OlmoEarth v1.3: State-of-the-art embeddings for the whole planet

Why it matters

The team behind OlmoEarth v1.3 has released an open satellite embedding model that condenses a year of radar and multispectral satellite time series into a single vector for each patch of ground on Earth. The model scored highest overall against competing embedding models from AlphaEarth Foundations and Tessera v2 across a broad evaluation suite spanning six continents, while also being cheaper to train than earlier OlmoEarth generations. Alongside the model, the team published precomputed annual embeddings for the entire planet and a browser-based tool, the OlmoEarth Embeddings Explorer, that lets anyone compare places across years, find visually similar locations, and train small classifiers without signing in or needing specialized hardware. The article demonstrates the approach through illegal gold mining in Peru's Madre de Dios region, where comparing pixel vectors across years reveals both the impact of a government crackdown and continued pressure on the landscape. Because weights, data, and code are all open, the release lowers the barrier for researchers and organizations doing land-cover mapping, change detection, and environmental monitoring.