Ai2 introduces OlmoEarth to analyze Earth with AI

Ai2 has introduced OlmoEarth Platform, an open platform that turns satellite imagery and other Earth data into useful maps and analysis for agriculture, conservation, wildfire prevention, and deforestation monitoring.

The platform launches in early access with a specific goal: making planetary analysis less dependent on large laboratories with specialized teams. Ai2 says OlmoEarth was trained on millions of Earth observations, around 10 terabytes of data, and can be adapted to different uses without creating a separate model for every problem.

One model for multiple tasks

Until now, an organization wanting to detect forest loss, classify crops, and monitor land-use changes would typically have to combine several models, data sources, and tools. That means higher costs, more technical work, and more time before results are available.

OlmoEarth works as a foundation model: a system trained in advance that can be fine-tuned with local data for specific tasks. For example, an agency could use it to analyze imagery from a region, add its own labels, and adapt it to identify crops or deforested areas.

The advantage is reusing the same foundation across multiple projects. Ai2 says its evaluations show solid performance across different tasks without sacrificing accuracy, although results will depend on the data and configuration of each application.

Three pieces for turning data into a map

The platform brings several tools together in one environment:

• OlmoEarth Studio lets you upload imagery and labeled data, coordinate annotation teams, and fine-tune models in a few steps. It can also request imagery from sources such as Sentinel-1, Sentinel-2, and Landsat based on the selected area and period. It is available at no cost.
• OlmoEarth Viewer is used to explore and publish maps from a browser. It includes map comparisons, a timeline for reviewing changes, and data such as the percentage of forest or crops visible in an area, along with confidence levels.
• OlmoEarth Run manages large training, fine-tuning, and inference jobs. It splits tasks, assigns CPU or GPU resources, logs errors, and can automatically retry failed processes.

There are also APIs for importing datasets, launching analysis from other systems, and downloading predictions. This lets organizations integrate the results into the tools they already use instead of requiring them to work entirely within the platform.

Real-world use cases

Ai2 has already tested the technology with organizations working on specific problems. In collaboration with NASA's Jet Propulsion Laboratory, OlmoEarth uses field samples and radar and optical imagery to estimate the moisture content of vegetation fuel, a factor that helps calculate wildfire risk.

In Nandi County, Kenya, the International Food Policy Research Institute adapted a model to update crop maps more frequently than the historical five-year cycle. That information can help anticipate needs for seeds, fertilizer, and technical assistance.

The platform has also been used in projects tracking mangroves, deforestation in the Amazon, and ecosystem mapping. In these cases, updating data more frequently could allow governments and conservation organizations to respond sooner to habitat loss or direct their resources more effectively.

What changes for you

If you work in agriculture, forest management, biodiversity, or disaster response, the main change is that you can start with an already trained model and adapt it to your territory instead of building everything from scratch. The process includes collecting imagery, labeling it, training, validating, and publishing results.

For the general public, OlmoEarth Viewer can turn technical analysis into maps that show how a place changes over time. For researchers and developers, Ai2 publishes code, documentation, and fine-tuned models in a public GitHub repository, which can also run offline in their own environments.

The platform is in its initial rollout, so its performance outside associated projects and the conditions for using it at larger scale still need to be assessed. The important direction is clear: turning the data already captured by satellites and sensors into more frequent, concrete decisions about crops, wildfires, forests, and ecosystems.