Abstract. Snow-avalanche hazard in mountainous areas may change in frequency and severity due to climatic change, especially in Arctic regions such as northern Norway where temperature change is amplified. Expanding earlier work, we train machine-learning (ML) models on dynamically downscaled reanalysis data including snow-cover simulations to predict avalanche danger for the Troms county in northern Norway. In contrast to earlier work, the trained ML models can distinguish between avalanche types, in particular those of dry and wet snow. Due to insufficient avalanche observations, we construct a binary metric (avalanche day/non-avalanche day) based on the avalanche danger warnings published in the Norwegian avalanche bulletin. The ML models provide a hindcast of the frequency of avalanche days for the period 1970–2024 (based on reanalysis) and a projection into the future for the 21st century (based on climate model simulations). Over the historical period the results confirm earlier studies showing that while multi-decadal linear trends are marginal, the interannual variability of the avalanche-day frequency is linked to the Arctic Oscillation. The projected future changes indicate a general decrease of avalanche danger, especially for dry-snow avalanches. In contrast, wet-snow avalanche danger exhibits changes dependent on elevation, increasing at all elevations until mid-century, but thereafter continuing the increase only at higher elevation, while at lower elevation reversing to a decrease. Our results are in line with an emerging consensus of an overall decline of avalanche danger in the 21st century and a shift in avalanche characteristics towards fewer dry and more wet-snow avalanches. Such a shift can be challenging for avalanche-prone populations as the current knowledge of the local avalanche conditions may become less relevant and increasingly fail to provide protection from the avalanche hazard.
Eiselt et al. (Wed,) studied this question.