Chesapeake Conservancy data scientists taught a new AI tool to map wetlands more accurately

The Annapolis nonprofit’s data science team has developed a new artificial intelligence deep learning model for mapping wetlands, with a reported 94% accuracy. This “AI wetlands model” shows hope for what has been decades of challenges for protecting and conserving wetlands, VP for Climate Strategy Susan Minnemeyer told Technical.ly.

What are wetlands? Quite literally, they’re areas where water covers soil. The soppy nature of these geographical terrains has always proved tricky to map for organizations like the Chesapeake Conservancy.

“Wetlands are called the kidneys of a landscape or nature’s supermarket because they provide so many ecological services,” said Dr. Kumar Mainali, a lead member of the Chesapeake Conservancy-housed Conservation Innovation Center team responsible for this new and highly accurate deep learning model. “Unfortunately, in North America, we’ve lost about 36.5% of our wetlands since 1900. … Wetlands are important because they clean polluted waters, protect sewer lines, recharge groundwater, stabilize water supplies, regulate climate, provide food, water, timber, and they interestingly mitigate both floods and drought.”

National Wetlands Inventory data hasn’t been comprehensively updated for many years. Mainali said modeling an approach to wetland mapping that can use training data of varying geographies will be useful in modernizing wetland mapping where it is most needed.

This image shows areas where the training data (green) did not map a wetland, but the Chesapeake Conservancy model correctly mapped where a wetland is present. “This is likely an area here a wetland has been restored,” per the org. (Courtesy image)

The results of this research were recently published in the peer-reviewed Science of the Total Environment. Much of the Chesapeake Conservancy data team’s work was funded by EPRI, or Electric Power Research Institute.

“The institute has had big challenges around wetlands whenever they have new infrastructure projects, since there’s a lot of infrastructure growth in cities, including solar development and renewable energy projects,” Minnemeyer said. “They want to reduce the risk of stumbling upon an environmentally sensitive area before getting to the field work portion of development.”

According to the researchers, unmapped wetlands are a result of old data being manually drawn using aerial photographs. That’s where an AI model comes in handy.

Ultimately, we think we could be using the model to map wetlands nationally.

The wetland model was trained with the decades-old National Wetlands Inventory dataset and recent satellite and aerial imagery data. Mainali said developing this model took about a year, but once such a tool is up and running, it’s able to consistently give a desirable output at a very small fraction of cost compared to other tools.

“It’s definitely an upfront investment,” Minnemeyer said, “but once the model is trained, it’s as close to push button as this type of product can get.”

The team found that the data improved the local accuracy of wetland mapping by 10% compared to predictions before training. The project was piloted in three geographically diverse counties in different regions of the United States: Kent County, Delaware; Mille Lacs County, Minnesota; and St. Lawrence County, New York.

“The first three counties for the project were chosen in part because they had high-quality data,” Minnemeyer said. But the team also wanted to learn whether the model could succeed in mapping wetlands where the data hadn’t been updated in an even longer period, which led them to a fourth site: Lincoln, Nebraska, where wetlands are an important issue “because they have unusual saline wetlands with endangered, unusual or endemic species.”

The National Wetlands Inventory data for that region, like many, dated back to the 1980s, and included inaccuracies like noting a wetland area where there’s now a suburban development or highway. Even so, the AI model achieved 90% accuracy.

The outdated training data is shown in green. The model prediction is in purple, overlaid over recent satellite imagery. (Courtesy image)

The performance of the model is promising for the usefulness of this new AI approach. Minnemeyer said the team has aspirations of scaling the model.

“Being the Chesapeake Conservancy, we do a lot of work with the Chesapeake Bay Program, and the Bay Program has a goal to restore our lost wetlands across our region,” she said. “They’ve had a very hard time achieving that goal.”

The team is hopeful of the role of their data in providing better information on the current status of wetlands in the Chesapeake Bay watershed for the enjoyment, education, and inspiration of this and future generations.

AI has the ability to learn deeply hidden patterns in data, per Mainali, so the model gets really good when researchers have a big data set and lots of information for the model to learn from. Thus, the Chesapeake Conservancy team and their colleagues at organizations that support the work believe that this AI model is only going to get more accurate and become more powerful with every project: “Ultimately,” the researcher said, “we think it could be using the model to map wetlands nationally.”