Waiting for the snow to melt…
Snow-dominated mountain basins: Upper Genil basin (Sierra Nevada, southern Spain); Uncompahgre basin (Rocky Mountains, Colorado – USA); Mapocho basin (Andes, Chile); Adda basin (Alps, Italy); Nenskra basin (Caucasus Mountains, Georgia); Indrawati basin (Himalaya, Nepal). Snow-dominated mountain regions are highly sensitive to climate variability and act as early indicators of climate change impacts. Accurate prediction of snow cover dynamics is essential for assessing implications for water resources and mountain ecosystems. However, widespread snow cover decline and increasingly nonlinear relationships between snow evolution and climate drivers have reduced the reliability of linear modeling approaches. This study proposes a hybrid framework combining Nonlinear Autoregressive Networks with Exogenous Inputs (NARX) and Long Short-Term Memory (LSTM) networks to predict snow cover extent. The NARX-LSTM model was designated to be applicable in data-scarce regions, particularly in mountainous and alpine environments where in situ observations are often sparse, discontinuous, or unavailable. This framework was evaluated across six snow-dominated basins representing contrasting climatic regimes: semi-arid (Sierra Nevada, Spain; Rocky Mountains, USA; Andes, Chile) and humid (Alps, Italy; Caucasus, Georgia; Himalaya, Nepal). Historical climate drivers were derived from ERA5-Land reanalysis, while MODIS snow cover area were used for model training and validation. Future projections were generated for mid-century (2051–2070) and late-century (2081–2100) under two emission scenarios (SSP2–RCP4.5 and SSP5–RCP8.5), using statistically downscaled climate forcing from the LARS-WG weather generator. The hybrid NARX–LSTM model effectively captures complex nonlinear snow-climate interactions, demonstrating robust predictive performance across diverse climatic conditions (R² ≈ 0.9; NSE = 0.68–0.96; KGE = 0.73–0.97). Future projections indicate a consistent warming signal across all basins, while precipitation changes remain spatially heterogeneous. Significant snow cover reductions are projected for all sites, ranging from 9.4 to 50.3% under SSP2–RCP4.5 and 13.1–73% under SSP5–RCP8.5, intensifying toward the end of the century. Semi-arid region basins emerge as particularly vulnerable, experiencing the largest snow cover declines. In addition, the snow season is projected to shorten by up to 1–3 months due to earlier melt and delayed accumulation. These findings highlight increasing hydrological sensitivity in mountain regions and the need for advanced modeling tools to support climate adaptation strategies. • A novel hybrid NARX–LSTM neural network framework is proposed for predicting snow cover area time series. • The hybrid NARX-LSTM demonstrates robust transferability and temporal stability across diverse basins. • Future projections from downscaled CMIP6 models indicate substantial declines in snow cover area across all basins. • Snow season duration is projected to shorten by 1–3 months by 2100 under the most pessimistic emission scenario (SSP5-RCP8.5). • Semi-arid basins exhibit heightened sensitivity to climate change impacts on snow cover area.
Items connected by shared entities, co-authorship, citations, or semantic similarity.