A semi-supervised total electron content anomaly detection method using LSTM-auto-encoder

Abstract

Total electron content (TEC) is one of the important features used in studying of ionospheric properties. In seismo-ionospheric studies, variations/anomalies of the vertical total electron content over a seismic event are used to study the manifestation of lithospheric in the ionosphere. It is significant to design a robust TEC anomaly detection algorithm, which is capable of detecting non-spurious Seismo-induced TEC variations. In this paper, the long short term memory (LSTM) based auto-encoder network is presented for the detection of TEC anomalies recorded by GNSS receivers. The LSTM-auto-encoder is applied as a semi-supervised scheme to learn TEC responses in quiet solar and geomagnetic conditions. It then uses the learned TEC features to identify anomalies in a given TEC input data. The method is implemented to detect TEC anomalies recorded by three different GNSS-TEC receivers (ankr, ista, and tubi) in Türkiye. The model is also used to verify results from a recently published study on Mexico earthquake (magnitude 7.4). In each case, plausible results are obtained, and the relationship between detected anomalies with some lithosphere-atmosphere processes are discussed. The method highlights the significance/applications of AI in studying ionospheric variations.