Análisis comparativo entre aprendizaje supervisado y aprendizaje semi-supervisado para la clasificación de señales sísmicas vulcanológicas del volcán Cotopaxi

Abstract

Inside of Cotopaxi volcano certain events called vulcanological microseisms occur, it is essential to identify these events to know what we are facing in the face of a possible eruption. However, tagging events is a difficult task that is time consuming and requires expert volcanologists to do this job. To alleviate the need to require abundant tagged data to train a supervised classification algorithm, we propose a semi-supervised approach using a self-training algorithm. First, we extract several relevant characteristics of the microseisms contained in the records of the database provided by the Geophysical Institute. Then we apply principal component analysis to reduce the redundancy present in the characteristics and finally classify the microseisms using an SVM algorithm, in turn, we perform the optimization of SVM hyperparameters to train the algorithm with optimal learning values according to the basis of data available. As a result of this methodology, we show that although the accuracy of the use of a supervised scheme is still better than a semi-supervised one, if we allow a 10% false positive rate, our approach achieves a performance similar to that of supervised techniques. This demonstrates the potential of semi-supervised schemes.