Seasonal forecasting of Alpine snow depth: evaluation of a climate service prototype

Mountain glacier shrinking, seasonal snow cover reduction and changes in the amount and seasonality of meltwater runoff are already affecting water availability for both local and downstream uses. Water is needed by different competing sectors including drinking water supply, energy production, agriculture, tourism, and extremely dry seasons can lead to economic losses. Reducing potential impacts of changes in water availability involves multiple time scales, from the decadal time scale for the realization of water management infrastructures to the seasonal scale, to plan the use of water resources and allocate them with some lead time.

In the framework of the MEDSCOPE ERA4CS project we focused on the seasonal time scale and we developed a climate service prototype to estimate the temporal evolution of snow depth and snow water equivalent with up to 7 months lead time. Forecasts are initialized on November 1^st and run up to May 31^st of the following year. The prototype has been co-designed with and tailored to the needs of water and hydropower plant managers and of mountain ski resorts managers. 

We present the modelling chain, based on the seasonal forecasts of ECMWF and Météo-France seasonal prediction systems, made available through the Copernicus Climate Change Service (C3S). Seasonal forecasts of precipitation, near-surface air temperature, radiative fluxes, wind and humidity are bias-corrected and downscaled to three high elevation sites in the North-Western Italian Alps, and finally used as input for a physically-based multi-layer snow model (SNOWPACK). The RainFARM stochastic downscaling procedure adapted for mountain regions is used for downscaling precipitation data, and stochastic realizations are employed to estimate the uncertainty due to the downscaling method.

The skill of the prototype in predicting the monthly snow depth evolution from November to May in each season of the hindcast period 1995-2015 is demonstrated using station observations as a reference. We show the correlation between forecasted and observed snow depth and we quantify the forecast quality in terms of reliability, resolution, discrimination and sharpness using a set of probabilistic measures (Brier Skill Score, Area Under the ROC Curve Skill Score and Continuous Ranked Probability Skill Score). We finally discuss implications of the forecast quality at different lead times as well as the added value of bias-correction and downscaling of precipitation data on snow depth forecasts. Real-time snow forecasts for the current season (2021-2022) and for earlier ones are available at this link: http://wilma.to.isac.cnr.it/diss/snowpack/snowseas-eng.html