We used the presence/absence data for 108 European tree and tall shrub species (Jalas & Suominen 1972–1996
), split into five classes (= groups of chorotypes from the dataset): EURopean (43, temperate/pan); MEDiterranean (30); ALPine (19); BOReo-ALPine (9); and BOReal (7). We used two sets of bioclimatic data. The first (‘TRAD’) included only traditionally used variables: annual aridity (equilibrium evapotranspiration minus precipitation); annual precipitation; winter precipitation; mean annual temperature; mean winter temperature; and growing degree days until April and August. The bioclimatic variables were derived from a high-resolution (10′
km) climate grid of Europe, including the recent past (1901–2000) and future scenarios (2001–2100; Mitchell et al. 2004
). We averaged the bioclimatic variables for two time frames (1971–2000 and 2050–2080) and two emission scenarios (A2 and B1; Nakicenovic & Swart 2000
) based on the HadCM3 global circulation model (Mitchell et al. 2004
In the second dataset (‘CO2
’ for short, though it includes vegetation effects), aridity and precipitation were replaced by the average fraction of plant-available soil water-holding capacity (fAWC) during the growing season (daily temperature more than 5°C) in two soil layers (0–0.5 and 0.5–1.5
m). This was simulated with the LPJ-GUESS vegetation dynamics model (Smith et al. 2001
), using parameters for European potential natural vegetation and including both CO2
and vegetation effects (appendix I in the electronic supplementary material).
Using the BIOMOD framework (Thuiller 2003
), implemented in the R software (R_Development_Core_Team 2004
), we fitted six models (classification tree analysis, generalized additive model, generalized boosted model, generalized linear model, mixture discriminant analysis and randomForest) relating tree species distributions to climate, using a random subset (70%) of the data. We used the remainder to evaluate the models using Cohen's κ
and the area under the curve (AUC) of a receiver-operating characteristic plot.
We projected the species current and future potential habitats with/without CO2 and with the A2/B1 scenarios. We transformed the current occurrence probabilities into presences/absences, using the thresholds that maximize both the percentage of presences/absences correctly predicted and Cohen's κ statistic; we used the same thresholds to convert future occurrence probabilities. There are 48 model combinations per species: six models×two evaluation methods×two climate datasets×two scenarios. We calculated SHC as the proportion of new habitat respective to current habitat.
To constrain model uncertainty, we calculated a consensus projection (Araújo & New 2007
) by stacking, that is, average of the 12 model combinations per climate datasets (TRAD and CO2
) and scenarios (A2 and B1), weighted according to each model's AUC score.