Now an artificial intelligence revolution is enabling the USGS to accomplish a lot more—and at a much faster pace.

For decades after artificial intelligence (AI)—which refers to machines or computers designed to think, learn and make decisions like humans—was introduced in the 1950s, it was mainly used to follow a set of rules to solve specific problems. Later, a new type of AI called machine learning emerged. This approach relies on algorithms that detect patterns and uses them to make predictions, enabling systems to not only address specific issues but also to learn from their experiences and apply that knowledge in solving new problems. 

Media

Sources/Usage: Public Domain. View Media Details

USGS EROS Center Director Pete Doucette is enthusiastic about the potential AI now offers to meet complex science and data needs at the USGS and EROS. Doucette has served on high-profile AI panel discussions, and in a recent two-episode Eyes on Earth podcast series about AI, he compared earlier and modern AI to different ways to learn to ride a bike.

“You can learn to ride a bike from a manual, a user’s guide, on how to ride a bike. That's kind of a knowledge-based way of thinking about the early days of AI,” Doucette said. 

“A more effective way to learn to ride a bike is to actually get on it and go through trial and error. Balancing and falling down over and over until you figure out how to balance, which is more in line with how machine learning works by processing information and learning from it in real time.”

Of course, machine learning can tackle far more complex topics than riding a bicycle, with outcomes that have far-reaching benefits.

Annual NLCD is not the only science work at EROS based on Landsat and reliant on AI. Another example is the Rangeland Condition Monitoring Assessment and Projection (RCMAP), which provides land cover detail specifically for rangeland in the western United States. AI helps make predictions about the amount of different types of land cover, including shrubs and grasses, across the landscape.

Other science efforts are showing interest in Annual NLCD’s AI methods, too, Sohl said. But there’s an even more advanced AI approach that EROS is exploring called a foundation model, which once set up could be adapted more easily to other projects. Foundation models are large-scale models trained on vast amounts of data that can be applied to a wide variety of tasks.

Media

Sources/Usage: Some content may have restrictions. View Media Details

USGS EROS Research Physical Scientist Neal Pastick, who also participated in the podcast episodes, has been working to develop a suite of foundation models pretrained on large amounts of geospatial information such as imagery from Landsat and other satellites. 

“The idea is akin to how large-language models operate on text,” Pastick said. ChatGPT, for example, reviews content on the internet and other places to learn from context and get better at predicting appropriate words in order to create new content. “But instead, we’re working with remote sensing data. So we’re trawling the entire archive, trying to exploit patterns therein and using that information to better understand the landscape around us.”

The results, which can range from mapping fire scars and monitoring snow depth to forecasting invasive plant species, could be applied to a wide variety of uses for land and water resource managers. 

“Once we stand these models up, researchers and users don't have to start from scratch,” Pastick said. “They can leverage the weights within these models to kind of speed up deployment of solutions. In short, we're basically making a cost-effective tool for gaining insights from geospatial information.”

AI could also offer advancements in data integration, Doucette said. Combining remote sensing data from different sensor types can provide even more information than either sensor alone could. For example, Landsat data has been harmonized with similar data provided by the European Sentinel-2 satellites. This ultimately provides more frequent imagery of the Earth, which is helpful for monitoring rapidly changing land conditions.

Media

Sources/Usage: Public Domain. View Media Details

To take this idea a step further, AI techniques might be able to help integrate data from dissimilar sensors, such as the light energy of Landsat with the radio waves of radar. Doucette compares the potential to how the human brain can blend information from two senses, seeing and hearing, “in an almost seamless way” and put it to use. 

Doucette sees the big challenge to this “next frontier” as being the large amount of data and computing power required to train an algorithm. 

And the data just keeps coming. The trio of satellites in Landsat Next, for example, planned for launch in late 2030/early 2031, will capture far more detail about features of the Earth’s surface more frequently than current Landsat satellites. Landsat Next would be expected to produce about 20 times more data than its predecessor, Landsat 9.

“As we start moving in that direction of having more data that we have to comb through, I think it's even more important to try to develop these types of models to exploit the data for better managing our world,” Pastick said.