For Chesapeake Bay anglers, an algae bloom is a foreboding sight. It indicates that bad days for fishing probably lie ahead.
That’s because when the plant-like organisms die, they rain down through the water column, sucking up oxygen along the way. At lower depths, this biological process often leads to massive “dead zones,” tracts of water that are all but devoid of life.
For years, scientists in the region have been trying to improve forecasts of when and where these oxygen-starved waters are likely to occur. Now, they’re turning to artificial intelligence.
What they’ve come up with — a hybrid system that blends AI-powered analysis with a conventional computer modeling tool — shows promise, said Guangming Zheng, a remote-sensing scientist at the University of Maryland who worked on training the AI for the project.
“Overall, I think the model shows some very positive improvements,” he said.
Zheng and his colleagues have dubbed their model “HypoxAI,” a nod to the term “hypoxia,” which is the scientific name for a lack of oxygen. The conventional modeling system that Bay researchers currently use to forecast hypoxia is called the Chesapeake Bay Environmental Forecast System (CBEFS).
The effort tested how well each system could predict the timing and location of hypoxic water, using measurements gathered from 2019-2020. During the study, the researchers had the systems produce more than 10,000 hypoxia forecasts at various locations in the Bay. Because there is already a record of the ebbs and flows of dead zones for those two years, the team could check the predictions against what really happened.
HypoxAI proved to be about 16% more accurate than CBEFS alone, the study found.
Still, AI probably won’t replace traditional modeling anytime soon, said Stephanie Schollaert Uz, the project’s lead scientist and an applied sciences manager at NASA’s Goddard Space Flight Center in Greenbelt, MD. HypoxAI, after all, incorporates CBEFS data into its analysis.
“This could not have been as successful without CBEFS,” Schollaert Uz said. “But what it’s showing is a well-tuned AI model has skill and can complement a physical model like CBEFS.”
The study was published in the summer of 2024 in the journal Artificial Intelligence for the Earth Systems. Zheng presented the findings this August before a scientific panel representing the Chesapeake Bay Program, the multistate and federal partnership that oversees the estuary’s cleanup.
Developed by the Virginia Institute of Marine Science (VIMS), CBEFS uses a computer model to forecast not only dead zones but also salinity, water temperature and acid concentrations in the Bay at various depths.
The system issues forecasts over three separate timescales: the current day, two days into the future and a picture of trending conditions between those two points. But CBEFS is designed to follow the rules of nature, trying to predict how one force impacts another — for example, nutrient pollution and dissolved oxygen concentrations.
“There are no mechanisms in the CBEFS to artificially bring the model closer to reality,” wrote Zheng and his colleagues, “even when observations are available or when systematic biases are present.”
This isn’t AI’s first foray into environmental forecasting. Elsewhere, researchers have developed AI-based tools to predict the weather, levels of air pollution and the whereabouts of wildlife.
Where there are algae blooms, hypoxic conditions typically follow. So, the team theorized that the forecasts could be improved by incorporating data on what those blooms are doing at any given time. NASA satellites have been beaming that information back to Earth for more than 20 years with imaging systems that capture light waves invisible to the human eye.
But incorporating that imaging into a computer model hasn’t been easy. “Seeing” algae pigments from space involves filtering out a lot of “noise,” such as clouds, sea salt, human-generated emissions and gases emitted by plants, Schollaert Uz said. Organic matter in the water also can make it harder to detect algae, she added.
“What you see in the water,” Zheng said, “is misleading for a lot of algae.”
AI, though, can be trained to sort out the signals created by algae blooms from everything else. That takes a lot of computational power; the researchers in this case accomplished that feat with high-performance NASA computers. But once the training is done, the amount of computing power needed is “orders of magnitude less,” compared with conventional systems, Zheng said.
“Once AI is trained well, it’s efficient,” Schollaert Uz said.
The satellite imaging wasn’t as helpful as the researchers expected. The CBEFS model for water temperature turned out to be the best predictor for incoming hypoxia, the study found. Zheng said he hopes that Earth-imaging satellites put into orbit in more recent years, aided by modern AI algorithms, will yield better results and enable scientists to wean the forecasts off CBEFS.
For its part, VIMS is experimenting with AI as well. Its website has gone live with AI-powered models for forecasting harmful algae blooms and Vibrio, which is a suite of about a dozen water-borne bacteria species that can cause illness or even death. Two of the six co-authors on the AI-hypoxia study are affiliated with VIMS.
Dante Horemans, a VIMS researcher who wasn’t involved in the study, said he was impressed by the AI’s performance in the experiments. “With more information, you expect a better model result,” he said.
It will be critical in coming years for scientists to update their models to stay in lockstep with climate change, he added. If models can’t keep up with rising temperatures and stronger thunderstorms, their forecasts will degrade in quality over time.