This repository contains the machine learning code used in the paper: Thunderstorm nowcasting with deep learning: a multi-hazard data fusion model, submitted to Geophysical Research Letters.

You need NumPy, Scipy, Matplotlib, Seaborn, Tensorflow (2.6 used in development), Numba, Dask and NetCDF4 for Python.

Clone the repository, then, in the main directory, run

$ python setup.py develop

(if you plan to modify the code) or

$ python setup.py install

if you just want to run it.

The dataset can be found at the following Zenodo repositories: https://doi.org/10.5281/zenodo.6325370 and https://doi.org/10.5281/zenodo.7157986. Follow the instructions there on where to place the data.

The pretrained models are available at https://doi.org/10.5281/zenodo.7157986. Unzip the files models-*.zip found there to the models directory and results.zip to the results directory..

Go to the scripts directory and start an interactive shell. There, you can find training.py that contains the script you need for training and plots_sources.py that produces the plots from the paper.