NH-SDN: A Deep Learning-Derived Daily 0.25° Snow Density Dataset for the Northern Hemisphere from 1982–2025

Snow density is a key parameter in the physical properties of the snowpack. Accurately mapping snow density is essential for water resources, avalanche hazards, and snowpack thermal processes. However, large-scale snow density datasets that provide reliable accuracy, spatial and temporal continuity, and comprehensive coverage across the Northern Hemisphere remain unavailable. To bridge this gap, we develop SDN-Net, a spatiotemporal framework integrating vision Transformers and convolutional neural networks, to capture global dependencies and local heterogeneity for snow density estimation. The framework incorporates multisource environmental variables and leverages adaptive feature refinement mechanisms to better characterize temporal variations in snow density. Using this framework, we develop NH-SDN, a daily snow density dataset for the Northern Hemisphere at a spatial resolution of 0.25° from 1982–2025. Compared with reanalysis datasets covering the Northern Hemisphere, NH-SDN shows higher accuracy across all months, particularly from March to May, where RMSE decreases from 0.130 g cm⁻³ in reanalysis products to 0.053 g cm⁻³ in NH-SDN. It also performs well across different snow classes and elevations, with particularly notable improvements in mountainous regions above 2-km elevation, where systematic bias in reanalysis data is effectively reduced, and RMSE decreases from 0.100 to 0.042 g cm⁻³. In addition, NH-SDN reduces the excessive smoothing of increase-stage variability and the overestimation of melt-stage trends in reanalysis products, allowing improved representation of interannual snow density variations. It also provides a continuous snow density dataset that closely agrees with in-situ measurements. The dataset provides a reliable basis for large-scale snowpack characterization and may support applications in snow water equivalent estimation, hydrological and cryospheric modeling, and snow thermal process studies. The dataset is available at https://doi.org/10.5281/zenodo.23768969 (Zhang et al., 2026).