A Multi-Network based Hybrid LSTM model for ionospheric anomaly detection: A case study of the Mw 7.8 Nepal earthquake

Abstract

We propose a Multi-Network-based Hybrid Long Short Term Memory (N-LSTM) model for ionospheric anomaly detection to forecast highly irregular data of the ionospheric Total Electron Content (TEC). Previously purposed models were suffering from the problems of vanishing gradient, exploding gradient, uncertainty, and parameter bias. LSTM is used to overcome the vanishing and exploding gradient problems and the remaining two (uncertainty and parameter bias) are subjugated by N-LSTM. We have also discussed the implementation of the model by using the Mw 7.8 Nepal Earthquake (EQ) that occurred on April 25, 2015. The proposed model detects a significant negative anomaly about 14 days before the impending EQ. Furthermore, we find gradual increments in the TEC time-series, where the TEC values gradually increased, in terms of positive anomalies, for four continuous days (April 21–24, 2015) till the mainshock. Further analysis revealed that the nighttime TEC enhancements are more dominant than the daytimes, which persisted for 10 h (22:00–08:00 LT) in the absence of the solar flux, we named them the absolute seismic precursors. These gradual enhancements establish a perfect correlation with several atmospheric anomalies, reported in the previous studies, where the variations in different atmospheric parameters like Outgoing Longwave Radiation (OLR), air temperature, Ozone, and sub-ionospheric Very-Low Frequency (VLF), were reported for 4 consistent days till the EQ day. The N-LSTM result was compatible with a 30-day running median, which is the classical method of ionospheric anomaly detection. We also examined the planetary K-index (Kp), disturbance storm-time (Dst), solar radio flux (F10.7), and solar wind speed (VSW) indices to check the possible attribution of space weather on the ionosphere during the analysis. The results showed that the anomalies of the ionospheric TEC were more dominantly caused by the EQ phenomenon than the space weather during the nighttime hours. Also, N-LSTM provided sufficient short-term prediction performance of the ionospheric TEC and anomaly detection.