Probabilistic species models for Montenegro

Methodology Habitat suitability modelling was performed using the Random Forest (RF) algorithm in the R environment. The following input data were used: a multi-layer raster of environmental predictors (including CLC – Corine Land Cover), species occurrence data (presence points) from an Excel database, the boundary of Montenegro to spatially constrain the analysis. Data preparation All spatial layers were projected to a common coordinate reference system and clipped to the study area (Montenegro). Urban CLC classes (111, 112, 121, 122) were excluded by masking the raster. Species records were filtered to include only species with ≥25 presence points. Coordinates were converted to spatial format and intersected with raster data to retain valid observations. Habitat class definition For each species, core CLC classes were defined as those containing at least 25 presence records. These classes represent primary habitats and were used to filter presence data. Pseudo-absence generation Pseudo-absence points were randomly sampled from areas outside the core classes. The number of absence points was equal to the number of presence points to ensure a balanced dataset. Modelling For each species: data were split into training (70%) and testing (30%) subsets, a Random Forest model (500 trees) was trained, model performance was evaluated using AUC (Area Under the Curve). Spatial prediction The trained models were applied to the full raster stack to produce continuous habitat suitability maps. Predictions outside core classes were penalized (multiplied by 0.5) to reduce suitability in less relevant habitats. Outputs For each species, the following outputs were generated: habitat suitability maps (GeoTIFF format), a summary table of AUC values (Excel format).