First 10m global farm boundary map from Sentinel-2 U-Net
3.17 billion polygons across 241 countries using AI segmentation.
A team led by Caleb Robinson (Microsoft) and collaborators from academic institutions has created the first openly available, globally consistent agricultural field boundary map at 10m resolution. Published on arXiv (2605.11055), the work applies a U-Net segmentation model to cloud-free Sentinel-2 satellite mosaics, using the Fields of The World training dataset. The result is 3.17 billion polygons (1.62 billion for 2024, 1.55 billion for 2025) spanning 241 countries and territories. This addresses a critical gap: previously, field-level data existed only for specific regions like Europe or individual countries, while global remote-sensing products offered only pixel-level analysis.
Validation against ground truth from 24 countries yielded a mean pixel-level recall of 0.85, with 14 countries exceeding 0.90. Full-country evaluations in Austria, Latvia, and Finland produced F1 scores of 0.89, 0.88, and 0.74 respectively. To account for incomplete reference data, the team also provides a 500m confidence layer to indicate reliability. The dataset is released as three global maps: a confidence-thresholded default, the full unfiltered set, and a continuous confidence raster. This resource enables field-level crop monitoring, food security assessments, and downstream agricultural science at an unprecedented scale.
- 3.17 billion polygons (2024-2025) covering 241 countries at 10m resolution
- U-Net model trained on Fields of The World dataset applied to Sentinel-2 mosaics
- Mean pixel recall 0.85 across 24 countries; F1 scores of 0.89 (Austria), 0.88 (Latvia), 0.74 (Finland)
Why It Matters
Enables global field-level crop monitoring and food security analysis for the first time.