Snotelier is a system for inferring avalanche danger from snow telemetry (SNOTEL) sites in Washington and Oregon. These are remote weather stations run by the USDA Natural Resource Conservation Service which measure temperature, snow depth, snow water content, and cumulative precipitation for the year. Trends in these measurements tell a story of what the snowpack is like, and certain patterns (e.g., rapid accumulation of snow) can be indicative of instability.
In Oregon, SNOTEL data are commonly used by recreationalists to assess avalanche danger, since expert avalanche advisories are not available between Mount Hood and Mount Shasta. However, assessing risk directly from SNOTEL measurements requires significant domain knowledge. Snotelier aims to make this assessment more straightforward by encoding this knowledge in a machine learning model.
When you select a point on the map on the main interface, Snotelier pulls histories from the five nearest SNOTEL sites and offers its estimation of the avalanche danger at each one on a simple 2-point scale: "Relatively High" or "Relatively Low". These estimations are general guidance based solely upon recent data from the SNOTEL site, and will be subject to variation in elevation, aspect, future weather, and other factors.
Snotelier distinguishes troublesome conditions from relatively safe ones by recognizing a set of patterns, or "engineered features", deemed to be indicative of avalanche danger. For example, it quantifies the maximum rate of recent snowfall, and the temperature while that snowfall was occurring. These choices were informed by the chapter on Class III factors described in the Avalanche Handbook. Snotelier then applies a statistical model to this feature space to determine the probability that the snowpack is dangerous.
The parameters of this model were determined by fitting a logistic regression to approximately 300 labeled example histories. These example histories were selected from SNOTEL sites in one of the NWAC advisory regions, on occasions when either a "Low" or "High" rating was issued for that region. These sites were all located at approximately 5000 feet above sea level.
SNOTEL features are sufficient to distinguish these two classes of NWAC rating with an accuracy of about 88% and an AUC score of greater than 0.9. These metrics have held up for labeled examples from the 2016 snow year, which was held out from the training set for validation.
Perfect accuracy is not expected, since avalanche experts are generally smarter and have access to more information than Snotelier. It's important to keep in mind that Snotelier makes its assessment only from SNOTEL data - it has no knowledge of wind, qualitative weak layer observations, or the weather forecast.
I'm Tom Baldwin, an Insight Data Science fellow. I completed my PhD in physics at the University of Oregon in 2015 and my level 1 avalanche certification in 2016. Snotelier is my 3-week project for the Insight fellowship.
For more about me, see my website.