Global Fields of The World (FTW)

Global mosaics and agricultural field predictions for 2024 and 2025, from Taylor Geospatial.

Accompanying paper: https://aka.ms/ftw-global-paper

Data Products

• features (version alpha)
  • Planting and harvest median composites over 5–10 Sentinel-2 scenes.
  • Distributed as COGs and as a single EPSG:4326 Zarr V3 mosaic.
• predictions (version alpha)
  • Outputs of the PRUE model run on the features above.
  • Distributed as Zarr V3, GeoParquet v1.1.0 (with covering), and PMTiles.

---

Features (COGs)

Location: s3://us-west-2.opendata.source.coop/ftw/global-data/features/cogs/alpha/.

Features are defined by selecting DOY ranges as planting/harvest heuristics and computing the median of masked pixels across ~5–10 scenes. See Appendix for the heuristics and masking details.

• s2med_harvest/*:

  • Contains bands ["B02", "B03", "B04", "B08", "N_VALID_PIXELS"]
  • Each band is the median over the respective Sentinel-2 scenes.
  • N_VALID_PIXELS is the number of valid scenes after quality-flag masking.

• s2med_planting/*:

  • Contains bands ["B02", "B03", "B04", "B08", "N_VALID_PIXELS"]
  • Each band is the median over the respective Sentinel-2 scenes.
  • N_VALID_PIXELS is the number of valid scenes after quality-flag masking.

• index.parquet:

1import geopandas as gpd
2
3gpd.read_parquet("s3://us-west-2.opendata.source.coop/ftw/global-data/features/cogs/alpha/index.parquet")
1import geopandas as gpd
2
3gpd.read_parquet("s3://us-west-2.opendata.source.coop/ftw/global-data/features/cogs/alpha/index.parquet")

---

Features (Zarr)

Location: s3://us-west-2.opendata.source.coop/ftw/global-data/features/zarr/alpha/global.zarr

All feature COGs are reprojected and resampled to EPSG:4326 at 8.983119e-5° (~10 m at the equator) using GDAL cubic resampling, producing a single Zarr mosaic with dimensions (time, band, y, x).

1import rasterix
2import xarray as xr
3
4features = xr.open_zarr(
5    "s3://us-west-2.opendata.source.coop/ftw/global-data/features/zarr/alpha/global.zarr"
6).pipe(rasterix.assign_index)
7features
1import rasterix
2import xarray as xr
3
4features = xr.open_zarr(
5    "s3://us-west-2.opendata.source.coop/ftw/global-data/features/zarr/alpha/global.zarr"
6).pipe(rasterix.assign_index)
7features

---

Predictions (Zarr)

Location: s3://us-west-2.opendata.source.coop/ftw/global-data/predictions/zarr/alpha/global.zarr

The PRUE model is run over features/zarr/alpha/global.zarr to produce a Zarr dataset with bands [non_field_background, field, field_boundaries]. Feature and prediction Zarrs share the same grid, so they are stackable.

1predictions = xr.open_zarr(
2    "s3://us-west-2.opendata.source.coop/ftw/global-data/predictions/zarr/alpha/global.zarr"
3).pipe(rasterix.assign_index)
4predictions
1predictions = xr.open_zarr(
2    "s3://us-west-2.opendata.source.coop/ftw/global-data/predictions/zarr/alpha/global.zarr"
3).pipe(rasterix.assign_index)
4predictions

Features and Predictions

We can inspect inputs and outputs side-by-side since they are stackable. This enables researchers to validate inputs to better understand their influence in model outputs.

