AI for avalanche warning?
From a skiing perspective, the question naturally arises: What can artificial intelligence tell us about the avalanche danger? Do the computers know something that the avalanche warning services don't? Not at the moment - although they may be able to recognize multidimensional correlations that are too complex for the human brain. Operational applications are certainly conceivable in the future. Two current SLF preprints are moving in this direction (preprints are studies submitted to scientific journals that are accessible online but have not yet undergone a full peer review process). In both cases, a so-called random forest model is used. The "random forest" is a "forest" of many decision trees and is often used for the automated classification of large data sets.
In "A random forest model to assess snow instability from simulated snow stratigraphy" it looks like this: Over 700 snow profiles serve as a database, for which, among other things, the result of a landslide block test is available. Depending on the slide block, the profiles are manually classified as stable, unstable or "something in between". With the help of weather data, all profiles were also simulated with a snowpack model - so there is an actual profile dug in the snow and a simulated profile. For each simulated profile, relevant weak layers and the overlying snow slab are described using 34 "features", including, for example, the grain size of the layer, the difference in cone size to the next layer, the shear strength, viscosity, density, etc.
The Random Forest algorithm then receives manually classified training data, i.e. profiles and corresponding "features" that have been identified as stable or unstable. The algorithm then "learns" which features are decisive for stability or instability and sorts the simulated profiles into the "stable" or "unstable" category depending on the properties of the weak layer and the overlying board. The comparison with the "real" profile and avalanche data shows that instability is recognized fairly reliably. The authors of the study see potential to incorporate such methods into operational avalanche warning in the near future.
Interestingly, instability is usually recognized even though the simulated profiles of the snowpack model are generated with spatially interpolated weather data, i.e. they are somewhat blurred and do not reflect small-scale, microclimatic differences. The weather factors that are most decisive for weak layers are apparently nevertheless captured.
In addition, it is interesting to see which of the 34 characteristic features are most important for the stable/unstable classification. In the Random Forest algorithm, the following 6 are the most central:
The viscous deformation rate
The "critical cut length" (length that is sawn in a propagation saw test until breakage occurs)
Skier penetration depth (strongly dependent on the density of the top 30cm of the snowpack)
The grain sphericity in the weak layer (how round are the grains?).)
The ratio of the average snow slab density to the grain size
The grain size in the weak layer
All other features have little or no influence on the result. This will certainly lead to many more research questions. The Random Forest may be "intelligent", but it cannot explain exactly why it chooses these features. Other approaches are needed to explain this in terms of snow physics.
The second current SLF preprint, "Data-driven automated predictions of the avalanche danger level for dry-snow conditions in Switzerland", also uses a Random Forest classification. Here, the avalanche danger level is determined automatically based on weather data and snowpack model calculations. This is successful in 70 to almost 80% of the cases examined. However, there are certain regional differences and the whole thing works somewhat less well if an old snow problem prevails. Wet snow situations were excluded due to the special avalanche processes. The training data with which the model "learns" includes the winters 1997-98 to 2017/18, the model was tested with data from the last two winters (2018/19 and 2019/20)
Once again, there are numerous features and characteristic parameters that are available to the Random Forest in order to assign the appropriate hazard level to a day. And again, it turns out that only relatively few features significantly influence the classification: various new and drift snow parameters, the snowfall rate, the skier penetration depth, the "critical cut length", as well as the relative humidity, the air temperature and stability indices. In principle, exactly the same things that are also heavily incorporated into the "manual" avalanche forecast.
