nikolkoo/SatelliteSegmentation

Satellite Segmentation

Summary

This dataset contains paired Sentinel-2 RGB tiles and corresponding land-cover masks (derived from ESA WorldCover) prepared for semantic segmentation. Each example has:

• image: RGB image (PNG)
• mask: integer-labelled mask (PNG, uint8)

Key details

• Source imagery: Sentinel-2 L2A via Microsoft Planetary Computer (STAC)
• Land-cover masks: ESA WorldCover (derived)
• Spatial resolution: 10 m (aligned to Sentinel-2 grid)
• CRS: EPSG:4326
• Number of samples: 790
• Train/validation split: Train

Provenance & license

This dataset was derived from third‑party datasets:

• Sentinel‑2 (Copernicus)
• ESA WorldCover

The user of this dataset must respect the original licenses and terms of use. The repository contains derived files (tiles).

Data format

• Images: PNG, RGB, 3 channels
• Masks: PNG, integer values representing classes (do not normalize/convert to RGB)
• Filenames: {prefix}.png and {prefix}_mask.png (paired by prefix)

How to load

Example (datasets library):

from datasets import load_dataset
ds = load_dataset("nikolkoo/SatelliteSegmentation")

Example evaluation/training snippet

Use CrossEntropyLoss with logits and integer masks:

# pseudocode
images = batch["image"]  # (B,H,W,3) -> to tensor & permute
masks  = batch["mask"]   # (B,H,W) ints
logits = model(images)
loss = torch.nn.CrossEntropyLoss()(logits, masks)

Citation

If you publish results using this dataset, cite the original data providers (Copernicus / ESA / Microsoft Planetary Computer) and this dataset repo.

Contact

Feel free to add a comment in the Community 🤗