Automatic Classification of Seismo-Volcanic Signals with Deep Learning: The Case of Turrialba Volcano

Monitoring and surveillance of volcanic activity is crucial to properly forecast the associated hazards. In this context, the analysis of volcanic seismicity plays a fundamental role. There are different seismic signals associated with volcanic activity, such as: volcano tectonic earthquakes (VT), long period earthquakes (LP) or low frequency earthquakes, earthquakes associated with explosions, hybrid earthquakes, deep LP earth-quakes and tremors. These are originated due to the movement or pumping of magma, fracturing of the rock under the surface, sound vibrations in the emission conduits, magma gasification processes, collapse of the magmatic chamber and explosions originated by the eruption. The identification and classification of these signals is a complex and time-consuming proces due to the lack of applicability of conventional tectonic earthquake location procedures and the difficulty experienced by expert operators during periods of high volcanic activity. In recent years, numer-ous research works have been carried out proposing approaches based on machine learning techniques, specifically in the field of deep learning, for the automatic classification of seismo-volcanic events. However, due to the intrinsic variability of seismo-volcanic signals and the heterogeneity of the characteristics of volcanic buildings, which greatly influence the waveforms of these signals, a concrete and definitive method for their characterization has not yet been established. In this paper we show how a convolutional neural network (CNN) can be used to classify seismic-volcanic signals at Turrialba volcano, located in Costa Rica. To train this CNN we use a transfer learning approach on 3 different pre-trained model architectures to correctly identify 12 event categories. We evaluate the performance of our proposal with 1941 data collected from seismo-volcanic events of Turrialba volcano. The results show that our approach achieves an accuracy of over 80 % in event classification.