What the study found
Geospatial foundation-model embeddings for satellite data improved tree species classification in temperate mountain forests compared with conventional satellite composites. The study found that the embeddings also reached high accuracy with relatively little training data and that a nonlinear classifier worked best.
Why the authors say this matters
The authors conclude that geospatial foundation models may shift tree species mapping away from manual feature engineering and toward the availability, quality, and timing of ecological reference data. They also say this opens new opportunities for scalable biodiversity monitoring and the analysis of ecological change.
What the researchers tested
The researchers evaluated two geospatial foundation-model embeddings, AlphaEarth and Tessera, for tree species classification in the Trentino region of northern Italy. They used parcel-level forest inventories as reference data for 18 species and species groups, and compared the embeddings with conventional Sentinel-1+2 satellite composites across controlled experiments on accuracy, label efficiency, classifier complexity, robustness to label impurity, and temporal transferability.
What worked and what didn't
The foundation-model embeddings outperformed the composite-based baseline, with weighted F1 of 0.83 versus 0.80 and macro F1 of 0.55 versus 0.50. They reached near-asymptotic accuracy with as little as 5% of the available training parcels, and a compact neural network performed better than Random Forest and as well as deeper neural networks; a linear classifier on the embeddings underperformed a neural network on conventional composites. Ancillary environmental covariates did not add classification benefit, moderate label impurity did not substantially harm accuracy, soft labels from parcel-level species proportions improved peak performance, and temporal transfer across years reduced performance, especially for rare species.
What to keep in mind
The reported advantages were strongest under the study's tested setting in the Trentino region and for the two embeddings examined. Performance declined when models were transferred across years, and the abstract notes disproportionate losses for rare species. The available summary does not describe other limitations beyond these points.
Key points
- AlphaEarth and Tessera outperformed conventional Sentinel-1+2 composite baselines for tree species classification.
- The best reported scores were weighted F1 0.83 versus 0.80 and macro F1 0.55 versus 0.50.
- Near-asymptotic accuracy was reached with as little as 5% of the available training parcels.
- A compact neural network worked better than Random Forest on embedding-based models.
- Temporal transfer across years reduced performance, with larger losses for rare species.
Disclosure
- Research title:
- Geospatial foundation models improved tree species mapping accuracy
- Authors:
- James GC Ball, Jana Annika Wicklein, Zhengpeng Feng, Jovana Knezevic, Sadiq Jaffer, Anil Madhavapeddy, Clement Atzberger, Michele Dalponte, David Coomes
- Institutions:
- Conservation Leadership Programme, Fondazione Edmund Mach, University of Trento, University of Cambridge, Life Cycle Engineering (United States)
- Publication date:
- 2026-02-24
- OpenAlex record:
- View
Get the weekly research newsletter
Stay current with peer-reviewed research without reading academic papers — one filtered digest, every Friday.