 and E-CHAIM obtain R² values of –0.08 and 0.34, respectively. These results indicate that the model can provide reliable, continuous electron-density estimates at high latitudes, even under auroral conditions. This methodology can be extended to develop empirical ionospheric models for other regions with historical ISR data and to invert ionograms to electron-density profiles when ISR observations are unavailable. A similar approach could also be applied in short-term forecasting of the ionospheric electron density.