Seismic precursors to the Whakaari 2019 phreatic eruption are transferable to other eruptions and volcanoes

The problem

Steam-driven (phreatic) eruptions usually occur with little warning. In 2014 an eruption at Mt Ontake in Japan killed 60 people (Yamaoka et al., 2016). In 2019 an eruption at Whakaari (White Island) in New Zealand killed 22 (Dempsey et al., 2020). Even with real-time monitoring, recognising an imminent eruption at Whakaari remains difficult. This is partly because pressurisation and eruption triggering occur within a few tens to hundreds of metres of the surface (Montanaro et al., 2016; Jolly et al., 2020).

Our goal was to find a seismic precursor that (1) recurs before multiple eruptions, (2) is rare when the volcano is not erupting, and (3) appears at more than one volcano.

What we did

We analysed volcanic tremor recorded before 18 eruptions at six volcanoes: Whakaari, Ruapehu and Tongariro in New Zealand, Veniaminof and Pavlof in Alaska, and Bezymianny in Kamchatka, Russia. We processed a combined 50 years of data, from one seismometer near the vent of each volcano.

We measured tremor in three frequency bands, 2 to 5, 4.5 to 8 and 8 to 16 hertz (vibrations per second). Our main quantity is the displacement seismic amplitude ratio (DSAR), the ratio of ground displacement in the medium band to that in the high band. High-frequency shaking is attenuated more strongly as it travels up from below the surface. A change in the ratio may therefore reflect a change in the shallow rock, such as sealing (Caudron et al., 2019).

We normalised the records, divided them into overlapping 48-hour windows, and computed 524 features for each window of the three bands and DSAR, about 2000 in all. For each feature, we correlated the four weeks before each eruption with those before every other eruption, and shortlisted the features with the highest average correlation.

To test each candidate, we correlated the four weeks before one eruption, the template, with every four-week segment of the record. A Kolmogorov–Smirnov test then compared correlations before eruptions with those during repose. We repeated this with each eruption as the template, shifted back at random by up to 7 days to allow for differences in timing, and compared the results with a benchmark built from templates taken at random from repose.

What we found

Two features passed these tests. The first is the nDSAR median, the median of normalised DSAR over 48 hours. Before several eruptions it rose steadily to a peak about 2 to 4 days beforehand. The second feature, nDSAR rate variance, measures how much DSAR changes between consecutive values. It was persistently high 5 to 10 days before six major eruptions, one at each volcano.

With Whakaari 2019 as the template, the nDSAR median before each of the other four Whakaari eruptions correlated more closely than 90% of four-week periods from more than 40 years without eruptions. The p-value was 0.00015. This suggests that the correlations before eruptions come from a different distribution from those during repose. Eruptions at Ruapehu, Tongariro and Veniaminof also ranked highly, with larger p-values of 0.017, 0.066 and 0.072. A third candidate, a 75-minute oscillation in high-frequency tremor, was strong before six eruptions. However, it also appeared to occur frequently during repose. We rejected it as spurious.

In repeated tests, both features distinguished the five Whakaari eruptions from the benchmark. They did so less clearly when steam-driven eruptions at Ruapehu and Tongariro were added. When Whakaari eruptions were pooled with magmatic eruptions at Veniaminof, Pavlof and Bezymianny, the results were not obviously different from the benchmark. We suggest that eruption style limits transferability. This would be expected if the precursor records a physical process specific to steam-driven eruptions.

With surface observations, we identified five phases before the 2019 Whakaari eruption. We interpret sustained tremor from 10 to 23 November as magma interacting with groundwater (phase 1). From 23 November to 2 December, day-long tremor cycles accompanied frequent gas emissions and water jetting (phase 2). We interpret each cycle as pressure building below a weak seal until the seal failed.

After 3 December, medium-frequency tremor rose steadily, high-frequency tremor did not, and the nDSAR median increased. No sulfur dioxide emissions or unusual surface activity were reported. We hypothesise that this rise records a stronger seal forming and the groundwater beneath it pressurising (phase 3). This is because a seal reduces fluid release at the surface and with it the high-frequency tremor. Continued fluid entry below the seal strengthens the medium-frequency tremor. In the final days the nDSAR median declined. This could indicate that the groundwater beneath the seal had reached an equilibrium pressure with the fluid supply from depth (phase 4). About 16 hours before the eruption, tremor increased in all bands (phase 5). We interpret this as cracking of the seal that eventually allowed the pressurised fluid to escape explosively.

The 16-hour burst was recognised as a precursor in earlier work (Dempsey et al., 2020). Under this interpretation, we show that pressurisation can be recognised much earlier, in the week before the eruption.

Limitations

Our analysis is retrospective. The correlation method is also sensitive to differences in the timing and duration of a pattern. The random shifts correct for timing only in part. Other methods may find patterns that ours missed. The mechanism that formed the seal in 2019 is uncertain.

Real-time monitoring for these precursors may improve short-term eruption warnings at some volcanoes. However, forecasting skill has to be demonstrated in prospective experiments.

Glossary

A steam-driven (phreatic) eruption is an explosion driven by hot, pressurised fluid flashing to steam when it is suddenly released.

A magmatic eruption is one in which magma itself is erupted.

Volcanic tremor is continuous low-level shaking of the ground at an active volcano.

The displacement seismic amplitude ratio (DSAR) is the ratio of tremor displacement in the 4.5 to 8 Hz band to that in the 8 to 16 Hz band. The prefix n means the values were normalised.

Normalisation rescales records from different volcanoes to a common scale so that they can be compared.

A hydrothermal seal is a low-permeability layer that restricts the flow of gas and water to the surface.

A precursor is a signal that occurs before eruptions and is rare at other times, so that it could be used to warn of them.

A feature is a single number that summarises a segment of the tremor record, such as its median.

A correlation coefficient measures how similar two curves are in shape, equal to 1 for the same shape and near 0 for unrelated curves.

The Kolmogorov–Smirnov test checks whether two sets of values are likely to come from the same distribution. A small p-value means they are unlikely to.

Repose is a period when a volcano is not erupting.

A prospective experiment tests forecasts that are issued before the outcome is known.