Globally Distributed Sentinel-2 cloud and cloud shadow Dataset (GDSD)

This Globally Distributed Sentinel-2 cloud and cloud shadow Dataset (GDSD) was derived by refining the existing publicly available Sentinel-2 cloud and cloud shadow dataset and augmenting with our own annotations (Luo et al. 2025). Refinement of existing dataset is needed as the definition about cloud and cloud shadow and quality of the dataset are various (Aybar et al., 2023). Augmentation was necessary due to extensive evaluations revealing suboptimal model performance across certain land cover conditions when trained without augmented data. Consequently, the dataset contains three subsets, CESBIO dataset (Baetens et al., 2019), Sentinel-2 cloud mask catalogue dataset (hereafter as zenodo) dataset (https://zenodo.org/records/4172871), and the augmented SDSU-MSU dataset.

We provided both the original image version and the clipped patch version of these datasets. The clipped dataset consists of 512 × 512 patches without any fill values. Both versions have TOA reflectance and annotated labels. 

Annotated label: 0 FILL VALUE, 64 CLOUD SHADOW, 128 CLEAR, 255 CLOUD

1. Original image version

data_CESBIO2024.tar.gz

data_zenodo.tar.gz

data_SDSU.tar.gz

2. Clipped version

The clipped dataset consists of 512 × 512 patches without any fill values.

PATCH.stl2b13c3.0.zip

PATCH.stl2b13c3.1.zip

PATCH.stl2b13c3.2.zip

PATCH.stl2b13c3.3.zip

PATCH.stl2b13c3.4.zip