An algorithm for predicting snow quality?

[blasta34]

Hey all,

I'm trying to develop an algorithm to provide a snow quality rating (from 1 - 5) for ski resorts on a per-day basis, both current day and upcoming (within the next 5 days). The objective is to quickly provide a perspective on if we're looking at concrete or fluffy conditions at Resort X, as well as to potentially help decide between different resorts: say resort X is supposed to get 8 wet inches but resort Y (30 minutes further) is supposed to get 8 soft inches, it's probably worth the extra drive time.

Data I have available per resort per day:
-temperature min/max
-freezing level (and the resort min/max elevations)
-snowfall / projected snowfall
-wind strength and direction
-whether it will be sunny/snowy/cloudy

I have some thoughts on how I could use these to create an algorithm, for example a great snow quality rating would have most of the following: cold temperatures (10 deg F and less), low wind (< 5mph), good FL, recent snow
A poor rating would be the opposite (say above 40 deg F, > 30mph winds, poor FL, no snow for > 3 days).

In an ideal world I could get a shitload of sample data where people have rated the snow quality 1-5 and run some linear regression on all the factors above and come up with a great algorithm, but for now I'm stuck to guesstimating. My plan is to come up with an initial model, then test and tweak throughout the season.

Any thoughts or feedback? What other weather factors do you use when predicting if the snow is going to be good or not?

Thanks

[neck beard]

Different to your objectives, however possibly of related interest: ASARC in Calgary have some research into computer modelling of surface hoar formation and also building a snowpack from weather inputs [for areas where there is no one to take local observations.]

[buildakicker]

This would be cool. There are so many variables to try and get computed in. I'd love to see the equation.

You probably know about StackExchange, but if you don't I bet this Answers based forum could be helpful: http://math.stackexchange.com/.

Have you talked to some ski patrollers about sample data?

I like the site you have created. Good idea!

[blasta34]

Thanks for the replies, will definitely check these out. At this point I'm probably going to go with simple vs complex and then tweak/adjust after some field testing.

[klar]

how complicated do you want this to be? in terms of input, atmospheric soundings/ information on layering might be nice to have. also, temperature during snowfall?

for algorithm ideas look at the literature on snowpack modelling and see how specific models deal with snow density? try a google scholar search for SNOWPACK (swiss model) or CROCUS (french). there is also some literature on wind drift but it seems difficult to model. might find some ideas about wind effects there though.

you say you have snowfall/projected snowfall data. how is the projected snowfall projected? weather models do not predict depth of new snow, they predict water (i.e. how much would it rain if it was warmer). forecasters then apply algorithms to estimate snow depth (i.e. density) from that. what are those algorithms? your projected snow fall already contains information on snow density (fluffy or cement) if you know the amount of water that the snow depth value was calculated from. what kind of forecast data are you using in general? is it an automated forecast generated from one of the big providers or is there a person that looks at it? how detailed is the model topography?

are you planning to have a spatial dimension, i.e. exposition/radiation input?

[wcf3]

Until you get some data on the dependent variable (snow quality) it will be tough to make much progress. One thing you might try to use as a surrogate for this is change in snow depth from day to day which may at least give you something to start modeling with. I would assume that on days where considerable depth is lost due to wind, heat or radiation, subjective quality will go down, while when depth is increasing, quality will go up.

Instead of dumping everything into a linear regression, it's usually a good idea to do some exploratory data analysis: Plot the dependent variable against each potential explanatory variable and see if there is much of a relationship and whether it appears to be linear or not. If one sees a strong non-linear relationship, try and transform the explanatory variable to produce a more linear relationship.

You may also want to try and reshape the distributions of any variables to get them closer to a normal distribution before putting everything into a linear regression. Otherwise the results can be pretty strange. (Recover the actual forecast then by inverting the transformation on the forecast values). Do some histograms of the variables in your data set to see how weird they might be.

You also want to look for outliers in your data (they will show up in histograms) and clip them or ditch them entirely before doing any regressions, since they will tend to throw everything off.

Good luck!

[blasta34]

Thanks for the excellent feedback klar and wcf! Good idea on temperature during snowfall, I do have that as well.

As for forecast data, I'm using a combination of NOAA and snow-forecast.com. Maybe not always the best, but they are consistent - format/frequency of update/etc. I'm looking to be simplistic - basically trying to auto-compute some numbers a person might do in their head when trying to answer a question of "who is going to have better snow?". (Spring/Corn conditions not in scope). Admittedly, it's a bit of a fools errand as conditions within even a single resort/area can vary greatly.

I'm going to try playing around with some models and then explore the results as wcf suggested. Hopefully I can get something promising that makes sense. Will report back with anything of value.

[klar]

maybe not relevant for your purpose but just to be clear: snow-forecast uses GFS data, which is a NOAA product. GFS has a horizontal resolution of 28km globally (better in some parts of North America, I think), snow-forecast downscales the forecast with a 1km resolution digital elevation model. You are using two versions of the same raw data, unless when you say NOAA you mean a local forecast that a human has made, based on NOAA data. The GFS forecast maps are all publicly available btw.

I think wcf is right that you need some sort of measureable proxy for snow quality, like change in snow depth. If you have either manually gathered or automatic weather station data (do some sort of quality control for aws data!), you can validate your model.

For a simplistic approach I would use the amount of predicted precipitation during timewindow X and temperature at the surface during the precip event as the main drivers, with a temp dependant rain-snow conversion. I would try to figure out constants to apply as scale factors if windspeed is over some threshhold or there is a large temperature swing after the snowfall.

Fun project, keep us updated.

[IVplay]

Snow quality can vary largely from one slope/aspect of a resort to another. For example slope x that faces south will become baked later in the year, whereas slope y that is north facing will preserve better snow. Also, I would weigh some categories higher than other, for example FL can be directly tied to the temperature. Recent snow by itself it not a good a good barometer either, for example I would much rather prefer 3 cold days without snow where the terrain remained soft, then getting snow two days ago and then going through a freeze/thaw cycle.

The way I guess snow quality is rather simple-- if the conditions the days leading up to the day you weren't skiing were not terrible, and it is currently below 15F and snowing, you're going to have a good time.

[blasta34]

Agreed on your points. I think I'm going to try a simplistic approach based on Joel Gratz' video "Finding Perfect Powder": http://www.youtube.com/watch?v=FG2k1fi0Zy8
Basically: cold temps + less wind = better snow, which is consistent with what you and klar have suggested. Now to try some test data runs and play with the weighting/constants of the factors. One problem is that I only have wind for data from mid September 2013 and later, I don't have wind data from last year so my test data set is smaller than I'd like at the moment.

Snow quality can vary largely from one slope/aspect of a resort to another. For example slope x that faces south will become baked later in the year, whereas slope y that is north facing will preserve better snow. Also, I would weigh some categories higher than other, for example FL can be directly tied to the temperature. Recent snow by itself it not a good a good barometer either, for example I would much rather prefer 3 cold days without snow where the terrain remained soft, then getting snow two days ago and then going through a freeze/thaw cycle.

The way I guess snow quality is rather simple-- if the conditions the days leading up to the day you weren't skiing were not terrible, and it is currently below 15F and snowing, you're going to have a good time.