IceCloudNet

This is the official repository for the manuscript "IceCloudNet: 3D reconstruction of cloud ice from Meteosat SEVIRI"

Clouds containing ice remain a source of great uncertainty in climate models and future climate projections. IceCloudNet overcomes the limitations of existing satellite observations and fuses the strengths of high spatio-temporal resolution of geostationary satellite data with the high vertical resolution of active satellite retrievals through machine learning. With this work we are providing the research community with a fully temporal and spatial resolved 4D dataset of cloud ice properties enabling novel research ranging from cloud formation and development to the validation of high-resolution weather and climate model simulations.

📜 Cite as

@article{jeggle2024icecloudnet3d,
      author = "Kai Jeggle and Mikolaj Czerkawski and Federico Serva and Bertrand Le Saux and David Neubauer and Ulrike Lohmann",
      title = "IceCloudNet: 3D Reconstruction of Cloud Ice from Meteosat SEVIRI",
      journal = "Artificial Intelligence for the Earth Systems",
      year = "2025",
      publisher = "American Meteorological Society",
      address = "Boston MA, USA",
      volume = "4",
      number = "4",
      doi = "10.1175/AIES-D-24-0098.1",
      pages=      "240098",
      url = "https://journals.ametsoc.org/view/journals/aies/4/4/AIES-D-24-0098.1.xml"
}

💾 Dataset Access at WDC Climate

Inference

This repo contains a pretrained model that can be used to create 3D cloud structures from MeteoSat Seviri input.

Steps:

1. Download SEVIRI input data: src/download_data/MSG_eunmdac_satpy.ipynb
2. Unzip with src/unzip_seviri.sh
3. Create input data for ML model: src/inference/CreateSeviriWholeAreaTimeSeries.ipynb
4. Specify DATA_DIR variable in src/inference.py and set up directory structure as described there
   • mv helper_files/* DATA_DIR
5. Run inference: run_inference.sh

Training

Download input data

Steps:

1. Download SEVIRI input data: src/download_data/MSG_eunmdac_satpy.ipynb
2. Unzip with src/unzip_seviri.sh
3. Download DARDAR-Nice from [ICARE data servers](earlier version: https://www.icare.univ-lille.fr/data-access/data-archive-access/?dir=CLOUDSAT/DARDAR-NICE_L2-PRO.v1.00/) → login required.
   • Currently only DARDAR Nice v1 available on ICARE. Contact the author for v2 (which is used in the paper).

Preprocess and co-locate SEVIRI and DARDAR-Nice

python data_preproc.py YYYYMMDD YYYYMMDD n_workers

Run experiment

We are using Comet ML to track experiments. Setup up account and insert your credentials in the necessary places. Setup up environ variables alternatively.

Steps:

• mv helper_files/* /path/to/your/data/directory
• specify filepaths in run_experiment.py and then run

Evaluation

Run EvaluateModel.ipynb. Requires that co-located patches were created already

Setup

There are two conda environment files:

• satpy.yml used for everything that requires satpy
• pytorch.yml used for everything that involves pytorch