Autoencoders as a Characterization Technique and Aid in the Classification of Volcanic Earthquakes

Abstract

Volcanic seismicity is one of the most relevant parameters for the evaluation of volcanic activity and consequently the prognosis of eruptions. Earthquakes of volcanic origin are of different classes, directly related to the physical process that generates them. The distribution of the data between classes of seismic-volcanic signals generally presents an unbalanced profile (imbalanced datasets), which can hinder the performance of the classification in machine learning models. Therefore, this research presents a characterization technique (feature extract) that, in addition to reducing the dimension of each seismic record, allows a representation of the signals with the most relevant and significant information. This work proposes the use of a Dual Feature Autoencoder (DAF), which is compared with conventional characterization techniques such as Linear Prediction Coefficients (LPC) and Principal Component Analysis (PCA). The training of the model was performed with a dataset containing volcano-tectonic earthquakes (VT), long period events (LP) and Tornillo-type events (Tor) of the Galeras volcano, one of the most active volcanoes in Colombia. The classification results reach 99% of the classification of the mentioned classes.