Clay

Clay reads a square of satellite or aerial imagery together with where and when it was taken, and returns an embedding: a list of 1,024 numbers that summarises what is on the ground in that square at that moment.^1,8 Squares that look alike to the model get similar numbers, so the embeddings can be searched by similarity or fed to small classifiers.^1,9 The model is a vision transformer, the same family of network used for photographs, trained without labels by hiding three quarters of each image and learning to fill it back in.^1,3

Versions v0 to v1.5 were a fiscally sponsored project of the non-profit Radiant Earth; Clay is now a program of Renaissance Philanthropy.¹ The current release, v1.5, was trained in September 2024 on 70 million image chips chosen to cover the world's land cover types.^3,4 Its sources were Sentinel-2, Sentinel-1 radar, Landsat 8 and 9 and MODIS worldwide, plus NAIP aerial photos of the USA and LINZ aerial photos of New Zealand, with ground detail from under half a metre to 500 metres per pixel.⁴ Instead of a fixed sensor, the model is told the wavelength of each band, so it accepts any number of bands from many kinds of sensor.³ The documentation credits Prithvi among the models whose design it drew on.^3,4

The code installs as a Python package from GitHub and the v1.5 weights download from Hugging Face; the encoder alone is 1.25 GB.^1,3 It has three main uses:¹

• finding features such as surface mines, aquaculture or feedlots by similarity¹
• training a small model on the embeddings to classify crop type or land cover, estimate biomass, or detect change such as deforestation, wildfire, flooding or new building¹
• using Clay as the backbone of a larger model^1,6

Each embedding describes one place at one time, so following a place through the seasons means comparing embeddings from several dates.³ Some are already computed and published on Source Cooperative.^7,8,9 Clay used v1.5 to embed the whole NAIP aerial archive of the USA from 2010 to 2021.⁸ The company lgnd.ai publishes monthly v1.5 embeddings of Sentinel-2 worldwide from January 2017 to April 2026, under CC BY 4.0.⁹ Embeddings for all the training data of the older v0 model are there too, under ODC-BY.^1,7

Code and weights are under Apache 2.0 and the documentation under CC-BY 4.0.^1,2 The training data covered only land and coastal water, left out the poles, open ocean and night-time scenes, did not set out to include extreme events, and had at most six dates per location.³