Cooler but still well above normal this week

MPR News helps you turn down the noise and build shared understanding. Turn up your support for this public resource and keep trusted journalism accessible to all. 

Turn Up Your Support

It's spring flooding season in Minnesota. And while there's a low risk for serious floods this year, hydrologists are still closely watching more than 300 river gauges across Minnesota that relay water levels and how fast the rivers are flowing. 

However, the National Weather Service only issues flood forecasts for about a third of those locations. And many communities have no river gauges, which means there are many places all around the state without a localized flood forecast.

Those forecasts are critical, answering questions such as: How high to build walls of sandbags to protect homes? Which neighborhoods might need to be evacuated? Or even simply, where is a safe place to go fishing? 

Take Le Sueur, a small community of about 4,000 people along the Minnesota River, southwest of the Twin Cities. There are gauges with flood forecasts upstream in Mankato, and downstream in Henderson — but nothing specifically for Le Sueur. 

City officials do everything they can to make up for that gap. They fly a drone to monitor river conditions. They closely watch the forecasts up and downstream. They also rely on extensive local knowledge of past flooding events to know when key resources are at risk. 

“We have some visual high-water marks in reference to where our infrastructure is, how it's going to impact transportation, our business community and private property,” explained city manager Joe Roby. 

But he concedes it’s a very unscientific process. 

“If we compare it to weather forecasting, we don't rely upon a weather radar that's accurate every 45 miles when there's a risk of severe weather,” Roby said. 

That’s why he’s very excited about the potential of a new flood forecasting model developed at the University of Minnesota using artificial intelligence, that could bring localized flood forecasts to places like Le Sueur that don’t currently have them. 

“These artificial intelligence models can give you a better model accuracy score than physics-based models that are specifically calibrated to river basins,” said Zac McEachran, a research hydrologist at the University of Minnesota Climate Adaptation Partnership who developed the model. 

That was a “crazy” thing to see, admits McEachran, who before coming to the U worked as a hydrologist and flood forecaster for the National Weather Service. 

“I've devoted my life to this stuff, and now this text generation model is doing amazing,” he said. 

Still, humans maintain a critically important role, he says. Currently, expert forecasters use real-time field observations to modify physics-based computer models. They may tweak the levels of soil moisture or snowpack levels in the models, for example. That produces very accurate forecasts. In fact, they're better than AI models. 

But expert human forecasting is labor-intensive and hard to scale up. And with climate change already increasing the risk of severe flooding in many areas of the state, the demand for more forecasts is only going to grow. 

McEachran's approach involves combining artificial intelligence with hydrology and river science — kind of a third way he calls “knowledge guided machine learning.”

The research, published recently in two peer-reviewed papers, shows very promising results. McEachran believes his model could improve the accuracy of forecasts produced by humans, as well as provide forecasts to places that currently don’t have them. 

“So it's really an excellent opportunity to hopefully enhance the hydrological warning enterprise as it currently exists, but also expand it really rapidly.”

The new model could be especially helpful in places like the North Shore of Lake Superior, where rivers that drain the steep hillside can rise rapidly and violently, but then quickly recede. There's currently only one active forecast point there. 

Lack of forecast points can make life difficult when it rains for fly fishers and guides who target steelhead, brook trout and other fish in North Shore streams. 

“It's a lot of phone calls between me and other guides and my clients to try to figure out whether the river is going to blow out and be at an unfishable level or not,” said Jason Swingen, a Duluth-based guide and president of the local chapter of Trout Unlimited. “Sometimes that happens and sometimes it doesn't, it is really hard to predict.”

River forecasts along the North Shore are challenging to produce, because the hydrology of the region is so complex and difficult to model using current methods. 

“There's just a lot of nuance because of the way that things can compound so quickly and become so complicated so quickly, that it is a big lift to get those official forecast points,” explained Ketzel Levens, a meteorologist who co-manages the hydrology program at the National Weather Service office in Duluth.

Levens says the Weather Service has worked hard to expand flood forecast points in the past decade, including adding a forecast at Knife River. 

Three forecasts were also added to the Rainy River Basin near the international border after 

record flooding inundated homes for months on end

 in 2022 along the shores of Rainy Lake and Kabetogama Lake. 

Those efforts paid off the following year. 

“We were working with record amounts of snow water, and we were able to accurately outlook the really high risk of spring flooding months in advance,” Levens said. 

Placing forecast points took a lot of work. Levens is still 

recruiting volunteer precipitation observers

 to report rain and snowfall, snow depth and other measurements so forecasters can better predict flooding in a very complex and remote watershed. 

McEachran’s model, Levens said, could be beneficial in reaching more places. 

“Trying to bring more services any way we can, I think, is a pretty powerful thing, especially if it's proven to be accurate.”

The U's Climate Adaptation Partnership is seeking funding to launch a statewide river forecast system based on McEachran's modeling approach. Initially, the program would add forecast capabilities at existing river gauges currently without forecasts. Researchers are working on increasing the accuracy of models for communities that don’t have gauges. 

In the long term, McEachran envisions a phone app where users can check a river forecast before they go camping or fishing, just as they might check a weather app today. 

Just as there are weather geeks today, he wants to create more hydrology enthusiasts. 

“We really would benefit from this information,” he said. “And it's kind of overdue in a way, to get this information out to people.”

Environmental News

Lifestyle

Outdoors

NewsPartners

Homepage

Dan Kraker

Author

A new model developed by a U of M researcher that blends artificial intelligence with more traditional forecasting techniques could bring flood forecasts to many parts of the state that lack them. 

There are gaps in Minnesota's flood forecast network. Could AI help?

Go Deeper.

Like this?

News you can use in your inbox