Statistical forecasting of regional avalanche danger using simulated snow-cover data

Abstract In the past, numerical prediction of regional avalanche danger using statistical methods with meteorological input variables has shown insufficiently accurate results, possibly due to the lack of snowstratigraphy data. Detailed snow-cover data were rarely used because they were not readily available (manual observations). With the development and increasing use of snow-cover models this deficiency can now be rectified and model output can be used as input for forecasting models. We used the output of the physically based snow-cover model SNOWPACK combined with meteorological variables to investigate and establish a link to regional avalanche danger. Snow stratigraphy was simulated for the location of an automatic weather station near Davos, Switzerland, over nine winters. Only dry-snow situations were considered. A variety of selection algorithms was used to identify the most important simulated snow variables. Data mining and statistical methods, including classification trees, artificial neural networks, support vector machines, hidden Markov models and nearest-neighbour methods were trained on the forecasted regional avalanche danger (European avalanche danger scale). The best results were achieved with a nearest-neighbour method which used the avalanche danger level of the previous day as additional input. A cross-validated accuracy (hit rate) of 73% was obtained. This study suggests that modelled snow-stratigraphy variables, as provided by SNOWPACK, are able to improve numerical avalanche forecasting.