Report-back:
The search for the perfect “recipe” for rain-snow partitioning models
The search for the perfect “recipe” for rain-snow partitioning models
January 2023 - Led by Keith Jennings
Written by Sonia Nieminen
How your observations help assess the performance of various rain-snow partitioning models.
In the same way that a baker might try multiple recipes to bake the perfect loaf of bread, scientists often consider how different models perform for analyzing data based on how the data that is used, how the models are constructed, and how the end product shapes up. Instead of flour and yeast, models have data as “ingredients”, and the “recipe” involves analyzing the data in a way that best represents the real world.
Just like the quest for the perfect loaf, many hydrologists are seeking the best model to predict the temperature at which rain transitions to snow in a way that is specific to region and storm conditions. As we explain here, the rain-snow transition has big implications for water budgets, safety, and other societal impacts.
The wealth of Mountain Rain or Snow observations that you have submitted in current and previous seasons serves as an excellent standard against which to compare the performance of various models.
Here is an overview of what we did.
We compared 2,248 of your ground-based observations from the northern Sierra Nevada from fall 2019 to early 2021 to model-based predictions of precipitation phase. Understanding variation across different models is like comparing the taste, texture, and aroma of different bread recipes. For each observation of rain, snow, or mixed precipitation, we matched certain types of data - like temperature and elevation - entered this information in each model, and then checked how well the results matched what was actually happening on the ground.
In total, we compared fourteen models, which fall into a few different categories: temperature threshold, temperature range, and a binary logistic regression model (an equation that uses a few different variables). The “ingredients'' used by the models included air temperature, wet bulb temperature, and dewpoint, which are three different ways of measuring how much energy is in the air. Only dewpoint and wet bulb temperature take humidity into account. Temperature thresholds used a single temperature for prediction (above, rain; below, snow), while temperature range models also predict mixed precipitation between certain temperatures.
With these “recipes” for rain-snow partitioning, we determined the variability, success rates, and bias for each model (bias is how the phase’s predicted frequency compared to observations of that phase). In addition, we analyzed ground-based radar that looks at falling precipitation to find where its reflectivity changes. Finally, we utilized advanced precipitation measurement tools that belong to the NASA Global Precipitation Monitoring (GPM) mission.
How the models measured up to your observations.
Jennings, K.S., Arienzo, M.A., Collins, M., Hatchett, B., Nolin, A.W., and Aggett, G.R. (In review). Crowdsourced Data Highlight Precipitation Phase Partitioning Variability in Rain-Snow Transition Zone. Earth and Space Science.
The partitioning methods varied in how well they matched the reports of precipitation phase on the ground. The average snow frequency for all fourteen models was 60.7% (observers reported snow 64% of the time) with a standard deviation of 18% (the minimum snow prediction frequency was 23.5%, and the maximum was 84.7%). This represents a lot of variability – the models did not match each other well in terms of how often they predicted snow for the same variables.
The figure to the left shows the success rates for each partitioning model (abbreviated on the y-axis) in context of the air temperature on the ground (x-axis). The right panel of this figure shows the success rates when mixed precipitation was classified as rain. Darker colors correspond to poorer success rates. The success rates drop at and just above freezing temperatures (0.0°C to 10°C).
The figure to the right compares the performance of two partitioning techniques: IMERG, a product of the NASA GPM mission (blue) and 0.5°C wet bulb temperature threshold (mint green). The actual snowfall frequency as computed from observations is shown in a black dotted line. Notice that the success rates of IMERG and wet bulb (solid blue and green lines) decreased for air temperatures between 0-10°C, and the snowfall frequencies of both (dashed blue and green lines) were lower than the observed snowfall frequency within this range as well, hence underpredicting snow.
We also found that the method with the highest success rate (71%) uses the dewpoint temperature of 0.0°C as the threshold for predicting snow. Overall, the fourteen models tended to overpredict rain. Interestingly, the analyzed NASA GPM product that predicts precipitation phase from meteorological information also overpredicted rain, forecasting liquid precipitation with a bias of +17.9%. This research also showed that models that consider humidity (like dewpoint and wet bulb temperature) are more effective than using the air temperature alone.
Jennings, K.S., Arienzo, M.A., Collins, M., Hatchett, B., Nolin, A.W., and Aggett, G.R. (In review). Crowdsourced Data Highlight Precipitation Phase Partitioning Variability in Rain-Snow Transition Zone. Earth and Space Science.
Why does it matter that there is so much variability in rain-snow partitioning models?
The fact that these models’ precipitation phase partitioning results are variable highlights the inadequacies in our current rain-snow partitioning methods. In areas like the Sierra Nevada, recognizing the tendency for models to predict rain when it is actually going to snow is crucial for assessing the risks posed by storms. In addition, analyzing model performance helps scientists learn about how the region’s unique meteorological characteristics impact the results of modeling techniques.
In order to further improve our methods of determining whether rain or snow will fall, more observations are needed. Each observation of precipitation phase helps us to more accurately assess the performance of rain-snow partitioning. Crucially, observations with greater spatial and temporal variety are needed so that scientists can perform similar analyses of rain-snow partitioning techniques for other regions. This season of Mountain Rain or Snow will help us take one step closer to doing so!
This article is a summary of recent research conducted by our team of Mountain Rain or Snow scientists. The background information, data, and figures are from:
Jennings, K.S., Arienzo, M.A., Collins, M., Hatchett, B.J., Nolin, A.W., and Aggett, G.R. (2023). Crowdsourced data highlight precipitation phase partitioning variability in rain-snow transition zone. Earth and Space Science, 10(3). https://doi.org/10.1029/2022EA002714