Semi-supervised Classification of Lower-Ionospheric Perturbations using GNSS Radio Occultation Observations from Spire Global’s Cubesat Constellation

Abstract

In this study, we present a new methodology to automatically classify perturbations in the lower ionosphere using GNSS radio occultation (RO) observations collected using Spire’s constellation of CubeSats. This methodology combines signal processing techniques with semi-supervised machine learning by applying spectral clustering in a metric space of wavelet spectra. A “bottom-up” algorithm was applied to extract E layer information directly from Spire’s high-rate (50 Hz) GNSS-RO profiles by subtracting the effect of the F layers. This processing algorithm has been implemented in our ground segment to operationally produce high rate sTEC profiles with a vertical resolution of better than 100 m. The key idea behind the semi-supervised classification is to produce a database of labeled clusters that can be used to classify new unlabeled data by determining which cluster it belongs to. A dataset of more than 12000 GNSS-RO profiles collected in 2019 containing sTEC perturbations is used to find the initial clusters. This dataset is used as a representation of the climatology of ionospheric perturbations, such as MSTIDs and sporadic Es. The wavelet power spectrum (WPS) is computed for these profiles, and a metric space is defined using the Earth mover's distance (EMD) between the WPS. A self-tuning spectral clustering algorithm is used to cluster the profiles in this metric space. These clusters are used as a reference database of perturbations to classify new sTEC profiles by finding the cluster of the closest profile of the clustered dataset in the EMD metric space. This new methodology is used to construct an automated system to monitor ionospheric perturbations on a global scale.