The timing and processes that govern the end of volcanic eruptions are not yet fully understood, and there currently exists no systematic definition for the end of a volcanic eruption. Currently, end of eruption is established either by generic criteria (typically 90 days after the end of visual signals of eruption) or criteria specific to a given volcano. We explore the application of supervised machine learning classification methods: Support Vector Machine, Logistic Regression, Random Forest and Gaussian Process Classifiers and define a decisiveness index D to evaluate the consistency of the classifications obtained by these models. We apply these methods to seismic time series from two volcanoes chosen because they display contrasting styles of eruption: Telica (Nicaragua) and Nevado del Ruiz (Colombia).We find that, for both volcanic systems, the end-date we obtain by classification of seismic data is 2 -4 months later than end-dates defined by the last occurrence of visual eruption (such as ash emission). This finding is in agreement with previous, general definitions of eruption end and is consistent across models.Our classifications have a higher correspondence of eruptive activity with visual activity than with database records of eruption start and end. We analyse the relative importance of the different features of seismic activity used in our models (e.g. peak event amplitude, daily event counts) and find little consistency between the two volcanic systems in terms of the most important features which determine whether activity is eruptive or non-eruptive. These initial results look promising and our approach may offer a robust tool to help determine when an eruption has ended in the absence of visual confirmation.

Understanding the timing of critical changes in volcanic systems, such as the beginning and end of eruptive behavior, is a key goal of volcanic monitoring. Traditional approaches to forecasting these changes have used models motivated by the underlying physics of eruption onset, which assume that geophysical precursors will consistently display similar patterns prior to transition in volcanic state. We present a machine learning classification approach for detecting significant changes in patterns of volcanic activity, potentially signaling transitions during the onset or end of volcanic activity, which does not require a model of the physical processes underlying critical changes. We apply novelty detection, where models are trained only on data prior to eruption, to the precursory unrest at Augustine Volcano, Alaska in 2005. This approach looks promising for geophysically monitored volcanic systems which have been in repose for some time, as no eruptive data is required for model training. We compare novelty detection results with multi‐class classification, where models are trained on examples of both non‐eruptive and eruptive data. We contextualize the results of these classification models using constraints from petrological, satellite and visual observations from the 2006 eruption of Augustine Volcano. The transition from non‐eruptive to eruptive behavior we identify in mid‐November 2005 is in agreement with previous estimates of the initiation of dike intrusion prior to the 2006 eruption. We find that models which include multiple types of data (seismic, deformation, and gas emissions) can better distinguish between non‐eruptive and eruptive data than models formulated on single data types.

Volcano-seismic event classification represents a fundamental component of volcanic monitoring. Recent advances in techniques for the automatic classification of volcano-seismic events using supervised deep learning models achieve high accuracy. However, these deep learning models require a large, labelled training dataset to successfully train a generalisable model. We develop an approach to volcano-seismic event classification making use of active learning, where a machine learning model actively selects the training data which it learns from. We apply a diversity-based active learning approach, which works by selecting new training points which are most dissimilar from points already in the model according to a distance-based calculation applied to the model features. We combine the active learning with an existing volcano-seismic event classifier and apply the model to data from two volcanoes: Nevado del Ruiz, Colombia and Llaima, Chile. We find that models with data selected using an active learning approach achieve better testing accuracy and AUC (Area Under the Receiver Operating Characteristic Curve) than models with data selected using random sampling. Additionally, active learning decreases the labelling burden for the Nevado del Ruiz dataset but offers no increase in performance for the Llaima dataset. To explain these results, we visualise the features from the two datasets and suggest that active learning can reduce the quantity of labelled data required for less separable data, such as the Nevado del Ruiz dataset. This study represents the first evaluation of an active learning approach in volcano-seismology.

No abstract