Preprocessing cuts uninterpretable water-gauge photos from 17% to 2%

A volunteer sends a photograph of a vertical ruler beside a stream. For Christopher Lowry, the founder of the University at Buffalo’s CrowdHydrology project, that image once meant another manual check before the water level could enter the database. Now, an artificial-intelligence workflow helps process it: the share of photos the system could not interpret fell from 17% to 2%.

CrowdHydrology began in May 2011 and has grown into a network of more than 8,000 citizen scientists. They have submitted more than 20,000 water-level observations from more than 200 monitoring stations across 26 states. The measurements fill gaps around smaller streams that are not routinely monitored, helping researchers study how those waterways respond to changing environmental conditions and affect downstream ecosystems.

The challenge arrived as volunteers began sending photos instead of text messages. Lowry and Abhinna Manandhar, then a graduate student in UB’s computer science department, found that asking Google’s Gemini large language model to interpret a whole image was unreliable when pictures were blurry, shadowed or taken from several feet away. They first processed each image to correct uneven light, convert it to black and white and isolate the darker waterline from the gauge, then passed the focused information to Gemini. Monitoring station IDs were correctly identified about 98% of the time.

The approach matters because the gauge itself is only about 3 feet (0.9 meters) tall. With unprocessed images, some readings missed the mark by about a foot (0.3 meters)—large enough to undermine the observation. The study, published in Hydrology, shows how a relatively simple image-processing step can make a general-purpose AI model more useful for a narrowly defined scientific task.

The extra review step keeps the data reliable while reducing the amount of manual work involved. Volunteers receive the AI-generated level and can adjust it when it is incorrect. The system supports human judgment rather than replacing it. 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.