This repository aims to provide a benchmark dataset for statistical downscaling of meteorological fields with deep neural networks.
The work was initiated in scope of the MAELSTROM project which aimed to develop new machine learning applications for weather and climate under the cooridniantion of ECMWF.
The benchmark dataset is based on two reanalysis datasets that are well established and quality-controlled in meteorology: The ERA5-reanalysis data serves as the input dataset, whereas the COSMO-REA6 datasets provides the target data for the downscaling. With a grid spacing of 0.25° of the input data compared to 0.055° of the target data, the downscaling factor is set to 4 in the benchmark dataset. Furthermore, different specific downscaling tasks adapted from the literature are provided. These pertain the following meteorlogical quantities:
- 2m temerature
- 100m wind speed
- solar irradiance