Covariance-Boosted GP Achieves 99.9% Integrity for Satellite Navigation

Jeremy Ovadia introduces the Covariance-Boosted Gaussian Process (CBGP) framework, designed to address nonstationary spatiotemporal irregularities in high-stakes environments like satellite navigation. Traditional nonstationary GPs are prone to overfitting and overconfident uncertainty estimates when data is sparse and covariance structures are heavily parameterized. CBGP tackles this by iteratively boosting covariance priors using gradient descent-like updates on partially whitened observations, then imposing restrictions and inflating posterior uncertainties to prevent overfitting and overconfidence.

The method is specifically motivated by ionospheric modeling for Satellite-Based Augmentation Systems (SBAS), which must deliver extremely reliable corrections for aviation and other critical applications. Ovadia demonstrates CBGP's efficacy through out-of-sample testing on both synthetic data and a real extensive ionospheric storm dataset over South America. The model consistently meets a three-nines (99.9%) integrity standard—a key safety metric. Compared to currently-operating SBAS methods that rely on local fitting, CBGP provides more informed and responsive regional models, even in the most challenging space weather environments.