This repository applies the pre-trained Deep Generative Model of Rainfall (DGMR) developed by Google DeepMind to Meteo-France rainfall data. The goal of this project is to provide real-time, high-quality short-term rainfall forecasts (nowcasting) for France.

• Features
• Installation
• Usage
• Data Sources
• Model Architecture
• Licenses
• Contributors

• Download radar data through the Meteo-France public data API
• Short-term rainfall forecasting: max lead time 2h, timestep 5min, resolution 1km
• Utilizes the DGMR pre-trained (UK data) model for realistic forecasts
• Visualizations of rainfall predictions with GIFs

git clone https://github.com/meteofrance/dgmr
cd dgmr

With micromamba or conda:

micromamba create -f environment.yml

micromamba activate dgmr

If you don't already have an account on the Météo-France Public Data portal, please create one.

1. Once connected, subscribe to the Radar data API (not the "Radar Package" one).

2. Then click on Configure the API.

3. Select API Key as the token type, enter the duration (in seconds), and Generate Token.

4. Copy the token; you will need it in the next step.

Go to this Google Cloud Storage link and download the whole folder dm-nowcasting-example-data/tfhub_snapshots/1536x1280 (this folder contains the pre-trained model with a size of 1536x1280).

Note that the AI model is licensed under Creative Commons BY 4.0.

Create a .env file in the root directory and add your configuration settings:

METEO_FRANCE_API_KEY="<your_meteo_france_api_key>"
METEO_FRANCE_DATA_PATH="<path_to_save_the_downloaded_data>"  # If empty, downloaded data will be saved in ./data
DGMR_MODEL_PATH="<path_to_the_pretrained_dgmr_folder>/1536x1280/"
DGMR_PLOT_PATH="<path_to_save_the_plots>"  # If empty, forecast GIFs will be saved in ./plot

Simply run the script ./dgmr/download_data.sh to download the latest radar images.

For real-time forecasting, we need to download the radar data every 5 minutes.

To automate the data downloading process using a cron job, follow these steps:

1. Open your terminal and type crontab -e to edit the cron table.

2. Add a new line at the end of the file to define the cron job. The general syntax for a cron task is:

   */5 * * * * <ABSOLUTE_PATH_TO_REPO>/dgmr/download_data.sh

To make a real-time rainfall nowcast, use:

python main.py

This script expects that you have already downloaded the 4 latest radar images.

It saves the forecast as a GIF in the plot folder defined by the DGMR_PLOT_PATH env variable.

This project uses rainfall data from Meteo-France.

The H5 files contain 5-minute cumulated rainfall data over France, with a 500m resolution.

To match the pre-trained model inputs, we convert the data to instantaneous rainfall (mm/h) and downgrade the resolution to 1km.

The DGMR model is a state-of-the-art generative model designed for weather nowcasting. It leverages a combination of convolutional neural networks (CNNs) and generative adversarial networks (GANs) to produce high-resolution, realistic rainfall forecasts.

For more details, please see Ravuri, S., Lenc, K., Willson, M. et al. Skilful precipitation nowcasting using deep generative models of radar. Nature 597, 672–677 (2021).

A notebook is also available as a supplement to the article: notebook.

The DGMR model is under the terms of the Creative Commons Attribution 4.0 International License.

• The DGMR pre-trained model is licensed under the Creative Commons Attribution 4.0 International License.
• Radar data from Météo-France are licensed under the Open Licence.
• The code in this repository is licensed under the Apache License, Version 2.0.

• From Météo-France's AI Lab: Théo Tournier, Léa Berthomier, Frank Guibert, Bruno Pradel
• From the French National School of Meteorology: Flora Chapelle, Hugo Germain, Paul Lacotte & Ariadna Perello