OlmoEarth brings AI to Earth observation

Ai2 has introduced OlmoEarth, a family of open AI models that analyzes satellite imagery to detect changes on Earth and turn them into useful information for agriculture, wildfire response, conservation, and maritime surveillance.

The system does not work from a single photograph. It combines images collected over months by different sensors, including optical and radar satellites, along with land-cover maps, OpenStreetMap data, and vegetation-height measurements.

Four sizes for different uses

OlmoEarth is available in four versions, all using the same architecture and training method:

• OlmoEarth-v1-Nano, with around 1.4 million parameters, for fast, low-cost analysis at scale.
• OlmoEarth-v1-Tiny, with around 6.2 million parameters, also designed for efficient deployment.
• OlmoEarth-v1-Base, with around 90 million parameters, as a balance between speed and accuracy.
• OlmoEarth-v1-Large, with around 300 million parameters, for more demanding tasks.

Parameters are essentially the values the model learns during training. More parameters do not guarantee better results on their own, but they usually allow the model to represent more complex patterns at the cost of requiring more resources.

Ai2 says its models outperform several recent alternatives, including Meta's DINOv3, IBM's Prithvi and Terramind, as well as academic models such as CROMA and Panopticon. The evaluation covered image classification, object and change detection, segmentation, and predictions based on individual images and time series.

It also compared OlmoEarth with Google DeepMind's AlphaEarth Foundations. Because Google published processed representations but not the complete model, the comparison was not identical. On three tasks, OlmoEarth matched or exceeded AlphaEarth when using those representations directly. After fine-tuning it for each task, it outperformed AlphaEarth by a wide margin, according to Ai2.

What it can do in practice

The model has been tested on concrete problems, such as identifying the causes of deforestation in the Amazon, locating mangrove loss, estimating wildfire risk in North America, identifying smallholder crops in sub-Saharan Africa, and detecting vessels at sea.

That can lead to faster decisions and more accurate maps for:

• Knowing what is being grown and where, so agricultural aid can be directed more effectively.
• Estimating vegetation dryness and improving wildfire-risk maps.
• Reducing false alarms and undetected vessels in maritime operations.
• Prioritizing areas for restoration, habitat protection, or on-the-ground inspections.

The benefit is not only improved accuracy. Ai2 says that, by competing with larger specialized models, OlmoEarth can reduce analysis costs and allow organizations to update their maps more frequently.

Trained on incomplete real-world data

OlmoEarth was trained on approximately 10 terabytes of data and millions of samples from different regions around the planet, adjusted to a uniform resolution of 10 meters per pixel. It can use up to 12 monthly observations, although many samples do not include every month or sensor.

To learn, the model hides parts of the images and tries to reconstruct them using the visible information. Combined with map data and less precise labels, this method allows it to learn patterns without relying on every image having been manually reviewed and classified.

Ai2 has released the models, code, and training and evaluation materials as open resources. It also offers OlmoEarth Platform, a platform that lets organizations adapt the models to their own problems without specialized AI teams.

Versions adapted to classify mangroves, crops, agricultural land, and wildfire-related vegetation fuel are already available. The next family is in development and, according to Ai2, will arrive next year with weather data and applications for humanitarian response.

For you, the important point is that AI is not limited to describing a satellite image: it can compare places and dates, combine different signals, and produce operational alerts. The next thing to watch is how these models perform outside their test sets and how much it costs to keep them updated in regions where data is scarce or irregular.