Import Copernicus Global Land Cover 100m (CGLS-LC100 v3, 2019)

[cboettig]

Dataset

Copernicus Global Land Service — Land Cover 100m Collection 3 (CGLS-LC100)

• Source: Zenodo — https://zenodo.org/records/3939050
• Resolution: 100 m, WGS84 (EPSG:4326)
• Temporal coverage: Annual epochs 2015–2019 (start with 2019 baseline)
• License: CC-BY 4.0
• Classes: ~23 discrete classes + continuous fractional layers (tree cover %, grass cover %, etc.)
• Accuracy: ~80% overall — highest of major global land cover products
• Format: Single global GeoTIFF per epoch (~1.7 GB for discrete classification)

Why CGLS-LC100 over alternatives

Dataset	Resolution	Accuracy	Direct download?
CGLS-LC100 (recommended)	100 m	~80%	✅ Single GeoTIFF
ESA WorldCover 2021	10 m	~75%	⚠️ 2,651 tiles (VRT needed)
ESA CCI LC	300 m	~72%	⚠️ NetCDF, CDS API
MODIS MCD12Q1	500 m	~73%	⚠️ HDF4 sinusoidal
Dynamic World	10 m	~75%	❌ GEE only

100 m resolution is the sweet spot for H3 resolution 8 (~74 source pixels per cell — good for mode aggregation). Single global WGS84 GeoTIFF means zero tiling/reprojection complexity before COG conversion.

Download URL (2019 discrete classification)

https://zenodo.org/api/records/3939050/files/PROBAV_LC100_global_v3.0.1_2019-nrt_Discrete-Classification-map_EPSG-4326.tif/content

Deliverables

• Upload raw GeoTIFF to s3://public-land-cover/raw/
• Convert to COG (gdal_translate with DEFLATE/ZSTD)
• Run cng-datasets raster-workflow at H3 resolution 8, --value-column lc_class
• STAC collection with class legend documented in table:columns

Future extension

Epochs 2015–2018 could be added later for temporal analysis (same Zenodo record, same URL pattern). The discrete classification band uses the FAO land cover class codes — document all values in STAC.

[boettiger-lab-llm-agent]

Processing note: When hexing this dataset, include parent resolutions up to h5 (e.g. --parent-resolutions "5,6,7,8,9,0" or whatever subset applies). Parent resolutions are cheap to compute alongside the base resolution and significantly speed up spatial joins at coarser scales. For lower-resolution datasets (e.g. base res h6), still include h5 as a parent. Minimum: always include h5 and h0.