Machine Learning Analysis of Time- and Frequency-Domain VLF Signal Variations Associated With Seismic Activity

This contribution investigates the relationship between very low frequency (VLF) signal noise reduction and seismic activity using machine learning methods applied to VLF amplitude measurements. The problem is formulated as a binary classification task distinguishing earthquake-related intervals from non-seismic periods, using features derived from both time and frequency domains. Time-domain models show moderate performance, with the best results achieved by Support Vector Machines (AUC ≈ 0.76). In contrast, frequency-domain representations substantially enhance discriminative capability especially for the Deep Learning neural networks. Spectral features corresponding to wave periods between 0.2 and 6 s yield the strongest performance, with F1-scores up to 0.89 and AUC values reaching 0.94, while longer periods remain informative but less effective. These results provide quantitative evidence that VLF signal variations contain seismic-related signatures and demonstrate the effectiveness of spectral analysis combined with machine learning for characterizing earthquake-associated VLF anomalies.