Analysing the sensitivity of pollen based land-cover maps to different auxiliary variables

Abstract. Realistic depictions of past land cover are needed to investigate prehistoric environmental changes and anthropogenic impacts. However, observation based reconstructions of past land cover are rare. Recently Pirzamanbein et al. (2015, arXiv:1511.06417) developed a statistical interpolation method that produces spatially complete reconstructions of past land cover from pollen assemblage. These reconstructions incorporate a number of auxiliary datasets raising questions regarding both the method's sensitivity to the choice of auxiliary data and the unaffected transmission of observational data.

Here the sensitivity of the method is examined by performing spatial reconstructions for northern Europe during three time periods (1900 CE, 1725 CE and 4000 BCE), based on irregularly distributed pollen based land cover, available for ca. 25 % of the area, and different auxiliary datasets. The spatially explicit auxiliary datasets considered include the most commonly utilized sources of the past land-cover data – estimates produced by a dynamic vegetation (DVM) and anthropogenic land-cover change (ALCC) models – and modern elevation. Five different auxiliary datasets were considered, including different climate data driving the DVM and different ALCC models. The resulting reconstructions were evaluated using deviance information criteria and cross validation for all the time periods. For the recent time period, 1900 CE, the different land-cover reconstructions were also compared against a present day forest map.

The tests confirm that the developed statistical model provides a robust spatial interpolation tool with low sensitivity to differences in auxiliary data and high capacity to un-distortedly transmit the information provided by sparse pollen based observations. Further, usage of auxiliary data with high spatial detail improves the model performance for the areas with complex topography or where observational data is missing.