WeatherFusionNet: Predicting Precipitation from Satellite Data

This is our solution to the Weather4cast 2022 competition, we achieved 1st place in the Core Challenge. For more info about the competition, description of the used data and a baseline starter kit, please see iarai/weather4cast-2022. The related paper is WeatherFusionNet: Predicting Precipitation from Satellite Data at https://arxiv.org/abs/2211.16824.

Usage instructions

• Download the data (see link above) and extract it into the data subfolder, or edit models/configurations/config.yaml to point to the right folder.
• Install dependencies with

  conda env create -f environment.yml
  conda activate weather4cast
  

• Download trained weights from Releases.

Generating a submission

The predict-submission.py script generates a submission zip file for a given challenge and split. For example:

python predict-submission.py --challenge core --split test --gpus 0

or

python predict-submission.py --challenge transfer --split heldout --gpus 0

You can find the result in submission/submission.zip.

See python predict-submission.py --help for more info on the arguments.