Bayesian Networks to Forecast Eruptions at Frequently Active Volcanoes from Monitoring Data and Expert Elicitation: Application to Mt Ruapehu

Bayesian networks (BNs) are probabilistic graphical models that have been applied to model complex systems and their uncertainties in many different domains. They provide a flexible framework to incorporate conceptual understanding of a system, learn from data when available and incorporate expert knowledge when data are sparse. The graphical component of the BN visually represents the input variables and helps users to gain a clear understanding of the modelled system; robust statistics describe the joint probability distribution of all variables. Applications include modelling multiple data streams for eruption forecasting and volcanic hazard and risk assessments.  

Despite offering many benefits, BNs are not widely used for volcano monitoring or eruption forecasting yet. We have developed a method, based on the process flow of risk management by the International Organization for Standardization, to build a BN model to support decision-making. We have worked with the New Zealand Volcano Monitoring Team (VMG) to apply this method to develop an eruption forecast model for Mt Ruapehu, Aotearoa New Zealand.  

The model structure was built with expert elicitation, based on the conceptual understanding of Mt Ruapehu, and with reliance on the long eruption catalogue and long record of monitoring data. The model was parameterised partly by data learning, complemented by expert elicitation. The model forecasts the probability of volcanic eruptions that impact beyond the crater rim within the next 28 or 91 days. The VMG has regularly estimated this probability since 2014. The retrospective BN model forecasts agree well with the VMG elicitations. The BN model is now implemented as a software tool to automatically calculate daily forecast updates.