Benchmarking Ten Machine-Learning Algorithms for Mediterranean Forest Mapping with Sentinel-2

Accurate forest-cover mapping in fragmented Mediterranean landscapes remains challenging because forests, croplands, grasslands, and sparsely vegetated surfaces often exhibit overlapping spectral responses. This study benchmarks ten supervised machine-learning classifiers for regional forest-cover mapping in the Beni Mellal-Khenifra region of Morocco, evaluates the contribution of spectral, textural, and topographic predictors, and assesses the consistency of the resulting forest map with the Hansen Global Forest Change product. A reference dataset of 2800 samples representing forest, crops/grass, bare ground, urban areas, and water was combined with an annual 2025 Sentinel-2 median composite, spectral indices, NDVI-derived gray-level co-occurrence matrix textures, and SRTM-derived elevation and slope. The samples were divided once into a stratified 70% training set and a 30% test set common to all classifiers. Redundancy, correlation (|r| > 0.80), and variance inflation factor (VIF < 10) screening, fitted on the training set only, reduced 56 candidate predictors to 16, and hyperparameters were tuned by five-fold grid-search cross-validation within the training set. On the common test set (n = 840), CatBoost, XGBoost, and the multilayer perceptron each achieved an overall accuracy of 95.95% (Cohen’s κ = 0.947), followed by Gradient Tree Boosting (95.83%); McNemar tests with Holm correction detected no significant differences among these four models. Spatial block cross-validation reduced overall accuracy by 0.3–2.7 percentage points, and CatBoost retained the highest spatially blocked accuracy (95.51% with 10 km blocks). SHapley Additive exPlanations (SHAP) identified slope, SATVI, NDWI, and the blue band as the dominant contributors to CatBoost forest predictions. The leading maps estimated 859,057–870,663 ha of forest. CatBoost captured 90.9% of the Hansen forest extent (≥30% canopy cover) but mapped a much larger forest area; blind interpretation of stratified samples showed that 90.5% of the CatBoost-only forest was woody vegetation, mainly open woodland and matorral. These findings show that boosting ensembles combined with a compact, non-redundant predictor set provide accurate and spatially robust forest-cover maps and that differences from global products largely reflect forest definitions.