Climate-driven changes in avalanche activity in the Haute-Maurienne valley (French Alps) over the period 1950–2100 based on machine-learning modelling

Snow avalanches in alpine environments are influenced by climate change. Assessing their long-term evolution remains difficult due to the scarcity of temporally consistent observation records. This lack of homogeneous data complicates the attribution of past changes to climate drivers and the development of credible future projections based on climate scenarios. To address these limitations, we develop a machine learning model that links avalanche observations with simulated meteorological and snowpack properties at a daily scale. A gradient-boosting regression model is trained to estimate daily avalanche counts, using weather and snow conditions derived from simulations. The analysis is conducted for a well-instrumented alpine catchment, the upper Haute-Maurienne valley in the French Alps, where approximately one hundred avalanche paths are monitored on a daily basis. We focus on avalanche events that reach predefined observation thresholds located at elevations of around 1800 m a.s.l. We show that particular attention to data quality and consistency has to be paid: accounting for uncertainties in avalanche release dates and restricting the training phase to a recent period with homogeneous observations are key prerequisites to obtain consistent results. Once trained, the model is first applied to reconstruct avalanche activity over the period 1958–2023 using reanalysed snow and meteorological data. It is then used to compute the evolutions of the avalanche activity between 1950 and 2100 using a downscaled ensemble of snow–climate simulations. Changes in avalanche activity are assessed using three complementary indicators corresponding to annual, monthly and weekly time scales. The reconstructed historical time-series indicates a marked decline in avalanche activity, with an average reduction of about 9 % per decade in the annual number of avalanches since 1958, mainly due to a decrease in the spring avalanche activity, while extreme avalanche cycles exhibit a more moderate decline. Future projections suggest a continued downward trend. Under the RCP4.5 and RCP8.5 scenarios, annual avalanche counts are projected to decrease by roughly 5 % and 9 % per decade, respectively, again largely driven by changes in spring conditions. Extreme avalanche activity is also expected to weaken, although at slower rates, with projected decreases in the 30-year return level of about 2 % per decade for RCP4.5 and 5 % per decade for RCP8.5. These climatic trends are associated with climate-induced changes in snowpack and meteorological variables through the use of machine-learning interpretation approaches. Overall, this study provides a quantitative assessment of climate-driven changes in avalanche activity for a representative alpine valley, combining machine-learning approaches with physically based snow-climate simulations.