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

Code for the paper (preprint).

Authors: James GC Ball, Jana Annika Wicklein, Zhengpeng Feng, Jovana Knezevic, Sadiq Jaffer, Anil Madhavapeddy, Clement Atzberger, Michele Dalponte, David Coomes

Data: Zenodo (10.5281/zenodo.18877941)

Overview

This repository contains all code for classifying 18 tree species and species groups at 10 m resolution across the Autonomous Province of Trento (Italian Alps) using geospatial foundation-model embeddings (AlphaEarth, Tessera) compared against conventional Sentinel-1/2 composites.

Repository structure

trentino-trees/
├── src/
│   ├── common/             Shared utilities (data loading, fold generation, metrics)
│   ├── model/              MLP classifiers, fraction/purity curves, temporal transfer
│   ├── preprocessing/      NPZ export, parcel merging, fractions lookup
│   ├── evaluation/         Cross-year transfer evaluation
│   ├── plotting/           UMAP, fraction curves, purity curves, temporal transfer plots
│   └── data_acquisition/   Google Earth Engine scripts, raster download utilities
├── scripts/                Experiment launchers and figure generation scripts
├── environment.yml         Conda environment specification
└── LICENSE

Setup

git clone https://github.com/PatBall1/trentino-trees
cd trentino-trees
conda env create -f environment.yml
conda activate trent
export PYTHONPATH="$(pwd):$PYTHONPATH"

Data

Download the data from Zenodo.

Place the NPZ files into data/labels_merged/, renaming to remove the data__ prefix:

data__pre_distilled_labels_tessera2018.npz  →  data/labels_merged/pre_distilled_labels_tessera2018.npz
data__parcel_fractions_lookup.npz           →  data/labels_merged/parcel_fractions_lookup.npz
data__class_registry_jana.csv               →  data/labels_jana/class_registry_jana.csv

Reproducing the results

Table 2 and Figure 5 — Label efficiency curves

bash scripts/rerun_fraction_curves_v2.sh
python scripts/plot_fraction_curve_comparison_v3.py

Table 3 and Figure 6 — Nested cross-validation and confusion matrices

python -m src.model.mlp_nested_cv_sklearn \
    --npz data/labels_merged/pre_distilled_labels_tessera2018.npz \
    --out results/mlp_nested_cv_sklearn/tessera_merged_balanced

Repeat for each representation NPZ.

Figure 7 — Purity filtering and soft-label experiments

bash scripts/rerun_hard_label_purity_curves_v5.sh
python scripts/plot_purity_curve_comparison.py

Table 4 — Temporal transfer

bash scripts/rerun_temporal_transfer_v3.sh

Figure 8 — Wall-to-wall species map

python scripts/predict_landscape_combined.py

Regenerating input rasters from scratch

Not required if using the Zenodo NPZ files, but for full end-to-end reproducibility:

• Sentinel composites: Run src/data_acquisition/s1s2annual.js and src/data_acquisition/s1s2seasonal.js in the Google Earth Engine Code Editor
• AlphaEarth embeddings: Run src/data_acquisition/alphaearth.js in the GEE Code Editor
• Tessera embeddings: Download via src/data_acquisition/geotessera_pull.sh
• NPZ generation from rasters: python -m src.preprocessing.export_label_raster_npz --feature-raster <raster.tif> --parcel-path <parcels.gpkg> --out-npz <output.npz>

Citation

If you use this code, please cite the accompanying preprint and dataset:

@article{ball2026geospatial,
  title={Geospatial foundation models enable data-efficient tree species mapping
         in temperate mountain forests},
  author={Ball, James GC and Wicklein, Jana Annika and Feng, Zhengpeng and
          Knezevic, Jovana and Jaffer, Sadiq and Madhavapeddy, Anil and
          Atzberger, Clement and Dalponte, Michele and Coomes, David},
  journal={bioRxiv},
  year={2026},
  doi={10.64898/2026.02.23.707022}
}

License

MIT