Combining Eigenanalysis with Machine Learning Techniques to Identify Ionospheric Storm Drivers in Large Space Weather Data Sets

Abstract:

The authors present the utility in using exploratory data analysis (EDA) techniques to prepare space weather data sets for ionospheric storm classification using machine learning (ML) techniques. Support vector machines (SVMs) are a type of ML technique that seeks to draw intelligent quantitative boundaries between sets of data based on their membership in two or more classes. The basic idea is to take data sets of low dimensionality and transform them to a hyperspace (i.e., a geometric basis of higher dimension) where the separation of distinct populations is subsequently made by hyperplanes. These and other ML methods have been used successfully in a number of classification schemes, but in order to make them useful for ionospheric storm classification, it is necessary to precondition the space weather data sets. In particular, we use common factor analysis to simplify solar and geomagnetic activity indices into “S” and “G” macro-indices, and principal components analysis to simplify electron density profiles (EDPs) into empirical orthogonal functions and weighted coefficient time series. We then demonstrate how outlier analysis can be used to identify solar, geomagnetic, or combination storm influence on the ionospheric EDPs. Results are compared with similar eigen-decomposition of EDPs from the International Reference Ionosphere (IRI) model, showing high agreement in most cases, but with some notable exceptions, especially during post-sunset hours at equatorial latitudes. The eigenanalysis data can be used by a human researcher to identify outliers that signify storm influence on the EDPs, but ML is needed to study large sets of data for this purpose. Thus, the work concludes with a description of an on-ramp for these parameters to be incorporated into an SVM paradigm for storm classification.