Statistical evaluation of local to regional snowpack stability using simulated snow-cover data

Snow stability, or the probability of avalanche release, is one of the key factors defining avalanche danger. Most snow stability evaluations are based on field observations, which are time-consuming and sometimes dangerous. Through numerical modelling of the snow cover stratigraphy, the problem of having sparsely measured regional stability information can be overcome. In this study we compared numerical model output with observed stability. Overall, 775 snow profiles combined with Rutschblock scores and release types for the area surrounding five weather stations were rated into three stability classes. Snow stratigraphy data were then produced for the locations of these five weather stations using the snow cover model SNOWPACK. We observed that (i) an existing physically based stability interpretation implemented in SNOWPACK was applicable for regional stability evaluation; (ii) modelled variables equivalent to those manually observed variables found to be significantly discriminatory with regard to stability, did not demonstrated equal strength of classification; (iii) additional modelled variables that cannot be measured in the field discriminated well between stability categories. Finally, with objective feature selection, a set of variables was chosen to establish an optimal link between the modelled snow stratigraphy data and the stability rating through the use of classification trees. Cross-validation was then used to assess the quality of the classification trees. A true skill statistic of 0.5 and 0.4 was achieved by two models that detected “rather stable” or “rather unstable” conditions, respectively. The interpretation derived could be further developed into a support tool for avalanche warning services for the prediction of regional avalanche danger.