Assessing Machine Learning Techniques for Identifying Field Line Resonance Frequencies From Cross‐Phase Spectra

Abstract Monitoring the plasmasphere is an important task to achieve in the Space Weather context. A consolidated technique consists of remotely inferring the equatorial plasma mass density in the inner magnetosphere using Field Line Resonance (FLR) frequencies estimates. FLR frequencies can be obtained via cross‐phase analysis of magnetic signals recorded from pairs of latitude separated stations. In the last years, machine learning (ML) has been successfully applied in Space Weather, but this is the first attempt to estimate FLR frequencies with these techniques. We survey several supervised ML algorithms for identifying FLR frequencies by using measurements of the European quasi‐Meridional Magnetometer Array. Our algorithms take as input the 2‐hour cross‐phase spectra of magnetic signals and return the FLR frequency as output; we evaluate the algorithm performance on four different station pairs from L  = 2.4 to L  = 5.5. Results show that tree‐based algorithms are robust and accurate models to achieve this goal. Their performance slightly decreases with increasing latitude and tend to deteriorate during nighttime. The estimation error does not seem to depend on the geomagnetic activity, although at high latitudes the error increases during highly disturbed geomagnetic conditions such as the main phase of a storm. Our approach may represent a prominent space weather tool included into an automatic monitoring system of the plasmasphere. This work represents only a preliminary step in this direction; the application of this technique on a more extensive data set and on more pairs of stations is straightforward and necessary to create more robust and accurate models.