1planting_bands = ["s2med_planting:B04", "s2med_planting:B03", "s2med_planting:B02"]
2harvest_bands = ["s2med_harvest:B04", "s2med_harvest:B03", "s2med_harvest:B02"]
3xmin, ymin, xmax, ymax = -93.75, 41.5, -93.5, 41.75
4year = "2024"
5
6fig, axs = plt.subplots(1, 3, figsize=(15, 5))
7
8features["variables"].sel(
9    band=planting_bands, time=year, y=slice(ymax, ymin), x=slice(xmin, xmax)
10)[0].plot.imshow(robust=True, ax=axs[0])
11axs[0].axis("off")
12axs[0].set_title("Planting")
13
14features["variables"].sel(
15    band=harvest_bands, time=year, y=slice(ymax, ymin), x=slice(xmin, xmax)
16)[0].plot.imshow(robust=True, ax=axs[1])
17axs[1].axis("off")
18axs[1].set_title("Harvest")
19
20predictions["variables"].sel(band="field", time=year, y=slice(ymax, ymin), x=slice(xmin, xmax))[
21    0
22].plot.imshow(ax=axs[2])
23axs[2].axis("off")
24axs[2].set_title("Predictions")
25
26plt.tight_layout()
27plt.show()
1planting_bands = ["s2med_planting:B04", "s2med_planting:B03", "s2med_planting:B02"]
2harvest_bands = ["s2med_harvest:B04", "s2med_harvest:B03", "s2med_harvest:B02"]
3xmin, ymin, xmax, ymax = -93.75, 41.5, -93.5, 41.75
4year = "2024"
5
6fig, axs = plt.subplots(1, 3, figsize=(15, 5))
7
8features["variables"].sel(
9    band=planting_bands, time=year, y=slice(ymax, ymin), x=slice(xmin, xmax)
10)[0].plot.imshow(robust=True, ax=axs[0])
11axs[0].axis("off")
12axs[0].set_title("Planting")
13
14features["variables"].sel(
15    band=harvest_bands, time=year, y=slice(ymax, ymin), x=slice(xmin, xmax)
16)[0].plot.imshow(robust=True, ax=axs[1])
17axs[1].axis("off")
18axs[1].set_title("Harvest")
19
20predictions["variables"].sel(band="field", time=year, y=slice(ymax, ymin), x=slice(xmin, xmax))[
21    0
22].plot.imshow(ax=axs[2])
23axs[2].axis("off")
24axs[2].set_title("Predictions")
25
26plt.tight_layout()
27plt.show()

---

Predictions (GeoParquet)

Location: s3://us-west-2.opendata.source.coop/ftw/global-data/predictions/vectors/alpha/results/

A GeoParquet vector dataset is derived from the prediction Zarr by thresholding the softmax outputs for [non_field_background, field, field_boundaries] at 0.5 and polygonizing.

Files follow the GeoParquet v1.1.0 spec: ~8.2B rows across 1,001 files, ~629 GB on S3, with the following schema:

1┌─────────────┬────────────────────────────────────────────────────────────┬─────────┬─────────┬─────────┬─────────┐
2│ column_name │                        column_type                         │  null   │   key   │ default │  extra  │
3│   varchar   │                          varchar                           │ varchar │ varchar │ varchar │ varchar │
4├─────────────┼────────────────────────────────────────────────────────────┼─────────┼─────────┼─────────┼─────────┤
5│ geometry    │ GEOMETRY                                                   │ YES     │ NULL    │ NULL    │ NULL    │
6│ time        │ TIMESTAMP                                                  │ YES     │ NULL    │ NULL    │ NULL    │
7│ label       │ VARCHAR                                                    │ YES     │ NULL    │ NULL    │ NULL    │
8│ bbox        │ STRUCT(xmax DOUBLE, xmin DOUBLE, ymax DOUBLE, ymin DOUBLE) │ YES     │ NULL    │ NULL    │ NULL    │
9└─────────────┴────────────────────────────────────────────────────────────┴─────────┴─────────┴─────────┴─────────┘
1┌─────────────┬────────────────────────────────────────────────────────────┬─────────┬─────────┬─────────┬─────────┐
2│ column_name │                        column_type                         │  null   │   key   │ default │  extra  │
3│   varchar   │                          varchar                           │ varchar │ varchar │ varchar │ varchar │
4├─────────────┼────────────────────────────────────────────────────────────┼─────────┼─────────┼─────────┼─────────┤
5│ geometry    │ GEOMETRY                                                   │ YES     │ NULL    │ NULL    │ NULL    │
6│ time        │ TIMESTAMP                                                  │ YES     │ NULL    │ NULL    │ NULL    │
7│ label       │ VARCHAR                                                    │ YES     │ NULL    │ NULL    │ NULL    │
8│ bbox        │ STRUCT(xmax DOUBLE, xmin DOUBLE, ymax DOUBLE, ymin DOUBLE) │ YES     │ NULL    │ NULL    │ NULL    │
9└─────────────┴────────────────────────────────────────────────────────────┴─────────┴─────────┴─────────┴─────────┘

Query with DuckDB and visualize with Lonboard:

