Predicting Avalanche Danger in Northern Norway Using Statistical Models

Abstract. Snow avalanches are one of the most impactful natural hazards in mountainous areas. Thus, the assessment and forecasting of avalanche danger is of great importance for the protection of life and property. A changing climate may lead to changes in avalanche danger although the manifestation is unclear. Since climate change is regionally different, an assessment of potential avalanche danger changes should be conducted on a regional basis. Here the focus is on avalanche danger in northern Norway, i.e., a region in the Arctic. We utilise regional expert avalanche-danger level (ADL) assessments together with the 3-km Norwegian reanalysis (NORA3) to estimate the linkage between avalanche danger in the Troms region of Norway and the prevailing weather conditions represented by NORA3 as well as snow-cover information from the snow model seNorge. Both a binary and 4-level case are considered. Two random forest (RF) models are optimised and trained, one for the binary case and one for the 4-level case.

The binary-case RF model exhibits a much higher overall accuracy (76 %) than the 4-level case RF model (57 %), which is due to the latter often confusing ADLs 1 and 2 and ADLs 3 and 4. Still, the missclassification difference is almost never larger 1 ADL and the distribution of the frequencies of the different ADLs is reproduced. The most important predictive features of avalanche danger found here are broadly consistent with earlier studies and are mostly related to new snow and wind accumulated and averaged over several days. The binary-case RF model is used to hindcast binary-case avalanche activity (BCA) from 1970 to 2023. In this period, the spring season (Mar–May) shows a small but in most regions significant increase in BCA, whereas the winter season (Dec–Feb) exhibits mostly non-significant negative trends. Moreover, BCA is found to be correlated with the Arctic Oscillation (AO) index especially in winter, although this correlation may have deteriorated in recent years. Given recent advances in skill of representing the AO in decadal prediction systems this is an encouraging result for the predictability of future avalanche danger tendencies in northern Norway.

The methodology presented here may be generally applied to link climate indicators to numerical climate model output, enabling their prediction for future climate change scenarios.