Use this repository to create semantic segmentation datasets made up of multiple different modalities into a unified multi-channel datasets. E.g by combining imagery and LiDAR data.

Data and labels are cut into patches and dataset is divided into train and test/valid subsets while taking geographical overlap into consideration. Code supports conversion of labeled polygons stored in GeoPackage files into GeoTIFF label images.

The resulting datasets can be used for training and inference with ML_sdfi_fastai2.

The accompanying "example_dataset" combines the following georeferenced layers:

• Orthophoto: GeoDanmark Orthophoto
• Oblique Camera (Ortho Version): LOD Images
• Digital Terrain Model (DTM): Danmarks Højdemodel – DHM Raster Download
• Digital Surface Model (DSM): Danmarks Højdemodel – DHM Raster Download

training_dataset/
  labels/
    large_labels/
      image-x.tif
  data/
    original_data/
      image-X_rgb.tif
      image-X_cir.tif
      image-X_OrtoRGB.tif
      image-X_OrtoCIR.tif
      image-X_DSM.tif
      image-X_DTM.tif
    rgb/
      image-X.tif
    cir/
      image-X.tif
    OrtoRGB/
      image-X.tif
    OrtoCIR/
      image-X.tif
    DSM/
      image-X.tif
    DTM/
      image-X.tif

Images located in the original_data folder will be renamed and distributed into the appropriate subfolders (rgb, cir, OrtoRGB, OrtoCIR, DSM, DTM).
If original_data is empty, the tool will use existing images from these subfolders.

Labels should be provided as GeoPackages containing polygon features marking different semantic areas.
These will be rasterized into GeoTIFF label images during dataset creation.

Use conda or mamba (Miniforge includes conda; mamba is optional). From this repository root (or from a parent folder where all four shared-env repos are cloned as siblings):

conda env create --file environment.yml
conda activate ML_sdfi
pip install --pre --no-build-isolation -r requirements_pip.txt

This installs PyTorch nightly with CUDA 12.8 (for NVIDIA Blackwell / RTX 50-series / sm_120 GPUs), fastai, git-based deps, and this package in editable mode.

To install the other shared-env repos and extra deps, from the project root (parent of all four repos):

cd ML_Production && bash install_local_repos.sh && pip install -r requirements_extra.txt && cd ..

Other GPUs: To use stable PyTorch instead of nightly (e.g. cu121), after the steps above run:

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121

(Adjust cu121 to your CUDA version; see pytorch.org/get-started/locally.)

Use conda's libstdc++ (Linux): On some Linux systems, set this before running Python:

export LD_LIBRARY_PATH=$CONDA_PREFIX/lib:$LD_LIBRARY_PATH

Verify CUDA support:

python -c "import torch; print('CUDA available:', torch.cuda.is_available()); print('Device:', torch.cuda.get_device_name(0) if torch.cuda.is_available() else 'N/A')"

Windows: After the steps above, run once: pip install --force-reinstall pillow rasterio so PIL and rasterio use pip's Windows wheels.

Pull the prebuilt shared image and run with this repo as working directory:

docker pull rasmuspjohansson/kds_cuda_pytorch:latest

docker run --gpus all --shm-size=100g -it \
  -v /path/to/your/projects:/home/projects \
  -w /home/projects/multi_channel_dataset_creation \
  rasmuspjohansson/kds_cuda_pytorch:20260302 \
  bash

To have all four shared-env repos installed in the container, run once from ML_Production (e.g. with -w /home/projects/ML_Production and sh install_local_repos.sh && pip install -r requirements_extra.txt). Then use -w /home/projects/multi_channel_dataset_creation for this repo.

Example after setup:

python src/multi_channel_dataset_creation/create_dataset.py --dataset_config configs/create_dataset_example_dataset.ini

A small example dataset is included with this repository.
You can generate a dataset using the example configuration:

python src/multi_channel_dataset_creation/create_dataset.py   --dataset_config configs/create_dataset_example_dataset.ini

To see all available options:

python src/multi_channel_dataset_creation/create_dataset.py -h

Creating label Images from a GeoPackage can be done with

Example with not labeled areas marked up as ignore_label (background_value == 0)
python src/multi_channel_dataset_creation/geopackage_to_label_v2.py   --geopackage example_dataset/labels/example_dataset_ground_surface.gpkg   --input_folder example_dataset/data/rgb/   --output_folder example_dataset/labels/large_labels/   --attribute ML_CATEGORY --background_value 0

Example with all polygons interpreted as label 2 (value_used_for_all_polygons == 2) and unlabeled areas interprted as background class 1 (background_value == 1)

python src/multi_channel_dataset_creation/geopackage_to_label_v2.py   --geopackage example_dataset/labels/example_dataset_buildings.gpkg   --input_folder example_dataset/data/rgb/   --output_folder example_dataset/labels/large_labels   --background_value 1 --value_used_for_all_polygons 2

Cleaning labels can be done by

1. create labels based on geopackage older than the data
2. create labels based on geopackage newer than the data
3. create cleaned labels based on the old and new labels python src/multi_channel_dataset_creation/data_cleaning_based_on_newer_ground_truth.py --old_labels dir_with_olod_labels --new_labels dir_with_new_labels --output dir_with_cleaned_labels Labels that have changed in this time interval should not be trrusted and are set to ingore value (0)

Do the installation acording to instructions above run

python src/multi_channel_dataset_creation/create_dataset.py --dataset_config configs/create_dataset_example_dataset.ini

There should be no error messages in output

• The tool is designed for geospatial datasets with consistent coordinate systems.
• Each channel (RGB, CIR, OrthoRGB, OrthoCIR, DSM, DTM) should be aligned and georeferenced properly before processing.