Spatiotemporal Implicit Neural Representation for Ionospheric Tomography With Multi-LEO Occultation Data
Fangsong Yang, Wang Li, Jiale Li, Xiaoqing Zuo, Dongsheng Zhao, Kefei Zhang
Abstract
Current ionospheric tomography faces a bottleneck: heavy reliance on empirical background fields leads to poor spatiotemporal accuracy and robustness. To resolve this, we propose an inversion strategy fusing interpretability analysis, Spatio-Temporal Implicit Neural Representation (ST-INR), and the Multiplicative Algebraic Reconstruction Technique (MART). By training a physics-informed ST-INR model on multi-LEO satellite occultation data, we overcome the spectral bias of traditional networks to generate high-fidelity prior fields with fine physical gradients. These priors serve as initializations for MART-based refinement using ground-based GNSS Slant Total Electron Content (STEC), ensuring alignment with real-time observations. Validation during an independent test period featuring a severe geomagnetic storm (Dst approximately -150 nT) reveals significant improvements. Compared to the IRI2020 model, our method reduces the 3D electron density reconstruction RMSE by 71.9% (quiet) and 61.2% (storm), while improving key physical parameters (NmF2 and hmF2) by 45.0% and 40.5%, respectively. It also significantly outperforms the traditional Back-Propagation (BP) neural network. Comparisons with Global Ionospheric Map (GIM) products from the International GNSS Service (IGS) further verify high horizontal consistency, particularly in regions lacking station coverage. This study confirms that the proposed method effectively mitigates reconstruction challenges in data-sparse areas, significantly enhancing physical realism under complex space weather conditions.