A Benchmark Dataset for Statistical Downscaling of Meteorological Fields with Deep Neural Networks

Obtaining localized and high-resolution information on the atmospheric state is of particular relevance in several domains such as agriculture, the renewable energy sector or natural hazard management. Complementary to running costly numerical simulations at high resolution, statistical downscaling with deep neural networks has recently gained momentum. While various studies have approached or even surpassed the accuracy of classical statistical downscaling methods, intercomparison of the approaches is impeded due to a large variety of methods and deployed datasets.
Inspired by the available benchmark datasets for various computer vision tasks and for weather forecasting (e.g. WeatherBench and WeatherBench Probability), we provide a benchmark dataset for statistical downscaling of meteorological fields. We choose the coarse-grained ERA5 reanalysis (Δx≅30 km) and the fine-scaled COSMO-REA6 (Δx≅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 ready-to-use data for three standard downscaling tasks, that are downscaling of the 2m temperature, the near-surface wind field and the irradiance. Along with the dataset, baseline neural network architectures from the literature such as U-Nets, GANs and a Swin-Transformer network are provided which have been evaluated with a set of suitable task-specific metrics. The provided source-code for evaluating trained neural networks facilitates benchmarking of new downscaling methods and is suitable for a quick experimenting cycle.
Relaxing the obstacle to retrieve data with the benchmark dataset as well as enabling systematic comparison is considered to be fundamental for the development of deep learning methods for statistical downscaling. Furthermore, fair intercomparison between different approaches enhances confidence and transparency of novel deep learning methods for statistical downscaling which, in turn, is believed to foster the advance of deep learning in Earth system Science in general.