“Is there enough water? What’s happening with climate change? How do I operate my hydro plant? In a world of water problems, there are problems that we need a good model to solve,” says Colin Gleason, a hydrologist in the Riccio College of Engineering at the University of Massachusetts Amherst. But the models are only as good as the math that they’re based on, and, short of measuring every river in the world, there is no way to check this on a global scale.

Image

Until SWOT.

Now, a new study, led by Gleason, uses data from the NASA-launched Surface Water and Ocean Topography (SWOT) satellite to determine where state-of-the-art, machine learning computer model river estimates are accurate and where they are not.

“In this study, we learned where the models are right and where the models are wrong. In turn, anywhere they’re wrong means their climate change predictions are wrong or their irrigation forecasts are wrong,” he says.

They found that hardest to model are rivers that are dammed, rivers in arid climates (especially in highly populated areas), and Arctic rivers. While less than 10% fall into the category of “serious error,” these reaches are among the most sensitive and important for water resources. This poses a challenge when one of the main uses of river modeling is to create accurate predictions for developing hydropower, water use planning (for people and agriculture), and preparing for climate change in regions with limited or no ground data.

By comparison, models are good at estimating “normal” rivers—not glacially fed, single thread, no estuaries and no dams. “It turns out it’s only 11% of the rivers on Earth that fit that definition of ‘regular,’ so it also kind of challenges our understanding of what a river is,” Gleason said. “A weird river is the norm. A damned river is the norm. A multichannel, complicated planform river is the norm, not the exception.”

The accuracy of modeling dammed rivers has been a known issue, but this is the first time the scope of the challenge has been definitively demonstrated. Gleason explains the challenge using as an example the Connecticut River, which contains a pumped hydro reservoir that can change the river’s depth by more than a meter per day, beyond fluctuations of the natural water cycle.

In order to accurately predict the amount of water in the Connecticut River, a model would have to know that the pumped hydropower plant exists, the price of electricity, and the cost threshold used by the electric company’s pumping operation plan.

Gleason also highlighted that models struggle in arid areas like Australia, Central Asia, the southwestern U.S. and Mexico. This is likely because these locations rely on groundwater, which affects the rivers, but in non-obvious ways and with a delayed response to pumping.

The Arctic is also poorly modeled, largely because the datasets to build machine learning models are lacking. He points to Iceland as an example, which has a higher elevation than the rest of the Arctic, as well as an ocean current.

“Iceland is highly geologically active, it’s full of snow and ice, it has glaciers, and it does not have a lot of data. So, if you’re a machine learning model, you’d ask yourself: what other places on the planet are like Iceland that I can learn from? Just parts of New Zealand. That’s pretty much it. Machine learning does really well at replicating patterns it can find in the data, but if there’s no data, it can’t find any patterns,” he said.

For Gleason, this study, published in AGU’s Geophysical Research Letters, points to where researchers need to focus their attention. “What this work shows us is ‘where are the areas that global hydrology can’t give you a good starting point for that information?’”

He also sees this research as justification for where models could be jettisoned altogether in favor of simply using SWOT data. “In many ways, we’re thinking of SWOT kind of like an early microscope,” he says. “We want to return to that way of thinking about rivers: Let’s measure them first, and let’s trust the measurements rather than the models.”

This research was covered in a recent story in Science Magazine on SWOT.