How to Export Custom Geospatial Embeddings via OlmoEarth Studio

The Allen Institute for AI recently updated the OlmoEarth Studio platform to support the direct extraction of embedding vectors for satellite imagery. This new OlmoEarth embeddings export feature bypasses the computational overhead of full model fine-tuning. You can generate compact numerical representations of Earth observation data for downstream tasks like similarity searches, segmentation, and unsupervised data exploration.

Model Architecture and Efficiency

The embeddings are generated using the OlmoEarth v1.2 foundation model family. This update introduced Rotary Positional Embeddings (RoPE) to replace the absolute positional embeddings (APE) used in earlier architectures.

Switching to RoPE eliminates the striping and tiling artifacts common in traditional vision transformers applied to geospatial data. Continuous spatial relationships are preserved across chunk boundaries. This architectural shift also brings significant efficiency gains over the original v1.0 release. The v1.1 and v1.2 models deliver a 1.7x reduction in GPU training hours and a 2.9x reduction in multiply-accumulate operations (MACs) during inference.

These models are part of the broader OlmoEarth Platform designed for mission-driven organizations tracking climate change, deforestation, and food security. Ai2 recently expanded its partnership with Hugging Face, securing nearly two petabytes of storage to house the underlying datasets and model weights.

Configuring Custom Exports

You can trigger embedding extraction through either the OlmoEarth Studio web interface or the API. The extraction pipeline requires four primary configuration parameters.

Parameter	Description
Area of Interest (AOI)	The geographic bounding box defining the spatial extent of the export.
Time Range	The temporal window for the imagery, filtering out data outside the specified dates.
Imagery Source	The target satellite or sensor constellation, such as Sentinel-2.
Encoder Variant	The specific OlmoEarth v1.2 model size used to generate the numerical representations.

Selecting a smaller encoder variant accelerates the export process but may reduce accuracy on complex segmentation tasks. Review the OlmoEarth API documentation for the specific JSON payload structures required for automated batch exports.

Handling Cloud-Optimized GeoTIFF Outputs

OlmoEarth Studio formats all exported embeddings as Cloud-Optimized GeoTIFFs (COGs). COGs maintain complete spatial referencing while optimizing file structure for cloud-native workflows.

When working with large regional datasets, standard GeoTIFFs force your local machine to load gigabytes of pixel data into memory before running similarity searches. The COG format structures the embedding vectors internally by tiles and overview levels. Because COGs support HTTP GET range requests, geospatial software can read specific portions of the file over a network without downloading the entire dataset.

This format is natively compatible with standard Earth observation libraries like rslearn. You can stream these lightweight files directly into your analysis pipelines or vector databases. Your downstream Python scripts can query only the specific bounding box required for a given inference step, dramatically reducing local memory pressure.

Performance and Tradeoffs

Using frozen embeddings rather than fine-tuning a domain-specific embedding model saves substantial compute. Direct embedding exports isolate the heavy lifting on Ai2’s infrastructure. You receive the processed COGs and can immediately run clustering algorithms or train lightweight linear classifiers on top of the vectors.

According to standard industry benchmarks, OlmoEarth embeddings offer state-of-the-art performance across 24 geospatial tasks.

Evaluation Mode	Tasks Won	Competitive Models Outperformed
Embeddings Only	15 out of 24	Meta DINOv3, IBM/NASA Prithvi, DeepMind AEF
Fully Fine-Tuned	19 out of 29	Meta DINOv3, IBM/NASA Prithvi, DeepMind AEF

While the base embeddings win the majority of standard evaluations, full fine-tuning still yields superior results on 19 out of 29 extended benchmark tasks. Applications requiring highly localized pixel-level classification may still benefit from fine-tuning the base weights.

If your mission involves highly specialized sensor data or extremely subtle spectral signatures not well-represented in the baseline Sentinel-2 training data, the exported embeddings might underperform. In these scenarios, you will need to provision your own GPU instances and fine-tune the v1.2 foundation models directly on your custom imagery.

Configure your initial AOI constraints in the OlmoEarth Studio console to validate the embedding quality against a small subset of your target geospatial data before scaling up to regional exports.