PRESENTATION: Developing Data-Driven Earthquake Forecasts Using Inlabru
Here is the presentation from Dr Kirsty Bayliss in our group at the Seismological Society of America this year. She is funded by the Real-time earthquake rIsk reduction for a reSilient Europe project (H2020 RISE).
ABSTRACT: To develop robust and reliable earthquake forecasts, we must have a good understanding of spatial patterns of seismicity. Log-Gaussian Cox processes with a spatially varying, random intensity field may be used to flexibly model the spatial pattern formed by the locations of earthquakes. Using a Bayesian approach implemented with the R pacakge inlabru, we fit models that use different combinations of spatial covariates that might help describe observed seismicity, including different geophysical and geological observations such as past seismicity, strain rate, mapped fault information, and slip rate on individual faults, and derived products such as distance from nearest fault. The models include a random field component to model remaining spatial structure that cannot be explained by the included components. These models have the advantage of not requiring expert judgement in defining seismic zones using similar input data, while greatly reducing the number of free model parameters. They also allow a finite likelihood of having an event in areas where seismicity is low or absent, hence allowing for ‘surprises’. The flexibility of the <>inlabru approach makes it possible to easily construct and compare models containing different combinations of potentially useful spatial covariates. This allows us to assess the extent to which individual components and combinations of these improve a spatial model of observed seismicity. The best-performing spatial models can then be extended to time-independent forecasts by considering the magnitude distributions of modelled event rates. We demonstrate this method with application to California data. We find that fault distance maps and strain rate data can be particularly useful for constraining spatial seismicity and that fault slip rate models can be useful for describing the spatial distribution of large earthquakes.