Semi-Supervised Learning for Volcanic Seismic Event Classification: A Comparative Study of Self-Training and Label Spreading

This paper presents a comprehensive analysis of the impact of the proportion of labeled data on the performance of semi-supervised learning models for the classification of volcanic micro-seismic events. This work utilizes a dataset of over 22,000 micro-earthquake records from the Cotopaxi Volcano, with extracted features from time, frequency, and scale domains. Two semi-supervised approaches are implemented: Self-Training, using Support Vector Machine (SVM), Random Forest (RF), and Naive Bayes (NB) as base classifiers; and L abel Spreading, based on graph-based label propagation. Model performance is evaluated using two primary metrics: the Area Under the Receiver Operating Characteristic Curve (AUC-ROC) and F1-score. The SVM Self-Training model achieved the highest results, with an AUC of 0.9575 and an F1-score of 0.9472 when trained with 90% of labeled data. The RF model also performed robustly, particularly in noisy or imbalanced scenarios, reaching an AUC of 0.9505 and an F1-score of 0.9436. In contrast, NB showed limited gains, and the Label Spreading model failed to improve with more labeled data, stabilizing at an AUC around 0.88. These findings highlight the effectiveness of SVM and RF in leveraging unlabeled data for seismic event classification under varying label scarcity conditions.