Machine Learning for Detecting Time-Transient Phenomena in the Ionosphere and Correlation with Seismo-Induced Events.

The ionosphere, a crucial interface between Earth’s atmosphere and space, demonstrates complex temporal dynamics influenced by both terrestrial and extraterrestrial factors. This research investigates the potential of detecting precursors to seismic events by analyzing transient phenomena within the ionosphere. We utilized machine learning algorithms to process and analyze extensive VLF electromagnetic spectrum data gathered by the Demeter satellite over the period from 2005 to 2010. During this five-year duration, approximately 8000 earthquakes with magnitudes of 5.0 or higher were recorded.

We employed a grid-based method, segmenting the Earth's surface into 20x20 degree grids and examining eleven low-frequency bands of electric field data. A time-series dataset was developed from the power spectrum by deriving the feature of interest from each frequency band. Our approach utilized an LSTM Autoencoder model trained to identify anomalies in the time-series data from daytime orbital observations. The model demonstrated effective generalization, successfully detecting a high proportion of seismic-correlated anomalies. In the frequency-based analysis, more feature-specific significance was identified, further enhancing detection accuracy across various frequency bands. The model outperformed random sampling methods, underscoring the reliability of the detected anomalies.

These findings highlight the model's proficiency in detecting ionospheric anomalies, thereby enhancing the broader understanding of ionosphere-lithosphere interactions. Incorporating machine learning techniques into ionospheric research marks a significant advancement in the detection of ionospheric disturbances, providing a robust framework for correlating ionospheric disturbances with seismic events.