Sentinel-2 Cloud Mask Catalogue: 513 Subscenes with Surface and Cloud Tags

513 1022-by-1022 pixel subscenes sampled from the 2018 Level-1C Sentinel-2 archive provide pixel-wise cloud masks. The dataset was annotated semi-automatically using the IRIS toolkit by Alistair Francis and John Mrziglod from University College London. It includes binary classification tags for each subscene covering 11 surface types and 7 cloud types.

Use Cases

• Train cloud segmentation models based on pixel-wise CLEAR, CLOUD, and CLOUD_SHADOW masks.
• Evaluate algorithm performance on specific surface types based on the 11 provided surface categories.
• Analyze failure modes of cloud masks in relation to cloud structure using the 7 cloud type tags.
• Benchmark shadow detection algorithms using the subset of 424 subscenes with annotated shadows.

Strengths

• 513 subscenes at 20m resolution provide a substantial sample for model training.
• Semi-automatic annotation using a dynamically trained Random Forest allowed processing many more masks than manually possible.
• Includes binary tags for 11 surface types and 7 cloud types to facilitate targeted testing.

Limitations

• Cloud shadows were not annotatable for 89 subscenes due to ambiguous boundaries or terrain, limiting shadow detection training.
• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.

Provenance

Source

University College London

Collection Method

Semi-automatic annotation using the IRIS toolkit, with final manual changes by annotators.

Time Range

2018

Geography

Global (sampled randomly from the Sentinel-2 archive)