tessera-eval

Evaluate land-cover / habitat classifiers on Tessera satellite embeddings.

Tessera is a geospatial foundation model that produces a 128-dimensional embedding for every ~10 m pixel of the Earth's surface, per year. tessera-eval is a small, framework-independent Python library for the question that immediately follows: how well can you map a category of interest (a habitat, crop, or land-cover class) from those embeddings, given some labelled polygons?

It handles the unglamorous-but-fiddly parts end to end:

• Loading + dequantizing embeddings from the formats Tessera tooling emits (GeoTessera int8 × per-pixel-scale tiles, and TEE per-dim uint8 vector directories).
• Rasterizing a labelled shapefile/GeoJSON onto the embedding pixel grid with stable class IDs.
• Training + scoring a panel of classifiers/regressors (k-NN, random forest, MLP, spatial MLP, optional XGBoost, optional U-Net) with learning curves, k-fold cross-validation, and spatial hold-out splits.
• An optional local compute server (tee-compute) so you can run the ML on your own machine while pulling tiles/UI from a hosted service.

The library core (data, rasterize, classify, evaluate) is pure NumPy / scikit-learn / rasterio and has no web-framework or hosting dependency. The compute server reads zarr embeddings through geotessera's own GeoTesseraZarr interface, falling back to NPY tiles when the store lacks coverage.

Install

pip install tessera-eval                 # core library
pip install "tessera-eval[geotessera]"   # + tile access (load_embeddings_for_shapefile)
pip install "tessera-eval[server]"       # + the tee-compute local server
pip install "tessera-eval[all]"          # geotessera + xgboost + matplotlib

Optional extras: geotessera (fetch tiles), xgboost (gradient-boosted models), torch (the U-Net), plot (matplotlib), server (Flask compute server), dev (pytest/ruff/mypy). Python ≥ 3.10.

Quickstart

Cross-validate a classifier on labelled polygons, pulling embeddings tile-by-tile. This runs as-is against the bundled example (examples/austria_crops.geojson — 349 field parcels near Vienna labelled by crop type):

import geopandas as gpd
from geotessera import GeoTessera
from tessera_eval import load_embeddings_for_shapefile, run_kfold_cv

# 1. Labelled polygons (any CRS — reprojected internally) with a class column.
#    (a repo checkout has this at examples/austria_crops.geojson)
gdf = gpd.read_file(
    "https://raw.githubusercontent.com/ucam-eo/tessera-eval/main/examples/austria_crops.geojson"
)

# 2. Pull a 128-d embedding for every pixel under the polygons (memory-bounded:
#    one GeoTessera tile at a time, keeping only labelled pixels).
gt = GeoTessera()
vectors, labels, class_names, stats = load_embeddings_for_shapefile(
    gdf, field="crop", year=2024, gt_instance=gt
)
print(f"{stats['total_pixels']:,} labelled pixels over {stats['n_classes']} classes")

# 3. 5-fold cross-validation of a random forest and a nearest-neighbour baseline.
for event in run_kfold_cv(vectors, labels, ["rf", "nn"], k=5):
    if event["type"] == "aggregate":
        for name, m in event["models"].items():
            print(f"{name:>4}: macro-F1 {m['mean_f1']:.3f} ± {m['std_f1']:.3f}")

For this window that prints roughly rf: macro-F1 0.80 ± 0.00, nn: macro-F1 0.76 ± 0.00 (10-way crop classification from embeddings alone).

run_kfold_cv also does regression — pass task="regression" and the _reg model names (rf_reg, nn_reg, xgboost_reg, mlp_reg); the aggregate event then carries mean_r2 / mean_rmse / mean_mae (± std) and a pooled predicted-vs-actual scatter. It also covers the Spatial MLP models (spatial_mlp, spatial_mlp_5x5) when you pass their neighbourhood features via spatial_vectors= / spatial_labels=.

Already have a TEE vector directory on disk? Load it directly:

from tessera_eval import load_tee_vectors

vectors, coords, metadata = load_tee_vectors("/path/to/vectors/aoi/2024")
# vectors: float32 (N, 128); coords: int32 (N, 2) pixel (x, y); metadata: dict

See the tutorial for the full workflow (labels → learning curve → confusion matrix → interpretation), and CHANGELOG.md for the release history.

Command-line interface

The workflow covered in the tutorial can also be run through the command line.

First, install the extras the load step needs:

pip install "tessera-eval[geotessera]"   # tile access
pip install "tessera-eval[xgboost]"      # optional, for xgboost models

Download the Tessera embeddings for your labelled ground truth, and save the result to a file (vectors.npz by default, change the name with argument --output). --data accepts either a shapefile/GeoJSON of labelled polygons or a GeoTIFF of an already-rasterized reference layer.

