AI helps turn citizen photos into water-level data for UB researchers

BUFFALO, N.Y. – For the past 15 years, Christopher Lowry, PhD, has led CrowdHydrology, a University at Buffalo citizen-science project that relies on thousands of volunteers to collect water-level observations from streams and other waterways that are not routinely monitored. This information helps researchers understand how smaller streams respond to changing environmental conditions and how they affect downstream waters and ecosystems.

What began as a small effort has grown into a nationwide network of more than 8,000 citizen scientists who have submitted over 20,000 water‑level observations from more than 200 monitoring stations across 26 states.

Originally, Lowry asked the citizen scientists to text the water level shown on staff gauges, which are vertical rulers installed along waterways. But as smartphones became more common, he began receiving photos of the gauges instead of text messages with the readings. Although the photos contained the same information, they created an unexpected challenge for him.

“I was suddenly accumulating all of these photos, and each one required manual review before I could enter the water level into the database. Ensuring accuracy made the process even more time‑intensive,” said Lowry, a hydrogeologist and professor in UB’s Department of Earth Sciences who founded CrowdHydrology in May 2011.

Looking for a way to automate the photo review process, Lowry approached Abhinna Manandhar to determine whether artificial intelligence could do the work. At the time, Manandhar was a graduate student in UB’s Department of Computer Science and Engineering who had been helping with coding tasks for CrowdHydrology.

Manandhar proposed using Google's Gemini large language model instead of building a custom AI system from scratch. Together, the two designed a workflow that combines image-processing techniques with AI to automatically interpret photos submitted through CrowdHydrology.

The project highlights one of the University at Buffalo's strengths by bringing together experts from different disciplines to solve practical problems. Lowry said UB's depth in AI and computer science helped transform an idea into a working solution.

“I’m a hydrologist, not an AI expert, so having access to UB’s computer science community was essential,” Lowry said. “That’s what makes UB such a powerhouse. When I had questions about AI, I could consult directly with people who work in that field. Abhinna’s knowledge of AI made this project possible. It's the kind of collaboration that allows us to do research at UB that would be difficult to do at other institutions.”

Early in their research, Lowry and Manandhar realized that asking AI to interpret an entire photograph wasn’t effective, especially when volunteers occasionally sent images that were blurry, shadowed, or taken from several feet away. They began preprocessing each image to guide the model toward the part that mattered. By converting the photos to black and white, they made it easier for the system to distinguish the darker water from the staff gauge and pinpoint the waterline before sending the information to Google’s Gemini model.

Lowry compares the approach to searching for a friend in a crowded stadium.

“Think about it like this. You’re at a Buffalo Bills game, and you’re looking for your friend somewhere in the stadium,” he said. “You can look all around the stadium and spend a lot of time scanning. But if I tell you that your friend is in Section 8, you can focus on that one section and find them. That’s what we’re doing with AI. We’re giving it a little help by saying to look at this area and tell us what’s there.”

That simple guidance made a big difference. When they sent raw images directly to the AI model, some water level readings were off by about a foot, which is a significant error for staff gauges since they are only about three feet tall.

After preprocessing the images to focus the model’s attention on the relevant area, measurement errors dropped substantially. The percentage of images the system could not interpret fell from 17% to 2%, and monitoring station IDs were correctly identified about 98% of the time.

“Our findings show that how information is presented to an AI model can be just as important as the model itself,” Lowry said. “A relatively simple preprocessing step made a remarkable difference in performance.”

Lowry and Manandhar wanted to ensure that AI supports human judgment without replacing it. After the system reviews a submitted photo, the citizen scientists receive the AI‑generated water level and can adjust it if it’s incorrect. That extra step keeps the data reliable while reducing the amount of manual work involved.

“Moving forward with artificial intelligence, having humans in the loop is a critical factor,” Lowry said. “We need humans in the loop to make sure we’re producing something that is accurate and useful.”

Although the project focused on hydrology, Lowry sees broad potential for the approach.

Researchers in other fields could adopt a similar strategy, combining existing AI models with relatively simple image-processing techniques to automate time-consuming tasks rather than building specialized AI systems from scratch.

“We've shown a methodology that allows someone who is not an expert in AI or computer vision but is still a scientist to use these tools effectively with a fairly low barrier of entry,” Lowry said.

Lowry and Manandhar’s research was recently published in the journal Hydrology.