Analytical dynamic Spatial Bayesian models for the interface between computer codes Seminar

A class of Gaussian dynamic spatial Bayesian (DSB) models is proposed for the interface between complex deterministic mathematical models implemented as computer codes. They are analytical state-space models that coherently combine the outputs from a computer code, with available measured data to provide inputs to another code in a system. The interface provided not only performs data assimilation in space and time but also explicitly handle uncertainties associated with (i) the outputs from the computer code, (ii) data measurements, (iii) spatial interpolations and (iv) temporal evolution of the underlying process. For being analytical, it can perform calculations relatively quickly when compared with other computer intensive methods being thus appropriate for real time applications.

An analysis of the predictive performances of two Gaussian DSB models (an exponential and an inverse Wishart) is performed in an application using data from the Chernobyl nuclear accident in southern Germany. The models combined predictions of radioactivity ground deposition from a long-range atmospheric dispersal model with measurements of deposited activities to produce statistical inference to be used by a food chain contamination model.