Seismic background level growth can reveal slowly developing long-term eruption precursors

Abstract

The accelerating growth of seismic unrest before eruptions has been observed at many volcanoes and utilized for eruption forecasts. However, there are still many eruptions for which no precursory unrest has been identified, even at well-monitored volcanoes. The recent eruptions of Shinmoe-dake, Japan, had been another negative example of this kind. Here we present seismological evidence that the eruption preparation had been ongoing at the shallow depths beneath Shinmoe-dake for several months to a year. We investigated the seismic background level (SBL) of eleven-year data recorded around the volcano, including two stations about 1 km from the eruptive crater. We searched for persistent weak signals, focusing on low-amplitude time windows recorded during quiet nighttime. Then the spectra of daily background noise were classified by clustering analysis. The SBL analysis successfully revealed very weak precursory tremors from more than several months before the eruption, and residual tremors to the end of the eruptive period. The precursory signals grew acceleratory in a similar way as is assumed in the material failure forecast method applied to eruption forecasts. However, their growth was significantly slower and longer compared to previous cases. Such slow and quiet preparations would not be captured by conventional seismological methods but could be a common feature at volcanoes with developed hydrothermal systems. We expect that long-term SBL analyses on proximal seismic data will help detect early precursors, even at seismically quiet volcanoes, and will also help judge the end of an eruptive period.

Machine Learning Analysis of Seismograms Reveals a Continuous Plumbing System Evolution Beneath the Klyuchevskoy Volcano in Kamchatka, Russia

“…We can exclude the oceanic microseisms or large‐scale meteorological phenomena for this behavior since the center frequencies of the first‐order wavelets do not cover frequencies below 0.78 Hz. Other studies have shown that the signal properties of the ambient seismic noise can change due to volcanic activity (Glynn & Konstantinou, 2016; Ichihara et al., 2023).…”

“…Due to advancements in information processing and the introduction of continuous recordings, some studies have shown that continuous seismograms contain relevant information beyond classical event catalogs. The entropy of the seismic noise and the seismic background level seems to vary significantly prior to eruptions (Glynn & Konstantinou, 2016; Ichihara et al., 2023; Rey‐Devesa et al., 2023). Machine learning strategies, in particular unsupervised learning, provide a promising approach for automatically analyzing large amounts of continuous seismograms and inferring such patterns without requiring predefined labels (e.g., Holtzman et al., 2018; Jenkins et al., 2021; Köhler et al., 2010; Ren et al., 2020; Seydoux et al., 2020; Steinmann, Seydoux, Beaucé, & Campillo, 2022; Steinmann, Seydoux, & Campillo, 2022; Zali et al., 2023).…”