Choosing an Earth embedding

An embedding is a list of numbers that a trained model gives to a piece of the Earth, chosen so that places that look alike get similar lists. With embeddings in hand, mapping crops or finding every quarry in a region comes down to labelling some examples and training a small classifier, which TESSERA's makers say runs on an ordinary CPU.² Several groups now publish open models that make embeddings, and some publish the numbers already computed.^1,2,3 Four questions decide between seven of them.

Do you want the numbers ready-made, or will you run a model? Ready-made is far easier. TESSERA publishes yearly embeddings for the whole land surface from 2017 to 2025 under CC0, free to download.² Google's AlphaEarth publishes yearly embeddings inside Earth Engine.³ The company lgnd.ai publishes monthly Clay embeddings of Sentinel-2 for the world from January 2017 to April 2026,¹¹ and Clay has embedded the US NAIP aerial archive from 2010 to 2021.¹² Ai2 publishes yearly OlmoEarth embeddings on Source Cooperative, worldwide for 2023 to 2025 and for some countries back to 2017, with the earlier years still being computed.¹ TerraMind has to be run on imagery you choose.^6,13

What does one embedding describe? TESSERA, AlphaEarth and OlmoEarth's published set give one per 10 metre square.^1,2,3 Clay and TerraMind give one per patch of pixels, so the ground each covers depends on the imagery and the settings, and so does OlmoEarth when you run it yourself.^1,10,14 SatCLIP is different in kind: it takes a longitude and latitude, needs no image at all, and describes the broad character of a region, so it suits regional questions such as temperature or population density rather than one field.^7,15

One moment or a whole year? TESSERA folds a year of radar and optical passes into each embedding, so a field that greens and browns is told apart from evergreen forest.² AlphaEarth is also annual.³ Clay describes one place at one time, so seasons come from comparing dates.⁴ OlmoEarth reads a time series of monthly mosaics, and its published set is one embedding per year.¹ TerraMind reads a single date unless wrapped for time series.⁶

What does the licence allow? TESSERA's embeddings are CC0 and its code MIT.² Clay's weights and TerraMind's are Apache 2.0, and SatCLIP's MIT, though its model cards say it is for research, not production.^4,6,7,15 OlmoEarth's own licence, which covers its published embeddings as well as the weights, lets anyone use and change them free of charge.^1,9 It bars military, surveillance and policing work and the extraction of oil, gas and minerals.⁹ Anything built from OlmoEarth carries the same limits.⁹

	Ready-made embeddings	One embedding per	Time	Numbers	Licence
TESSERA	yes, global, 2017 to 2025	10 m square	a year	128	MIT model, CC0 embeddings2
AlphaEarth	yes, in Earth Engine	10 m square	a year	64	see Earth Engine3
Clay v1.5	yes: S2 monthly global, NAIP US	patch of 8 pixels	one date	1,024	Apache 2.0; embeddings CC BY4,10,11,12
OlmoEarth	yes, global 2023 to 2025, some countries earlier	10 m square	a year	128 (or the first 64)	own licence, with use limits1,9
TerraMind	no	16 by 16 pixel patch	one date	768 (base)	Apache 2.013,14
SatCLIP	not needed	a coordinate	fixed	varies	MIT, research use7,15

To compare models on one place, ESA's Φ-lab runs Major TOM. It cuts the world into cells about 10 km apart.⁸ Imagery, elevation, night lights and the embeddings of several models, among them Clay, SatCLIP and a subset of AlphaEarth, are all keyed to the same cells.^8,16,17

TerraMind does one thing the others do not: from an ordinary Sentinel-2 picture it can generate a radar image or a land-cover map, and use the imagined layer to improve its answer.^6,14 Its model card warns against reading a generated layer as a reconstruction.¹³

TESSERA's Python library can download a year for a small box and paint three of its 128 bands as a colour map.¹⁸ Looking at that map shows quickly whether fields, roofs or wetlands in your area stand apart.