Precipitation in Germany

Image source: kulkann/Getty Images


The goal of our second challenge is to estimate precipitation rates (in mm/h) in Germany based on Meteosat Second Generation (MSG) data.

To this end, we provide the following preprocessed data sets from December 2017:

RADOLAN is a radar based product with precipitation rates in mm/h generated by the German Weather Service (DWD). You can find more information here.

MSG data are already clipped to Germany. RADOLAN data are reprojected to the same domain. The MSG data have a temporal resolution of 15 minutes. The Radolan data have a temporal resolution of 5 minutes and the units are mm/h. For the competition, the corresponding hourly scenes were compiled from both data sets.

Below you find a visualization of both data sets for an exemplary time slot. Black areas around the domain boundaries are due to missing radar data in these regions.

RADOLAN data are provided for 595 training time slots. 149 additional RADOLAN scenes that were randomly selected are kept back for testing the submissions. The train/test split looks like this:


You should hand in the prediction results for all 149 test scenes (173x233 px) as one netcdf-file. As a guide, you can use this template file and fill it with your predictions.

Like the RADOLAN data used in training, your data should be submitted in units of mm/h.

You can submit your results a maximum of 3 times. Only the best submission will be used for your ranking score.

Results must be submitted via Ilias. Please also stick to the following file name format:

so for instance:


January 30, 2022

Results handed in after this deadline will not be rated.


The rating will be based on the R² score.

It will be applied to all “precipitating pixels” in the 149 test scenes. This means that pixels which were not detected as precipitating (0.0 mm/h) by the radar network of the DWD will not be included in the scoring procedure.


The first place will be rewarded with a 20€ “Marburg Gutschein”.