Convolutional Neural Networks for Time Series Classification of Ionospheric Scintillation

In global navigation satellite systems (GNSS), the amplitude scintillation index (??4) and the phase scintillation index (????) indices are well-established to monitor ionospheric scintillation activity. However, these indices are not able to distinguish between the scintillation-induced propagation effects from the line-of-sight (LOS) dynamic. To circumvent this limitation, Kalman-based algorithms capable of separating the scintillation effects from the LOS dynamics have enabled the estimation of the scintillation phase and amplitude directly from the postcorrelated data. A remaining challenge of both scientific significance and practical concern is to properly classify the estimated time series among different scenarios of scintillation intensity. In this article, a time series convolutional neural network (CNN) is employed to classify the estimated amplitude and phase scintillation signal. The computer-generated scintillation data is based on the Cornell scintillation model (CSM), where different values of decorrelation time and ??4 index are used to build a multiclass dataset from which the CNN extracts the underneath features. Performance results, obtained via computer simulations, show that the proposed model can maintain a reasonable accuracy for certain levels of estimation noise and can therefore provide a temporal characterization of the estimated scintillation signal.