Evaluating eruption uniqueness within a large group of volcanoes using feature analysis of seismic data: implications for eruption forecasting and risk management

In volcanology, there is a tension between the perception of a volcano as a unique entity and the desire generalize understanding across many volcanoes. On the one hand, much can be learned by studying the granular detail of individual eruptions, emphasizing their unique character or 'personality'. However, it is also common to classify eruptions under broad categories of style and type. Thus, there is an apparent conundrum: are volcanoes unique individuals or do they share common processes that are readily generalized (Cashman and Biggs, 2014; https://doi.org/10.3389/feart.2014.00028). Whichever is the dominant viewpoint adopted has significant implications for volcano eruption forecasting and risk management. For instance, in assuming that volcanoes are unique, a tailored forecasting system needs to be developed for each system. On the other hand, assuming that pre-eruptive processes of volcanic eruptions are common allows sharing of information between systems. This helps build a large databank of case studies from which eruption forecasting models can develop generalized approaches. As eruptions are rare events at most volcanoes, the second approach is desirable if it is feasible.

In a recent study, we have found some commonalities between eruptions at three New Zealand volcanoes by processing pre-eruption seismic data using feature time-series analysis (Ardid et al., 2022; https://doi.org/10.1038/s41467-022-29681-y). Here, we expand this analysis to 18 volcanoes and more than 40 eruptions that cover several regions of the world. We extend our machine learning approach to explore a larger range of input signals, searching for seismic commonalities prior to eruptions that are differentiable from background non-eruptive signals. We explore time scale similarities and clustering of volcanos and eruptions. Similarities between eruptions are conceptualized as precursor that potentially reveal common physical processes. Real-time monitoring of located precursors may improve short-term eruption warning systems, and could potentially be used to train machine learning forecasting models, as precursory information is essential for forecast models whose goal is to quantify the probability of eruption.