SPADA

Dataset and code for the paper Land Cover Segmentation with Sparse Annotations from Sentinel-2 Imagery (IGARSS 2023).

Note

Dataset available at hf.co/datasets/links-ads/spada-dataset.

---

Bring me to the important bits

The main components of the training loop are located in:

• custom_dacs.py

Installation

First, create a python environment. Here we used python 3.9 and torch 1.9, with CUDA 11.1. We suggest creating a python environment, using venv or conda first.

The repository is based on mmsegmentation. Follow their installation instructions, or launch the following commands:

pip install -r requirements.txt -f https://download.pytorch.org/whl/torch_stable.html
pip install mmcv-full -f https://download.openmmlab.com/mmcv/dist/cu111/torch1.9/index.html
pip install -e .

Training

Once the dataset is downloaded, create a soft link inside the data directory, it should be named FuelMap. You can use a custom name if you prefer, in that case however you will need to update the configuration files.

Once ready, you can launch a training with the following commands:

$ CUDA_VISIBLE_DEVICES=... python tools/train.py [CONFIG_PATH]

For instance:

$ CUDA_VISIBLE_DEVICES=... python tools/train.py configs/landcover/da_segformer_b5_lucas_dacs_v2.py

Inference

To produce inference maps, run something like the following:

$ CUDA_VISIBLE_DEVICES=... python tools/test.py <path_to_workdir_config.py> <path to the checkpoint> --show-dir <output path> [--eval mIoU]

If you have large images, use the prepare_mosaic.ipynb to tile the large files into 2048x2048 tiles.

Citing this work

@inproceedings{galatola2023land,
  title={Land Cover Segmentation with Sparse Annotations from Sentinel-2 Imagery},
  author={Galatola, Marco and Arnaudo, Edoardo and Barco, Luca and Rossi, Claudio and Dominici, Fabrizio},
  booktitle={IGARSS 2023-2023 IEEE International Geoscience and Remote Sensing Symposium},
  pages={6952--6955},
  year={2023},
  organization={IEEE}
}