Note
This is a Hugging Face dataset. For large datasets, ensure huggingface_hub>=1.1.3 to avoid rate limits. Learn more in the Hugging Face integration docs.
Dataset Card for Fields of the World#
This is a FiftyOne dataset with 70,484 samples (chips), 140,968 total samples across the window_a/window_b group slices.
Installation#
If you haven’t already, install FiftyOne:
pip install -U fiftyone
Usage#
import fiftyone as fo
from fiftyone.utils.huggingface import load_from_hub
# Load the dataset
# Note: other available arguments include 'max_samples', etc
dataset = load_from_hub("Voxel51/fields-of-the-world")
# Launch the App
session = fo.launch_app(dataset)
Dataset Details#
Dataset Description#
Fields of The World (FTW) is a global machine learning benchmark dataset for agricultural field instance segmentation, spanning 25 countries across four continents (Europe, Africa, Asia, and South America). Each chip pairs two-date, four-band (red, green, blue, near-infrared) Sentinel-2 satellite imagery with instance and semantic (2-class and 3-class) field-boundary segmentation masks. FTW is an order of magnitude larger than prior field-boundary benchmarks such as AI4Boundaries, AI4SmallFarms, PASTIS, and PASTIS-R, with 70,462 sample chips (70,484 in this FiftyOne parsing after including a small number of “none”-split rows) covering roughly 166,293 km² and 1.63 million labeled field polygons.
This FiftyOne dataset was parsed from the official FTW archive distributed via Source Cooperative and the ftw-baselines repository’s ftw CLI tool.
Curated by: Hannah Kerner, Snehal Chaudhari, Aninda Ghosh, Caleb Robinson, Adeel Ahmad, Eddie Choi, Nathan Jacobs, Chris Holmes, Matthias Mohr, Rahul Dodhia, Juan M. Lavista Ferres, and Jennifer Marcus (Arizona State University, Microsoft AI for Good Research Lab, Taylor Geospatial Institute, Washington University in St Louis, Taylor Geospatial Engine)
Funded by: Taylor Geospatial Engine and a NASA Supplemental Open Source Software Award
Shared by: The Fields of The World project (data hosted on Source Cooperative; this FiftyOne parsing shared by Harpreet Sahota)
Language(s): Not applicable (satellite imagery and geospatial vector/raster data)
License: There is no single license for the dataset as a whole. FTW aggregates 24 source field-boundary datasets from different countries, each retaining its own original license (e.g. CC-BY, open government data licenses, etc.) — see the FTW paper’s Table 1 and each country’s README on Source Cooperative for the specific license and source link. The
ftw-baselinescodebase itself is released under the MIT License.
Dataset Sources#
Uses#
Direct Use#
FTW is intended for training and evaluating machine learning models for agricultural field instance and semantic segmentation from multi-temporal, multi-spectral satellite imagery, including: region-specific (per-country) model evaluation, cross-region transfer learning / fine-tuning, and zero-shot generalization studies across diverse global field morphologies and agro-climatic conditions.
Out-of-Scope Use#
The dataset excludes permanent/perennial crops (e.g. orchards, vineyards, forestry) and non-crop agricultural land uses (pasture, fallow land, grazing) — it is not intended for delineating those land-cover types. Per the original authors’ ethics statement, digitized field boundary data could inadvertently expose the practices and characteristics of individual land parcels; the dataset is not intended for uses that could infringe on landowner privacy or marginalize vulnerable individuals such as smallholder farmers. It is also not designed to determine legal land ownership or boundaries.
Dataset Structure#
This is a grouped FiftyOne dataset (media_type="group"). Each group corresponds to one 1536m Ă— 1536m sample chip (identified internally by an aoi_id, not stored as a field on this parsing), with two slices:
window_a(default slice) — early-season Sentinel-2 capturewindow_b— contrasting later-season Sentinel-2 capture
There are 70,484 groups (140,968 total samples across both slices) spanning 25 countries: Austria, Belgium, Brazil, Cambodia, Corsica, Croatia, Denmark, Estonia, Finland, France, Germany, India, Kenya, Latvia, Lithuania, Luxembourg, Netherlands, Portugal, Rwanda, Slovakia, Slovenia, South Africa, Spain, Sweden, and Vietnam. (The paper reports 24 countries plus Corsica treated as a separate region from mainland France, totaling 25 parsed regions here.) 333 of the 70,817 chips listed in the source parquet indices were not included because their underlying raster files were missing from the downloaded archive.
