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

Key findings from this study

• The study found that foundation model embeddings outperformed conventional satellite composites for tree species classification, achieving weighted F1 of 0.83 versus 0.80 and macro F1 of 0.55 versus 0.50.
• The authors report that near-asymptotic classification accuracy required only 5% of available training parcels when using foundation model embeddings.
• The researchers demonstrate that realizing foundation model advantages requires nonlinear classifiers, with compact neural networks matching deeper architectures while linear models underperform.
• The study found that training with soft labels based on species proportions achieved higher peak performance than hard labels without requiring purity filtering.
• The authors report that temporal transfer across years degraded weighted F1 by 9% for Tessera and 15% for AlphaEarth, with disproportionate losses for rare species.

Overview

This study evaluates two geospatial foundation models, AlphaEarth and Tessera, for tree species classification in heterogeneous temperate mountain forests in northern Italy. Using parcel-level forest inventories covering 18 species and species groups, the authors compare foundation model embeddings against conventional Sentinel-1+2 satellite composites. The research examines classification accuracy, label efficiency, classifier architecture requirements, robustness to label impurity, and temporal transferability. The study addresses challenges in satellite-based species mapping including environmental gradients, mixed stands, limited high-purity training labels, and illumination-angle effects. Foundation models provide learned representations from large multi-sensor archives, offering an alternative to manual feature engineering for downstream classification tasks.

Methods and approach

The authors tested AlphaEarth and Tessera foundation model embeddings for classifying 18 tree species and species groups in the Trentino region using parcel-level forest inventory reference data. They compared foundation model performance against Sentinel-1+2 satellite composite baselines through controlled experiments. Classification architectures ranged from linear models and Random Forest to compact and deeper neural networks. Experiments systematically varied training data quantity, label purity levels, and temporal alignment between training and test datasets. The authors evaluated both hard labels and soft labels based on parcel-level species proportions. Performance metrics included weighted and macro F1 scores and Proportion L1 error for mixed-composition parcels.

Results

Foundation model embeddings consistently outperformed conventional satellite composites, achieving weighted F1 of 0.83 versus 0.80 and macro F1 of 0.55 versus 0.50. Near-asymptotic accuracy required only 5% of available training parcels, demonstrating substantial label efficiency. However, realizing this advantage depended critically on classifier architecture: a compact neural network performed as well as deeper networks and outperformed Random Forest, while linear classifiers on embeddings underperformed neural networks trained on conventional composites. Ancillary environmental covariates provided no additional benefit when added to embedding-based models. Classification accuracy proved robust to moderate label impurity, allowing retention of mixed parcels without substantial performance penalties.

Training with parcel-level species proportions as soft labels achieved higher peak performance than hard labels, with macro F1 reaching 0.586 for Tessera and 0.589 for AlphaEarth, along with lower Proportion L1 error. This approach maximized value from the full range of input data without requiring purity filtering. Temporal transfer across years revealed performance degradation, with weighted F1 declining by 9% for Tessera and 15% for AlphaEarth. Rare species experienced disproportionate accuracy losses during temporal transfer. The embeddings preserved ecologically meaningful structure aligned with functional and taxonomic groupings, supporting interpretable species relationships.

Implications

The results demonstrate that geospatial foundation models shift the primary constraint in species mapping from feature engineering toward the availability, quality, and temporal alignment of ecological reference data. This transition has practical consequences for biodiversity monitoring workflows: data collection and labeling strategies become more critical than satellite data processing pipelines. The label efficiency of foundation models enables species classification in data-scarce regions, though the requirement for nonlinear classifiers adds computational overhead and modeling complexity. The finding that ancillary environmental covariates add no value suggests embeddings already capture relevant environmental gradients, simplifying model architecture decisions.

The temporal transfer degradation, particularly for rare species, indicates that foundation models do not fully resolve phenological and interannual variability challenges. Operational mapping systems must address temporal alignment between training labels and target mapping periods. The effectiveness of soft labels based on species proportions offers a pathway to incorporate mixed-composition parcels without discarding valuable training data, expanding the usable inventory base. These findings support scalable biodiversity monitoring applications while highlighting that label quality and temporal consistency remain critical factors determining classification performance. The preservation of ecologically meaningful structure in embeddings suggests potential for transfer learning across regions with similar forest types.

Scope and limitations

This summary is based on the study abstract and available metadata. It does not include a full analysis of the complete paper, supplementary materials, or underlying datasets unless explicitly stated. Findings should be interpreted in the context of the original publication.