Paul Allen’s AI nonprofit unveils satellite data platform

A new open AI platform from the nonprofit created by the late Microsoft cofounder Paul Allen aims to make satellite imagery and other data about the earth more available and useful.  

The Allen Institute for Artificial Intelligence (Ai2) on Tuesday unveiled the OlmoEarth Platform, backed by a family of AI foundation models trained on roughly 10 terabytes of data derived from millions of observations of the planet, including satellite images, radar readings, and existing maps of features like forest cover.

The OlmoEarth models can then be fine-tuned for specific purposes, like detecting changes in vegetation, with the help of a companion software tool called OlmoEarth Studio. Details on model performance are included in a new scientific paper.  

Ai2 has already been working with a variety of organizations harnessing the AI, including groups looking to better assess and respond to wildfire risk. The International Food Policy Research Institute is using the technology to more frequently update maps of crops grown in one region of Kenya. Amazon Conservation is using the AI system to quickly spot deforestation. And a project called Global Mangrove Watch is harnessing the technology to more comprehensively track mangrove populations and quickly detect threats to those critical coastline trees. 

Patrick Beukema, lead researcher on Ai2’s OlmoEarth team, says the project grew out of a realization that while AI can help put earth imagery and data to use by quickly analyzing both new and historic images, actually deploying the technology could be a challenge for many organizations, including government agencies and nonprofits doing important work. 

“I think there’s a recognition that this kind of technology can be very valuable, but it’s so difficult to use, so we haven’t seen widespread adoption of these kinds of models, as we’ve seen, for example, within natural language processing or with [large-language models],” he says.  “And we haven’t really seen the transformative power of artificial intelligence within this domain.” 

State-of-the-art

To fill that gap, Ai2 created not only what Beukema calls state-of-the-art models, built using vision transformer technology similar to the large-language models that power tools like ChatGPT and Claude, but a set of companion tools making them practical to use. Those include OlmoEarth Studio, which simplifies the process of training the models for specific tasks by uploading human-labelled sample data showing relevant features like lands growing specific crops, areas of mangroves, or bits of forest vulnerable to wildfire.  

Once OlmoEarth models are fine-tuned, they can be used to analyze areas of the earth at a particular moment in time, selected as easily as finding a neighborhood on Google Maps and scheduling an event on a calendar app.  

“They can just tell the system, ‘I want mangroves, I want them in Indonesia over the last six months,’ or ‘I want a global inference over the last four years,’” Beukema says. “The idea is to build in that flexibility so that users can choose whatever they need.” 

Then, users can publish or privately share maps illustrating their findings, which can be viewed in an OlmoEarth Viewer app that can support interactive maps with options to select places and time ranges. The Studio and Viewer tools can be used without the need to write any code, though Ai2 also released a suite of automation tools and APIs for easy programming of its technology.  

Providing the tools

The organization also released documentation and examples on its GitHub page for the project, along with existing fine-tuned models that can immediately be put to use or even run offline on an organization’s own computers. And OlmoEarth follows other recent software releases from Ai2, including a package of science-focused AI tools called Asta released in August, and a set of language models known as OLMo, for open language models. 

Overall, Beukema says, the goal with OlmoEarth is to give organizations free technology that compares favorably to existing commercial and academic AI projects, letting them efficiently analyze and visualize planetary data they often already deeply understand, even when they don’t have the resources to build their own AI models from scratch. 

“These people are often experts, so they know exactly what they’re looking for,” he says. “They just don’t necessarily have, or want to build, these complicated foundation models that are expensive to train, expensive to inference, expensive to really work with.” 

Global Mangrove Watch—which tracks those coastal trees that are environmentally important as fish habitats, carbon stores, and as barriers to erosion, storm surge, and even tsunamis—is already working to improve its mapping and analysis processes with OlmoEarth. An existing machine learning and mapping system could already track mangroves with relatively high accuracy, but organizing training data and verifying the output still requires a lot of manual labor, says Lammert Hilarides, senior technical officer at Wetlands International, one of the organizations behind Global Mangrove Watch. 

A plan to scale up

Hilarides says OlmoEarth should allow the organizations to spend more time on other tasks, including working with governments and organizations around the world that are working to preserve mangroves and protect them from often-illegal deforestation. Critically, it will allow the project to update mangrove loss maps more quickly and let them cover a greater extent of the planet, catching disturbances to mangroves faster and more comprehensively.  

“We really hope that as of next year, we can scale up our work from covering not just half the world’s mangroves but all of the world’s mangroves,” he says. 

Ai2 plans to make OlmoEarth accessible to a wide range of organizations, with most features free for anyone to use, though a few features like fine-tuning elements will generally require groups to coordinate with Ai2 to make sure the product isn’t used for harmful purposes. Beukema says the nonprofit institute encourages organizations that think the technology could be useful to be in contact. 

“If you think this tech is going to help you accelerate your mission, please reach out,” he says. “We really want to help you.”

View the full article