1import duckdb
2from lonboard import Map, PolygonLayer
3from lonboard.basemap import CartoStyle
4
5con = duckdb.connect()
6con.execute("INSTALL spatial; LOAD spatial;")
7con.execute("INSTALL httpfs; LOAD httpfs;")
8
9q = """
10SELECT geometry AS geometry, time, label, bbox
11FROM read_parquet('s3://us-west-2.opendata.source.coop/ftw/global-data/predictions/vectors/alpha/results/*.parquet')
12WHERE label = 'field'
13  AND struct_extract(bbox, 'xmax') >= -93.71488
14  AND struct_extract(bbox, 'xmin') <= -93.06492
15  AND struct_extract(bbox, 'ymax') >= 41.78201
16  AND struct_extract(bbox, 'ymin') <= 42.09459
17"""
18
19layer = PolygonLayer.from_duckdb(q, con, get_fill_color=[0, 0, 139])
20m = Map(layer, height=700, basemap_style=CartoStyle.Positron)
21m
1import duckdb
2from lonboard import Map, PolygonLayer
3from lonboard.basemap import CartoStyle
4
5con = duckdb.connect()
6con.execute("INSTALL spatial; LOAD spatial;")
7con.execute("INSTALL httpfs; LOAD httpfs;")
8
9q = """
10SELECT geometry AS geometry, time, label, bbox
11FROM read_parquet('s3://us-west-2.opendata.source.coop/ftw/global-data/predictions/vectors/alpha/results/*.parquet')
12WHERE label = 'field'
13  AND struct_extract(bbox, 'xmax') >= -93.71488
14  AND struct_extract(bbox, 'xmin') <= -93.06492
15  AND struct_extract(bbox, 'ymax') >= 41.78201
16  AND struct_extract(bbox, 'ymin') <= 42.09459
17"""
18
19layer = PolygonLayer.from_duckdb(q, con, get_fill_color=[0, 0, 139])
20m = Map(layer, height=700, basemap_style=CartoStyle.Positron)
21m

---

Predictions (PMTiles)

Location: s3://us-west-2.opendata.source.coop/ftw/global-data/predictions/vectors/alpha/global.pmtiles

PMTiles built from the GeoParquet above for scalable browser-side visualization.

1import leafmap.foliumap as leafmap
2
3url = "s3://us-west-2.opendata.source.coop/ftw/global-data/predictions/vectors/alpha/global.pmtiles"
4
5metadata = leafmap.pmtiles_metadata(url)
6m = leafmap.Map(center=[0, 20], zoom=2, height="800px")
7m.add_basemap("SATELLITE", visible=True)
8style = leafmap.pmtiles_style(url, layers="field-2024-01-01 00:00:00", opacity=0.7)
9m.add_pmtiles(url, style=style)
10m
1import leafmap.foliumap as leafmap
2
3url = "s3://us-west-2.opendata.source.coop/ftw/global-data/predictions/vectors/alpha/global.pmtiles"
4
5metadata = leafmap.pmtiles_metadata(url)
6m = leafmap.Map(center=[0, 20], zoom=2, height="800px")
7m.add_basemap("SATELLITE", visible=True)
8style = leafmap.pmtiles_style(url, layers="field-2024-01-01 00:00:00", opacity=0.7)
9m.add_pmtiles(url, style=style)
10m

---

Appendix

Season Heuristics (alpha)

Relevant Sentinel-2 scenes are selected by acquisition DOY using the following functions:

1def tile_to_planting_season_heuristic_doy(point: Point) -> tuple[int, int]:
2    lat = point.y
3    abs_lat = abs(lat)
4    if abs_lat > 45:
5        start_doy, end_doy = (91, 151) if lat > 0 else (274, 334)  # Apr-May or Oct-Nov
6    elif 20 < abs_lat <= 45:
7        start_doy, end_doy = (60, 120) if lat > 0 else (244, 334)  # Mar-Apr or Sep-Nov
8    elif 5 < abs_lat <= 20:
9        start_doy, end_doy = (121, 212) if lat > 0 else (305, 365)  # May-Jul or Nov-Dec
10    else:  # Equatorial
11        start_doy, end_doy = (60, 121)  # Approx. Mar-Apr and Sep-Oct (simplified)
12    return start_doy, end_doy
13
14
15def tile_to_harvest_season_heuristic_doy(point: Point) -> tuple[int, int]:
16    lat = point.y
17    abs_lat = abs(lat)
18    if abs_lat > 45:
19        start_doy, end_doy = (244, 304) if lat > 0 else (60, 151)  # Sep-Oct or Mar-May
20    elif 20 < abs_lat <= 45:
21        start_doy, end_doy = (213, 304) if lat > 0 else (32, 120)  # Aug-Oct or Feb-Apr
22    elif 5 < abs_lat <= 20:
23        start_doy, end_doy = (274, 365) if lat > 0 else (91, 181)  # Oct-Dec or Apr-Jun
24    else:  # Equatorial
25        start_doy, end_doy = (182, 243)  # Approx. Jul-Aug and Jan-Feb (simplified)
26    return start_doy, end_doy
1def tile_to_planting_season_heuristic_doy(point: Point) -> tuple[int, int]:
2    lat = point.y
3    abs_lat = abs(lat)
4    if abs_lat > 45:
5        start_doy, end_doy = (91, 151) if lat > 0 else (274, 334)  # Apr-May or Oct-Nov
6    elif 20 < abs_lat <= 45:
7        start_doy, end_doy = (60, 120) if lat > 0 else (244, 334)  # Mar-Apr or Sep-Nov
8    elif 5 < abs_lat <= 20:
9        start_doy, end_doy = (121, 212) if lat > 0 else (305, 365)  # May-Jul or Nov-Dec
10    else:  # Equatorial
11        start_doy, end_doy = (60, 121)  # Approx. Mar-Apr and Sep-Oct (simplified)
12    return start_doy, end_doy
13
14
15def tile_to_harvest_season_heuristic_doy(point: Point) -> tuple[int, int]:
16    lat = point.y
17    abs_lat = abs(lat)
18    if abs_lat > 45:
19        start_doy, end_doy = (244, 304) if lat > 0 else (60, 151)  # Sep-Oct or Mar-May
20    elif 20 < abs_lat <= 45:
21        start_doy, end_doy = (213, 304) if lat > 0 else (32, 120)  # Aug-Oct or Feb-Apr
22    elif 5 < abs_lat <= 20:
23        start_doy, end_doy = (274, 365) if lat > 0 else (91, 181)  # Oct-Dec or Apr-Jun
24    else:  # Equatorial
25        start_doy, end_doy = (182, 243)  # Approx. Jul-Aug and Jan-Feb (simplified)
26    return start_doy, end_doy

