Research

OlmoEarth v1.3: Precomputed planetary embeddings democratize satellite analysis

Allen Institute releases 128-dimensional satellite embeddings for the entire planet, outperforming competitors on 20 of 27 benchmarks and enabling zero-code geospatial tasks.

Last verified:

Precomputed embeddings eliminate infrastructure barriers for geospatial AI

According to Hugging Face, the Allen Institute has released OlmoEarth v1.3, a satellite embedding model that shifts the computational burden upstream by precomputing dense vector representations for every 10-meter pixel across the entire planet. The release bundles 128-dimensional and 64-dimensional annual embeddings derived from a full year of Sentinel-1 radar and Sentinel-2/Landsat multispectral time-series data. The key innovation is accessibility: rather than requiring end-users to operate GPU-backed fine-tuning workflows, researchers and practitioners can now train compact classifiers on precomputed embeddings and run inference locally, eliminating both GPU dependency and the need to download or store global satellite imagery.

Benchmark leadership across tropical and global tasks

OlmoEarth v1.3 was evaluated against two competitive baselines—AlphaEarth Foundations (AEF) and Tessera v2—across 27 benchmarks spanning six continents. According to the Hugging Face Blog, the 128-dimensional embeddings ranked highest on 20 of 27 evaluations, and the 64-dimensional variant (at half the vector size) still led on 18 of 27. Notably, the evaluation suite includes PLANTEUR, a newly introduced crop-type classification benchmark designed for smallholder farming in tropical regions. The embeddings demonstrated strong generalization to geographies unseen during classifier training, addressing a critical deployment challenge in regions where labeled training data is sparse or absent.

Open infrastructure and practical Earth-observation tasks

The release emphasizes transparency: model weights, pretraining data, and pretraining code are all public. According to Hugging Face, precomputed embeddings are distributed via Source Cooperative, enabling downstream tasks—crop mapping in Kenya, brick-kiln detection near Dhaka, fire-damage assessment in Indonesia—to become “lightweight modeling problems” solvable with small classifiers running on commodity hardware. A companion browser-based tool, the OlmoEarth Embeddings Explorer, removes the final technical barrier: users can explore and customize global maps without authentication or command-line tools.

Why This Matters

The embedding-as-infrastructure model reframes satellite geospatial analysis from a GPU-scarce, data-intensive research workflow into a democratized, compute-once-reuse-many-times service. For organizations in the Global South—where GPU capacity and satellite-data storage are bottlenecks—precomputed embeddings lower the entry cost for crop-yield forecasting, land-use monitoring, and disaster-impact assessment. The PLANTEUR benchmark validates that the embeddings transfer to underrepresented geographies and crop types, reducing the risk that foundation models trained on satellite data perpetuate coverage gaps. If the 64-dimensional variant sustains performance across downstream tasks as adoption scales, it will establish a cost-efficient standard for embedded geospatial inference, reshaping how agricultural advisory services and climate-resilience tools are built in data-constrained environments.

Frequently Asked Questions

What is OlmoEarth v1.3 and how does it differ from the previous version?

OlmoEarth v1.3 is a satellite embedding model that converts raw satellite imagery into dense vector representations. Unlike earlier OlmoEarth versions requiring fine-tuning in OlmoEarth Studio with GPU access, v1.3 precomputes embeddings for the entire planet, allowing users to train lightweight classifiers locally without downloading imagery or GPU hardware.

What satellite data sources does OlmoEarth v1.3 use?

The model ingests a full year of radar data from Sentinel-1 and multispectral time-series data from Sentinel-2 and Landsat, producing one 128-dimensional (or 64-dimensional) embedding vector per 10-meter ground pixel globally.

How does OlmoEarth v1.3 perform compared to competitors?

According to Hugging Face, OlmoEarth v1.3's 128-dimensional embeddings score highest on 20 of 27 evaluations across six continents against AlphaEarth Foundations and Tessera v2. The 64-dimensional variant leads on 18 of 27 benchmarks at half the size.

Can I use these embeddings without technical expertise?

Yes. The OlmoEarth Embeddings Explorer is a browser-based tool requiring no sign-in or advanced technical knowledge to explore and customize maps based on precomputed global embeddings.

#earth-observation #embeddings #geospatial-ai #open-source #foundation-models