Statistical Analysis of Large Volcano Seismology Datasets to Determine Patterns of Volcanic Behaviour.

Abstract:

Consistent and efficient analysis of both real-time volcano seismology data and of data catalogues from key eruptions is important to characterise seismicity surrounding eruptions and to determine patterns of volcanic behaviour. Seismicity patterns can be characterised by a variety of metrics, such as spectral analysis, identification of repeating waveform families and event classification, and temporal changes in these seismicity patterns can indicate changes in volcanic behaviour. Data catalogues from key eruptions can be large, and during seismic crises real-time monitoring of volcano seismicity can be overwhelming. This highlights the need for simple, rapid and effective analysis of such datasets, however, large-scale or rapid analysis of this type has been hindered by computational limitations associated with cross-correlation of large datasets and the labour intensive nature of waveform classification. Consistent waveform classification remains a challenge during both real-time analysis and retrospective analysis. In real-time many events are classified by an analyst, but during seismic crises there may be hundreds to thousands of events to analyse per day and this can rapidly become unfeasible. Automated classification allows consistent classification of waveforms but often requires an extensive training period. We have developed a fast-approximation method, peakmatch, to cross-correlate large seismic data catalogues for rapid analysis of repeating waveforms, and we use machine-learning techniques to automatically classify seismic waveforms with minimal training data.