Towards a benchmark dataset for statistical downscaling of meteorological fields

The representation of the atmospheric state at high spatial resolution is of particular relevance in various domains of Earth science. While global reanalysis datasets such as ERA5 provide comprehensive repositories of meteorological data, their spatial resolution (∆x≥25 km) is too coarse to capture relevant local features, mainly over complex terrain (e.g. cold pools in valleys, low-level jets, local heavy precipitation events).
Recently, various studies have started to apply deep neural networks adapted from computer vision to increase the spatial resolution of meteorological fields. Although these studies reveal great potential in the domain of statistical downscaling, intercomparison of the approaches is impeded due to a large variety of methods and deployed datasets. Comparisons to classical downscaling methods developed for decades in the meteorological community are also often underrepresented.

Inspired by the available benchmark datasets for various computer vision tasks and for weather forecasting (e.g. WeatherBench and WeatherBench Probability), our study aims to provide a benchmark dataset for statistical downscaling of meteorological fields. We choose the coarse-grained ERA5 reanalysis (∆x_ERA5≃30 km) and the fine-scaled COSMO-REA6 (∆x_CREA6≃6km) as input and target datasets. Both datasets enable the formulation of a real downscaling task: super-resolve the data and correct for model biases.
The benchmark dataset provides a collection of predictors and predictands for a couple of standard downscaling tasks. These comprise downscaling of the 2m temperature, the surface irradiance, the near-surface wind field and precipitation. Along with the dataset, benchmark deep neural networks, namely variants of U-Nets and GANs, will be provided. Well-chosen sets of evaluation metrics including baseline scores of the benchmarked deep neural networks are presented to enable comparison between different methods.
The envisioned benchmark dataset will provide a comprehensive basis for comparing neural network approaches on statistical downscaling of meteorological fields. This, in turn, is considered to enhance confidence and transparency in the application of deep learning methods on Earth system problems.