Classification of pre-eruption and non-pre-eruption epochs at Mount Etna volcano by means of artificial neural networks

• Published in: Geophysical research letters Vol. 34; no. 10; pp. L10311 - n/a
• Main Authors: Castellaro, Silvia, Mulargia, Francesco
• Format: Journal Article
• Language: English
• Published: Washington, DC American Geophysical Union 01-05-2007; Blackwell Publishing Ltd
• Subjects: artificial neural networks; Computational Geophysics; Earth sciences; Earth, ocean, space; eruptions; Europe; Exact sciences and technology; Geographic Location; Mathematical Geophysics; Neural networks, fuzzy logic, machine learning; pattern recognition; Seismology; Time series analysis; Volcanic hazards and risks; Volcano seismology; Volcanology; Southern Europe; Italy; Sicily Italy; Mount Etna; artificial neural networks; eruptions; pattern recognition; algorithms; Europe; neural networks; volcanoes; tectonics; classification; multivariate analysis
• ISSN: 0094-8276; 1944-8007
• DOI: 10.1029/2007GL029513

We apply artificial neural networks to the classification of pre‐eruption time epochs of Mount Etna volcano on the basis of variables depending on tectonics and on the volcano ‘recharging system’. We consider time‐epochs from 7 to 30 days and train the supervised nets, with the aim of recognizing the time epochs preceding summit eruptions, lateral eruptions and not preceding any eruption. Tested on a number of independent data sets, these patterns are found to be efficient (75 ± 10% success) in recognizing pre‐summit eruption epochs, while distinguishing pre‐lateral from non‐pre‐eruption epochs is impossible. We then apply non‐supervised algorithms to the whole set of data obtaining a confirmation of the findings of supervised nets. This difficulty in recognizing patterns characteristic of pre‐lateral eruption epochs is at odds with all previous work and seems to depend on the small size of the eruptive series, which makes unstable the results of any multivariate analysis.