Towards Improved Volcano Monitoring via Automatic Classification of Continuous Tremor

Abstract

Machine Learning provides new means to detect previously unknown patterns in volcano-seismic signals. Some of these patterns may be precursors to eruption, or change of volcanic activity state. This is particularly interesting for ‘wet’ volcanic systems such as Whakaari (White Island), which regularly produce steam-explosions that are difficult to predict. An unsupervised classifier based on Self-Organising Maps was trained on tremor data recorded at Whakaari to understand changes in seismicity throughout different phases of volcanic unrest. The prototype classifier distinguishes between quiescent and unrest signals, focussing the temporal resolution of periods of enhanced hazard. This provides insights into the mechanisms driving individual unrest episodes. The analysis also highlighted the presence of subtle seismic patterns, which were interpreted to reflect a magma intrusion event and subsequent growth of a lava dome. Visualisation of the periodic characteristics of the signal allowed earlier detection of this than most other current monitoring techniques. This discovery may benefit monitoring at other volcanoes where similar subtle, periodic signals are observed. In order to improve our understanding of the link between parameters that best describe continuous tremor and volcanic processes, their temporal evolution (‘feature dynamics’) are extracted from a high-resolution 100 Hz seismic signal. This was achieved via High Performance Computing. Different visualisation approaches highlight the numerical existence of pre-defined classes of tremor. Hypothesis and correlation testing helped to extract a small number of interpretable features that efficiently describe tremor classes. This is an improvement on classifiers that work with unknown parameters, or approaches using reduced seismic data as input. Automatic selection of tremor classes was coherent with volcanic regimes manually defined in the literature and also expands existing classification schemes for transient volcanic earthquakes. This study provides a promising start on a journey toward identifying and systematically categorising continuous tremor observed at volcanoes worldwide.