Seismo-ionospheric precursory detection using hybrid Bayesian-LSTM network model with uncertainty-boundaries and anomaly-intensity

Several efforts have been made to understand the complex physical processes involved in a seismic process, but the findings are vague considering prediction capabilities. Nevertheless, recent seismo-ionosphere precursory research has enlightened new pathways toward building an earthquake (EQ) forecasting system. Previously, some conventional mathematical/statistical approaches have been proposed for detecting an anomalous value as a potential precursor. We propose a hybrid Bayesian-based Long Short-Term Memory (B-LSTM) Network model to forecast the Total Electron Content (TEC) data. We applied B-LSTM on different Vertical TEC (VTEC) datasets of two EQs (Mw7.7 Awaran EQ and Mw7.1 Van EQ) by forecasting VTECs with Normalised Root Mean Square Error (NRMSE) scores of 0.15 and 0.10, respectively. We calculated errors and estimated the 99 % confidence interval to extract the VTEC anomalies. The model detects VTEC anomalies successfully but these anomalies may still have some biases and may lead to a false alarm. In order to minimize possible false alarms, we calculated the intensity of each anomaly and found a strong anomaly that occurred 2–3 days before the EQs. To strengthen the relationship of the detected VTEC anomalies with the investigated earthquakes, we examined the state of space weather conditions during and before the event. Our analysis expands the use of deep learning methods in EQ prediction and VTEC forecasting that can be used for various applications e.g. space weather and navigation.