Dataset Card for South Africa Crop Type Clouds

This dataset contains the cloud masks generated and used for the paper KAN You See It? KANs and Sentinel for Effective and Explainable Crop Field Segmentation.

Curated by: Daniele Rege Cambrin
License: OpenRAIL

Uses

The dataset will provide a quality assessment for Sentinel-2 images of the South Africa Crop Type dataset. Since MSI is ineffective through clouds, it was used to exclude samples that contain a large portion of the area of interest covered by clouds.

Dataset Structure

The dataset has the following structure:

{
  "__key__": FileName,
  "__url__": OriginTarFile,
  "tiff": PILImage
}

where FileName is also the name of the folder in the South Africa Crop Type Dataset.

IMPORTANT: Remember to convert the PIL Image to an array to obtain the probability mask of clouds.

Dataset Creation

The masks are created automatically using the s2cloudless library using the algorithm presented by Sergii Skakun et. al.

The missing bands are replaced with a channel with no-data value (0) to avoid the algorithm relying on this channel for the prediction.

Bias, Risks, and Limitations

Since no human expert is involved in the process, some annotations could be inaccurate or unreliable. The masks are intended to exclude samples that could be under a certain degree of uncertainty noise, and that cannot be annotated by a human expert, too. They should not be used outside this scope.

Citation

If you use this dataset in your work, consider citing our work.

BibTeX:

@misc{cambrin2024kanitkanssentinel,
      title={KAN You See It? KANs and Sentinel for Effective and Explainable Crop Field Segmentation}, 
      author={Daniele Rege Cambrin and Eleonora Poeta and Eliana Pastor and Tania Cerquitelli and Elena Baralis and Paolo Garza},
      year={2024},
      eprint={2408.07040},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2408.07040}, 
}