Blurring the line between rain and snow: the limits of meteorological classification

A new study led by the University of Vermont (UVM) uncovers a critical challenge in accurately classifying precipitation as rain or snow using surface weather data.

Published in Nature Communications, the research evaluates the performance of both traditional precipitation phase partitioning methods and advanced machine learning models, revealing that near-freezing temperatures create an inherent limitation in distinguishing between rain and snow, restricting the accuracy of these approaches.

Accurately identifying precipitation phase is critical for weather forecasting, hydrologic modeling, and climate research, with significant implications for transportation safety, air travel, infrastructure operations, and water resources management.

This is especially true in mountain regions, where accurate distinctions between rain and snow help natural resource managers to better predict and mitigate threats and challenges. While a storm consisting of mostly snowfall may benefit ski areas and water resources, a rain-dominated event may cause devastating flooding and infrastructure damage.

Due to the scarcity of direct rain and snow observations, most of which come from airports and rarely reflect the complex weather patterns of nearby mountain regions, scientists and forecasters rely on mathematical techniques that use weather data to estimate the precipitation phase.

These precipitation phase partitioning methods—thresholds, ranges, and statistical models—utilize data such as air temperature, humidity, and pressure. However, most of these methods only perform well in either cold or warm temperatures when snowfall or rainfall, respectively, are nearly certain. At temperatures near freezing, however, all traditional methods struggle to accurately predict rain and snow due to the meteorological similarity of the two phases.

This study figured out why.

“The challenge is at those temperatures near freezing, the air and wet bulb temperature distributions of rain and snow overlap heavily” said Dr. Keith Jennings, the study’s lead researcher.  “This means the traditional partitioning methods cannot consistently separate rain from snow. What surprised us is that the machine learning models did not perform much better. Even by using more data and complex mathematics, they are still trying to tease apart the same information, and they’re seeing rain and snow with almost the exact same meteorological properties.”

To perform this study, the Director of Research at UVM’s Water Resources Institute, Dr. Keith Jennings, partnered with scientists at Lynker, the Desert Resource Institute (DRI), the

Cooperative Institute for Research in the Atmosphere, the University of Nevada Reno, and Utah State University. They mined two unique datasets of precipitation phase: nearly 40,000 crowdsourced observations across the United States from the NASA-funded Mountain Rain or Snow participatory science project and over 17 million synoptic weather reports from across the Northern Hemisphere.

The research team used these datasets to evaluate different methods for classifying precipitation as rain, snow, or mixed. These techniques included a selection of high-performing traditional methods as benchmarks (temperature thresholds and a statistical model) and three machine learning (ML) models: random forest, XGBoost, and an artificial neural network (ANN).

While the ML models provided negligible improvements over the best benchmarks, increasing accuracy by up to 0.6%, they still struggled to correctly classify precipitation in the near-freezing range (1.0°C–2.5°C) and failed to consistently identify mixed precipitation and sub-freezing rainfall events.

The study uncovered a primary obstacle: the natural overlap in meteorological conditions between rain and snow makes classification difficult when solely relying on surface weather data. Despite advancements in machine learning, there is a limit to how well precipitation phase can be predicted with only these meteorological inputs.

Dr. Jennings suggests that researchers should switch their focus from marginally improving the inherently limited rain-snow partitioning methods using surface weather data to creating new techniques that assimilate novel data sources. These alternatives include crowd-sourced observations, such as those from the Mountain Rain or Snow project, weather radars, and satellite  precipitation products.

As climate change drives more frequent rain-on-snow events and alters precipitation patterns, balancing the risks to life, property, and ecosystem function will only become more difficult. However, leveraging multi-source data integration rather than relying on surface weather data alone may offer improvements.

If you want to volunteer your weather observing skills, sign up for Mountain Rain or Snow today. To get alerts on incoming storm events that researchers are studying in New England, text NorEaster to 855-909-0798.