Foundation models for tree species mapping, now in Science of Remote Sensing
Our paper on using geospatial foundation models to map tree species in temperate mountain forests is out in Science of Remote Sensing:
There is also a phys.org write-up that captures the practical pitch well.
What we asked
Can pre-trained satellite embeddings — specifically TESSERA and Google’s AlphaEarth — map tree species in messy, topographically complex forests better than conventional Sentinel composites, and can they do it with far fewer training labels?
We worked in Trentino (Italian Alps) with 18 species and species groups, parcel-level inventory labels that are often impure, and the usual mountain-forest complications: steep terrain, mixed stands, and scarce clean training data.
What we found
- Foundation-model embeddings outperform conventional multispectral baselines (weighted F1 about 0.83 vs 0.80).
- Most of the accuracy is already there with roughly 5% of the training parcels — label efficiency is the headline result.
- A compact neural net on the embeddings matters; a linear classifier on the same embeddings underperforms a neural net on ordinary composites.
- Temporal transfer across years still hurts, especially for rare species.
“Tessera is a democratizing force that brings satellite data to the masses.”
That line from the press piece is what I care about operationally: groups without huge compute can train on forest inventories against embeddings that already exist, rather than rebuilding a full EO pipeline from scratch.
Why it matters for ecology
Species mapping has long been bottlenecked by feature engineering and label scarcity. Foundation models shift the bottleneck toward the quality and design of reference data — which is where ecologists actually have leverage. That is the strand I want to push next: crown- and plot-level traits as training signal for landscape-scale transfer.
Huge thanks to Jana Wicklein, Michele Dalponte, Zhengpeng Feng, Jovana Knezevic, Sadiq Jaffer, Anil Madhavapeddy, Clement Atzberger, David Coomes, and the wider TESSERA team.