Manually Labelled Sea Ice Masks for Sentinel-1 SAR Imagery in the Canadian Arctic (2022–2023)

1. Overview

This dataset contains manually labelled sea-ice masks for a set of Sentinel-1 Extra-Wide (EW) swath scenes acquired in the Canadian Arctic during 2022–2023.
Labels were produced for use in sea-ice segmentation experiments comparing supervised UNet models, a BYOL-pretrained UNet, Random Forest classification, and the Segment Anything Model (SAM).

The masks provide pixel-level classification of:

• 2 = Sea ice

• 3 = Land

These labels can be used directly for training and validation of machine-learning models, or as a reference dataset for benchmarking segmentation methods and lead-detection approaches.

2. Source Imagery

The labels correspond to Sentinel-1 SAR scenes (HH and HV, EW mode) processed through ESA SNAP.
Raw satellite images are not included here due to ESA licensing restrictions.

They can be downloaded from:

Copernicus Browser
https://browser.dataspace.copernicus.eu/

Each mask is named according to the corresponding Sentinel-1 SAFE product.

3. Preprocessing (Summary)

All SAR scenes were processed in ESA SNAP and exported as terrain-corrected, georeferenced GeoTIFFs.
The pixel spacing is approximately 80 m, consistent with EW mode multilooking and terrain correction.

4. Labelling Procedure

Labels were digitised manually in QGIS using processed HH + HV composites as visual guidance.

• Digitisation performed per scene.

• Visual interpretation based on SAR backscatter texture, tone, and contextual patterns.

• Land pixels were labelled explicitly as 3.

• Output resolution matches the SAR grid (∼80 m).

5. File Contents

Each labelled scene is provided as:

Sentinel_1_file_name_labels.tif

A GeoTIFF containing:

• 2 = Sea ice

• 3 = Land

All files include georeferencing information and match the spatial extent and resolution of the corresponding Sentinel-1 scene.

6. Coordinate Reference System

All masks use the Arctic polar stereographic projection:

• EPSG:3995

• Datum: WGS84

• Units: metres

• Pixel size: ~80 × 80 m

7. Usage Notes

• The dataset is suitable for semantic segmentation, active learning, lead detection, and self-supervised SAR benchmarking.

• If any scene contains NoData areas (e.g., sensor padding or terrain-correction gaps), these are encoded as -9999 and should be excluded in training pipelines.

• Sea ice is consistently labelled as 2, enabling straightforward class-mapping across experiments.

8. Citation

If you use this dataset, please cite the Zenodo DOI and the associated research article (when available):

Seston, J., Harcourt, W.D., Leontidis, G., Rea, B., Spagnolo, M., & McWhinnie, L. (2025).
Manually Labelled Sea Ice Masks for Sentinel-1 SAR Imagery in the Canadian Arctic (2022–2023).
Zenodo.

9. Contact

For questions about the dataset or related research:

Jacob Seston
School of Geosciences, University of Aberdeen
Email: j.seston.23@abdn.ac.uk