Geospatial foundation models enable data-efficient tree species mapping in temperate mountain forests

Accurate mapping of tree species from satellite data remains challenging in heterogeneous mountain forests due to environmental gradients, mixed stands, and limited training labels. Geospatial foundation models (GFMs) learn rich representations from large multi-sensor archives, but their utility for species-level mapping remains unclear. Here, we evaluate two GFM embeddings, AlphaEarth and Tessera, for tree species classification in a demanding mountain landscape (Trentino, Italy; 18 species and groups), versus conventional Sentinel-1+2 composites, across experiments spanning classification accuracy, label efficiency, classifier complexity, environmental covariates, label impurity, and temporal transferability. GFM embeddings consistently outperform conventional baselines (weighted F1 = 0.83 vs. 0.80; macro F1 = 0.55 vs. 0.50), approaching saturation with only 5% of training parcels while organising species into ecologically meaningful taxonomic and functional groupings. Realising this advantage requires a nonlinear classifier: a compact neural network yields the largest single gain, whereas a linear classifier on GFM embeddings underperforms a neural network on conventional composites. Classification is robust to moderate label impurity, and training with parcel-level species proportions as soft labels improves minority-species discrimination (macro F1 = 0.586 –0.589) without purity filtering; adding terrain-derived covariates provides no further benefit. Temporal transfer across years degrades performance: weighted F1 falls from 0.84 to 0.77 ( 9%) for Tessera and from 0.85 to 0.73 ( 14%) for AlphaEarth, with disproportionate losses for rare species. These results show that GFMs shift the primary bottleneck in species mapping from feature engineering toward the availability, quality, and temporal alignment of reference data