Reconstruction of Past European Land Cover Based on Fossil Pollen Data: Gaussian Markov Random Field Models for Compositional Data

Research output: ThesisDoctoral Thesis (compilation)

625 Downloads (Pure)


The aim of this thesis is to develop statistical models to reconstruct past land cover composition and human land use based on fossil pollen records over Europe for different time periods over the past 6000 years. Accurate maps of past land cover and human land use are needed when studying the interaction between climate and land surface, and the effects of human land use on past climate. Existing land cover maps are mainly simulations from dynamic vegetation models and anthropogenic land cover change scenarios. Pollen records is an alternative to existing land cover estimates that might give better insight into past land cover. The pollen counts are extracted from lake and bog sediments and used to estimate the three land cover compositions; coniferous forest, broadleaved forest, and unforested land for grid cells surrounding the lakes and bogs.

In this thesis, first, a statistical model is developed to interpolate transformed
pollen based land cover compositions (PbLCC) with spatial dependency modelled
using a Gaussian Markov random Field (GMRF). The mean structure is modelled using a regression on different sets of covariates including elevation and model based vegetation estimates. The model is fitted using Integrated Nested
Laplace Approximation. The results indicated the existence of spatial dependence structure in the PbLCC and the possibility of reconstructing past land cover from PbLCC. If the compositional data is over-dispersed, the transformed Gaussian model might underestimate the uncertainties. To capture the variation in the composition correctly, a Bayesian hierarchical model (BHM) for Dirichlet observations of a GMRF is developed. The model is estimated using MCMC with sparse precision matrix of the GMRF being used for computational efficiency. Comparison between the Dirichlet and Gaussian models showed the advantages of the Dirichlet in describing the PbLCC. The large discrepancies in the model based estimates used as covariates could affect the Dirichlet models ability to reconstruct past land cover. To assess this concern a sensitivity study was performed, showing that the results are robust to the choice of covariates. Finally, the BHM is extended to reconstruct past human land use by combing the PbLCC with anthropogenic land cover change estimates. This extension aims at decomposing the PbLCC into past natural land cover and human land use.
Original languageEnglish
Awarding Institution
  • Faculty of Science
  • Lindström, Johan, Supervisor
Award date2016 Dec 19
Place of PublicationLund, Sweden
ISBN (Print)978-91-7753-076-3
ISBN (electronic) 978-91-7753-077-0
Publication statusPublished - 2016 Nov

Bibliographical note

Defence details
Date: 2016-12-19
Time: 09:15
Place: Annexet, lecture hall MA:04, Sölvegatan 20, Lund
External reviewer(s)
Name: Illian, Janine
Title: Senior Lecturer, Dr.
Affiliation: University of St Andrews, United Kingdom

Subject classification (UKÄ)

  • Environmental Sciences
  • Probability Theory and Statistics
  • Climate Research

Free keywords

  • Spatial Statistics
  • Adaptive Markov Chain Monte Carlo
  • Dirichlet Observation
  • Confidence Region
  • Palaeoecology
  • Past Human Land Use
  • Stochastic Partial Differential Equation


Dive into the research topics of 'Reconstruction of Past European Land Cover Based on Fossil Pollen Data: Gaussian Markov Random Field Models for Compositional Data'. Together they form a unique fingerprint.

Cite this