Modeling future snow cover dynamics under SSP scenarios with LSTM: Insights from Northern Finland (2025–2100)

Study region Oulujoki River basin, an inland sub-Arctic catchment in northern Finland. Study focus This study projects daily snow cover fraction in the study area during 2025–2100 under four Shared Socioeconomic Pathways using a hybrid Long Short-Term Memory–Multilayer Perceptron neural network trained on ERA5-Land reanalysis data (2000–2019) and driven by bias-corrected downscaled CMIP6 temperature and precipitation projections. Key snow cover features, including onset (SC onset ), rise (SC rise ), fall (SC fall ), and end (SC end ), were extracted and analyzed using Gaussian Mixture Models for distributional shifts, Mann-Kendall tests, Sen's slopes, and Pettitt tests for trends, and a novel non-stability index for variability. Moreover, spatial clustering identifies three regions in Oulujoki with distinct snow regimes enabling the assessment of snow cover dynamics. New hydrological insights for the region The model accurately captures historical snow dynamics, with accumulative mean absolute errors below 5% in validation. Results reveal delayed accumulation and advanced melt, reducing snow duration, with melt phases more sensitive than accumulation. High-emission scenarios show pronounced changes, including steep trends (e.g., −8 days/decade for SC fall ), a post-2050 decline of more than 50% in full snow cover days, and increased instability in snow cover. Northeastern regions retain snow longer than southwestern ones, yet projected changes could alter hydrological regimes, and increase flood risks, emphasizing the urgency of emission mitigation and regional climate adaptation to preserve boreal snow-dependent systems.