Deep Learning Based Super-Resolution Spatial Downscaling of Snow Cover in the Eastern Black Sea Region

Snow cover distribution in mountainous regions is characterized by high spatio-temporal variability, particularly in the form of ephemeral snow cover, which undergoes rapid accumulation and ablation cycles. In the Eastern Black Sea region of Turkey, these dynamics are increasingly influenced by climate change, leading to significant shifts in snowmelt timing and accelerated early-season runoff. Accurately monitoring these processes is constrained by the spatial resolution of regional reanalysis datasets. While the CERRA-Land dataset provides a robust and high-quality long-term climatic record, its 5.5 km spatial resolution is primarily optimized for regional-scale dynamics, which inherently limits the characterization of sub-grid orographic effects and the precise delineation of snow lines in topographically complex terrains. This study implements a novel deep learning framework based on Enhanced Super-Resolution Generative Adversarial Networks (ESRGAN) to downscale CERRA-Land snow cover data to a 500 m resolution. The model is trained using multi-platform MODIS (Terra and Aqua) satellite imagery as the high-resolution ground truth, enabling the capture of diurnal variations and reducing cloud-cover interference. A fundamental aspect of this methodology is the integration of auxiliary topographic features, including elevation, slope, and aspect derived from Digital Elevation Models (DEM). Given the extreme vertical gradients of the Eastern Black Sea, these topographic variables are essential for the model to learn altitude-dependent snow distribution patterns and correct for complex shading effects. Quantitative evaluations demonstrate that the proposed framework significantly outperforms traditional spatial interpolation methods. Specifically, the topography-informed ESRGAN model reduced the Root Mean Square Error (RMSE) from a baseline of 49.20% to 17.40% and improved the Peak Signal-to-Noise Ratio (PSNR) from 6.16 dB to 15.20 dB, successfully reconstructing sharp textural details. Beyond its performance in spatial reconstruction, this method offers significant computational efficiency. By providing a high-fidelity and rapid downscaling mechanism, the framework can be seamlessly applied to historical climate reconstructions and future climate change simulations. Consequently, this research provides a robust foundation for high-resolution hydrological modeling, enabling better-informed water resource management and flood risk assessment under shifting climatic conditions.