부산대학교 도서관

A short‐term (30 days before an earthquake) prediction of an earthquake is a big challenge in seismology. As a first step, we apply deep learning to the ionospheric total electron content (TEC) data between 2003 and 2014 to detect the seismo‐ionospheric precursors of M ≥ 6.0 earthquakes in Taiwan. The bidirectional Long Short‐Term Memory (Bi‐LSTM) network is employed to use observed input data (features) to obtain the sequential TEC variations. The five input features are sequential vectors of TEC, the geomagnetic index Dst, the solar activity index F10.7, sunspot number (SSN), and solar emission index Lyman‐α. The daily values of F10.7, SSN, and Lyman‐α are converted into hourly values, depending on the solar elevation angle. The calculated hourly TEC variations can be more precisely predicted with this data conversion. We calculate the normalized difference of errors between two 15‐day adjacent stages as the "relative error". Three trained models with the best discrimination between the relative errors of earthquake and no‐earthquake cases are chosen as classifiers. These three classifiers are then used to have a majority vote to declare whether the 30‐day period is related to the preparation of an earthquake or not. The results show that all 22 positive cases (earthquakes) are successfully predicted, giving a true positive rate of 100%. Among the 19 negative cases (normal cases), 10 of them are true negative." Overall, a high accuracy of 78.05% is obtained. Plain Language Summary: A short‐term (30 days before an earthquake) prediction of earthquake is a big challenge in seismology. As a first step, we apply deep learning to the ionospheric total electron content (TEC) data, solar activity and geomagnetic data between 2003 and 2014 to detect the seismo‐ionospheric precursors of magnitude M ≥ 6.0 earthquakes in Taiwan. The deep learning method is employed to use the running 7‐day observed data to predict the next hour TEC value. The seismo‐ionospheric precursors can be identified by the normalized difference between the predicted TEC and the observed TEC data. The results show that all 22 positive cases with M ≥ 6.0 earthquakes are successfully predicted, while 10 of 19 negative cases without M ≥ 5.3 earthquakes are correctly predicted "No". A high accuracy of 78.05% is obtained. Key Points: Deep learning is applied to the detection of ionospheric precursors associated with M ≥ 6.0 earthquakes between 2003 and 2014 in TaiwanAll positive cases are successfully predicted, while 10 of 19 negative cases are predicted "NO". A high accuracy of 78.05% is obtainedThere is no occurrence of M ≥ 5.3 earthquake when the trained models predict "NO". The false rate is 0% [ABSTRACT FROM AUTHOR]