Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Fields of The World: A Field Guide for Extracting Agricultural Field Boundaries

Isaac Corley, Hannah Kerner, Caleb Robinson, Jennifer Marcus

TL;DR

The paper addresses the need for scalable, accurate field boundary extraction to support agricultural monitoring and deforestation screening. It introduces the Fields of The World (FTW) benchmark, comprising 1.6M field polygons across 24 countries, pre-trained segmentation models using bi-temporal Sentinel-2 imagery, and open-source inference tools. Two notebooks illustrate end-to-end workflows: local-scale boundary inference with crop classification and forest loss attribution, and country-scale exploration of pre-computed predictions stored in cloud-native Zarr archives. The results show macro-F1 scores of 0.65–0.75 for crop-type classification with limited labels, and demonstrate per-field forest loss attribution and efficient year-over-year change detection at continental scales, enabling practical, scalable agricultural monitoring and policy analysis.

Abstract

Field boundary maps are a building block for agricultural data products and support crop monitoring, yield estimation, and disease estimation. This tutorial presents the Fields of The World (FTW) ecosystem: a benchmark of 1.6M field polygons across 24 countries, pre-trained segmentation models, and command-line inference tools. We provide two notebooks that cover (1) local-scale field boundary extraction with crop classification and forest loss attribution, and (2) country-scale inference using cloud-optimized data. We use MOSAIKS random convolutional features and FTW derived field boundaries to map crop type at the field level and report macro F1 scores of 0.65--0.75 for crop type classification with limited labels. Finally, we show how to explore pre-computed predictions over five countries (4.76M km\textsuperscript{2}), with median predicted field areas from 0.06 ha (Rwanda) to 0.28 ha (Switzerland).

Fields of The World: A Field Guide for Extracting Agricultural Field Boundaries

TL;DR

The paper addresses the need for scalable, accurate field boundary extraction to support agricultural monitoring and deforestation screening. It introduces the Fields of The World (FTW) benchmark, comprising 1.6M field polygons across 24 countries, pre-trained segmentation models using bi-temporal Sentinel-2 imagery, and open-source inference tools. Two notebooks illustrate end-to-end workflows: local-scale boundary inference with crop classification and forest loss attribution, and country-scale exploration of pre-computed predictions stored in cloud-native Zarr archives. The results show macro-F1 scores of 0.65–0.75 for crop-type classification with limited labels, and demonstrate per-field forest loss attribution and efficient year-over-year change detection at continental scales, enabling practical, scalable agricultural monitoring and policy analysis.

Abstract

Field boundary maps are a building block for agricultural data products and support crop monitoring, yield estimation, and disease estimation. This tutorial presents the Fields of The World (FTW) ecosystem: a benchmark of 1.6M field polygons across 24 countries, pre-trained segmentation models, and command-line inference tools. We provide two notebooks that cover (1) local-scale field boundary extraction with crop classification and forest loss attribution, and (2) country-scale inference using cloud-optimized data. We use MOSAIKS random convolutional features and FTW derived field boundaries to map crop type at the field level and report macro F1 scores of 0.65--0.75 for crop type classification with limited labels. Finally, we show how to explore pre-computed predictions over five countries (4.76M km\textsuperscript{2}), with median predicted field areas from 0.06 ha (Rwanda) to 0.28 ha (Switzerland).
Paper Structure (13 sections, 5 figures)

This paper contains 13 sections, 5 figures.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methods
  4. Data Pipeline.
  5. Field Boundary Extraction.
  6. Crop Classification with MOSAIKS.
  7. Forest Loss Monitoring.
  8. Country-Scale Inference.
  9. Results
  10. Crop Classification.
  11. Forest Loss Attribution.
  12. Country-Scale Predictions.
  13. Conclusion

Figures (5)

  • Figure 1: FTW field boundary predictions for MGRS tile 14TPN in Iowa, USA. Blue polygons show extracted agricultural parcels overlaid on Sentinel-2 true-color imagery.
  • Figure 2: Crop classification results on Iowa fields: (left) PCA of 256-dim MOSAIKS embeddings colored by CDL crop type showing separable clusters; (center) macro F1 score versus training data fraction demonstrating few-shot capability; (right) confusion matrix revealing high accuracy for dominant crops (corn, soybeans) with more errors on minority classes.
  • Figure 3: Country-scale FTW inference for Japan. Left: Sentinel-2 RGB mosaic. Right: per-pixel field probability (softmax output). Data accessed via cloud-optimized Zarr archives.
  • Figure 4: Forest loss attribution for MGRS tile 21LXF in Mato Grosso, Brazil. Colors indicate mean deforestation year per field polygon from Hansen GFC data. Pink and Purple colors denote recent clearing (2012--2024), relevant to EUDR compliance thresholds.
  • Figure 5: Year-over-year change detection from FTW field probability maps. Panels show 2023 and 2024 field probability, detected changes ($\ge$99th percentile), and a change overlay on the 2023 map.