Sentinel-2 Source

All Sentinel-2 scenes were sourced from s3://sentinel-cogs/sentinel-s2-l2a-cogs.

Sentinel-2 SCL Flags

Pixels with the following SCL values are masked out before taking the median:

• 0: No Data
• 1: Saturated Defective
• 3: Cloud Shadow
• 7: Cloud Low Probability / Unclassified
• 8: Cloud Medium Probability
• 9: Cloud High Probability
• 10: Thin Cirrus

License

CC-BY-4.0

Contact

isaac.corley@taylorgeospatial.org

Cite

If you use this dataset in your research please cite the following papers:

1@inproceedings{kerner2025fields,
2  title={Fields of the world: A machine learning benchmark dataset for global agricultural field boundary segmentation},
3  author={Kerner, Hannah and Chaudhari, Snehal and Ghosh, Aninda and Robinson, Caleb and Ahmad, Adeel and Choi, Eddie and Jacobs, Nathan and Holmes, Chris and Mohr, Matthias and Dodhia, Rahul and others},
4  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
5  volume={39},
6  number={27},
7  pages={28151--28159},
8  year={2025}
9}
1@inproceedings{kerner2025fields,
2  title={Fields of the world: A machine learning benchmark dataset for global agricultural field boundary segmentation},
3  author={Kerner, Hannah and Chaudhari, Snehal and Ghosh, Aninda and Robinson, Caleb and Ahmad, Adeel and Choi, Eddie and Jacobs, Nathan and Holmes, Chris and Mohr, Matthias and Dodhia, Rahul and others},
4  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
5  volume={39},
6  number={27},
7  pages={28151--28159},
8  year={2025}
9}

1@article{muhawenayo2026prue,
2  title={PRUE: A Practical Recipe for Field Boundary Segmentation at Scale},
3  author={Muhawenayo, Gedeon and Robinson, Caleb and Khanal, Subash and Fang, Zhanpei and Corley, Isaac and Wollam, Alexander and Gao, Tianyi and Strnad, Leonard and Avery, Ryan and Estes, Lyndon and others},
4  journal={arXiv preprint arXiv:2603.27101},
5  year={2026}
6}
1@article{muhawenayo2026prue,
2  title={PRUE: A Practical Recipe for Field Boundary Segmentation at Scale},
3  author={Muhawenayo, Gedeon and Robinson, Caleb and Khanal, Subash and Fang, Zhanpei and Corley, Isaac and Wollam, Alexander and Gao, Tianyi and Strnad, Leonard and Avery, Ryan and Estes, Lyndon and others},
4  journal={arXiv preprint arXiv:2603.27101},
5  year={2026}
6}

1@article{corley2026fields,
2  title={Fields of The World: A Field Guide for Extracting Agricultural Field Boundaries},
3  author={Corley, Isaac and Kerner, Hannah and Robinson, Caleb and Marcus, Jennifer},
4  journal={arXiv preprint arXiv:2602.08131},
5  year={2026}
6}
1@article{corley2026fields,
2  title={Fields of The World: A Field Guide for Extracting Agricultural Field Boundaries},
3  author={Corley, Isaac and Kerner, Hannah and Robinson, Caleb and Marcus, Jennifer},
4  journal={arXiv preprint arXiv:2602.08131},
5  year={2026}
6}