Annual, 10 m, per-pixel embeddings pre-computed with Spheer FM Albatross, a self-supervised geospatial foundation model trained on Sentinel-2 time series.

Spheer FM is deliberately specialised: it is trained on European Sentinel-2 time series, with a focus on nature and biodiversity. Unlike foundation models built primarily around spatial image structure, Spheer FM places temporal land-surface dynamics at the centre of its representation. Its temporal representations also perform strongly on agricultural tasks, including crop classification.

To learn more about our model:

Datasets at a glance

We publish 3 datasets: nl-tiles, Pastis Tiles, and Treesat.

nl-tiles: MGRS tiles fully covering the Netherlands

This is the entirety of the Netherlands from 2017-2020.


Model	Spheer FM (version: `albatross`)
Source imagery	Sentinel-2 timeseries
Embedding dimension	100
Ground sampling distance	10 m (one embedding per Sentinel-2 pixel)
Temporal aggregation	One embedding per pixel per calendar year
Years (in this release)	2017, 2018, 2019, 2020
Spatial coverage	10 MGRS tiles covering the Netherlands (all within MGRS grid zone 31U)
Format	Cloud-Optimized GeoTIFF, `float32`, 100 bands, EPSG:32631
License	CC BY-NC-SA 4.0 (Commercial licenses are available)

Pastis Tiles: 4 MGRS tiles fully covering the Pastis benchmark

Since the Pastis benchmark has a lot of little snippets over 4 MGRS tiles, we supply the entire tiles.


Model	Spheer FM (version: `albatross`)
Source imagery	Sentinel-2 timeseries
Embedding dimension	100
Ground sampling distance	10 m (one embedding per Sentinel-2 pixel)
Temporal aggregation	One embedding per pixel per calendar year
Years (in this release)	2018, 2019
Spatial coverage	4 MGRS tiles in France covering the Pastis benchmark
Format	Cloud-Optimized GeoTIFF, `float32`, 100 bands, EPSG:32631
License	CC BY-NC-SA 4.0 (Commercial licenses are available)

Treesat: Shapes fully covering the TreesatAI benchmark

The treesat benchmark makes use of point data. Instead of inferencing individual points we made shapes that together cover all of them. These shapes can be seen in the geoparquet.

The treesat benchmark also includes data from before 2017, which we cannot provide embeddings for since Sentinel 2 has only been fully functional since 2017. We suggest either snapping all earlier data to 2017 or dropping it all together.


Model	Spheer FM (version: `albatross`)
Source imagery	Sentinel-2 timeseries
Embedding dimension	100
Ground sampling distance	10 m (one embedding per Sentinel-2 pixel)
Temporal aggregation	One embedding per pixel per calendar year
Years (in this release)	2017, 2018, 2019, 2020
Spatial coverage	Shapes in Germany fully covering the TreesatAI benchmark
Format	Cloud-Optimized GeoTIFF, `float32`, 100 bands, EPSG:32632
License	CC BY-NC-SA 4.0 (Commercial licenses are available)

Other years and regions: Spheer FM is trained on and for Europe, with embeddings available or generatable for 2017–2025 across Europe. We are also developing regional versions for areas including Africa, Southeast Asia, and the United States. For access to additional years or regions, get in touch.

Intended use and limitations

These embeddings are intended as general-purpose geospatial features for downstream tasks such as habitat and vegetation mapping, land-cover and crop classification, biodiversity modelling, change analysis, and other Earth-observation applications.

Spheer FM Albatross was trained specifically on European Sentinel-2 time series, with a focus on nature and biodiversity.

Each pixel is represented by a 100-dimensional learned embedding. Individual embedding dimensions have no direct spectral or ecological interpretation, and the embeddings should not be used as a substitute for Sentinel-2 reflectance where absolute spectral values are required. Downstream models should generally use the full 100-dimensional vector.

Each annual embedding summarizes the temporal behaviour observed during that calendar year. Its quality therefore depends on the availability and quality of the underlying Sentinel-2 observations.

Embeddings produced by different Spheer FM model versions occupy different representation spaces and should not be mixed or directly compared within the same analysis.

As with any learned representation, downstream performance depends on the task, training data, geographic area, and modelling approach. Users should validate performance on representative local data before relying on the embeddings for operational or decision-making applications.

Repository layout

albatross-EU-v2025/
└── nl-tiles/
    └── 31UGV/
        ├── 2017.tif
        ├── 2018.tif
        ├── 2019.tif
        └── 2020.tif

albatross-EU-v2025/ — the model version that produced the embeddings. Vectors from different model versions are not comparable; pick one version and stay on it for a given analysis.
nl-tiles/ — Dataset. Also present are pastis-cogs/ and treesat-cogs/.
31UGV/ — Sentinel-2 MGRS tile ID in case of pastis-cogs and nl-tiles, and shape id in case of treesat-cogs.
<year>.tif — one COG per calendar year: 100 bands, float32.

