The official code of the following paper: https://arxiv.org/abs/2302.04102
WF-UNet and Core U-Net precipitation prediction examples for multiple timesteps ahead. The images in the left side are generated with the test set from the dataset containing at least 20% of rain pixels (EU-20). The images in the right side are generated with the test set from the dataset containing at least 50% of rain pixels (EU-50).
Average test MSE values over all timesteps ahead of the tested models for the EU-50 and EU-20 datasets.
The required modules can be installed via:
pip install -r requirements.txt
After placing the the raw .nc datasets in the "dataset" folder, the following command will create the final dataset in a HDF5 format, contain the training and test set ready to use for modelling.
python create_datasets.py
The models which use only precipitation images can be trained running the following command:
python training_rain.py
while the WF-UNet model with the following command:
python training_rain_wind.py
To evaluate the models and visualize some predictions, please run:
python evaluation_and_predictions.py
- The scripts contain the models, the generators, the training files and evaluation files.
In order to download the data please email to the following address:
The data must be downloaded and unzipped inside the 'dataset/' directory as indicated in the txt files inside them.
If you use our data and code, please cite the paper using the following bibtex reference:
@article{kaparakis2023wfunet,
title={WF-UNet: Weather Fusion UNet for Precipitation Nowcasting},
author={Kaparakis, Christos and Mehrkanoon, Siamak},
year={2023},
eprint={2302.04102},
archivePrefix={arXiv},
primaryClass={cs.LG}
}