Weather observation is an important part of monitoring the state of our planet. Beyond the immediate values of weather forecasts and weather warnings, it informs about continuing changes of our environment. Weather observation systems rely on meteorological satellite images. Geostationary satellites provide complete spatial coverage of a specific region and a frequent update cycle. From raw satellite radiances, many atmospheric parameters can be derived, known as weather products.

In recent years, rapid advances in communication and sensor and technologies and data storage allowed collecting unprecedented amount of data about the Earth. Analysis and interpretation of these high-dimensional heterogeneous data remains a major challenge. Application of the state-of-the art machine learning methods can help identify the patterns and underlying processes, which is critical for better understanding of our environment and mitigating climate change.

Meteosat 2nd generation satellite

Image source: eumetsat.int

Cloud mask image

Data example: cloud mask.

Convective rainfall rate image

Data example: convective rainfall rate.

The Data

The competition data are obtained from Meteosat geostationary meteorological satellites operated by EUMETSAT. A range of weather products are derived from the satellite data by EUMETSAT Satellite Application Facilities (SAF) units dedicated to Nowcasting (NWC SAF). The second generation Meteosat images are processed by NWC SAF software developed by a consortium of National Meteorological Services, including AEMET.

The following weather products are selected as target variables for the competition: temperature (on accessible surface: top cloud or earth), convective rainfall rate, probability of occurrence of tropopause folding, and cloud mask. We also provide additional weather products, such as cloud type. They don’t need to be predicted, but can be incorporated to improve the models. The weather products are encoded as separate channels in the weather images. Each weather image contain 256 x 256 pixels, each pixel corresponds to the area of about 4 km x 4 km. The images are recorded at 15 minute intervals throughout a year. The competition data are split into training, validation, and test sets.

The core challenge data contain the training, validation, and test sets for three regions:
R1 – Nile region (covering Cairo),
R2 – Eastern Europe (covering Moscow),
R3 – South West Europe (covering Madrid and Barcelona).
The resulting training set contains about 25 billion data points.

The transfer learning challenge data contain only the test set for three additional regions:
R4 – Central Maghreb (Timimoun),
R5 – South Mediterranean (covering Tripoli and Tunis),
R6 – Central Europe (covering Berlin).

For each day, the images are grouped into a separate folder, within which there are several folders containing the weather products. The training and validation sets contain up to 96 files (24 hours x 4 images/hour) in each folder.

Acknowledgement of the data source: the competition data contains modified AEMET/NWC SAF products from February 2019 to February 2021.

AEMET logo
NWC SAF logo

Weather4cast 2021 Challenge

Stage 1 of the competition comprises two challenges:

1) Core challenge: predict the target weather products in the three training regions R1, R2, R3. Get the data.

2) Transfer learning: predict the target weather products in the three additional regions R4, R5, R6. Get the data.

The goal is to predict the next 32 weather images (8 hours in 15 minute intervals) given 4 images (1 hour) . To solve each challenge, algorithms must predict the target weather products (temperature, convective rainfall rate, probability of occurrence of tropopause folding, and cloud mask) in the three respective regions for the test dataset. The submission format is an array of size (32, 4, 256, 256) per each day of the 36 test days provided.

Download also the Static Channels such as altitude, longitude and elevation (common files for both challenges). Get the data.

Participants can upload their predictions on the test dataset to the leaderboard of each competition until the submission deadline. Within one week after the deadline, participants are requested to provide additional predictions on a new dataset (called held-out dataset), together with an abstract explaining their methodology.

Clone our github repository to download the starting kit and start playing with the baseline models effortlessly! Write in the forums if you have any question or suggestion!

Storm image
regions stage 1

Regions in blue: core challenge; regions in orange: transfer learning.


The winners of the core competition and the transfer learning competition will be awarded the following prizes:

1st place –  a voucher or cash prize worth 5,000€ to the participant/team;

2nd place – a voucher or cash prize worth 3,000€ to the participant/team;

3rd place – a voucher or cash prize worth 2,000€ to the participant/team.

To receive the prizes, participants must submit the working code, the learned parameters, and a short scientific paper describing their approach.


All times and dates are Anywhere on Earth (UTC -12), unless specified otherwise.

Competition starts: April 1, 2021, 00:00 CET.

Leaderboard opens: April 1, 2021.

Competition ends (submission deadline): May 31, 2021, 23:59.

  • held-out dataset available: June 1, 2021, 00:00.

Abstract submission and held-out dataset prediction deadline: June 6, 2021, 23:59.

Announcement of the winners: June 21, 2021.


Imprint | Terms and conditions

Log in with your credentials


Forgot your details?

Create Account