Automatic detection and classification of seismic signals of the Popocatepetl volcano, Mexico, using machine learning methods.

Many studies highlight the benefits of using machine learning algorithms for the classification of volcano-seismic signals. However, when it comes to their widespread application, volcano observatories and researchers face two important challenges. i) The performance of these models highly depends on the size of the training set, where large amounts of labeled signals (thousands and sometimes even hundreds of thousands) are needed to get sufficient accuracy. ii) Most of them use data recorded by a single station and from only one component. This “master” station is generally one of the closest to the crater and, in volcanoes, it is common to face technical difficulties that interrupt the continuous recording, especially during periods of increased activity.

This strongly limits the possibility of applying machine learning approaches for efficient monitoring of volcanoes, especially during unrest periods.

Here, we show a simple method that addresses these difficulties using the information provided by the entire network of stations operating at Popocatepetl volcano (about 18 stations among permanent and temporal) and using all the components. Initially, we used a mid-size catalog of 507 single-channel labeled events recorded between 2019 and 2020. Later, to increase the size of our dataset and exploit the information provided by different channels, we added the signals of the three components of all the events, as well as signals of selected events recorded at different stations. This enlarged training set comprises 1725 signals of six classes: 345 noise, 324 explosions, 321 long periods (LP), 306 volcano-tectonics (VT), 264 tremors, and 165 regionals. To characterize the data, we used a previously proposed set of 102 features that describe the shape, statistics, and entropy of the signals. Then we applied two classification algorithms, random forest and support vector machines, to both our datasets. Our results show that the model of the enlarged dataset increases the overall accuracy by over 8% compared with the one produced using one station and only one component, with the additional benefit of guarantying continuous monitoring even when the “master” station is not working.