Machine learning-based LULC mapping with AlphaEarth foundation embeddings and Sentinel-2 multitemporal features: a comparative study focusing on hazelnut ( Corylus avellana L.) orchards

Accurate land use and land cover (LULC) mapping in heterogeneous agricultural landscapes remains challenging, particularly when discriminating spectrally similar perennial crops from surrounding vegetation classes. This study presents a systematic comparative evaluation of AlphaEarth Foundation (AEF) embeddings against conventional multitemporal Sentinel-2 spectral feature sets for 11-class LULC mapping in Sakarya Province, northwestern Türkiye—a region encompassing one of the world’s most extensive hazelnut (Corylus avellana L.) production areas. Twenty classification experiments were conducted by pairing four input feature configurations—AEF embeddings, multitemporal Sentinel-2 spectral bands, Sentinel-2 spectral indices, and their combination—with five machine learning algorithms: Random Forest, XGBoost, LightGBM, LinearSVC, and Decision Tree. AEF embeddings consistently outperformed all Sentinel-2-based configurations across every algorithm and evaluation metric. The best result, achieved by AEF + LightGBM, yielded an overall accuracy (OA) of 0.9579 and a weighted F1 score of 0.9566, surpassing the best conventional Sentinel-2 configuration (S2-All + LightGBM, OA = 0.9329) by 2.5 percentage points. Hazelnut orchard classification achieved an F1 score of 0.8926 with AEF + Random Forest, demonstrating the feasibility of fine-grained perennial crop delineation at 10 m resolution.