Detecting imminent eruptive activity at Mt Etna, Italy, in 2007–2008 through pattern classification of volcanic tremor data

Volcano monitoring aims at the recognition of changes in instrumentally observable parameters before hazardous activity in order to alert governmental authorities. Among these parameters seismic data in general and volcanic tremor in particular play a key role. Recent major explosive eruptions such as Okmok (Aleutians) and Chaitén (Chile) in 2008 and numerous smaller events at Mt Etna (Italy), have shown that the period of premonitory seismic activity can be short (only a few hours), which entails the necessity of effective automatic data processing near on-line. Here we present a synoptic pattern classification analysis based on Self Organizing Maps and Fuzzy Cluster Analysis which is applied to volcanic tremor data recorded during a series of paroxysmal eruptive episodes and a flank eruption at Etna in 2007–2008. In total, eight episodes were analyzed; in six of these significant changes in the dynamic regime of the volcano were detected up to 9 h prior to the onset of eruptive activity, and long before changes in volcanic tremor amplitude and spectral content became evident in classical analysis. In two cases, the state transition was < 1 h before the onset of eruptive activity, which we interpret as evidence for very rapid magma ascent through an open conduit. We further detected twenty failed paroxysms, that is episodes of volcanic unrest that did not culminate in eruptive activity, between March and April 2007. As the application of the software for this synoptic pattern classification is straightforward and requires only moderate computational resources, it was possible to exploit it in an on-line application, which was tested and now is in use at the Istituto Nazionale di Geofisica e Vulcanologia in Catania for the monitoring of Etna. We believe that the pattern classification presented here may become a powerful addition to the repertoire of volcano monitoring tools and early warning techniques worldwide.

Research Highlights
► The paper deals with unsupervised classification of patterns derived from volcanic tremor.
► Changes in signal characteristics before unrests are recognized at a very early stage.
► The robustness of the approach allows automatic on-line processing of the data.
► The pattern classifier is now in use at the operative control center of INGV Catania.