B1692
Title: Bayesian hierarchical models for uncertainty quantification in high-dimensional landscape problems
Authors: Oluwole Oyebamiji - Lancaster University (United Kingdom) [presenting]
Abstract: A Bayesian hierarchical model is tested for achieving dimension reduction in modelling of large dimensional spatial data. This being a step towards emulating a complex integrated model of land-use change. The method uses a combination of Bayesian principal component and Gaussian process based on nearest neighbour approximation. The approach is to first retrieve the low-dimensional underlying patterns from high-dimensional outputs using a Bayesian principal component analysis where the effective dimensionality of the latent space is determined automatically as part of the Bayesian inference procedure. This is followed by the emulation of the resulting low-dimensional data using a composite nearest-neighbour GP based on an assumption of conditional independence. This reduces model complexity and captures different aspects of the socio-economic scenarios. The approach is computationally efficient and improves the accuracy of estimating the parameters as well as incorporating various sources of uncertainty. The method is being applied to a dataset from the IMPRESSIONS Integrated Assessment Platform (IAP2) model, an extension of the CLIMSAVE IAP, which has been widely applied in climate change impact, adaptation and vulnerability assessment for robust policy analysis.