Each sample’s media is an 8-bit RGB JPEG preview, generated from the raw 4-band (R, G, B, NIR) uint16 Sentinel-2 GeoTIFF via a per-chip 2nd-98th percentile stretch (necessary since raw Sentinel-2 reflectance values only occupy a narrow low slice of the uint16 range and cannot be rendered directly). The original georeferenced GeoTIFFs and the NIR band are not included in this FiftyOne parsing.
Fields#
Field |
FiftyOne type |
Description |
|---|---|---|
|
|
Links the |
|
|
Country this chip belongs to (one of the 25 Fields of the World countries). Verbatim from source. |
|
|
Train/val/test split assigned in |
|
|
Which Sentinel-2 temporal window this slice captures: |
|
|
Start/end date (ISO format) of this window’s Sentinel-2 acquisition period, from |
|
|
Year the underlying field-boundary annotations were collected for this country, from |
|
|
Chip footprint centroid ( |
|
|
Fraction of pixels labeled “unknown” (unlabeled, no ground truth available) in |
|
|
Fraction of pixels labeled “field” in |
|
|
Binary field-extent mask: |
|
|
3-class field mask: |
|
|
Per-field instance segmentation: one |
3 (“unknown”) in both segmentation masks means “no ground truth available for this pixel,” not “confirmed background” — this distinction matters most for the 4 presence-only-labeled countries (brazil, india, kenya, rwanda), where only known field polygons were digitized and everything else is unconfirmed, versus the other 21 “full data” countries where backgrounds were exhaustively labeled.
Segmentation was used (rather than a flat integer field) because the source data is inherently a per-pixel raster mask; Detections (rather than a single multi-class Segmentation) was used for the instance layer because each field is a distinct, independently identifiable object with its own ID, area, and shape — matching FiftyOne’s standard instance-segmentation representation and enabling per-field browsing/filtering that a flat instance-ID raster would not.
dataset.mask_targets#
{
"ground_truth_2class": {0: "background", 1: "field", 3: "unknown"},
"ground_truth_3class": {0: "background", 1: "field", 2: "boundary", 3: "unknown"},
}
Saved views#
The dataset ships with 10 pre-built saved views surfacing its key structural distinctions: full_data_countries, presence_only_countries, mostly_unlabeled_chips, no_fields_detected, most_fields_per_chip, high_field_coverage, test_split, val_split, window_a_flat, and window_b_flat.
Parsing decisions#
Missing files: Chips whose parquet row lacked a complete set of 5 underlying raster files (2 Sentinel-2 windows + 3 masks) on disk were skipped, matching the same guard used in the official
ftw-baselinestraining pipeline.Per-grid season windows: 8 countries (belgium, cambodia, croatia, estonia, finland, france, germany, vietnam) store season date ranges as a list aligned to
grids[]rather than a single country-wide dict, since different grid cells were sourced/dated independently; these are matched to each chip via the grid ID prefix in itsaoi_id.Fields intentionally excluded from this parsing:
aoi_id(original chip identifier),raw_tif_path(path to the original raw GeoTIFF), andis_full_data(country-level presence/absence-labeling flag, superseded by the per-chipunknown_fractionfield and thefull_data_countries/presence_only_countriessaved views).RGB preview generation: a per-chip 2nd-98th percentile stretch was used (rather than a fixed intensity range) so that every chip is clearly visible in the FiftyOne App regardless of its absolute brightness; this stretched JPEG is a display-only asset and is not used for any quantitative analysis.
Dataset Creation#
Curation Rationale#
Crop field boundary datasets are foundational for agricultural monitoring, sustainable agriculture, and development applications, but existing ML-ready datasets were geographically concentrated (mostly in Europe) and too limited in scale and diversity to support models that generalize globally. FTW was curated to capture the dramatic diversity in field morphology, agricultural practices, and climate patterns across the globe (e.g. average field sizes range from 1.6 ha in Sub-Saharan Africa to 121 ha in North America), to be an order of magnitude larger than prior benchmarks, and to enable granular, country-scale evaluation of field-boundary segmentation models.
Source Data#
Data Collection and Processing#
The authors conducted a comprehensive search for field boundary polygons from government databases, published literature, and other sources, selecting 24 source datasets that had diverse geographic coverage, high-quality/trustworthy annotations, and licenses permitting reuse. Only annual (temporary) crop parcels were retained; permanent/perennial crops (orchards, vineyards) and non-crop agricultural land (pasture, fallow, forestry) were filtered out. For large, densely-annotated source datasets (mostly from EU government sources / EuroCrops), the authors sub-sampled by defining a grid over the source data’s bounding box and selecting 2-4 grid cells per country spanning a mix of high- and low-density agricultural areas; 8 source datasets with predefined grids, or too small/clustered to grid, were used without sub-sampling or via k-means clustering (Kenya, Brazil) or per-cluster chipping (India, Rwanda). Each sample grid was tiled into 1536m × 1536m chips.
