Mapping Threatened Native Grasslands Using AlphaEarth Satellite Embeddings:A Robust Alternative to Optical-Based Workflows in Fragmented Agro-Natural Landscapes

Accurate land-cover classification in heterogeneous and fragmented agro-natural landscapes represents a key challenge for environmental monitoring and territorial planning, particularly in complex ecosystems such as native grasslands. In this study, we evaluated the comparative performance of approaches based on multitemporal optical predictors derived from Sentinel-2 and satellite embeddings from AlphaEarth for supervised land-cover classification in the Tandilia mountain system, Argentina. We implemented reproducible and transparent workflows entirely within Google Earth Engine, using exclusively open-access satellite data. We defined eight predictor configurations combining optical, phenological, and topographic information with advanced satellite representations, and we trained Random Forest models using a spatial block–based design with stratified replicas. Model performance was evaluated using an external, spatially independent reference dataset, employing global and class-wise accuracy metrics, as well as analyses of model stability and map-level spatial consistency. Results show that, while Sentinel-2–based approaches achieved high accuracy levels (OA ≈ 89–93%), AlphaEarth-based models exhibited consistently superior performance (OA ≈ 94–96%), lower variability across runs, and a marked reduction in confusion for spectrally complex classes, particularly native grasslands and water bodies. These improvements were reflected in both class-wise accuracy metrics and enhanced spatial coherence, with reduced classification noise at the map level. Overall, our findings highlight the potential of satellite embeddings to improve land-cover classification in complex and fragmented landscapes, while also emphasising the trade-offs between biophysical interpretability and operational robustness in large-scale environmental monitoring and conservation-oriented mapping applications.