Investigating large-scale change in volcanic activity using machine learning

Eruption forecasting is a central goal of volcanology. Continued monitoring of volcanic systems has resulted in the identification of periodic behaviour occurring on multiple timescales. This cyclicity has been observed and well-characterised on multiple socially relevant timescales, and in several localities including Mount St Helens, Soufriere Hills Volcano and Santaguito (Barmin et al, 2002; Lamb et al, 2014), giving hope that this behaviour could be forecast. These patterns in behaviour are punctuated by change-points between effusive and explosive activity. Despite an ever-increasing bank of observations associated with volcanic activity, (such as seismicity, geodesy and gas measurements) the ability to pinpoint these change-points in near real-time currently remains beyond reach.
I aim to use machine learning methods to identify change-points in volcanic datasets. Seismic datasets are ideally suited to training a machine learning framework and due to the density of data available I will begin by analysing seismic datasets which span multiple phases of eruption. A pilot study of several machine learning techniques has successfully modelled volcanic behaviour using support vector machine and logistic regression methods. Later, it would be possible to incorporate other datasets into the analysis such as gas and deformation measurements. Comparison of results with non-machine learning analysis of the same datasets will identify the insights gained from new techniques into patterns of volcanic activity.
Moreover, knowledge of when change-points in volcanic activity occur is integral in informing the drivers of changing behaviour in volcanic systems. Machine learning methods have potential for informing physical models of eruption change-points. Currently, the end of a volcanic eruption is poorly defined. The mistaken identification of decreased activity as the end of eruption has been identified as one of the causes of loss of life during the El Chichon eruption in 1982 (Tilling, 2009). Hence, identification of when the end of eruption occurs using data-driven methods provides an independent constraint on timing, thus allowing a more informed approach towards analysis of multi-parameter datasets.