All label polygons were standardized to the fiboa (field boundaries for agriculture) specification, with id, determination_datetime, area, and geometry (WGS84/EPSG:4326) attributes. Sentinel-2 imagery was obtained via Microsoft Planetary Computer at Level 2A (bottom-of-atmosphere), using the red (B04), green (B03), blue (B02), and near-infrared (B08) 10m bands. For each country, two contrasting acquisition date ranges (“Window A” and “Window B”) were chosen based on that country’s crop calendar (planting/mid-season vs. harvesting/off-season), then iteratively adjusted for image contrast and cloud cover; for each chip ROI, the authors searched for scenes with <90% scene-level cloud cover in each window and selected the chip-level crop with the lowest computed cloud percentage (using the Sentinel-2 scene classification layer), discarding chips with no candidate below 10% cloud cover. Chips were resized to 256×256 pixels and stored as EPSG:4326 GeoTIFFs. Field label polygons were rasterized into binary field-extent masks, 3-class (interior/boundary/background) masks, and instance masks. Train/validation/test splits were assigned per country using a blocked random splitting strategy (chips grouped into 3×3 blocks, 80%/10%/10% split) to reduce spatial-autocorrelation leakage between subsets.
Who are the source data producers?#
Field boundary polygons were produced by a mix of national/regional government agricultural agencies (e.g. EuroCrops-sourced ministries of agriculture across many EU countries, the Slovak Republic Government, The Swedish Agency for Agriculture, the Finnish Food Authority, Denmark’s Ministry of Food, Agriculture and Fisheries, the Netherlands Enterprise Agency, and others), research/humanitarian organizations (NASA Harvest and Radiant Earth Foundation for Rwanda, Persello et al./AI4SmallFarms for Cambodia and Vietnam), and other data providers (Pula Advisors for Kenya, Oldoni et al. for Brazil, Wang/Waldner/Lobell for India, Planet/Radiant Earth/DLR for South Africa). See Table 1 in the paper for the full per-country source list.
Annotations#
Annotation process#
Underlying field polygons were originally created by the source data producers via farmer self-reporting, manual digitization on high-resolution satellite imagery using GIS software, or handheld GPS perimeter walking — methods vary by source dataset and are documented in each dataset’s own README on Source Cooperative. The FTW authors then performed semantic filtering (retaining only annual-crop parcels), spatial sub-sampling/gridding, chipping into 1536m × 1536m ROIs, pairing with cloud-filtered two-date Sentinel-2 imagery, and rasterization of the resulting polygons into binary, 3-class, and instance segmentation masks.
Who are the annotators?#
The original field boundary polygons were digitized/reported by the source data producers listed above (government agencies, NGOs, and research groups), not by the FTW authors themselves. The FTW authors performed quality review (author-reported quality assessment, prior ML use, and visual inspection for closed polygons and satellite-image consistency) when selecting which of the candidate source datasets to include, and excluded several candidate datasets (e.g. Zambia, Romania, an earlier Kenya dataset, GloCAB, USA/California) that did not meet their quality or licensing criteria.
Personal and Sensitive Information#
Per the original authors’ ethics statement: digitized field boundary data could inadvertently expose the practices and characteristics of individual land parcels, which could infringe on the privacy of local landowners who may be unaware of the digitization or its implications. There is also a risk that private or public entities could use digitized field boundary data in ways that marginalize vulnerable individuals such as smallholder farmers. The original authors recommend that downstream users work with local organizations and communities to build and release responsible field boundary datasets aligned with local needs and regulations.
Citation#
BibTeX:
@article{kerner2024fields,
title={Fields of The World: A Machine Learning Benchmark Dataset For Global Agricultural Field Boundary Segmentation},
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 Lavista Ferres, Juan M. and Marcus, Jennifer},
journal={arXiv preprint arXiv:2409.16252},
year={2024}
}
APA:
Kerner, H., Chaudhari, S., Ghosh, A., Robinson, C., Ahmad, A., Choi, E., Jacobs, N., Holmes, C., Mohr, M., Dodhia, R., Lavista Ferres, J. M., & Marcus, J. (2024). Fields of The World: A Machine Learning Benchmark Dataset For Global Agricultural Field Boundary Segmentation. arXiv preprint arXiv:2409.16252.
More Information#
Official code, data loaders, and pre-trained models: github.com/fieldsoftheworld/ftw-baselines. Project site: fieldsofthe.world.