For a shapefile/GeoJSON, --field is the column holding the class or target values (e.g. habitat):

tessera-eval load --data /path/to/habitats.geojson --field habitat --year 2024

For a GeoTIFF, --bbox is required, in EPSG:4326 (longitude,latitude in degrees), since a raster has no natural area boundary the way labelled polygons do. --nodata marks any missing-value codes:

tessera-eval load --data site_type.tif --bbox 27.1,67.75,27.2,67.85 --year 2024 --nodata 32766,32767

kfold and learning-curve reuse the cached vectors.npz automatically. Pass --vectors <path> to use a different cached file instead.

Run k-fold cross-validation and print accuracy per model.

tessera-eval kfold --models rf,nn,mlp      # for classification
tessera-eval kfold --models rf_reg,nn_reg  # for regression

The CLI covers the pixel models (nn/rf/xgboost/mlp and their _reg variants). The Spatial MLP models in k-fold need the neighbourhood-feature extraction that only the web Validation panel wires up.

Optional arguments:

# --k: number of cross-validation folds (default: 5)
# --seed: random seed for reproducible fold splits (default: 42)
# --confusion-matrix: also print the full confusion matrix (raw counts), on top of the summary shown by default (for classification only)
# --vectors: path to a different cached .npz (default: vectors.npz from `load`)
# --max-samples: cap the training set size per fold (random, not stratified by class) -
#   usually needed for raster-derived data, which labels every pixel, not a hand-picked subset
tessera-eval kfold --models rf --k 10 --seed 1 --confusion-matrix --max-samples 50000

Investigate how accuracy changes using different fractions of training labels (currently only for classification task).

tessera-eval learning-curve --models rf --training-pcts 1,5,10,30,50,80 --repeats 5  # --training-pcts: % of labels per step (default: 1,5,10,30,50,80); --repeats: random repeats per step (default: 5)

Since neighbouring pixels are usually very similar, a random split can overstate how accurate the model really is (see tutorial). For a more reliable estimate, train on one geographic half of your area and test on the other, splitting by longitude:

tessera-eval learning-curve --models rf --spatial-holdout

You can also choose exactly which region to hold out, by passing a bounding box (for a shapefile/GeoJSON, in the same CRS as your data; for a GeoTIFF, always EPSG:4326):

tessera-eval learning-curve --models rf --spatial-holdout --test-bbox 27.16,67.77,27.23,67.82

You can also use a completely separate file as the test set - a different region, a different year, or both. If you are evaluating on a different year, pass --test-year along with --test-data.

tessera-eval learning-curve --models rf --test-data /path/to/other_region.geojson --test-year 2023

For a GeoTIFF, --test-bbox is also required, defining the test region within that file:

tessera-eval learning-curve --test-data site_type_2023.tif --test-bbox 27.1,67.75,27.2,67.85 --test-year 2023 --models rf

This reuses the cached vectors.npz as the training data by default. To use a different training area or year instead, pass --data/--bbox/--year explicitly:

tessera-eval learning-curve --data site_type_2019.tif --bbox 27.1,67.75,27.2,67.85 --year 2019 --test-data site_type_2023.tif --test-bbox 27.1,67.75,27.2,67.85 --test-year 2023 --models rf

Run any command with --help for a list of all possible arguments.

Documentation

• Data formats — the Tessera embedding formats this library reads and the exact dequantization maths. Start here if you're wiring in your own data.
• API reference — every public function, with array shapes and dtypes.
• Tutorial — an end-to-end worked example.
• Compute server — running tee-compute (local ML, hosted data).
• examples/ — a small runnable dataset (austria_crops.geojson) and the quickstart snippet.
• CHANGELOG.md — release notes.

What's in the box

The compute server reads zarr embeddings directly through geotessera's GeoTesseraZarr interface as a fast path, probing coverage first and falling back to NPY tiles whenever the store is unavailable or lacks the requested region or year.

Available models: nn, rf, mlp, spatial_mlp, spatial_mlp_5x5, xgboost (if installed), unet (if torch installed); regressors nn_reg, rf_reg, mlp_reg, xgboost_reg. See available_classifiers() / available_regressors().

Design notes

• Class imbalance is expected and fine. Macro-F1 is reported alongside weighted-F1 precisely so rare classes are visible.
• Determinism. A single seed (default 42) threads through point sampling, fold/resample splits, and every estimator's own random_state. Same inputs + same seed → identical numbers.
• Spatial leakage. For honest accuracy on contiguous habitats, prefer the spatial hold-out (run_learning_curve(..., test_vectors=, test_labels=)) over a random pixel split — neighbouring pixels are highly autocorrelated.
• Memory. load_embeddings_for_shapefile streams one tile at a time and keeps only labelled pixels, so county/country-scale shapefiles are tractable.

Development

git clone https://github.com/ucam-eo/tessera-eval && cd tessera-eval
python -m venv .venv && source .venv/bin/activate
pip install -e ".[server,dev]"
ruff check . && ruff format --check . && pytest

See CONTRIBUTING.md.

Citing

If this is useful in academic work, please cite the Tessera model and link back to this repository. (A CITATION.cff will be added alongside the Tessera paper reference.)

License

MIT.