LSTM based Ionospheric Anomaly Detection: a study on Alaska (M8.2) and Turkey (M7.5) earthquakes.

Acoustic, acoustic-gravity, and gravity waves produced by large natural hazard events propagate into the ionosphere causing a disturbance known as a traveling ionospheric disturbance (TID). TIDs are observed by analyzing the temporal and spatial variations of ionosphere’s total electron content (TEC) that can be measured from GNSS receivers. Previous research has shown these TIDs induced by various large scale natural hazards such as the 2011 Tohoku earthquake, the 2022 Hunga Tonga volcanic eruption, and their respective tsunamis. Smaller natural hazards and space weather events have also been shown to produce TIDs that are easily missed due to background ionosphere and signal noise. Hence, past studies tend to rely on the prior knowledge of the event before commencing the search for induced TIDs. This research proposes to reverse this process by using deep learning neural networks, specifically a long short term memory (LSTM) neural network, to search for induced TID without prior knowledge of an event. This method draws from research in the field of deep learning anomaly detection and applies it to two geolocations and time frames. As a proof of concept, we performed two experiments using datasets during two earthquakes in 2021 and 2023. For the first test we processed over 6 months of GNSS observations across ten stations along the Alaskan peninsula and islands, and the second test examined over a 24-hour period of GNSS observations across seven stations in the Mediterranean. Slant TEC was passed through a Butterworth highpass filter of 1 mHz. This data was input to the LSTM in a sequence of 60 time steps for an output prediction of a single time step. Mean square error was calculated from these predictions and used for error thresholding and error count filtering process. Signals with high errors over threshold and with more than one error in a ten-minute window were sent to the final phase synchrony filter. This final step produces a pair-wise synchrony value between two stations that show flagged errors in the same time window. If phase synchrony exceeds the value of 0.9, this time window is classified as an anomaly. Our results show that using this deep learning based anomaly detection algorithm, trained on 6 months of TEC data, we were able to detect ionospheric anomalies after the 06:15 07/29/2021 UTC magnitude 8.2 earthquake south of Chignik, Alaska and after the 10:24 02/06/2023 UTC magnitude 7.5 earthquake in south-east Turkey. These results function as a proof of concept for improved methods of natural hazard detection and ionospheric monitoring in near-real time with GNSS.