Meteorology- and Topography-Constrained Downscaling of Snow Cover Fraction Over the Tibetan Plateau

Accurate monitoring of Snow Cover Fraction (SCF) over the Tibetan Plateau (TP) is critical for hydrological modeling and climate analysis, yet existing snow products are hindered by an inherent trade-off between spatial resolution and temporal continuity. Coarse-resolution reanalysis data (e.g., ERA5-Land) fail to resolve complex topographic heterogeneity, while high-resolution satellite retrievals (e.g., MODIS) are frequently hindered by cloud contamination. To bridge this gap, this study develops a physics-guided Convolutional Generative Adversarial Network (ConvGAN) framework to downscale daily SCF from 25-km to 5-km. The proposed model integrates coarse-resolution SCF with high-resolution topographic and meteorological variables to reconstruct fine-scale snow patterns. A hybrid loss function combining adversarial training with pixel-wise content loss is employed to ensure both spatial realism and numerical fidelity. Extensive validation against MOD10A1 reference data for the 2013–2022 period demonstrates that the ConvGAN framework achieves high spatial correlation (r > 0.81) and effectively reproduces the elevation-dependent snow distribution. Ensemble experiments reveal that integrating radiative forcing and atmospheric moisture variables substantially reduces errors in sparse and ephemeral snow zones at lower elevations, whereas high-elevation permanent snow is robustly captured across all configurations. The resulting high-resolution SCF dataset offers a spatially continuous, physically consistent solution for cryospheric studies in the Third Pole region.