nilsleh/tessera-embeddings-eval-datasets

TESSERA embeddings for geo-eval-harness evaluation datasets

Precomputed embeddings for the evaluation datasets of geo-eval-harness, stored as fixed records for KNN, linear and segmentation probes. Annual 128-dimensional, 10 m TESSERA v1.1 layers, read from the dClimate Icechunk store s3://tessera-embeddings/v1.1/dclimate.icechunk at the fixed snapshot recorded in each metadata.json, and nearest-neighbour resampled onto a grid of about 10 m cells covering each sample's footprint, in the sample's CRS.

Each collection is one self-contained folder with the same layout, named after the dataset's geo-eval-harness registry name. A <dataset>_pca16 folder holds a 16-component PCA projection of <dataset>.

<collection>/{train,val,test}-NNNNN.parquet
<collection>/metadata.json
<collection>/SHA256SUMS
<collection>/pca.safetensors     # PCA collections only

Use with geo-eval-harness

Evaluation reads collections from data/embeddings/<product>/<collection>. geo-eval-harness embeddings fetch downloads one at the revision pinned in geo-eval-harness and verifies every file's SHA-256:

geo-eval-harness embeddings fetch --product tessera --dataset treesatai treesatai_pca16
geo-eval-harness run --model tessera --dataset treesatai
geo-eval-harness run --model tessera-pca16 --dataset treesatai

Collections

Collection	Task	Classes	Codec	Embedding (C×H×W)	Target	Train	Val	Test	Empty (train/val/test)
benv2	classification	19	scaled-int8	128×120×120	label int64 19	20,000	4,000	4,000	2680/521/489
benv2_pca16	classification	19	float16	16×120×120	label int64 19	20,000	4,000	4,000	2680/521/489
dynamic_earthnet	segmentation	7	scaled-int8	128×154×154	mask uint8 512×512	700	100	200	0/0/0
dynamic_earthnet_pca16	segmentation	7	float16	16×154×154	mask uint8 512×512	700	100	200	0/0/0
fotw	segmentation	4	scaled-int8	128×154×88	mask uint8 256×256	4,000	1,000	2,000	0/0/0
fotw_pca16	segmentation	4	float16	16×154×88	mask uint8 256×256	4,000	1,000	2,000	0/0/0
kuro_siwo	segmentation	4	scaled-int8	128×130×130	mask uint8 224×224	4,000	1,000	2,000	2/0/123
kuro_siwo_pca16	segmentation	4	float16	16×130×130	mask uint8 224×224	4,000	1,000	2,000	2/0/123
spacenet7	segmentation	2	scaled-int8	128×210×211	mask uint8 512×512	3,500	652	1,152	16/0/0
spacenet7_pca16	segmentation	2	float16	16×210×211	mask uint8 512×512	3,500	652	1,152	16/0/0
treesatai	classification	15	scaled-int8	128×6×6	label float32 15	4,000	1,000	2,000	1410/148/1130
treesatai_pca16	classification	15	float16	16×6×6	label float32 15	4,000	1,000	2,000	1410/148/1130

Empty samples have no valid embedding cell (valid_fraction == 0), typically because the product does not cover the area or year, for example open sea. They are kept so that every split matches its source dataset.

Row schema

Each Parquet row is one sample, in the order of the source dataset split. Array columns are raw little-endian bytes; their dtype and shape are listed under arrays in metadata.json.

Column	Type	Meaning
index	int64	Sample index in the source dataset split
embedding	binary	Embedding (C, H, W) in the collection's codec
scale	binary	float32 (H, W) scale per cell (scaled-int8 only)
label	binary	Classification target: multi-hot (BigEarthNet v2 int64, TreeSatAI float32)
mask	binary	Segmentation target, uint8 (H, W) class indices on the source image grid
crs	string	CRS of the embedding grid
transform	6 × float64	Affine transform (a, b, c, d, e, f) of the embedding grid
height, width	int32	Embedding grid shape
product_year	int16	Annual layer the embedding was read from
observation_year	int16	Year of the source sample
valid_fraction	float32	Fraction of cells with a valid embedding

Decoding

scaled-int8: embedding holds int8 codes and scale one float32 scale per cell. A cell decodes to codes * scale, rounded to float16; a scale of 0 is a missing cell. PCA collections use float16, with NaN for missing cells.

