Your body mass index is a calculation of your health based on a simple calculation using your height and weight.
Now, imagine a far more complex bit of math. This one uses layers of equations to sort 30 million data points instead of just two. To help, you get to use powerful computer systems once available only to government agencies or to well-funded researchers.
Right now, an increasing number of scientists around the Chesapeake Bay are doing just that, using artificial intelligence to answer huge environmental questions that define the future of a watershed where more than 18 million live across six states.
Researchers at the Chesapeake Conservancy in Annapolis last month announced that they created a neural network that learned how to identify wetlands by sifting through satellite images — and spotting places covered by parking lots, roads and farm fields where they once existed.
A new paper from the Virginia Institute for Marine Studies uses similar deep learning to identify man-made shoreline structures around the bay. Scientists at the University of Maryland Center for Environmental Science are about to publish the results of an artificial intelligence study of state forestry lands in Pennsylvania that evaluates the most effective management techniques.
AI research is being published on determining water quality, studying fish populations and mapping land cover, to name just a few applications.
Artificial intelligence is changing environmental science, and that will affect everything from land use decisions to best farming practices across the bay. It will change how we deal with climate change, the impact of expanding development, and finally, perhaps, help create a healthy Chesapeake. As more and more groups utilize this technology over the next few years, this will be a revolution.
“Taking all this information and translating it in a way that communities can make informed decisions, that’s a direct product of artificial intelligence,” said Kandis Boyd, director of the EPA’s Chesapeake Bay Program.
AI is helping supercharge progress, she said, on reaching restoration goals in the 2014 Chesapeake Bay watershed agreement.
This moment didn’t just arrive. AI has been a topic of discussion in environmental science since the 1990s. Three recent developments, however, make this a turning point for researchers.
First, artificial intelligence has advanced to the point where it can accurately make predictions. There are a bunch of complex terms used in talking about this technology, including “machine learning” or “deep learning.” The bottom line is that humans can write programs that computers use to teach themselves how to accurately predict answers to complex questions.
Five or 10 years ago, you had to be affiliated with a research university or the federal government to have access to the kind of supercomputing clusters needed for this work. It involves processing huge amounts of complex data. Now, many smaller research groups have access.
Researchers can now use these tools to take advantage of an explosion in continually updated, remote sensing information — think satellite images — with detail so fine that the result is a level of unprecedented accuracy. Imaging that once detailed a city block can now zoom down to a square meter.
“Being able to process large quantities of images, for example, to better characterize the land use in the watershed, to better characterize where submerged vegetation is in the bay, an important environmental outcome — those to me are really going to be turning points ... into how we do science in the bay,” said Isabella Bertani, an assistant research scientist with the bay program.
That’s exactly what the Chesapeake Conservancy research project shows, demonstrating how small organizations can leverage the democratization of artificial intelligence, computing power and big data.
Here’s what researchers at the conservancy did.
They gathered free satellite imagery for three small areas: Mille Lacs County, Minnesota; Kent County, Delaware; and St. Lawrence County, New York. They filtered that data through something called a convolutional neural network — layers of algorithms or equations known as deep learning. It processed images of these three areas, assigned importance to some of the details and then separated them from others.
As it progresses and more data is added, the network learns what is accurate — teaching itself what a wetland area is through repetition.
“A neural network is stacking layers upon layers of those equations, and combining outputs from each equation as an input to the next set of equations,” said Mike Evans, senior data scientist at the conservancy. “When we layer these together and allow them to interconnect and interact, it is a system of equations that can learn way more complicated and intricate patterns in the data than you could by simply combining all your variables into one equation.”
The result, in this case, is maps that predicted the contours of wetlands with 94% accuracy. In some cases, the results corrected maps that hadn’t been updated in decades.
Humans have been mapping wetlands for years, but the process involves years of study, a lengthy process of creating a description of wetlands for comparison and then collecting data. Even then, there is no guarantee that the process won’t include a human error or miss some key information.
“What deep learning does is guaranteed to be better than that, as long as you train the model for an adequate number of addresses,” said Kumar Mainali, lead data scientist for the conservancy.
The concept is similar to what happens when you take photos with your cellphone. Over time, the camera recognizes your husband or your kids — or your dog — and labels photos. The more pictures, the more accurate the labels. In this case, the computer labeled wetlands.
It also identified historic wetland locations, determining where wetlands used to be. That’s something that doesn’t currently exist.
“We don’t really have any accurate maps of where wetlands have been lost,” said Susan Minnemeyer, lead climate researcher for the conservancy. “You could dig into the data and research it, but there’s no U.S. map of wetlands that are no longer there.”
The research is significant for a couple of reasons. Accurate maps make it easier to avoid paving over wetlands, which are a protected part of an environment that filters water pollution and provides habitat. That will be crucial in the expansion of solar panel farms, or the placement of new high-speed rail.
Scaled up to the Chesapeake Bay, the maps could help with the goal of restoring lost wetlands. They could also identify dry areas prone to flooding, as well as the potential for saltwater intrusion into drinking water sources as sea levels continue to rise because of climate change.
The project is one of several at the conservancy using AI. Researchers want to update the land use map the conservancy provides to the bay program, and Mainali is working on a plan to develop a detailed biodiversity map.
“Knowledge is power. So what we need is tools like this that will produce high-resolution, high-accuracy information that can be incorporated into the decision-making tools and processes within the states and federal agencies in the Chesapeake Bay watershed,” said Joel Dunn, CEO of the conservancy.
There’s more work to do, he added, but just cracking the code with the wetlands model is a giant first step.
The proliferation of AI environmental studies on the Chesapeake Bay won’t only influence the future of local restoration efforts, or land use decisions. Because environmental science done on the bay is a world leader in the field, what gains acceptance here will expand to scientists globally.
Boyd, who has often spoken on the significance of AI research during her career with NOAA and the EPA, said the bay program has grants, contracts and cooperative agreements that can advance initiatives in the field. The work done by AI-enhanced research will also filter through the advisory panels and decision makers the program supports.
“We’re getting the word out about artificial intelligence in so many ways, shapes and forms and different platforms,” she said. “And we’re going places that you haven’t seen before.”