Eruption Forecasting of Strokkur Geyser, Iceland, Using Permutation Entropy

Abstract

A volcanic eruption is usually preceded by seismic precursors, but their interpretation and use for forecasting the eruption onset time remain a challenge. A part of the eruptive processes in open conduits of volcanoes may be similar to those encountered in geysers. Since geysers erupt more often, they are useful sites for testing new forecasting methods. We tested the application of Permutation Entropy (PE) as a robust method to assess the complexity in seismic recordings of the Strokkur geyser, Iceland. Strokkur features several minute‐long eruptive cycles, enabling us to verify in 63 recorded cycles whether PE behaves consistently from one eruption to the next one. We performed synthetic tests to understand the effect of different parameter settings in the PE calculation. Our application to Strokkur shows a distinct, repeating PE pattern consistent with previously identified phases in the eruptive cycle. We find a systematic increase in PE within the last 15 s before the eruption, indicating that an eruption will occur. We quantified the predictive power of PE, showing that PE performs better than seismic signal strength or quiescence when it comes to forecasting eruptions.

Dynamical changes of seismic properties prior to, during, and after 2014–2015 Holuhraun Eruption, Iceland

“…The synthetic tests in subsection 3.1 show that PE, calculated using a fix m and τ , is higher for a signal with higher frequency content compared to one with lower frequencies. A signal with energy in a broad frequency bandwidth usually possesses a high PE (Dávalos et al, 2021;Sudibyo et al, 2022). Independent of its frequency content, PPE is more affected by the presence of non-zero-crossing oscillation in its signal.…”

“…However, by utilizing PE, PPE and mean IF, we could not observe any changes that could be related to pre-eruptive tremor Eruption forecasting is easier when the pre-eruptive process generates a pattern of distinct seismic events that are chronologically changing in time and depth. In the case of Strokkur Geyser, the PE variation can characterize the four different phases in the geyser's eruptive cycle (Sudibyo et al, 2022), which are eruption, conduit refilling, gas accumulation in the bubble trap, and the collapses of the bubble gas in the shallow conduit (Eibl et al, 2021). These different processes generate distinct signals with different complexity, thus resulting in distinct values of the corresponding PE.…”

“…Furthermore, each phase takes place in separated locations and depths (Eibl et al, 2021). Therefore, not only Sudibyo et al (2022) can observe a high correlation between PE and the hypocentral distance between seismic event's source and the seismic station, they can also use PE to accurately predict the geyser's eruptions. Konstantinou et al (2022) reported a consistently decreasing PE prior and during 3 eruptions in Shinmoedake, a stratovolcano in Japan, which is associated to the dominant occurrence of the pre-and eruptive tremor.…”

“…While PE can detect a seismic precursor occurring days before an eruption (Glynn and Konstantinou, 2016), PE is also reported to be sensitive to detect fast changes. Sudibyo et al (2022) calculated PE for 63 eruptions of Strokkur geyser in Iceland where the duration of a typical eruptive cycle is 3.7 ± 0.9 min (Eibl et al, 2020). The PE can resolve the typically observed four phases of Strokkur's eruptive cycle, that last several seconds to several minutes (Eibl et al, 2021), and 1 s long processes therein at high resolution (Sudibyo et al, 2022).…”

Generalized Eruption Forecasting Models using Machine Learning Trained on Seismic Data from 24 Volcanoes

“…To enhance accuracy, avenues for improvement include expanding precursory analysis by considering longer time periods (multi-time scale approaches; Ardid et al, 2023) and incorporating additional data types such as gas emission rate 59,60 , thermal anomalies 61 , and magnetotelluric data 62,63,64 . Also, application to geysers monitored with seismic data 65,66 given their frequent eruptive nature and lower time between eruption offers an opportunity to improve these methods with more balanced datasets between eruptive and non-eruptive data for the ML classification problem.…”