Smoothing Rough Waters: USU Data Scientists Use Deep Learning to Identify River Rapids

A chance encounter at a Utah State University Ecology Center seminar yielded a research collaboration that resulted in a comprehensive, year-long AI learning project for undergrads and grad students; a continental-scale, publicly available river image dataset, a peer-reviewed paper in a top journal and a presentation for USU’s upcoming 2026 Spring Runoff Conference.

“Talking at the seminar with a group of USU, National Park Service and U.S. Geological Survey researchers, including hydrologist Christy Leonard Stegman, we began to develop a real-world challenge, which was later presented to five USU undergraduates enrolled in an applied research in machine learning and AI class,” says USU statistician Brennan Bean. “With mentorship from two doctoral students in the same class, the scholars would tackle the problem as a hands-on semester project.”

He says the challenge was to determine whether or not AI tools could be used to identify certain kinds of rapids in satellite images of rivers.

“These rapids are important because of ongoing work by the USGS to use rapids to infer river flow,” says Bean, associate professor in USU’s Department of Mathematics and Statistics. “With the information, water managers could potentially infer river flow remotely in locations lacking physical streamgages.”

The class project began as a small-scale challenge to collect about 3,000 images and attempt to construct machine learning models, he says.

“I never dreamed it would develop into a sophisticated, year-long machine learning project that could be shared with river managers and the scientific community.”

The project team included Bean; Leonard Stegman, adjunct assistant professor in USU’s Department of Watershed Sciences, and Julie Bahr, both of the National Park Service; along with U.S. Geological Survey hydrologist Carl Legleiter and USU Mathematics and Statistics faculty member Kevin Moon, director of the USU Data Science and AI Center. But the professionals concede it was the students who drove the project, which is detailed in a paper published Jan. 22 in the journal Remote Sensing.

Lead author is USU undergraduate researcher Nicholas Brimhall. Co-authors, in addition to Bean, Moon, Leonard Stegman, Bahr and Legleiter, include current and former USU students:

• Kelvyn Bladen, doctoral student.
• Thomas Kerby, Ph.D.’ 25, assistant professor, Brigham Young University.
• Cameron Swapp, undergraduate.
• Hannah Fluckiger, undergraduate.
• Makenna Roberts, BS’ 25, master’s student, Duke University.
• Kaden Hart, undergraduate.

“Our students approached this project with a lot of energy and creativity, starting by exploring how to automatically scrape images from Google Earth and collecting more than 280,000 images,” Bean says.

From there, the students trained neural networks in AI to isolate rivers within a satellite image and predict the presence or absence of rapids.

“The class refined an image segmentation model that could isolate rivers in an image, along with a neural network to identify rapids in those images with fairly high accuracy,” Bean says. “The resulting dataset provides a framework to support a range of future hydrologic applications, including discharge estimation, habitat assessment, resource management and recreation planning.”

Moon says, to the team’s knowledge, no one has developed a river image dataset of this scale.

“The images span the continental U.S. and Alaska,” he says. “This is one of the first to specifically look at rapids.”

Bean says the undergrads and their grad student mentors exceeded expectations for the course.

“I thought the students did a great job of adapting to the realities of a real data analysis,” he says. “When something didn’t turn out as expected, they would ask themselves, ‘What’s not working and what do we do?’ We didn’t set up these challenges for them — we, as faculty, had no idea how things would go. The challenges they faced and the solutions they came up were really organic, and that made the learning process fun.”

Moon says the class developed into an effective experiential learning project for both the undergraduates and the graduate students.

“At the undergraduate level, the students identified the problem and developed the solutions,” he says. At the graduate level, the doctoral students did a great job of mentoring the undergraduates, including teaching students machine learning tools and providing project guidance. Having actual NPS and USGS agency personnel as advisors was also significant. Students confronted a real-world challenge with an unknown outcome.”

Bean says the class placed him in role of student rather than teacher.

“We didn’t teach the students: ‘Here’s the problem, go perform this solution,’” he says. “Instead, we asked questions and asked students to defend their findings. Had we scripted this class it wouldn’t have been nearly as beneficial, meaningful and exciting for the students.”

Statistics doctoral student Kelvyn Bladen will present the team’s research during USU’s 2026 Spring Runoff Conference, March 24-25, at the Cache County Event Center in Logan.