The Potential of Copernicus Satellites for Disaster Response: Retrieving Building Damage from Sentinel-1 and Sentinel-2

This repository contains the offical code and dataset for the paper "The Potential of Copernicus Satellites for Disaster Response: Retrieving Building Damage from Sentinel-1 and Sentinel-2". The dataset, referred to as xBD-S12, is derived from the original xBD dataset and includes co-registered Sentinel-1 and Sentinel-2 imagery.

Setup

Environment

We use uv to manage the Python environment. Simply install uv and run the following command to create and activate the environment with all necessary dependencies:

uv sync
source .venv/bin/activate

Dataset

The original xBD VHR images and labels cannot be redistributed here due to licensing restrictions. Please download them directly from xView2 by following their instructions. Place the resulting, extracted dataset in the data/ folder, e.g. in data/original_xbd folder. The expected structure after extraction should be:

data/
└── original_xbd/
    ├── hold/
    ├── test/
    ├── tier1/
    └── tier3/

From there, the xBD-S12 dataset can be generated by running the following command. [Optional:] Add the --create_hdf5 flag if you also want to create the HDF5 file for faster loading during training.

bash setup_dataset.sh --original_xbd_path="data/original_xbd" --num_workers=8 [--create_hdf5]

Alternatively, you can follow the steps below to create the dataset manually.

1. Download the xBD-S12 dataset from Zenodo at this URL.

cd data/
curl -L -O "https://zenodo.org/records/18960454/files/xbd_s12.tar.gz?download=1"
tar -xvzf xbd_s12.tar.gz # remove the v flag if you don't want to see the list of extracted files
rm xbd_s12.tar.gz

2. Create lightweight vrt files to wrap up the original raster files with corrected metadata under data/xbd_s12/xbd.

python src/data/create_aligned_vrt.py --original_xbd_path="data/original_xbd" --num_workers=8

3. Create the masks from the original xBD labels and store them under data/xbd_s12/masks/.

python src/data/create_masks.py --original_xbd_path='data/original_xbd'

4. [Optional] Create the HDF5 files for faster loading during training and inference. This will create one single HDF5 file containing all the Sentinel-1, Sentinel-2, and mask data chunked and compressed.

python src/data/create_hdf5.py --num_workers=8

The final dataset structure should look like this:

data/
└── xbd_s12/
    ├── masks/
    ├── s1/
    ├── s2/
    ├── s2_tci/
    ├── xbd/ # only contains the .vrt files, which reference the original .tif files in data/original_xbd
    ├── normalization.json
    └── xbd_s12_metadata.geojson

For a detailed description of the dataset, please refer to DATASET.md. To quickly explore it, refer to the notebook xbd-s12_exploration.ipynb.

---

Models & Inference

We provide the checkpoints of our best models trained on xBD-S12 on the Hugging Face Hub at this URL. The models can be easily loaded and used for inference using the InferenceFromHub class in src/inference/inference_from_hub.py. For example:

from src.inference.inference_from_hub import InferenceFromHub

inferor = InferenceFromHub()
preds = inferor.infer(
    fp_s1_pre='path/to/s1_pre.tif',
    fp_s1_post='path/to/s1_post.tif',
    fp_s2_pre='path/to/s2_pre.tif',
    fp_s2_post='path/to/s2_post.tif',
)

For a complete example, including how to download and prepare Sentinel data, please refer to the notebook inference_palisades_wildfires.ipynb.

Training

To train a model yourself, run the following command:

python src/training/main.py run_name="my_run_name"

All training artifacts, including model checkpoints and logs, will be saved under logs/my_run_name/. We use Hydra to manage the training configuration. To change the default configuration, you can either modify the config files under src/configs/ or override specific config values directly from the command line. For example:

python src/training/main.py run_name="my_run_name" data.which_split=xview2 data.modalities='[s1, s2_tci]' data.pixels_buffer_around_buildings=2 training.max_epochs=30 model=unet model.encoder_name=resnet50 model.encoder_depth=4

Citation

If you use this code or dataset in your research, please cite the paper as follows:

@misc{2511.05461,
Author = {Olivier Dietrich and Merlin Alfredsson and Emilia Arens and Nando Metzger and Torben Peters and Linus Scheibenreif and Jan Dirk Wegner and Konrad Schindler},
Title = {The Potential of Copernicus Satellites for Disaster Response: Retrieving Building Damage from Sentinel-1 and Sentinel-2},
Year = {2025},
Eprint = {arXiv:2511.05461},
}

Note: The paper will be presented at the ISPRS Congress 2026 and published in the ISPRS Annals. We will update the citation information once the proceedings are published.

License

The code in this repository is licensed under the MIT License. See LICENSE for details.