Characterization of Sources of Ionospheric Scintillation in High Latitudes through Machine Learning

Abstract:

Irregularities in Global Navigation Satellite System (GNSS) signals in high latitudes have different sources that cause scintillating signatures, which are mostly precipitation, gradient drift instability and Kelvin-Helmholtz-Instabilities. Those structures are only understood yet for certain events and case studies. In order to gain a deeper insight into scintillations in high latitudes and their sources, we are making use of the huge volume of recorded data over the last years and therefore investigate different machine learning approaches to classify and categorize scintillation events and draw conclusions about physical background processes. For the geomagnetic storm on the 9th of March 2012 we apply a simple machine learning approach to categorize the temporal scintillation signatures according to their geomagnetic source region. By use of a hierachical clustering analysis on high rate data in phase and amplitude we are able to distinguish manually selected events from stations inside the polar cap vs those from the auroral oval. Together with geomagnetic background data we are searching for input features and indices that can be used by our model to detect the scintillation signatures caused by particular irregularities and extract those candidate events to be further analyzed with inverse modeling. Based on the evolving database of sets of events we expect to understand more about the importance of the major sources of scintillation in each of the source regions. On top of that, we are investigating how to predict the physical characteristics of the irregularity and its source from the scintillation signatures we observe. We are applying a basic version of a CNN model to understand if we can automate the ML classification and clustering. This work is done to have a preliminary understanding of which method would work best for our requirement of analyzing high rate scintillation phase and amplitude data and classifying it based on scintillation signatures. This to our knowledge has not been attempted thus far.