import json
import numpy as np
import pyarrow.parquet as pq

meta = json.load(open("fotw/metadata.json"))
shape = meta["arrays"]["embedding"]["shape"]
row = pq.read_table("fotw/test-00000.parquet").slice(0, 1).to_pylist()[0]
codes = np.frombuffer(row["embedding"], np.int8).reshape(shape)
scale = np.frombuffer(row["scale"], np.float32).reshape(shape[1:])
values = (codes * scale).astype(np.float16).astype(np.float32)
values[:, scale == 0] = np.nan

Decoded values equal, bit for bit, the float16 values geo-eval-harness extracted from the source product.

Years

Each sample's embedding comes from the annual layer of its observation year. Samples observed before 2017, the first available year, use the 2017 layer; product_year and observation_year record both. Kuro Siwo uses the post-event year. DynamicEarthNet uses the year of the label month, so monthly samples of the same area and year share one embedding; its train split has 700 samples but only 195 distinct embeddings.

PCA collections

<dataset>_pca16 projects every cell of <dataset> onto 16 principal components. The PCA is fitted per dataset on the valid cells of the distinct training embeddings (duplicates removed), sampling at most 1,000,000 cells with seed 0, without whitening. pca.safetensors holds mean (C), components (16 × C), explained_variance and explained_variance_ratio (16); metadata.json records the fit and the digest of the source collection.

Integrity

Each collection's SHA256SUMS lists the SHA-256 of its files. metadata.json records the product source, codec, per-split counts and the geo-eval-harness version that exported it.

Licences

The embeddings are derived from TESSERA, licensed under CC BY-SA (as listed on the AWS Open Data registry; no version is stated). Attribution: "TESSERA (Temporal Embeddings of Surface Spectra for Earth Representation and Analysis), accessed from https://registry.opendata.aws/tessera."

Labels and masks come from the source datasets, as redistributed by GEO-Bench-2, and keep their licences:

Folder	Source dataset	Label licence	Reference
benv2, benv2_pca16	BigEarthNet v2 (reBEN)	CDLA-Permissive-1.0	https://arxiv.org/abs/2407.03653
dynamic_earthnet, dynamic_earthnet_pca16	DynamicEarthNet	CC BY-SA 4.0	https://doi.org/10.1109/CVPR52688.2022.02048
fotw, fotw_pca16	Fields of The World (GEO-Bench-2 subset)	CC BY 4.0 (GEO-Bench-2 includes only openly licensed countries)	https://arxiv.org/abs/2409.16252
kuro_siwo, kuro_siwo_pca16	Kuro Siwo	CC BY	https://arxiv.org/abs/2311.12056
spacenet7, spacenet7_pca16	SpaceNet 7 (MUDS)	CC BY-SA 4.0	https://spacenet.ai/sn7-challenge/
treesatai, treesatai_pca16	TreeSatAI	CC BY 4.0	https://doi.org/10.5194/essd-15-681-2023

Each folder is subject to both the product licence and its source dataset's licence.

Citation

Cite the product, GEO-Bench-2, and the source dataset of each collection you use.

@article{feng2025tessera,
  title={TESSERA: Temporal Embeddings of Surface Spectra for Earth Representation and Analysis},
  author={Feng, Zhengpeng and Atzberger, Clement and Jaffer, Sadiq and Knezevic, Jovana and Sormunen, Silja and Young, Robin and Lisaius, Madeline C. and Immitzer, Markus and Jackson, Toby and Ball, James and Coomes, David A. and Madhavapeddy, Anil and Blake, Andrew and Keshav, Srinivasan},
  journal={arXiv preprint arXiv:2506.20380},
  year={2025}
}

@article{simumba2025geobench2,
  title={GEO-Bench-2: From Performance to Capability, Rethinking Evaluation in Geospatial AI},
  author={Simumba, Naomi and Lehmann, Nils and Fraccaro, Paolo and Alemohammad, Hamed and De Mel, Geeth and Khan, Salman and Maskey, Manil and Longepe, Nicolas and Zhu, Xiao Xiang and Kerner, Hannah and Bernabe-Moreno, Juan and Lacoste, Alexander},
  journal={arXiv preprint arXiv:2511.15658},
  year={2025}
}