A New Approach for Low-Latitude Ionospheric Scintillation Prediction

Abstract

The prediction of ionospheric scintillation is a current research topic. Data-oriented models have been proposed since there is not a mathematical model able to simulate the complex ionospheric mechanisms for the prediction of scintillation. Data-oriented models employ machine learning algorithms that are trained with known, post-mortem data, and then perform predictions from new data. An ensemble method based on sampling boosting applied for a set of decision trees is employed for the prediction of scintillation in a single low-latitude location in Brazil. The prediction was performed considering two classes, occurrence or absence of scintillation. The method uses as input temporal series of the scintillation index S4, the total electron content (TEC), and some geomagnetic and solar indexes for that location. Data encompasses the summer months of the previous solar cycle years (2010-2018), and it was split into mutually-exclusive training and validation/test sets. Prediction performance was promising, showing a potential to be developed for operational use. Data limitations related to time series extension, or also to a balanced distribution of samples/instances covering both classes, since scintillation occurrences are usually scarce, can be further developed as new data are available.