AIToday
Open-Source AIHugging Face BlogPublished: Aug 13, 2026, 04:00 JST2 min read

OlmoEarth Studio adds custom embedding exports for Earth observation analysis

OlmoEarth Studio adds custom embedding exports for Earth observation analysis

Key takeaway

  • OlmoEarth Studio now lets users compute and export embedding vectors from satellite imagery, enabling fast, low-label analysis tasks like similarity search and land-cover mapping.

  • The embeddings—derived from the open-source OlmoEarth foundation models—encode rich spatial information and can be configured by area, time range, resolution, and imagery source, then downloaded as standard GeoTIFFs for use in any geospatial tool.

3 Key Points

  1. What happened

    OlmoEarth Studio, a platform for Earth observation models, now allows users to compute and export embedding vectors—compact numerical representations of satellite data. Users can configure parameters including area of interest, time range (1–12 monthly periods), encoder variant (Nano, Tiny, or Base), spatial resolution (10, 20, 40, or 80 meter per pixel), and imagery sources (Sentinel-2 L2A, Sentinel-1 RTC, or both), then download results as Cloud-Optimized GeoTIFFs.

  2. Why it matters

    Embeddings provide a fast, cost-effective way to leverage Earth observation data for downstream tasks—similarity search, land-cover segmentation, change detection, and unsupervised exploration—without requiring labeled training data or specialized infrastructure. The body demonstrates that a simple logistic regression classifier trained on just 60 labeled pixels can produce coherent land-cover maps with weighted F1 = 0.84, showing the embeddings encode rich ecological distinctions learned during pretraining.

  3. What to watch

    The source code and model weights are publicly available, so the community can inspect how embeddings are generated. For applications requiring higher performance beyond frozen embeddings, OlmoEarth Studio also supports supervised fine-tuning. Custom-computed embeddings are available now; users should reach out to gain access.

Ask the AI about this article →

Context & Analysis

OlmoEarth Studio positions embeddings as a practical entry point for Earth observation analysis, bridging the gap between raw satellite data and actionable insights. The platform's strength lies in its flexibility: users can query specific spatial and temporal windows without relying on pre-computed global archives, enabling seasonal dynamics to be captured rather than just annual snapshots. The body demonstrates this flexibility through concrete examples—change detection across monthly Sentinel-2 composites revealed the Park Fire burn scar in Butte County by comparing September 2023 and September 2024 embeddings—and shows that the learned representations already encode landscape structure without explicit task-specific training.

The few-shot segmentation example (60 labeled pixels yielding weighted F1 = 0.84 for mangrove, water, and other classes) illustrates a core value proposition: embeddings compress Earth observation data into vectors rich enough that simple linear classifiers can recover ecological distinctions. The body notes that accuracy saturates quickly—increasing from 30 to 300 labels barely improves results—because the embeddings are "doing most of the heavy lifting." This design philosophy treats embeddings as a frozen foundation layer suitable for rapid prototyping and resource-constrained environments, while the platform also supports supervised fine-tuning for applications requiring higher performance.

FAQ

What encoder options are available, and what do they differ in?
Three variants are available: Nano (128-dimensional, 1.4M parameters), Tiny (192-dimensional, 6.2M parameters), and Base (768-dimensional, 89M parameters). Larger variants encode richer representations but require higher compute and storage.
What imagery sources can be used to compute embeddings?
Users can choose Sentinel-2 L2A, Sentinel-1 RTC, or both. Sentinel-2 L2A imagery is accessed via Microsoft Planetary Computer.
How is the embedding data stored and formatted in the export?
Embeddings are exported as Cloud-Optimized GeoTIFFs (COGs) with one band per embedding dimension. Vectors are stored as signed 8-bit integers ranging from -127 to +127, with -128 reserved for nodata; floating-point vectors can be recovered using the dequantize_embeddings function.
Hugging Face BlogRead Original Article

Get the latest Open-Source AI news every morning

AI-summarized, only the topics you pick — one digest a day via Email, Slack, or Discord.

Free · takes 30 seconds · unsubscribe anytime

Ask AI

Ask AI anything about this article. Q&As are published on this page for other readers too.

Related Articles

Next articleAWS adds IAM cost tracking to Bedrock via Athena and CUDOS dashboards

The AI news that matters, in one minute each morning.

Sign up free