Introducing OlmoEarth embeddings: Custom embedding exports from OlmoEarth Studio for downstream analysis

2026-04-23 · Source: Ai2 Blog · Field: Science & Research — Environmental Science & Earth Systems, Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Intermediate, medium

Summary

OlmoEarth Studio now offers custom embedding exports, providing compact numerical representations of Earth-observation data from its open-source OlmoEarth foundation models. Released on April 23, 2026, these embeddings are available in Nano (128-dim, 1.4M params), Tiny (192-dim, 6.2M params), and Base (768-dim, 89M params) variants. Users can configure exports via UI or API, specifying area, time range (1-12 monthly periods), resolution (10 meter, 20 meter, 40 meter, 80 meter), and imagery sources (Sentinel-2 L2A, Sentinel-1 RTC). The output is a lightweight Cloud-Optimized GeoTIFF (COG) with vectors stored as signed 8-bit integers. These embeddings facilitate tasks like similarity search, few-shot segmentation (e.g., 60 pixels yielding weighted F1 = 0.84), change detection, and unsupervised exploration, offering a cost-effective entry point for Earth observation analysis. Supervised fine-tuning is also supported for higher performance.

Key takeaway

For Earth observation analysts or data scientists seeking efficient, label-light methods for geospatial analysis, OlmoEarth Studio's new embedding exports provide a powerful solution. You can quickly generate custom, high-fidelity representations of Earth data, enabling rapid similarity searches, few-shot segmentation, and change detection without extensive training data. Consider integrating these COG exports into your existing geospatial workflows to accelerate insights and reduce labeling efforts.

Key insights

OlmoEarth embeddings offer a cost-effective, label-efficient way to analyze Earth observation data for diverse downstream tasks.

Principles

• Embeddings capture rich information from many sensors.
• Similar surface characteristics yield similar vectors.
• Linear classifiers perform well on few-shot tasks.

Method

Configure model parameters (area, time, encoder, resolution, imagery) in OlmoEarth Studio UI or API, run the computation, then download the resulting Cloud-Optimized GeoTIFF (COG) with 8-bit integer vectors.

In practice

• Perform similarity search using cosine similarity.
• Train logistic regression for few-shot segmentation.
• Detect changes by comparing monthly embeddings.

Topics

• OlmoEarth Studio
• Earth Observation
• Geospatial Embeddings
• Foundation Models
• Remote Sensing
• Few-Shot Learning
• Change Detection

Code references

• allenai/olmoearth_pretrain
• allenai/rslearn
• allenai/olmoearth_projects

Best for: Machine Learning Engineer, Research Scientist, Data Scientist

Related on AIssential

• OLMo Earth — OLMoEarth democratizes Earth observation AI through open-source models and a user-friendly platform.
• OlmoEarth v1.1: A more efficient family of models — OlmoEarth v1.1 significantly cuts compute costs for satellite imagery processing by optimizing token design and positional embeddings.
• Enhancing the KidSat Model: Integrating Geographical Encoding and Data Quality Assessment for Childhood Poverty Prediction — Integrating geographic encoding and data quality assessment significantly improves satellite-based socioeconomic predictions.
• Honey, I Shrunk the Arc de Triomphe! — A new dataset and method mitigate scale-collapse in monocular depth estimation for distant, unconstrained scenes.
• GeoWeaver: Grounding Visual Tokens with Geometric Evidence before Scene Reasoning — GeoWeaver grounds visual tokens with token-adaptive geometric evidence as a representational prerequisite for spatio-temporal reasoning.
• Exploring the potential of AlphaEarth and TESSERA embeddings for Fine-scale Local Climate Zone Mapping: A case study across five cities in Switzerland — Precomputed embeddings from EO foundation models can upscale coarse LCZ maps to 10-m resolution effectively.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Ai2 Blog.