SafeNet: SwArm for Earthquake Perturbations Identification Using Deep Learning Networks

Abstract

Low Earth orbit satellites collect and study information on changes in the ionosphere, which contributes to the identification of earthquake precursors. Swarm, the European Space Agency three-satellite mission, has been launched to monitor the Earth geomagnetic field, and has successfully shown that in some cases it is able to observe many several ionospheric perturbations that occurred as a result of large earthquake activity. This paper proposes the SafeNet deep learning framework for detecting pre-earthquake ionospheric perturbations. We trained the proposed model using 9017 recent (2014–2020) independent earthquakes of magnitude 4.8 or greater, as well as the corresponding 7-year plasma and magnetic field data from the Swarm A satellite, and excellent performance has been achieved. In addition, the influence of different model inputs and spatial window sizes, earthquake magnitudes, and daytime or nighttime was explored. The results showed that for electromagnetic pre-earthquake data collected within a circular region of the epicenter and with a Dobrovolsky-defined radius and input window size of 70 consecutive data points, nighttime data provided the highest performance in discriminating pre-earthquake perturbations, yielding an F1 score of 0.846 and a Matthews correlation coefficient of 0.717. Moreover, SafeNet performed well in identifying pre-seismic ionospheric anomalies with increasing earthquake magnitude and unbalanced datasets. Hypotheses on the physical causes of earthquake-induced ionospheric perturbations are also provided. Our results suggest that the performance of pre-earthquake ionospheric perturbation identification can be significantly improved by utilizing SafeNet, which is capable of detecting precursor effects within electromagnetic satellite data.

Worldwide Statistical Correlation of Eight Years of Swarm Satellite Data with M5.5+ Earthquakes: New Hints about the Preseismic Phenomena from Space

“…Both the investigated time windows show alarmed times below the 50% threshold-in some cases, lower than 1%-which provide further evidence of the non-chance nature of the detected anomalies in relation to the following earthquakes. The accuracy is similar to that of a previous work performed on Swarm electromagnetic data and earthquakes but using a machine learning approach [40]. Furthermore, the overall performances reflected the previous considerations, with the scores for the magnetic field being generally better than those for the electron density.…”

Developing a Fuzzy Inference System Based on Multi-Sensor Data to Predict Powerful Earthquake Parameters

“…8 Nepal 2015 [11], the M6.0 and 6.5 Italy 2016 seismic sequence [12], the M7.5 Indonesia 2018 earthquake [13], and the M7.1 Ridgecrest, California 2019 earthquake [14]. Statistical proof of the existence of ionospheric precursors analysing DEMETER and Swarm satellites has also prompted some worldwide investigations of M5.5+ (or M4.8+ in the case of DEMETER) shallow earthquakes [15][16][17][18][19]. Creating cloud computing systems such as Google Earth Engine (GEE) has greatly helped to analyse various precursors without the necessity of downloading their corresponding raw data [20].…”

Quick Report of the ML=3.3 on 1 January 2023 Guidonia (Rome, Italy) Earthquake: Evidence of a Seismic Acceleration

“…Swarm mission is composed by three identical satellites in low Earth quasipolar orbits, called Alpha, Bravo and Charlie aiming to measure the geomagnetic field with the best precision available at the state of art [19]. The use of Swarm data to study the preparation of the earthquakes has been explored by several researches in the last years using a single case study approach [18,[20][21][22][23][24][25][26][27][28][29][30][31] or even statistically correlating the magnetic and electron density anomalies of Swarm with M5.5+ earthquakes occurred in the first 4.7 years of Swarm mission by De Santis et al [32], the first 8 years by Marchetti et al [33] or using Machine Learning by Xiong et al [34].…”