If we had known a year ago that this winter would be so dry, would we have conserved water more aggressively last summer? Would ski resorts have installed more snowmaking equipment? Would farmers buy different seeds to plant this spring?
These are among the tantalizing questions raised by a team of government and university scientists, who believe they have developed a tool to predict mountain snowpack in the West up to eight months in advance – long before the first winter snowflake has fallen.
The tool, a powerful computer model, is described in a new study recently published in the Proceedings of the National Academy of Sciences. It is still experimental, but it seems capable at this stage of giving a thumbs-up or -down signal about whether March 1 snowpack will be heavy. And it can do so at the scale of a particular mountain range, offering some indication about potential spring runoff for individual watersheds.
The one exception is the southern Sierra Nevada range in California, which presents unique forecasting challenges thanks to its extreme topography.
To understand this new forecasting tool a bit better, along with its potential to change water management in the West, Water Deeply talked to the study’s lead author, Sarah Kapnick, a research physical scientist at National Oceanic and Atmospheric Administration’s Geophysical Fluid Dynamics Laboratory at Princeton University in New Jersey.
Water Deeply: This seems like science fiction. How did you do it?
Sarah Kapnick: The prediction system we developed is a set of three global climate models. One has atmospheric and land surface resolution at 200km [124 miles], 50km [31 miles] and 25km [15 miles]. We put together what the actual state of the ocean is, the atmosphere and the land, using a snapshot of July 1. We start them on July 1 with all the available information we have. Then we run the model for a year. We run it 10 times for each surface resolution. What that creates is an ensemble of what multiple future paths of a March 1 snowpack will look like. We ask what is the average prediction of all those potential futures, and we use that information for our prediction.
Water Deeply: So you include lots of actual Earth observation data feeding into the model?
Kapnick: To start the model, we’re using satellite information that’s observing the state of the ocean, land surface and atmosphere. We’re also using Argo floats. The best way to describe them is like unmanned drones in the ocean; they go in the ocean and then dive to depth. They’re taking measurements of temperature and salinity and pressure as they go down and as they come back up, collecting data points across the ocean.
We have a separate model that takes all these data and combines them together into gridded information that can start the model. This is a global, fully coupled model that is modeling the ocean, atmosphere and land surface together using the physics of how all these factors interact.
And then to verify the model and the predictions, in this paper we are using point measurements of snowpack across the American West.
Water Deeply: And how accurate are the forecasts you’re generating?
Kapnick: We are working toward making probabilistic estimates, which can be described as estimates of likely ranges for potential futures. In the future we can give a range and use it to test the prediction skill.
We have all this modeling and it generates predictions of snowpack and, really, of climate in general. We actually also looked at temperature, precipitation and storm track. We cut up the western United States into tiny boxes and we produce predictions on these boxes. Then we cut up the regions into mountain ranges and tested our prediction skill over the different mountain ranges. What we find is, actually, the models were producing prediction skill over everywhere in the western U.S., all the mountain ranges, except for the southern Sierra Nevada.
For the snowpack predictions, we’ve produced “hindcasts,” where you reproduce what the predictions would have been in the past for the 1981 through 2016 March snowpack. The metric we use is a correlation, which gives a number between -1 and 1 of how well the modeled ensemble mean correlates with the observed snowpack (1 being perfect, -1 being perfectly in reverse). We find the predictions are positively correlated (above 0.4 in most of the West) and statistically significant (better than guessing).
The model does need work, clearly, since we don’t have the southern Sierra Nevada. There are certain aspects we think we can improve on.
The more pessimistic viewpoint is that perhaps the true prediction skill is not ever going to be perfect. Perhaps there are certain chaotic elements to predicting snowpack that will limit our prediction skill.
The study gives us a proof of concept that there is prediction skill in the western U.S. It’s a starting point. But now, going into the future, we need to develop these models further and really explore what the limits of prediction are.
Water Deeply: What’s the prediction problem in the southern Sierra Nevada?
Kapnick: South of about 39 degrees latitude – about where Stockton is – in that region of the Sierras you just have a lot more variability in precipitation and our model is not capturing the extreme variability year to year. In the southern Sierra, 50 percent of total precipitation in a year can occur in only five to 10 days of storms. Even one storm or two can actually generate the majority of the precipitation for a year. So it makes it more difficult to predict how much snowpack there will be because there are so few storms.
Water Deeply: Can these forecasts help us plan ahead for winter?
Kapnick: As of now, this paper is purely for research. I’m not using it for [forecast] operations and it’s not currently being transitioned to operations. But I work at NOAA, and I am doing the research, and our hope is to further develop the prediction system so these seasonal predictions will ultimately be transitioned to operations – either operational prediction of snowpack or operational forecasts in general. The major focus of my work right now is to improve these systems so they can be used operationally.
Water Deeply: How do you imagine it being useful, ultimately?
Kapnick: I have some ideas. Like, water managers will be able to use this information for managing year-to-year variability. But I’m sure there are lots of other uses. There’s a hope a farmer might be able to use information like this to determine what to plant when. If you have a particularly dry year coming, it might influence what you decide to plant.
The southern Sierra is also extremely narrow and also much higher in elevation. Our model, even at 25km [15-mile] resolution, may not have high enough resolution. There are also certain dynamics about what causes precipitation in really narrow mountain ranges that maybe isn’t captured in our model.