A Comparison of Some Machine Learning Classifiers Applied in Three Silvopastoral Systems of Greece.

Silvopastoral systems (SPS) provide critical ecosystem services across the Mediterranean basin; however, reliable classification models integrating machine learning and satellite imagery remain largely unexplored. Addressing this gap is critical for supporting agri-environmental policy within the EU Biodiversity Strategy 2030 and Common Agricultural Policy frameworks.

Extending a prior silvopastoral/non-silvopastoral classification phase, this study classifies silvopastoral extents into four vegetation categories: dominant tree species, other tree species, shrubs, and herbaceous vegetation. Sentinel-2 multispectral imagery was applied across three pilot areas in Greece, each representing a distinct SPS type -valonia oak (Quercus ithaburensis), oriental plane (Platanus orientalis), and walnut (Juglans regia)- with contrasting topographic and landscape characteristics. Three ML algorithms were comparatively evaluated: Random Forest (RF) and Support Vector Machine (SVM), recognized for their strong performance in remote sensing classification, and Naïve Bayes (NB), a computationally efficient classifier rarely applied in SPS contexts, yet showing effectiveness in landscapes with relatively distinct spectral class distributions. A 22-variable feature set was employed, comprising Sentinel-2 spectral bands, vegetation indices, and GLCM texture features. Model performance was assessed through AUC, overall accuracy, and Cohen's Kappa.

NB, outperformed RF and SVM in two study sites, suggesting its probabilistic approach is particularly suited to homogeneous and riparian SPS landscapes. SVM performed best in the semi-mountainous walnut system, where complex terrain favors non-linear decision boundaries. Among vegetation categories, dominant tree species achieved the highest separability, especially the evergreen valonia oak than deciduous walnut and plane tree. In cconstrast, shrubs exhibited the lowest accuracy due to persistent spectral confusion with adjacent categories. Our findings favor the application of NB as a classifier for homogeneous Mediterranean SPS landscapes, while RF /SVM performed better in structurally complex terrain. Further research is currently underway integrating the present framework with advanced classification methodologies to enhance vegetation mapping accuracy.