Climate-driven change in the Gulf of Maine is raising new threats that “red tides” will become more frequent and prolonged. But at the same time, powerful new data collection techniques and artificial intelligence are providing more precise ways to predict where and when toxic algae will bloom. One of those new machine learning prediction models has been developed by a former intern at Bigelow Labs in East Boothbay.
In a busy shed on a Portland wharf, workers for Bangs Island Mussels sort and clean shellfish hauled from Casco Bay that morning. Wholesaler George Parr has come to pay a visit.
“I wholesale to restaurants around town, and if there’s a lot of mackerel or scallops, I’ll ship into Massachusetts,” he said.
But business grinds to a halt, he said, when blooms of toxic algae suddenly emerge in the bay – causing the dreaded “red tide.”
Toxins can build in filter feeders to levels that would cause “Paralytic Shellfish Poisoning” in human consumers. State regulators shut down shellfish harvests long before danger grows acute. But when a red tide swept into Casco Bay last summer, Bangs Island’s harvest was shut down for a full 11 weeks.
“So when the restaurants can’t get Bangs Island they’re like ‘Why can’t we get Bangs Island?’ It was really bad this summer. And nobody was happy.”
As Parr noted, businesses of any kind hate unpredictability. And being able to forecast the onset or departure of a red tide has been a challenge — although that’s changing with the help of a type of artificial intelligence called machine learning.
“We’re coming up with forecasts on a weekly basis for each site. For me that’s really exciting. That’s what machine learning is bringing to the table,” said Izzi Grasso, a recent Southern Maine Community College student who is now seeking a mathematics degree at Clarkson University.
Last summer, Grasso interned at the Bigelow Laboratory for Ocean Sciences in East Boothbay. That’s where she helped to lead a successful project to use cutting-edge “neural network” technology that is modeled on the human brain to better predict toxic algal blooms in the Gulf of Maine.
“Really high accuracy. Right around 95 percent or higher, depending on the way you split it up,” she said.
Here’s how the project worked: the researchers accessed a massive amount of data on toxic algal blooms from the state Department of Marine Resources. The data sets detailed the emergence and retreat of varied toxins in shellfish samples from up and down the coast over a three-year period.
The researchers trained the neural network to learn from those thousands of data points. Then it created its own algorithms to describe the complex phenomena that can lead up to a red tide.
“Then we tested how it would actually predict on unknown data,” Grasso said.
Grasso said they fed in data from early 2017 — which the network had never seen — and asked it to forecast when and where the toxins would emerge.
“I wasn’t surprised that it worked, but I was surprised how well it worked, the level of accuracy and the resolution on specific sites and specific weeks,” said Nick Record, Bigelow’s big data specialist.
Record said that the network’s accuracy, particularly in the week before a bloom emerges, could be a game-changer for the shellfish industry and its regulators.
Once it’s ready, that is.
“Basically it works so well that I need to break it as many ways as I can before I really trust it.”
[iframe url=”https://www.documentcloud.org/documents/6629949-The-Hunt-for-Red-Tides-Deep-Learning-Algorithm.html” responsive=true height=”550″ width=”400″]
Still, the work has already been published in a peer-reviewed journal, and it is getting attention from the scientific community. Don Anderson is a senior scientist at the Woods Hole Oceanographic Institution who is working to expand the scope of data-gathering efforts in the Gulf.
“The world is changing with respect to the threat of algal blooms in the Gulf of Maine,” he said. “We used to worry about only one toxic species and human poisoning syndrome. Now we have at least three.”
Anderson noted, though, that machine-learning networks are only as good as the data that is fed into them. The Bigelow network, for instance, might not be able to account for singular oceanographic events that are short and sudden or that haven’t been captured in previous data-sets – such as a surge of toxic cells that his instruments detected off Cutler last summer.
“With an instrument moored in the water there, and we in fact got that information, called up the state of Maine and said ‘you’ve got to be careful, there’s a lot of cells moving down there,’ and they actually had a meeting, they implemented a provisional closure just on the basis of that information, which was ultimately confirmed with toxicity once they measured it,” Anderson said.
Anderson said that novel modeling techniques such as Bigelow’s, coupled with an expanded number of high-tech monitoring stations, like Woods Hole is pioneering in the Gulf, could make forecasting toxic blooms as simple as checking the weather report.
“That situational awareness is what everyone’s striving to produce in the field of monitoring and management of these toxic algal blooms, and it’s going to take a variety of tools, and this type of artificial intelligence is a valuable part of that arsenal.” Back at the Portland wharf, shellfish dealer George Parr said the research sounds pretty promising.
“Forewarned is fore-armed,” Parr said. “If they can figure out how to neutralize the red tide, that’d be even better.”
Bigelow scientists and former intern Izzi Grasso are working now to look “under the hood” of the neural network, to figure out how, exactly, it arrives at its conclusions. They say that could provide clues about how not only to predict toxic algal blooms, but even how to prevent them.
This article appears through a media partnership with Maine Public.