Bayesian Networks for Eruption Forecasting

Volcano monitoring observations range from high-frequency measurements, collected a hundred times per second, to manually collected data with a few data points per year. Searching for common patterns that may precede eruptions in such disparate data streams is challenging. This is especially so when eruptions are relatively rare events from which to learn about precursors from, no matter how well monitored. About nine unrest episodes at New Zealand volcanoes since 2000 did not result in an eruption (compared to 10 eruptive episodes over the same time period). This gives particular importance to volcanologists’ conceptual understanding of the physical and chemical processes inside the volcano that drive the unrest and whether it will tip into an eruption.

We propose the application of Bayesian Networks (BNs) to integrate monitoring data and expert knowledge to create probabilistic eruption forecasts. BNs are well suited to produce forecasts that reflect the, sometimes large, uncertainties in both data and conceptual models. The graphical presentation of variables and the option to model causal processes make BNs easy to understand and interpret, a key requirement for decision support tools.

We show preliminary results for eruption likelihood over a 30-day forecast window at Whakaari/White Island using seismic and gas-flux data since 2010. For five out of the six eruptions during that time, we observe an increase in eruption likelihood several weeks prior to eruption. However, increases are also observed at other times, sometimes for several months, possibly indicating heightened unrest that did not result in eruption.