Identification of the ionospheric disturbances prior to large earthquakes in Türkiye using Random Forests algorithm

Abstract

This study investigates the potential of Machine Learning techniques to detect ionospheric precursors of major earthquakes using Total Electron Content (TEC) data. A Random Forest classification algorithm is implemented to IONOLAB-TEC estimates from four significant earthquakes in Türkiye between 2020 and 2023, with magnitudes ranging from Mw 6.7 to 7.8. The algorithm distinguishes between four classes: geomagnetically quiet days, geomagnetically disturbed days, three days preceding earthquakes and earthquake days. Nine statistical features are extracted from TEC data to characterize temporal and spatial ionospheric variations. Results demonstrate that the classifier effectively differentiates between seismo-ionospheric anomalies and other ionospheric disturbances with Macro-Average Accuracy ranging from 93.36 % to 97.76 %. A clear correlation between earthquake magnitude and classification performance is observed with higher magnitude events (Mw 7.8 and 7.5) yielding substantially better classification metrics than lower magnitude events (Mw 6.7). 5-Fold Cross-Validation further confirms this pattern with mean Accuracy increasing progressively from 86.72 % for the lowest magnitude earthquake to 95.56 % for the highest. While the model successfully identifies pre-earthquake ionospheric variations with high Precision and Recall values, occasional misclassification between geomagnetic disturbances and pre-earthquake anomalies highlights the ongoing challenge of distinguishing seismic precursors from space weather effects. These findings demonstrate the potential of Machine Learning approaches for detecting earthquake precursors in ionospheric data and suggest possibilities for developing early warning systems through ionospheric monitoring.