spateGAN_ERA5

spateGAN-ERA5 is a deep learning framework designed for the spatio-temporal downscaling of ERA5 precipitation data. Utilizing a probabilistic conditional Generative Adversarial Networks (cGANs), it enhances the resolution of ERA5 data from 24 km and 1-hour intervals to 2 km and 10-minute intervals. This advancement enables high-resolution rainfall fields with realistic spatio-temporal patterns and accurate rain rate distributions, including extreme events. (https://www.nature.com/articles/s41612-025-01103-y)

---

Features

• Global Generalization
  Trained on German rain gauge adjusted radar data and validated in regions like the US and Australia.

• High-Resolution Output
  Converts coarse ERA5 precipitation data into 2 km × 10 min fields, essential for flood and hydrological risk assessments.

• Uncertainty Quantification
  Generates diverse ensembles of precipitation scenarios.

• User-Defined Inference Run spateGAN-ERA5 on custom regions and time periods using your own ERA5 netCDF files via a simple configuration-driven workflow.

---

Datasets

Training, model selection and evaluation datasets are available at https://doi.org/10.5281/zenodo.17417589

To run the model for downscaling precipitation data in inference, ERA5 data in hourly resolution can be downloaded here as a netCDF file. The variables "Convective precipitation" and "Large scale precipitation", a minimal spatial extent of 672kmx672km and a minimal time sequence of 16 hrs are required.

---

Getting Started

Installation

conda

mamba env create -f environment.yaml
or
conda env create -f environment.yaml

uv clone repo install uv: https://docs.astral.sh/uv/getting-started/installation/

cd spateGAN_ERA5
uv sync #creates .venv

use spateGAN_ERA5/.venv/bin/python3 as your interpreter

Downscaling

Update config.yaml to your downscaling domain (defined as centre lat lon coordinates), input path of ERA5 netCDF and output path.

Run

uv run main.py

High resolution precipitation fields are saved in UTM projection (2km, 10min resolution) and lat lon (0.018° and 10min resolution).

Probabilistic downscaling

For probabilistic downscaling seed and slideparameter in config.yamlcan be changed.

Evaluation example

As an evaluation example, a notebook is included in the repo:

downscaling.ipynb (and downscaling_example.py) — uses a small sample dataset to show spateGAN-ERA5 for downscaling ERA5 precipitation over Germany and compare it to radar observations.

License and co

The content of this repository is released under the terms of the MIT license.

Consider citing if you use spateGAN-ERA5:

@article{spateGAN_era5_2025,
	title = {Global spatio-temporal {ERA5} precipitation downscaling to km and sub-hourly scale using generative {AI}},
	doi = {10.1038/s41612-025-01103-y},
	journal = {npj Climate and Atmospheric Science},
	author = {Glawion, Luca and Polz, Julius and Kunstmann, Harald and Fersch, Benjamin and Chwala, Christian},
	year = {2025},
}