Abstract

Continuous seismic monitoring has achieved a key position in monitoring active volcanoes. However, it comes with the problem of a huge quantity of data difficult to handle. Automatic pattern recognition techniques have proven effective in seismic data processing and, consequently, have been increasingly implemented to solve different tasks. In this paper we investigate the development of the characteristics of the seismic signal on Mt Etna and its relation to regimes of volcanic activity. To this purpose we apply classification methods both with supervisor (Artificial Neural Networks, Support Vector Machine) and without supervisor (cluster analysis). The former "learn" from exemplar patterns, inferring rules to deal with new and/or noisy data to classify, whereas the latter seek for heterogeneities in the data set applying a specific metric. The choice of automatic classification methods is determined by the necessity to solve rather complex discrimination problems using as little a-priori information as possible. We focus on volcanic tremor recordings at Mt Etna in 2001, a time span where there is a wide variety of feature signals, encompassing periods of pre- and post-eruptive quiescence, episodes of lava fountains, and a 23 day-long effusive activity. We establish four target classes, i.e., pre-eruptive, lava fountains, eruptive, and post-eruptive. The a-priori information used for the classification with supervisor is based on volcanological reports, and therefore it does not directly depend on the characteristics of the seismic signal. We discuss performance and characteristics of the different techniques in light of an implementation to automatically analyze seismic data and reduce volcanic hazard.