Infrastructure·
OlmoEarth v1.3 Delivers Planet-Wide Satellite Embeddings with Minimal Compute
AI2 releases OlmoEarth v1.3, a vision-transformer-based satellite embedding model producing 128-dim global vectors at 10m resolution. Precomputed annual embeddings cover the entire planet and are freely accessible via Source Cooperative, enabling lightweight downstream tasks without GPU-heavy fine-tuning.
Lightweight Global Embeddings for Satellite Analytics
OlmoEarth v1.3 introduces a vision-transformer architecture that ingests a full year of Sentinel-1 radar, Sentinel-2 multispectral, and Landsat imagery, compressing each 10-meter pixel into a single 128-dimensional vector. The first 64 dimensions function as a compact standalone embedding, reducing storage and inference overhead for edge or low-resource deployments.
The model was evaluated across 27 benchmarks spanning six continents, outperforming AlphaEarth Foundations and Tessera v2 on 20 of 27 tasks at full dimensionality and 18 of 27 at half. A new crop-type benchmark, PLANTEUR, highlights strong generalization to unseen tropical regions, reinforcing the model’s utility in data-scarce environments.
All precomputed annual embeddings are hosted on Source Cooperative, allowing users to download, visualize, and classify land cover, track deforestation, or map smallholder farms without signing in or provisioning GPUs. A browser-based OlmoEarth Embeddings Explorer further lowers the barrier for non-technical users to customize planetary-scale maps.
Why it matters for GPU / AI infrastructure
By shifting from per-task fine-tuning to reusable embeddings, OlmoEarth v1.3 reduces the need for continuous high-throughput GPU workloads. Organizations can now run lightweight classifiers on precomputed features, minimizing cloud GPU spend and enabling scalable satellite analytics on modest hardware.
- aigpu
- ai gpu
- ai gpu cloud
- aigpu dubai
- satellite embeddings
- olmoearth
- vision transformer
- gpu infrastructure
By AiGpu Editorial · Editorial rewrite based on public reporting (Hugging Face Blog)
← All articles