Integrating Convolutional Neural Networks and Wave Hop Theory for Enhanced Sub-ionospheric VLF Remote Sensing via Satellite-Derived Terrain Mapping

Accurate characterization of the Earth–ionosphere waveguide (EIWG) is fundamental to Very Low Frequency (VLF) remote sensing of space weather variability and mesospheric lower ionospheric dynamics. Conventional wave-hop propagation models, however, are prone to parameter degeneracy, whereby uncertainties in assumed ground conductivity are offset by non-physical adjustments to ionospheric reflection parameters, undermining physical interpretability.

Here, we introduce a physically constrained modeling framework that combines deep learning (DL) based terrain classification with asymmetric ionospheric parameterization to improve the realism and identifiability of sub-ionospheric VLF simulations. High-resolution satellite imagery along the great-circle propagation path is classified into six terrain categories using a convolutional neural network based on the ResNet-50 architecture. Each terrain class is assigned a representative electrical conductivity, thereby replacing the common assumption of laterally homogeneous ground properties. In parallel, an asymmetric temporal ionospheric model driven by solar zenith angle is implemented to capture the hysteresis associated with unequal ionization and recombination rates across sunrise and sunset terminators.

Model performance is evaluated using narrowband observations from the AWESOME and WALDO receiver networks. Results demonstrate that incorporating spatially varying, AI derived ground conductivity substantially improves agreement between modeled and observed VLF amplitudes and phases. Importantly, although multiple parameter sets may reproduce similar signal amplitudes, only models constrained by physically realistic ground conductivity yield ionospheric reflection heights that remain within geophysical reasonable ranges. This approach mitigates long-standing identifiability issues in VLF propagation modeling and enhances the robustness of VLF-based diagnostics of lower ionospheric variability.