Comparing Gaussian Processes, Neural Processes and Neural Networks for Interpolation of Ionospheric ROTI

The Rate of Total Electron Content (TEC) Index (ROTI) is widely used as an indicator of small-scale ionospheric irregularities and GNSS signal disturbance risk. Unlike Vertical Total Electron Content (VTEC), ROTI reflects rapid spatio-temporal variability and is linked to degraded positioning performance. However, ROTI values estimated from ground-based GNSS are spatially sparse and unevenly distributed, limiting their use for global monitoring.

In this study, we investigate data-driven methods for spatio-temporal interpolation of sparse-observation ROTI values. We make use of the global International GNSS Service (IGS) station network with more than 400 stations to calculate a ROTI dataset using all available GPS satellites. Gaussian Processes (GPs), Neural Processes (NPs) and Neural Networks (NNs) are evaluated in controlled data gap scenarios, where entire regions are held out to mimic poorly covered areas. Performance is assessed in terms of interpolation accuracy, capturing the dynamic nature of ROTI. For the evaluation we also focus on the higher ROTI values that may be linked to degradation in GNSS positioning quality. By systematically comparing these kernel-based methods and neural approaches, we analyze their strengths and limitations in representing ROTI. Based on these results, we aim to identify a robust strategy for generating continuous ROTI products that complement existing global ionospheric maps and support GNSS reliability monitoring.