Sentinel-2 Water Edges Dataset (SWED)

Authors: Catherine Seale, Thomas Redfern, Paul Chatfield
Organsation: UK Hydrographic Office, Taunton, UK
Corresponding author: Catherine Seale
Contact Information: catherine.seale@ukho.gov.uk

Sentinel-2 Water Edges Dataset contains pairs of images and corresponding segmentation labels, intended for training image segmentation algorithms. The labels contain two classes indicated by binary values – water (1) and non-water (0). The creation of this dataset is part of an on-going effort by the UKHO Data Science team to promote the development of machine learning and deep learning for marine applications. The data is available under the Geospatial Commission Data Exploration license.

Download the full dataset

Note: this will download a large file (approx. 18Gb).

Download full dataset

Download a sample of the dataset

Choosing this option will download a smaller sample for evaluation purposes (approx. 1.36Gb)

Download sample

License

View the license terms.

News and updates

October 2021: Dataset published.

January 2022: Sample link added.

Dataset description

The Sentinel-2 Water Edges Dataset is based on Sentinel-2 imagery.

Images are of an array shape 256 x 256 x 12 – the 12 bands corresponding to Sentinel-2 bands in the order they are supplied by the European Space Agency. Labels have shape 256 x 256 x 1 – only one band is required for the segmentation mask.

Image and label pairs can be identified by their matching filenames as below:

S2A_MSIL2A_20190312T031541_N0211_R118_T48QYJ_20190312T083550_image_0_0.tif S2A_MSIL2A_20190312T031541_N0211_R118_T48QYJ_20190312T083550_label_0_0.tif

The dataset contains an independent test set to enable comparative study of deep learning models.

Information regarding tidal state of images in the test set can be obtained in this associated csv.

While you are here

Check out our blog for updates on our work!