Categorizing Characteristic Regions of High‐Latitude Scintillations Using a Combination of Isolation Forest and Neural Network Machine Learning Algorithms

Abstract Global Positioning System (GPS) scintillations are radio signal fluctuations due to ionospheric structures or irregularities. In this work, we utilize scintillation data to identify the source region (auroral oval vs. polar cap) of a high latitude irregularity responsible for any given scintillation instance, using a combination of ML and deep learning models. We used available high‐rate (50‐Hz) GPS data from three Canadian High Arctic Ionospheric Network stations for 2 yrs to develop a categorization methodology based on an unsupervised detection ML model called Isolation Forest (iForest), which detects scintillation instances to be fed to a supervised artificial neural network (ANN) model. The goal is to confidently categorize the high‐latitude ionospheric scintillations based on their regions of occurrence. Our method involves using low‐rate GPS scintillation indices to threshold high‐rate data to be fed to the iForest for event detection. We use satellite data‐derived auroral oval and polar cap boundaries to label our ANN data set. Our analysis of these preliminary data sets shows that the iForest algorithm detects events with high accuracy irrespective of the geomagnetic conditions and receiver location. The ANN consistently yielded an F1‐score close to 0.7 implying that the model can classify source regions based on the input data. The ANN‐based classification of source regions of scintillation events performed better during quiet times than stronger geomagnetic conditions. Our method can be applied to different ionospheric irregularity problems, making this a potentially useful tool for understanding scintillations and their relationship with different irregularity generation mechanisms.