Pollen-Based Spatial Reconstruction of Past Land Cover: Estimating Latent GMRFs with Dirichlet Observations

• Activity Number: 538
• Type: Contributed
• Date/Time: Wednesday, August 12, 2015 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistics and the Environment
• Author(s): Behnaz Pirzamanbein* and Johan Lindström and Marie-José Gaillard-Lemdahl and Anneli Poska
• Companies: Lund University and Lund University and Linnaeus University and Lund University
• Keywords: Pollen records ; Spatial statistics ; Gaussian Markov Random Field ; Dirichlet observation ; Past land cover ; Adaptive Metropolis adjusted Langevin

Abstract

When assessing past earth surface-climate feedbacks and past human impact on climate, accurate estimates of past land cover are required. Fossil pollen records extracted from lakes and bogs provide good insight into the local land cover around each site. For use in climate modelling these sparse observations have to be interpolated; creating continuous maps of past land cover at regional and sub-continental scales. We construct a hierarchical model based on latent Gaussian Markov Random Fields with Dirichlet observations. The model is used to reconstruct past land cover across Europe for five time periods - centred around 1900, 1725, 1425 CE and 1000, 4000 BCE, by combining estimates of past human land use and output from a dynamic vegetation model with pollen based local land-cover estimates. To estimate the model a block updated MCMC, which includes an adaptive Metropolis adjusted Langevin step, is used. Model results are evaluated by comparing results for the 1950-time period to a European forest map from 2006. The lack of good historic land-cover data makes the evaluation challenging.