Multi-Modal Foundation · allenai
OlmoEarth v1 Large
A 308M-parameter Vision Transformer foundation model for satellite imagery and Earth observation tasks.
What OlmoEarth v1 Large is
OlmoEarth v1 Large is a remote sensing foundation model developed by the Allen Institute for AI (Ai2). It serves as the largest checkpoint in the OlmoEarth series, created to process satellite observations for environmental, climate, and conservation research.
The model addresses the challenge of analysing diverse Earth observation data streams. By operating over multiple satellite modalities, it provides a base representation for tasks ranging from broad land cover mapping to fine-grained climate and habitat monitoring.
Researchers can use the model directly to extract numerical embeddings from satellite observations or fine-tune it on specialist datasets to support specific predictive and mapping tasks across Earth sciences.
How it works
OlmoEarth v1 Large is built upon a Vision Transformer Large (ViT-Large) backbone with 308 million parameters. Rather than relying on a single sensor format, the architecture is designed to ingest data from Sentinel-1 synthetic aperture radar, Sentinel-2 optical multispectral sensors, and Landsat missions.
The network can consume individual satellite scenes as well as temporal sequences of imagery, enabling the model to learn features across both spatial and temporal dimensions.
Key facts
- Architecture
- ViT-Large (Vision Transformer Large)
- Parameter count
- 308 million
- Supported sensors
- Sentinel-1, Sentinel-2, Landsat
- Input formats
- Single images and image time series
What you can use it for
- Land Cover Segmentation
- Fine-tuning the model to classify and map land cover categories across regions using Sentinel or Landsat imagery.
- Earth Observation Embeddings
- Extracting vector representations from satellite scenes or image collections to enable clustering, search, or downstream statistical modelling.
- Time Series Environmental Monitoring
- Processing repeated acquisitions over the same geographic area to observe environmental transitions, ecosystem disturbances, or climate impacts.
Run it here
We list this model but cannot run it yet: no pipeline_tag — needs a human.
Everything on this page still applies. Running it needs a loader for this repository, not a bigger machine.
Strengths and limits
Good at
- Operates across multiple sensor types, including Sentinel-1, Sentinel-2, and Landsat.
- Processes multi-temporal image sequences in addition to static satellite scenes.
- Can be deployed as a general feature extractor for embeddings or fine-tuned for specialised segmentation tasks.
Watch out for
- Designed specifically for Sentinel-1, Sentinel-2, and Landsat imagery, excluding other satellite sensor types without adaptation.
- Requires fine-tuning with external frameworks such as rslearn or olmoearth_run to solve downstream task-specific problems.
Inputs and outputs
You provide
Sequence or text
A protein or nucleotide sequence, or free text — whichever this model was trained on.
You get back
Mean-pooled representation, one vector per input.
Licence
Check the licence before commercial use · See model card
The See model card licence carries terms worth reading before you depend on it.
Questions
Can I use OlmoEarth v1 Large commercially?
Check the licence before commercial use. The See model card licence carries terms worth reading before you depend on it.
What sensor modalities can this model process?
The model is built to ingest imagery from the Sentinel-1, Sentinel-2, and Landsat Earth observation platforms.
Can OlmoEarth v1 Large process multi-temporal data?
Yes, the model accepts both single satellite scenes and multi-image time series.
What architecture does the model utilise?
It is based on a Vision Transformer Large (ViT-Large) architecture containing 308 million parameters.
How can researchers use the model for downstream tasks?
Researchers can either generate feature embeddings from images or fine-tune the backbone for tasks such as semantic segmentation.
Written from OlmoEarth v1 Large's own documentation (model card). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.