Deep Learning Model for Detecting Global Ionospheric Electric Field Perturbations and Seismic Correlation

Detecting ionospheric electric field anomalies that precede earthquakes is still an open scientific challenge. Progress is delayed by limitations in current observational methods and the lack of standardised, reproducible analysis and approaches. As a result, reported pre-seismic signatures often remain inconsistent and hard to validate across different events and datasets.

This study presents a data-driven deep learning (DL) approach that moves beyond traditional model-based frameworks by utilizing satellite  based ionospheric electric field measurements from the DEMETER mission (2005 - 2010). A grid based geospatial organization is applied to ensure consistent spatial mapping of the observed spectral electric field data, resulting in structured time series for analysis. The methodology focuses on lower frequency bands below 3 kHz, comprising calibrated data from 11 distinct frequency bands. The study adopts an iterative rolling-window strategy instead of the conventional fixed division of data into training and validation sets, with background corrections applied iteratively within each window. An unsupervised LSTM autoencoder is implemented and trained using this approach, preserving long term temporal nature of data. Anomalies detected by the model are subsequently examined for potential seismic associations and evaluated using statistical tests. A statistical investigation on spatial and temporal windows identifies an optimal configuration of a 22° x 22° spatial window along the orbital foot print, and a 48-hour temporal window for the study. Under this configuration, models trained solely on non-seismic data are able to detect anomalous sequences that show a statistically significant association with seismic events, exceeding the random baseline with a 2 - 3σ deviation range.

This study shows that modern deep learning methods, combined with flexible and adaptive training strategies and sound statistical analysis, can successfully extract useful information from satellite-based ionospheric electric-field data.