Accurate Monitoring and Timely Prediction of Ionospheric Scintillation Using Support Vector Machine

The monitoring and prediction of the ionospheric scintillation effect which is characterized by rapid fluctuations in the phase and/or amplitude of the satellite signals is considered to be a very important aspect of space weather monitoring and remote sensing. In addition, the ionospheric scintillation poses a significant threat to satellite-based navigation services. Typically, severe scintillation can result in the occurrence of cycle slip at the tracking stage or loss of lock events which can lead to degraded navigation performance. Thus, monitoring and timely prediction of ionospheric scintillation can help in improving the accuracy and reliability of navigation services, space weather monitoring, and remote sensing. Most of the traditional methods for scintillation monitoring are based on very basic event triggers which compare the scintillation values with fixed thresholds and then identify the scintillation event. However, these conventional methods having fixed thresholds may not be effective because the occurrence and severity of ionospheric scintillation is highly dependent on the geographical location of the receiver (i.e. latitude), time of the day, and geomagnetic and solar activity. In this paper, an efficient method for accurate, timely, and automatic prediction of amplitude scintillation events is proposed. Support vector machine (SVM) is the most efficient supervised machine learning method used for classification and regression, and it classifies the data based on kernel functions. In this paper, the proposed method applies the SVM on raw global navigation satellite system (GNSS) data to detect and classify the amplitude scintillation into three main categories, i.e. no Scintillation, moderate, and severe scintillation. For this purpose, raw GNSS data were acquired by deploying a Septentrio multi-frequency GNSS receiver at Sukkur, Pakistan. The results show that the machine learning algorithms can help in developing efficient scintillation prediction models that can accurately monitor scintillation occurrence patterns and predict the scintillation events. The proposed SVM-based model achieved an accuracy of 98% which is very close to manual classification driven by human. Moreover, the prediction responsiveness is also improved, allowing early scintillation alarms.