fatmur/dynamicworld-300

Dynamic World 300

A small land-cover segmentation dataset for learning and quick experiments: 300 expert-labelled Dynamic World tiles, each paired with the Sentinel-2 scene it was labelled on. It is the data of the GeoSave Engine example notebooks.

classes.yaml                                class names and colours
train/  images/<id>.tif  labels/<id>.tif   200 tiles
val/    images/<id>.tif  labels/<id>.tif    50 tiles
test/   images/<id>.tif  labels/<id>.tif    50 tiles

An image and its label share a file name and the same 510 × 510 grid of 10 m pixels, in the tile's own UTM zone.

Download

from huggingface_hub import snapshot_download

snapshot_download("fatmur/dynamicworld-300", repo_type="dataset", local_dir="data/dynamic_world")

Images

Cloud-optimized GeoTIFFs of Sentinel-2 Level-2A surface reflectance, four bands in this order:

Band	Light	Wavelength
B02	blue	490 nm
B03	green	560 nm
B04	red	665 nm
B08	near-infrared	842 nm

Pixels are uint16 digital numbers; reflectance is the number × 0.0001. Each file records that scale, nodata 0, which bands are red, green and blue, and its acquisition time, so GIS software and GeoSave read them correctly.

Labels

uint8 class indices 0–9, with 255 marking pixels a labeller left unmarked; ignore it in losses and scores (ignore_index=255). Each file records the class names and colours, and classes.yaml lists them for code that does not read the files' metadata, as class_map and color_map keyed by index:

Index	Class	Index	Class
0	water	5	scrub
1	trees	6	built area
2	grass	7	bare ground
3	flooded vegetation	8	snow and ice
4	crops	9	cloud

These are Dynamic World's codes 1–10 minus one; its code 0 (no data) is 255. Tiles were chosen with at most 1% cloud, so cloud is nearly absent.

How it was made

From the expert-labelled training tiles of the Dynamic World release (Tait et al., PANGAEA), keeping tiles at least 90% labelled and at most 1% cloud, spread evenly over the 14 RESOLVE biomes. Each label's Sentinel-2 scene of the same day was loaded from Microsoft Planetary Computer onto the label's grid. Tiles were split 4:1:1, each split spanning every biome. All tiles date from 2017 to 2019, before Sentinel-2 processing baseline 04.00, so no reflectance offset applies.

The script that builds it is examples/make_dataset.py.

Licence and attribution

Released under CC BY 4.0.

• Labels: Tait, A. M., Brumby, S. P., Hyde, S. B., Mazzariello, J., Corcoran, M. (2021). Dynamic World training dataset for global land use and land cover categorization of satellite imagery. PANGAEA, https://doi.org/10.1594/PANGAEA.933475. CC BY 4.0.
• Imagery: contains modified Copernicus Sentinel data (2017–2019), accessed through Microsoft Planetary Computer.

Brown, C. F., Brumby, S. P., Guzder-Williams, B. et al. Dynamic World, Near real-time global 10 m land use land cover mapping. Scientific Data 9, 251 (2022). https://doi.org/10.1038/s41597-022-01307-4