Deep-learning projections of snow-season dynamics in arid and semi-arid Northwest China during 2025–2045

Existing ski-climate risk studies, derived from humid or semi-humid mountains, have limited applicability to arid and semi-arid resorts. They use temperature, snow-cover duration, or suitability indices, rarely resolving natural snow depth, snowmaking windows, and interregional or inter-resort differences. In water-limited terrain, how temperature, moisture, and terrain regulate snow-resource risk remains insufficiently quantified. We developed SnowFormer, a physics-constrained hybrid Transformer integrating temporal encoding, terrain-conditioned parameterization, and rain–snow partitioning and melt constraints, to project snow depth, potential snowmaking days, and snow-season conditions across Northwest China and 216 ski resorts under CMIP6 scenarios for 2025–2045. SHAP diagnosed meteorological controls. Relative to the historical baseline, snow depth and potential snowmaking days decline by 26%–29.6% and 15.5%–18.6%. Snow-season timing is heterogeneous: median onset advances by 4.0–7.0 d, and cessation is delayed by 2–3 d in most scenarios. Some Tianshan resorts lengthen, whereas Altay and inland peripheral resorts shorten by 3.5–9.4 d. Tianshan extension is linked to high-elevation moisture compensation, whereby atmospheric moisture and terrain effects partly offset warming-induced melt, consistent with alpine snow responses to warming and wetting. Air temperature and specific humidity contribute 24.4% and 20.9%, indicating that moisture effects require cold conditions. Climate-resilient ski-tourism planning requires joint assessment of snow depth, snowmaking windows, and snow-season timing.