Synopsis of supervised and unsupervised pattern classification techniques applied to volcanic tremor data at Mt Etna, Italy

States of volcanic activity at Mt Etna develop in well-defined regimes with variable duration from a few hours to several months. Changes in the regimes are usually concurrent with variations of the characteristics of volcanic tremor, which is continuously recorded as background seismic radiation. This strict relationship is useful for monitoring volcanic activity in any moment and in whatever condition. We investigated the development of tremor features and its relation to regimes of volcanic activity applying pattern classification techniques. We present results from supervised and unsupervised classification methods applied to 425 patterns of volcanic tremor recorded between 2001 July and August, when a volcano unrest occurred. Support Vector Machine (SVM) and multilayer perceptron (MLP) were used as pattern classifiers with supervised learning. For the SVM and MLP training, we considered four target classes, that is, pre-eruptive, lava fountains, eruptive and post-eruptive. Using a leave one out testing scheme, SVM reached a score of 94.8 per cent of patterns matching the actual class membership, whereas MLP achieved 81.9 per cent of matching patterns. The excellent results, in particular those obtained with SVM, confirmed the reproducibility of the a priori classification. Unsupervised classification was carried out using cluster analysis (CA) and self-organizing maps (SOM). The clusters identified in unsupervised classification formed well-defined regimes, which can be easily related to the four a priori classes aforementioned. Besides, CA found a further cluster concurrent with the climax of eruptive activity. Applying a proper colour-coding to the microclusters (the so-called best matching units) identified by SOM, it was visually possible to follow the development of the characteristics of the tremor data with time, highlighting transitional stages from a regime of volcanic activity to another one. We conclude that supervised and unsupervised classification methods can be conveniently implemented as complementary tools for an in-depth understanding of the relationships between tremor data and volcanic phenomena.