Comparison of ML and DL methods in the problem of classifying TEC waveforms

The results of developing an algorithm for the automated detection of covolcanic ionospheric disturbances are presented. The algorithm employs supervised machine learning models, such as Random Forest and gradient boosting (XGBoost), as well as unsupervised approaches including PCA, KNN, Isolation Forest, and neural networks like FCN and InceptionTime, trained to classify ionospheric total electron content time series derived from GNSS data. XGBoost, trained on wavelet-based features, achieved 68.75% detection accuracy on the test set with an average of 0.15 false positives. When trained on Fourier transform-based features, the same model reached a detection accuracy of up to 81.25%, with an average of 0.27 false positives. While wavelet-based features demonstrated greater versatility, Fourier- based features provided higher accuracy due to their specificity. Neural networks (FCN and InceptionTime) achieved up to 98% detection rates but exhibited a higher false positive rate of up to 0.22 per file. Unsupervised methods showed a high false positive rate (up to 0.61 on average) while detecting up to 99% of disturbances, making them valuable for preliminary data annotation.