Modeling long-term volcanic hazards through Bayesian inference: An example from the Tohoku volcanic arc, Japan

The need to quantitatively estimate future locations of volcanoes in the long-term is of increasing importance, partly as a result of the requirement of constructing certain types of installations in regions of low geologic risk. The complex geological factors and natural processes controlling the locations of volcanoes make it problematic to estimate future patterns deterministically. Instead, the probabilistic approach can be developed with quite high levels of confidence; however, for regions with few or no volcanoes, there is a need to include additional geological and geophysical data that may indicate the likelihood of future volcanism. We achieve this using Bayesian inference in the Tohoku volcanic arc, Japan, in order to combine one or more sets of geophysical information to a priori assumptions of volcano spatiotemporal distributions yielding modified a posteriori probabilities. The basic a priori assumption is that new volcanoes will not form far from existing ones and that such a distribution ranges from Gaussian ( not so conservative) to Cauchy (conservative). Seismic tomographs are used as an indirect clue, and from this geophysical data a likelihood function is generated in the Bayesian context that updates or fine tunes the initial Gaussian or Cauchy kernels to better reflect the distribution of future volcanism. These models are evaluated using pre-100 ka volcanic events to forecast locations of subsequent events that actually formed from 100 kyr ago to present. Probabilities in Tohoku region range from 10(-10)/yr between clusters and up to 9.8 x 10(-6)/yr near the centers of clusters.