processing-chain-seed (Sentinel-2 NDWI chain)

In order to help you to industrialize your processing chain this repository provides a simple example.

As algorithme example with used a simple processing chain that takes two Sentinel-2 bands as input (green B03 and near-infrared B08 in GeoTIFF format) and produces an NDWI (Normalized Difference Water Index) product as output, used to delineate water surfaces (e.g. lake outline).

Project structure

.
├── Dockerfile              # docker image for the processing chain
├── ndwi-config.json        # NDWI output configuration
├── requirements.txt        # Python dependencies
├── src/
│   └── compute_ndwi.py     # NDWI computation script
├── cwl/
│   ├── ndwi.cwl            # CWL description of the CommandLineTool
│   └── ndwi-job.yml        # example job file (inputs)
└── data/
    ├── input/              # Input data
    ├── output/             # Output data

Getting Sentinel-2 bands

The script expects two single-band GeoTIFF files:

• B03.tif: green band
• B08.tif: near-infrared band (NIR)

These bands can be extracted from a Sentinel-2 L2A product (.SAFE folder) downloaded from the Copernicus Data Space Ecosystem. Place them for example in a data/ folder at the project root.

Run python code

python -m pip install -r requirements.txt
python src/compute_ndwi.py --product data/input/SENTINEL2B_20260829-104909-474_L2A_T31TDH_C_V4-0.zip --output data/output/ --config ndwi-config.json

Build the Docker image

docker build -t ndwi-processor:latest .

Run the chain directly with Docker

docker run --rm -v ${PWD}/data:/data -v ${PWD}:/cfg:ro cnes/processing-chain-seed:latest `
  --product /data/input/product.zip --output /data/output --config /cfg/ndwi-config.json

the same base name. Via CWL, both names are derived from the downloaded ZIP: --output sets the output directory. Both file names are derived by the Python script from the ZIP name: L2A is replaced by L2B, with .tif and .json extensions respectively.

Run the chain via CWL

The cwl/ndwi.cwl file describes the workflow (download from S3 + NDWI computation) and references the Docker image cnes/processing-chain-seed:0.0.1 (built in the previous step). The devcontainer installs cwltool, so after rebuilding the devcontainer, run:

cwltool --outdir data/output cwl/ndwi.cwl cwl/ndwi-job.yml

The S3 download steps (retrieve_s2 and retrieve_conf) require the S3_ENDPOINT_URL, S3_ACCESS_KEY and S3_SECRET_KEY environment variables. Pass them through to the containers with --preserve-entire-environment:

cwltool --preserve-entire-environment --outdir data/output cwl/ndwi.cwl cwl/ndwi-job.yml

Job inputs (bucket_name, l2a_path_s3_url, conf_path_s3_url, input_dir) can be adapted either by editing cwl/ndwi-job.yml or by overriding them directly on the command line, e.g.:

cwltool \
  --preserve-entire-environment \
  --outdir data/output \
  cwl/ndwi.cwl \
  --bucket_name larath-bucket \
  --l2a_path_s3_url SENTINEL2B_20260829-104909-474_L2A_T31TDH_C_V4-0.zip \
  --conf_path_s3_url ndwi-config.json \
  --input_dir data/input \

input_dir only controls the path (relative to the step's own temporary working directory) where the downloaded product and configuration are staged before being passed to the NDWI step; it does not persist them under the repository's data/input/ folder. With --outdir data/output, only the final workflow outputs are copied into data/output/ once the run succeeds. For example, SENTINEL2B_..._L2A_...zip produces SENTINEL2B_..._L2B_...tif and SENTINEL2B_..._L2B_...json.

Github Action build and push the image to CNES DockerHub

By default for each commit it will update the version cnes/processing-chain-seed:dev

If you create a tag and push it, the CD will build it with the tag version cnes/processing-chain-seed:x.x.x

git tag x.x.x
git push origin x.x.x

Manualy Build and push the image to CNES DockerHub

echo "$DOCKER_API_KEY" | docker login -u "$DOCKER_LOGIN" --password-stdin && docker build -t "cnes/processing-chain-seed:x.x.x" . && docker push "cnes/processing-chain-seed:x.x.x"

Interpreting the result

NDWI is positive over water surfaces and negative over vegetation/soil. To extract a lake outline, threshold the ndwi.tif raster (e.g. NDWI > 0) then vectorize the resulting binary mask (e.g. with rasterio.features.shapes or gdal_polygonize.py).