Forecasting volcanic eruptions on White Island in the next month with Bayesian network modelling: Preliminary results from an expert elicitation work

(Inc. GST)
(Ex. GST)
Write a Review

Christophersen, A. 2017 Forecasting volcanic eruptions on White Island in the next month with Bayesian network modelling : Preliminary results from an expert elicitation work. Lower Hutt, N.Z.: GNS Science. GNS Science report 2016/18 183 p.; doi: 10.21420/G20G9B

Abstract: This report presents preliminary findings of an expert elicitation workshop to estimate probabilities for a Bayesian Network (BN) model that forecasts volcanic eruptions on White Island. As well as describing the procedures this report includes all original answers as well as the preliminary analysis of the BN. It has the purpose to provide feedback to the participating experts and serve as a discussion document for future model development. The aim of the project is twofold: 1) to explore BNs as decision support tool for assessing volcanic hazard and risk, and 2) to introduce structured expert elicitation for capturing uncertainty in expert knowledge. BNs are graphical and probabilistic models that can be used to quantitatively model complex systems such as volcanoes. A recent application of a BN to analyse the 1975-1977 Guadeloupe volcanic crisis is used here to demonstrate how discrete and causal BNs work. The report also describes briefly the concepts of structured expert elicitation. A small team with expertise covering volcanology, social science, earthquake forecasting, neural networks and BN modelling modified the Guadeloupe BN for forecasting eruptions on White Island. The team also developed a questionnaire with all the necessary questions to elicit the probabilities for the BN. The head of department volcanology and the leader of the research programme in volcanology advised on experts to involve in reviewing the BN structure and the questions. In discussion with these experts the BN was modified to present the current conceptual mode of White Island and the monitoring data that are regularly collected. The questionnaire was updated and workshop notes prepared that included a brief description of the BN. A workshop was held over two half days at the beginning of December 2015 and included nine experts from within the GeoNet volcano monitoring team, one from the University of Auckland and one from Massey University. The workshop was interactive with discussions on structured expert elicitation and BN modelling. The experts estimated 120 conditional probabilities and their uncertainties and answered questions on thresholds for monitoring data as well as general comments on the BN and the process. For many conditional probabilities there is a wide spread of answers indicating large uncertainties about the variable in question. Overall, the probabilities for monthly eruption seem higher than observed. This might be a consequence of the challenge to estimate small probabilities. This challenge makes it even more important to estimate the probabilities within a model contest because the model can be tested and calibrated with observations. The experts had several suggestion to modify the BN by removing some of the components that were not clear, and clarifying others. General the feedback on both the BN and the structured expert elicitation was positive. Taking the feedback and findings into account, the model development is continuing. (aut h)