CryoSENSE-HK: Explainable Lead-Time Forecasting of ERA5-Land Snow State, Bounded by Station and Physically Based References (Hakkari, Türkiye)

In the semi-arid Cilo-Sat cryosphere of Hakkari Province, Türkiye, mountain snow is the dominant freshwater store, and a sparse ground network leaves reanalysis as the only practical source of data. This study introduces CryoSENSE-HK, an explainable machine-learning framework that forecasts ERA5-Land snow cover, snow water equivalent and snowmelt one to thirty days ahead. Every predictor is observed at or before the issue date, feature ranking uses training data alone, and the held-out period is used once. The framework reproduces the reanalysis with held-out coefficients of determination of 0.93, 0.91 and 0.73, respectively, at thirty days, and is favoured in 77 of 80 block-bootstrap paired tests against eight reference forecasts. Agreement with the ground is much weaker. Station records show the reanalysis carrying two to four times the observed snow depth, and satellite retrievals place its cover too high through the melt season. A reduced monthly emulator of it responds to warming far more weakly than physically based references do: 5.4% of snow water equivalent per degree of warming, against 21.3% for a degree-day balance and 23.4% for CMIP6 snow fields. Skill measured inside a reanalysis is in agreement with the product, not with the snowpack, and a framework of this class needs independent validation of its climate sensitivity before any scenario-driven use.