Hugging Face releases OlmoEarth v1.3 with global satellite embeddings

Until now, Hugging Face said, the main way to use OlmoEarth was fine-tuning in OlmoEarth Studio: training a model for each task and running it over satellite imagery to make a map. That approach requires GPU access and the ability to download and store imagery. Embedding models compress satellite data into reusable features that can be reused cheaply for many tasks, according to the blog post.

OlmoEarth v1.3 reads a year of radar data from Sentinel-1 and multispectral data from Sentinel-2 and Landsat and produces one 128-dimensional vector for every 10-meter pixel on Earth; the first 64 dimensions can also be used alone as a smaller embedding, Hugging Face said. A vision transformer reads the satellite data, an Attention Pooler pools it into one vector per patch of land, and a linear layer reduces it to 128 numbers.

The company computed the embeddings for the whole planet and shared them on Source Cooperative. With them, a crop map, a land-cover map or a change analysis becomes a lightweight modeling problem: a small classifier can be trained on top and run locally, according to the blog post.

Hugging Face evaluated the embeddings against AlphaEarth Foundations and Tessera v2 across 27 evaluations covering six continents and many different tasks. Its 128-dimensional embeddings scored highest on 20 of 27 evaluations and ranked first across the set, while the 64-dimensional version, at half the size, led on 18 of 27, the company said. It also introduced PLANTEUR, a crop-type benchmark for tropical farming, and said the embeddings generalized best to places a classifier had never seen. The model weights, pretraining data and pretraining code are open.

Hugging Face said a small classifier trained on the embeddings is enough to map what is on the ground, citing classifying smallholder crops in Kenya, detecting brick kilns west of Dhaka, assessing fire damage in Indonesia and tracking gold mining in the Peruvian Amazon.

OlmoEarth v1.3 was also the cheapest OlmoEarth model to train, using 53% of the computing time of v1.2 and 18% of v1, with 50% and 22% of the carbon emissions, according to the blog post. It remained a strong fine-tuning model, matching v1.2's performance across research and partner labels. Hugging Face said pretraining OlmoEarth v1.3 Base would cost roughly $2,720 of rented H100 time for the final run, and that full results and methodology are in its technical report.

In Peru's Madre de Dios region, illegal gold mining around La Pampa peaked in 2017 and 2018 and cleared more than 1,200 hectares of forest in the buffer zone of the Tambopata National Reserve, according to Amazon Conservation's MAAP monitoring program. Peru's Operation Mercurio shut most of it down in February 2019. Hugging Face said the embeddings help show whether it stayed down: comparing each pixel's vector across years without model training lights up the 2017-2020 spread to the southeast, while the 2020-2025 comparison shifts the signal. The crackdown is visible in the embeddings, it said. The second map is not dark, and according to MAAP, pockets in the northwest kept operating while the old mining zone kept changing.