Distribution modelling of vegetation types based on area‐frame survey data

AIM: Many countries lack informative, high‐resolution, wall‐to‐wall vegetation or land cover maps. Such maps are useful for land use and nature management, and for input to regional climate and hydrological models. Land cover maps based on remote sensing data typically lack the required ecological information, whereas traditional field‐based mapping is too expensive to be carried out over large areas. In this study, we therefore explore the extent to which distribution modelling (DM) methods are useful for predicting the current distribution of vegetation types (VT) on a national scale. LOCATION: Mainland Norway, covering ca. 324,000 km². METHODS: We used presence/absence data for 31 different VTs, mapped wall‐to‐wall in an area frame survey with 1081 rectangular plots of 0.9 km². Distribution models for each VT were obtained by logistic generalised linear modelling, using stepwise forward selection with an F‐ratio test. A total of 116 explanatory variables, recorded in 100 m × 100 m grid cells, were used. The 31 models were evaluated by applying the AUC criterion to an independent evaluation dataset. RESULTS: Twenty‐one of the 31 models had AUC values higher than 0.8. The highest AUC value (0.989) was obtained for Poor/rich broadleaf deciduous forest, whereas the lowest AUC (0.671) was obtained for Lichen and heather spruce forest. Overall, we found that rare VTs are predicted better than common ones, and coastal VTs are predicted better than inland ones. CONCLUSIONS: Our study establishes DM as a viable tool for spatial prediction of aggregated species‐based entities such as VTs on a regional scale and at a fine (100 m) spatial resolution, provided relevant predictor variables are available. We discuss the potential uses of distribution models in utilizing large‐scale international vegetation surveys. We also argue that predictions from such models may improve parameterisation of vegetation distribution in earth system models.