“Land surface models, radars, and satellite platforms tend to struggle with detecting or predicting rain versus snow in conditions like these,” said Dr. Keith Jennings, a hydrologist and water resources modeling expert.
Determining if rain or snow are falling seems like a problem with a simple answer: if the temperature is above freezing then it should be raining; at or below freezing, sleet or snow should be falling. Generally, modern land surface and hydrologic models and reanalysis products, such as those from NASA’s Global Precipitation Measurement (GPM) mission’s Integrated Multi-satellitE Retrievals for GPM (IMERG), successfully use temperature and other near-surface measurements to accurately predict precipitation type across temperature ranges.
But altitude, regional location, and other factors can influence the temperature near the land surface at which snow or rain falls and throw off predictions. When air temperatures are slightly above freezing, the best conventional methods can do is to accurately identify precipitation phase only 65% of the time.
How and why rain falls versus snow matters both for immediate and future reasons. In the immediate sense, being able to accurately predict precipitation type can improve forecasts and how people respond to a storm, such as applying salt and sand to roadways to prevent ice buildup. In the long term, researchers and decision-makers need to be able to look back at historical data and understand precipitation patterns in an area to properly develop infrastructure to manage rain or handle snow and other planning.
Mountain Rain or Snow Project Enlists Volunteer Observers
To improve precipitation predictions, detection, and modeling, Jennings and other scientists started a project called Mountain Rain or Snow funded by NASA's Citizen Science for Earth Systems Program (CSESP). The project has approximately 1,700 volunteer observers around the United States who report via a phone app when it’s raining or snowing to increase validated precipitation data for their area. The original thought behind the project was that gathering more datapoints from volunteers would be a great method of increasing overall precipitation measurement accuracy.
"The beauty in this study is not just in the algorithmic advancements but in the way that it clearly demonstrates the power of citizen science observations to enable these kinds of advancements with direct observations at scales not otherwise achievable," said CSESP Program Manager Dr. Gerald (Stinger) Guala.
To synthesize the satellite data and citizen science observations, the project scientists recently experimented with using artificial intelligence (AI) to classify precipitation data after learning that the technology has greatly improved hydrology models. The team trained three AI models — an artificial neural network, random forest, and XGBoost — with historical meteorology data and citizen scientist reports to see if computers could learn to make the correct guess about whether it was raining or snowing.
The data consisted of 38,500 Mountain Rain or Snow citizen scientist observations and 17.8 million synoptic meteorology reports detailing air temperature, wet bulb temperature, dew point temperature, relative humidity, and in some cases, surface pressure. The results of the experiment were recently published in the journal Nature Communications: Machine learning shows a limit to rain-snow partitioning accuracy when using near-surface meteorology.
Jennings and his colleagues found that when the AI systems churned on the data, the best they could manage was just a 0.6% overall improvement in accuracy in predicting precipitation type compared to conventional computer models.
“We were shocked to see that if you’re using any sort of meteorological measurement from the land surface you have the same exact problems whether you're using a traditional or machine learning model,” said Jennings. “We fed reliable data into the models using best practices and we didn't have any data leakage; it was just about as rigorous as you can get.”
According to Jennings, the crux of the problem appears to be that temperature, humidity, dew point, and other measurements can all be very similar yet produce different results in terms of rain or snow, fooling AI and conventional computer models. It could be that AI models may need more kinds of data than just near-surface measurements to more accurately identify precipitation type.
One possibility might be that the AI and other models need additional data on the physics of what the precipitation encounters throughout the atmosphere to determine its final form when it reaches the ground. To see if that's the case, the Mountain Rain or Snow team is looking into incorporating data from atmospheric air column measurements to see they improve accuracy.
Jennings and the team are also exploring creating a predictive model that plugs in citizen science data to understand how important volunteer observations made within miles of a storm might be in predicting the form of precipitation falling.
“When we started this project five years ago, we thought that by now there would be some new or novel numerical or AI method of predicting rain versus snow versus mixed precipitation that land surface or hydrologic models could use,” said Jennings. “This isn’t yet the case and this study shows how important the volunteers still are for recording and reporting these phenomena.”
Whatever the final answer to predicting precipitation type is, the immediate solution is clear: Direct observations from citizen scientists are still the best way to determine whether it is rain or snow falling from a storm, and when paired with near-surface and other measurements, provide a complete picture of the weather for current and historical analysis.
“There's nothing as good as volunteer scientists are at reporting precipitation type,” said Jennings. “We can’t replace them.”
Citizen Science at NASA
NASA’s Citizen Science projects invite the public (no U.S. citizenship required) to join collaborative research efforts in Earth science and other fields. Through these collaborations, citizen scientists help make important scientific discoveries. Subscribe to the Do NASA Science Newsletter to receive monthly updates and hear about new projects by sending an email to Do NASA Science Newsletter with “Subscribe” in the subject line.
To participate in the Mountain Rain or Snow project as a citizen scientist, visit the project signup page.
