Can AI Tell Rain From Snow? USU Researchers Find Key Limitations

It may seem that a forecast that accurately predicts whether precipitation will fall as rain or snow isn’t very important unless your plans include skiing, snowboarding or driving through a canyon.

In reality, when scientists develop forecasts or provide data that is used in making decisions about everything from agriculture to air traffic control, predicting when precipitation is either rain or snow is vital.

Forecasts and climate science rely on creating models from a mind-boggling amount of data. It’s the sort of “big data” problem that artificial intelligence tools should be great at solving, but a team of researchers found there are important limitations to what AI can achieve when it works mainly with meteorological data from sensors near the Earth’s surface.

The study, published in Nature Communications and led by the University of Vermont and with substantial contributions from Utah State University climate scientist Wei Zhang, evaluates traditional ways of knowing what phase of precipitation is falling at a particular location: liquid rain or solid snow.

The research team ultimately found that near-freezing temperatures create a limitation in knowing whether precipitation fell as rain or snow. The findings have critical implications for weather forecasting, climate research and developing mathematical models of the water cycle that inform decisions about managing water resources and predicting floods and drought. Those and other data-driven decisions have impacts on farming decisions, transportation safety, infrastructure management, and building codes.

"Snow has major impacts on water resources, food security, recreation, municipal and regional planning, and transportation safety," Zhang said. "For Utah and some other western states, over 95% of water resources come from snow."

Keith Jennings, director of research at the University of Vermont Water Research Institute and lead author of the study, explained in his Behind the Paper article that one reason models can’t consistently and reliably forecast rain or snow is that most precipitation data in the U.S. comes from sensors in low-lying valley areas and there are important differences between those spots and more mountainous areas.

This gap in accurately knowing where rain or snow is falling is especially important in mountain regions such as Utah’s Wasatch and Uinta Mountain ranges, where precipitation phase plays a crucial role in snowpack development. Rain on snow can accelerate snowmelt, increasing the risk of flooding, while snowfall contributes to long-term water storage.

"Like many other critical earth system variables, there are few direct observations of precipitation phase," Jennings wrote. "This means we often do not know whether it is raining or snowing in a given location, a challenge that is particularly acute in mountain regions at air temperatures near freezing."

Teaching Computers to Recognize Rain and Snow

Historically, scientists and forecasters have relied on mathematical techniques to estimate the precipitation phase, using variables like air temperature, humidity and pressure.

However, these models perform well only in extremely warm or cold conditions. In temperatures near freezing, traditional methods struggle due to the meteorological similarity between rain and snow.

To address this challenge, the researchers incorporated machine learning techniques to train computers to recognize patterns and make decisions about whether the variables traditionally used in forecasting accurately recognized rain or snow.

"On the surface, machine learning seemed well suited to this problem," Jennings said. "We had two large datasets — one from crowdsourced visual observations and another from weather reports. We expected machine learning models to improve accuracy by identifying complex patterns. However, the improvement was marginal. The best-performing machine learning model only improved accuracy by 0.6% compared to the best traditional method."

Zhang, an assistant professor in USU’s Department of Plants, Soils and Climate, who has studied artificial intelligence applications in climate science since 2010, emphasized that the study highlights the limitations of AI when using only surface meteorological data.

"AI models still struggle between 0 degrees C and 4 degrees C because rain and snow share nearly identical meteorological conditions at these temperatures," Zhang said. "We need more diverse data and physical variables to improve predictions. That’s why my student, Cody Ratterman, has been working with Apogee Instruments Inc. in Logan, to develop a sensor that can more accurately measure rain and snow."

The study highlights the value of citizen science and firsthand weather observations. To gather real-world data, the research team created the Mountain Rain or Snow Project, a NASA-funded initiative that enlists volunteers to report whether precipitation is falling as rain or snow, all through a smartphone app.

"No matter how much data we gathered — nearly 100,000 observations and counting — we kept running into the same problem: surface meteorological measurements alone cannot reliably distinguish rain and snow at near-freezing temperatures," Jennings said.

The Mountain Rain or Snow Project includes locations in Utah’s Great Basin and mountain ranges and opportunities to report from other areas as well. More information and participation details are available at rainorsnow.org.

"Observations are the most important source for advancing science," Zhang added. "USU is making progress in understanding precipitation phase, but we need data from multiple sources to continue improving our models. Citizen science plays a crucial role in that effort."