Improving numerical avalanche forecasting with spatial snow cover modeling

Abstract

Snow avalanches are a natural hazard in mountainous areas which endanger roads, villages and human lives. To inform the public on the current avalanche situation, avalanche warning services regularly publish avalanche bulletins in winter. However, forecasting snow avalanches is very challenging. Currently, it is not possible to predict the exact timing, location or size of snow avalanches. Avalanche forecasters therefore estimate the degree of avalanche danger at the scale of a region, by linking point observations of the snowpack, consisting of observations of snow stratigraphy and snow instability, with past and future weather. While snowpack observations are very time consuming and thus rather scarce, numerical snow cover models can considerably increase the spatial and temporal resolution of such data, especially if they provide information on snow instability. In our current understanding of avalanche formation, avalanche release is a fracture mechanical problem and snow instability is best understood in terms of failure initiation and crack propagation. Detailed snow cover models exists which can simulate snow stratigraphy, but snow instability information is partly missing or inaccurate. In view of applying snow cover models for avalanche forecasting, the aim of this thesis was to model spatially distributed snow instability. The snow cover model SNOWPACK simulates criteria for failure initiation and crack propagation for each snow layer, namely the stability index and the critical crack length. While the stability index had been validated with field observations and was related to the probability of skier triggering, this was not the case for the critical crack length, which was parameterized based on layer properties including density, shear strength and the elastic modulus. In a first step, we therefore validated the evolution of snow layer properties and the critical crack length in SNOWPACK with novel field measurements. Daily measurements with the snow micro-penetrometer allowed for direct comparison with SNOWPACK. Our results showed that the evolution of layer density was fairly well captured by the model, especially the first two months after deposition. For the validation of the critical crack length, we used results from the propagation saw test performed on a weekly basis over three winter seasons. A comparison to SNOWPACK highlighted some discrepancies, and we thus refined the parameterization with a fit factor depending on weak layer density and grain size. With the refined parameterization, spatially distributed modeling of snow instability in terms of failure initiation and crack propagation became tangible. Spatially distributed snow cover simulations require interpolation and downscaling of meteorological data, which may introduce uncertainties. How these uncertainties impact modeled snow stability remained mostly unknown. For the first time, we therefore investigated the sensitivity of modeled snow instability to meteorological input uncertainty with a global sensitivity analysis. Early in the season, during the period of weak layer formation, modeled instability metrics were mostly sensitive to air temperature and precipitation. After weak layer burial, during the period of slab formation, modeled instability metrics were mostly sensitive to precipitation. These results highlighted that accurate spatial snow depth distributions are required to obtain realistic snow instability patterns. In a last step, we used the distributed snow cover model Alpine3D to simulate snow instability for the region of Davos, Switzerland. Meteorological data from automatic weather stations were interpolated to a grid with 100 m resolution. Precipitation was scaled with highly resolved snow depth measurements from airborne laser scanning to account for realistic snow depth patterns. Modeled snow instability patterns were plausible, e.g south-facing slopes stabilized faster in spring. However, instability metrics were lower for south-facing slopes during the winter months, which was not in line with the forecasted avalanche danger level. While spatial patterns of modeled snow instability still have to be validated, spatially distributed snow cover modeling can greatly improve numerical avalanche forecasting. However, given a lack of accurate input data, simple virtual slopes instead of highly resolved modeling approaches is probably enough at this point in time.