Quickstart: read an area of interest

Using rioxarray / rasterio

This example fetches a region around the city of Groningen and displays the first three bands as a false color image.

import matplotlib.pyplot as plt
import rasterio
import rioxarray
from huggingface_hub import get_token

COG_URL = (
    "https://huggingface.co/datasets/spheer/spheer-fm-embeddings/resolve/main/"
    "albatross-EU-v2025/nl-tiles/31UGV/2020.tif"
)

# (minx, miny, maxx, maxy) in EPSG:4326
GRONINGEN_CITY = (6.45, 53.15, 6.70, 53.30)

GDAL_CLOUD_DEFAULTS = {
    "GDAL_DISABLE_READDIR_ON_OPEN": "EMPTY_DIR",
    "GDAL_HTTP_MERGE_CONSECUTIVE_RANGES": "YES",
    "GDAL_HTTP_MAX_RETRY": "10",
    "GDAL_HTTP_RETRY_DELAY": "0.5",
    "VSI_CACHE": "TRUE",
}

token = get_token()
assert token is not None, "No Hugging Face token found -- run `hf auth login` first."

with rasterio.Env(
    **GDAL_CLOUD_DEFAULTS,
    GDAL_HTTP_HEADERS=f"Authorization: Bearer {token}",
):
    da = rioxarray.open_rasterio(
        COG_URL, chunks={"x": 256, "y": 256}, parse_coordinates=True
    )

    # clip_box reprojects the bbox to the raster's CRS (EPSG:32631) for us.
    aoi = da.rio.clip_box(*GRONINGEN_CITY, crs="EPSG:4326")

    # Display first three embedding dimensions as false color image
    rgb = aoi.isel(band=slice(0, 3))
    (rgb / 10 + 0.5).clip(0, 1).plot.imshow()
    plt.show()

Notes:

chunks= makes the read lazy and requires dask. Drop it to load eagerly with numpy.
Keep all reads inside the rasterio.Env block. The auth header lives in the GDAL environment, so a lazy read triggered after the block exits will fail with a 401.
A full tile-year is ~50 GB. Try to clip / reduce before computing.

STAC catalog (experimental)

A static STAC catalog is included in this release. It does not currently expose a STAC API, so Client.search() is unavailable. A GeoParquet is provided which can be used to perform some search operations. Additionally, individual collections and items can still be accessed directly:

import odc.stac
import pystac_client
from huggingface_hub import get_token

# (minx, miny, maxx, maxy) in EPSG:4326
GRONINGEN_CITY = (6.45, 53.15, 6.70, 53.30)

CATALOG_URL = (
    "https://huggingface.co/datasets/spheer/spheer-fm-embeddings/resolve/main/"
    "albatross-EU-v2025/stac_catalog/catalog.json"
)

token = get_token()
assert token is not None, "No Hugging Face token found -- run `hf auth login` first."

catalog = pystac_client.Client.open(
    CATALOG_URL,
    headers={"Authorization": f"Bearer {token}"},
)
collection = catalog.get_collection("nl-tiles")
item = collection.get_item("nl-tiles-31UGV-2020")

odc.stac.configure_rio(
    cloud_defaults=True,
    GDAL_HTTP_HEADERS=f"Authorization: Bearer {token}",
)

# For these 100-band COGs, direct rasterio/rioxarray access is currently
# more convenient for small spatial windows.
ds = odc.stac.load(
    items=[item],
    bbox=GRONINGEN_CITY,  # EPSG:4326
    crs="EPSG:32631",
    resolution=10.0,
    chunks={"x": 256, "y": 256},
)

Searchable GeoParquet

A GeoParquet file is included in this release, which allows for efficient searching and filtering of the available items without requiring a full STAC API. One way to easily do this is through rustac.

from huggingface_hub import hf_hub_download
from rustac import DuckdbClient

catalog = hf_hub_download(
    repo_id="spheer/spheer-fm-embeddings",
    repo_type="dataset",
    filename="albatross-EU-v2025/stac_catalog/catalog.parquet",
)

client = DuckdbClient()

# This is a little square in Northern Germany that lies in both one of the treesat cogs and in one of the nl-tiles.
# Make sure this is in WGS84 (EPSG:4326)
example_geo_data = {
    "type": "Polygon",
    "coordinates": [
        [(7.08, 52.75), (7.08, 52.76), (7.09, 52.76), (7.09, 52.75), (7.08, 52.75)]
    ],
}

# We say we want data from 2019 that intersects with our little square
items = client.search(
    catalog,
    datetime="2019",
    intersects=example_geo_data,
)

for item in items:
    print(item["assets"]["embeddings"]["href"])

Output:

s3://spheer-fm-embeddings/albatross-EU-v2025/nl-tiles/31UGU/2019.tif
s3://spheer-fm-embeddings/albatross-EU-v2025/treesat-cogs/8/2019.tif

Using the Spheer app

If you are planning on doing machine learning on these embeddings, the Spheer app makes this a lot easier. At Spheer we make training AI models easy by offering a no-code platform. Please reach out if you are interested.

License

Released under CC BY-NC-SA 4.0: attribution required, non-commercial use only, derivatives under the same license. The Spheer FM model itself is proprietary and is not distributed here.

Sentinel-2 source imagery is provided by ESA / the Copernicus Programme.

Citation

@misc{spheer_fm_embeddings,
  title  = {Spheer FM Embeddings},
  author = {Spheer},
  year   = {2026},
  url    = {https://huggingface.co/datasets/spheer/spheer-fm-embeddings}
}

Contact

Spheer spheer.ai
Questions are also welcome in the Community tab

Downloads last month: 35

Total file size:

2.96 TB