November 2025 - Research by Keith Jennings, written by Luisa Ortega and Meghan Collins
Near the freezing point, it can often be difficult for weather monitoring technologies to track whether precipitation falls as rain or snow, especially in arid, mountain regions. If you have been involved in this project for long, you know that it is not uncommon to see snow falling at ground level at 34°F, 35°F, or even 36°F!
When it comes to forecasting the weather or managing regional water budgets, direct weather monitoring is just one piece of the puzzle. Scientists also rely on models to infer whether it is raining or snowing in places where weather stations or satellites cannot make direct observations. This is needed, for example, between weather stations that are 10 miles apart.
With the goal of improving how models infer rain vs. snow, the research team at Mountain Rain or Snow led by Keith Jennings, PhD wanted to know if machine learning can help differentiate precipitation type near freezing. This work was published in Nature Communications.
Machine learning is a process in which computer systems analyze patterns in data and improve over time. The goal of this paper is to understand if machine learning (ML) can improve rain-snow partitioning methods as compared to traditional or benchmark* methods. The crowdsourced dataset of rain, snow, and mixed precipitation observations offers a unique opportunity to test ML methods in using observations from climatologically diverse regions.
Interestingly, there are different approaches to machine learning, including random forest models, artificial neural networks, and XG Boost. Random forest and XG Boost are both “tree-based” ML techniques that combine many decision trees, each providing a prediction of the result. The model then calculates an average to make a more accurate and reliable result. Artificial neural networks (ANN) are programs that learn from data by finding patterns and using those patterns to make predictions – much like the human brain!
With these methods, our team tested whether advanced machine learning techniques could improve rain-snow predictions beyond traditional methods. We analyzed over 38,000 crowdsourced observations and nearly 18 million weather station reports, comparing standard forecasting approaches with the three machine learning models mentioned above.
Figure 1. Maps of observation locations for (a) Mountain Rain or Snow observations (crowdsourced dataset) and (b) weather station reports used in this study. The color of each hexagon corresponds to the number of observations over the study period. Note: the color scale differs between the two maps.
Overall, this study showed that machine learning hardly improved model accuracy or bias relative to traditional modeling methods, increasing accuracy by just 0.6% and reducing prediction biases by up to 4.7%.
All three machine learning methods (random forest, ANN, and XG Boost) slightly outperformed the benchmark methods. Also, increasing model complexity by adding more layers or ensembles to the model did not meaningfully improve accuracy for precipitation predictions.
Figure 2. Matrices showing the percentage of observations correctly and incorrectly predicted by the machine learning methods (from left to right: ANN, Random forest, and XG Boost).
To see the percentage of observations correctly predicted by machine learning methods, find the cell where the observed phase and predicted phase match. For example, if you line up mixed and mixed in the random forest panel, you see that 9.3% of observations were correctly predicted. To see the percentage of observations incorrectly predicted, line up an observed phase with a non-matching predicted phase. If you move to the cell where the observed phase is mixed and the predicted phase is rain in the random forest panel, then 32.9% of mixed precipitation was incorrectly predicted as rain by the random forest model.
At temperatures closer to freezing, performance declines for all ML and traditional methods, highlighting continued challenges in predicting precipitation phase in this important temperature range. The methods also failed to identify mixed precipitation or freezing rain.
Importantly, Dr. Jennings found that there is a sizable overlap in temperature ranges at which rain and snow occur: crowdsourced data show a 33.7% overlap in the air temperature distributions, and weather station data show a 16% overlap. There was a significant negative relationship between overlap and prediction accuracy—the more the temperature ranges overlap, the worse the predictions become.
Figure 3. Plots showing how the temperature ranges for rain vs snow for both crowdsourced and weather station datasets overlap significantly, especially near freezing. Overlap means that predictions of rain vs snow based only on near-surface temperature are prone to errors, especially in the near-freezing zone.
The take-home message for scientists from this work is that there is a limit to the improvements in prediction of rain vs. snow that use only near-surface metrology methods. In other words, the existing tools and datasets at scientists’ disposal can only get the field so far – novel approaches may be needed such as citizen science data, weather radar, satellite measurements, or upper-atmosphere conditions.
Jennings, K. S., Collins, M., Hatchett, B. J., Heggli, A., Hur, N., Tonino, S., Nolin, A. W., Yu, G., Zhang, W., & Arienzo, M. M. (2025). Machine learning shows a limit to rain-snow partitioning accuracy when using near-surface meteorology. Nature Communications, 16(1), 2929. https://doi.org/10.1038/s41467-025-58234-2
Benchmark methods are a set of formulas that use direct weather observations—like air temperature, wet bulb, or dew point—to decide whether precipitation will fall as rain or snow. These formulas are based on a type of math model called logistic regression, which is designed to handle yes or no questions (i.e., is precipitation falling as rain? Or, is precipitation falling as snow?). These traditional approaches have limitations, making it especially difficult to pinpoint phase changes in the 32°-39°F (0-4°C) range with air temperature alone.
The table compares different methods used in this study to decide whether precipitation is rain or snow, based only on near-surface weather data (like temperature and humidity). Each method was tested on two datasets — the crowdsourced data and the weather station reports — and then scored by how accurate it was.
For example, the benchmark method 'Ta' uses only air temperature. Following a simple rule: if the air temperature is above 0 °C, it’s rain; if it’s below 0 °C, it’s snow.
