Probabilistic, Multi-Sensor Eruption Forecasting at Remote Volcanoes

https://agu.confex.com/agu/agu24/meetingapp.cgi/Paper/1654818(external link)

Abstract:

Most volcanoes are sparsely monitored and even well monitored volcanoes often have relatively short data records compared to their eruption history. In addition, some data types are infrequently sampled, and sensor or infrastructure failures can lead to large data gaps, especially on remote island volcanoes and following eruptions.

We have developed a Bayesian Network-based multiple data stream eruption forecasting model tailored for Whakaari/White Island, an andesite island volcano in New Zealand. Our model utilizes seismic tremor recordings, earthquake rates, and gas emissions data (CO₂, SO₂, H₂S) to generate interpretable, probabilistic forecasts. The model demonstrates good predictive performance, showing increased eruption probabilities months to weeks prior to the biggest three eruptions in the past decade. More importantly, it can still provide useful forecasts even when some of the data streams are missing. We also observe that eruptions consistently occurred after the cumulative probability exceeded at least 80%.

Although initially developed for Whakaari/White Island, our model is easily adapted to other remote volcanoes, potentially complementing forecasting methods based on single data streams. While our model is purely trained on observational data, Bayesian Networks provide a natural way to combine expert knowledge and data to produce combined forecasts.