Bayesian Hierarchical Modelling of Soil Carbon Dynamics with LibBi
Dan Pagendam (CSIRO, Australia)
Friday 16th October, 2015 15:00-16:00 Maths 204
Deterministic mathematical models are commonplace for describing the dynamics of environmental processes. however, a common (and often challenging) request made of environmental modellers is to provide some quantitative measure of confidence in a model's predictive abilities (sometimes called uncertainty quantification). This is certainly true in models of soil carbon dynamics which are now widely used to fulfil the carbon accounting obligations of developed countries. In this talk, we demonstrate how Bayesian hierarchical modelling can be used for uncertainty quantification by embedding a soil carbon process model (FullCAM / RothC) within a rigorous statistical framework. This approach allows for the inclusion of three sources of error / uncertainty: (i) uncertainty in process dynamics; (ii) uncertainty surrounding model parameters; and (iii) measurement errors in observed data. Our application uses data from the Millennium Tillage Trial, New Zealand for which we model carbon dynamics in 42 fields, each having 6 dynamic carbon pools. Statistical analysis of this system is made computationally feasible through the use of a Particle Marginal Metropolis-Hastings (PMMH) algorithm implemented through the LibBi software for Bayesian State Space modelling on a machine equipped with GPUs. This talk will be of interest to those interested in environmental modelling, applied statistics and anyone interested in seeing a real-world application of the LibBi modelling language.