Sponsored by Taylor Geospatial, the Global Fields of The World (FTW) dataset provides global-scale estimates of agricultural fields for 2024–2025. The dataset includes both model inputs (Sentinel-2–derived median composites in COG and Zarr v3 formats) and outputs (in Zarr, GeoParquet and PMTiles).
Sentinel-2 planting- and harvest-season median composites at 10 m — the model-input features for the FTW Global field-boundary predictions. For each grid tile and season, ~5–10 quality-masked Sentinel-2 scenes (selected by latitude-based day-of-year heuristics) are reduced to a per-pixel median; bands are B02/B03/B04/B08 plus N_VALID_PIXELS, in the tile's native UTM zone. Part of Fields of the World. These composites feed the PRUE U-Net model (Muhawenayo et al. 2026; 76% IoU / 47% object-F1 on the Fields of the World benchmark of 70,462 Sentinel-2 samples across 24 countries, Kerner et al. 2025), whose outputs are the field-boundary predictions. The same composites are published in two equivalent formats: per-tile COGs (these per-year collections) and a single global EPSG:4326 Zarr mosaic (its own collection), stackable with the prediction Zarr.
Items & index: ~22.7k per-tile items (each with a planting and harvest COG asset) are indexed by the collection's STAC-GeoParquet asset (items.parquet); query that rather than enumerating item links.
Per-tile COGs are indexed by the STAC-GeoParquet item index (https://data.source.coop/ftw/global-data/features/2024/items.parquet); each item has a planting and a harvest 5-band COG (B02/B03/B04/B08/N_VALID_PIXELS) in its native UTM zone. The same data is also published as a global Zarr mosaic, and these composites are the model inputs for the field-boundary predictions.
CC-BY-4.0. Sentinel-2 imagery © Copernicus/ESA; composites by the Taylor Geospatial Institute and Microsoft AI for Good Research Lab.