Repeatable high-resolution statistical downscaling through deep learning

Abstract. One of the major obstacles for designing solutions against imminent climate crisis is scarcity high spatio-temporal resolution model projections variables such as precipitation. This kind information crucial impact studies in fields like hydrology, agronomy, ecology, and risk management. The currently highest spatial datasets on a daily scale projected conditions fail to represent complex local variability. We used deep-learning-based statistical downscaling methods obtain 1 km gridded data precipitation Eastern Ore Mountains Saxony, Germany. built upon well-established climate4R framework, while adding modifications its base-code, introducing skip connections-based deep learning architectures, U-Net U-Net++. also aimed address known general reproducibility issues by creating containerized environment with multi-GPU (graphic processing unit) TensorFlow's deterministic operations support. perfect prognosis approach was applied using ERA5 reanalysis ReKIS (Regional Climate Information System Saxony-Anhalt, Thuringia) dataset. results were validated robust VALUE framework. introduced architectures show clear performance improvement when compared previous benchmarks. best performing architecture had small increase total number parameters, contrast benchmark, training time less than 6 min one NVIDIA A-100 GPU. Characteristics models configurations that promote their suitability this specific task identified, tested, argued. Full repeatability achieved employing same physical GPU, which key build trust applications. EURO-CORDEX dataset meant be coupled trained generate high-resolution ensemble, can serve input multi-purpose models.