National Forest Inventory (NFI) data provided at the field plot-level represent a valuable data source for Species Distribution Modelling (SDM) as demonstrated by many studies. However, a stand-level forest inventory, conducted in some countries, including Bulgaria, is less appropriate because matching these data to spatial predictor datasets is not straightforward. This study aimed to propose and evaluate a methodology for tree species distribution modelling at a national scale driven by stand-level forest inventory data, focusing on Pinus sylvestris L. as a target species. The forest polygons were filtered using a set of attribute and geometric criteria and transformed to points to generate a presence/absence dataset. Based on the literature, four predictor variables were selected, including growing degree-days above 5° C (DD.5), August precipitation (PPT08), potential yearly global radiation (YRA), and soil volumetric coarse fragment content (CF60), all available as 250 m resolution rasters. Generalised Linear Models with different combinations of linear, quadratic, and interaction terms were tested. The effect of spatial errors in the presence/absence dataset was investigated by randomly shifting the points up to 1 km from their original position. Evaluation statistics were calculated by 5-fold spatial block cross-validation repeated 10 times. The basic model, which included only linear terms, performed the worst. The other models were more or less equal performers. The model, which included a quadratic term for DD.5, was therefore selected as the optimal model as it had the most parsimonious structure. The Area under the Receiver Operating Characteristic curve and the True Skill Statistic for that model were 0.80±0.04 and 0.49±0.08, respectively. The model was relatively robust to spatial errors in the occurrence data as simulated by randomly shifting the points. The modelled distribution was much wider than the actual natural distribution of the species, and all mountain areas were identified as potential distribution areas based on these four predictors. As drawbacks of the approach should be noted the assumption of homogeneity of forest stands, the unknown reliability of the inventory dataset, and the low prevalence. Nevertheless, the study demonstrated successfully the utility of the stand-level NFI dataset for SDM.
