Detecting Traveling Ionospheric Disturbances with LSTM Based Anomaly Detection

Abstract

Large natural hazards have been shown to induce traveling ionospheric disturbances (TIDs) caused by the acoustic, acousticgravity, and gravity waves produced from these events. These induced TIDs are observable in the total electron content (TEC) measures from dual-frequency GNSS receivers. Previous research has focused on methods to detect an event on a specific day. This research expands this objective to detect an event from an entire year of slant TEC (sTEC) data. There are several challenges of detecting a natural hazard induced TID across an entire year due to variations of day-to-day and seasonal ionospheric activity, and scintillation due to unknown sources. This work proposes the application of a deep learning neural network, specifically a long short-term memory (LSTM) neural network, combined with statistical analysis, to detect TIDs from detrended sTEC (dTEC) for the year 2021 from stations along the Alaskan peninsula and islands. The LSTM is a type of recurrent neural network (RNN) that outputs future dTEC predictions from past dTEC values. TID detections rely on errors calculated as the difference between ground truth observations and the LSTM predictions. These errors are filtered using a rolling mean error threshold to remove moderate to minor errors, as large errors indicate a significant and sudden change in the dTEC data as expected at the start of a TID induced by a significant natural hazard. Station dTEC on dates with large errors detected are then checked for moderate to strong cross-correlation with other stations. Date and times that show an overlap between a pair of stations with large errors and moderate to high correlation at the time of the error (i.e., show phase synchrony) are determined to be an ionospheric anomaly and potential TID. This research work serves as a new method of TID detection for the broader fields of natural hazard detection and ionospheric monitoring.