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

Technical Summary: Fields of The World (FTW)

Problem Statement
Agricultural monitoring at continental scales is critical for food security, land-use policy, and regulatory compliance (e.g., the EU Deforestation Regulation). However, the unit of analysis for tasks such as crop mapping, yield estimation, and subsidy allocation is the individual agricultural field. Manual digitization of field boundaries is not feasible at scale, necessitating automated approaches to extract these boundaries from satellite imagery. Existing methods often struggle with spectral variability and irregular parcel geometries, while previous deep learning models may lack the diversity required for global generalization.

Methodology
The paper introduces the Fields of The World (FTW) ecosystem, which integrates a benchmark dataset, pre-trained models, and command-line tools to facilitate end-to-end agricultural monitoring.

• Data Pipeline: The workflow utilizes bi-temporal Sentinel-2 imagery (Level-2A surface reflectance) representing planting and harvest windows. The system queries the Microsoft Planetary Computer STAC catalog to acquire cloud-free scenes. For each MGRS tile (100×100 km), seasonal windows are defined based on local crop calendars, and median composites are computed to reduce atmospheric artifacts.
• Field Boundary Extraction: The core segmentation models employ a U-Net architecture with an ImageNet-pretrained EfficientNet-B3 encoder. The models accept two 4-channel inputs (RGB + NIR at 10m resolution) from early and late-season observations to capture phenological differences. The training dataset spans 24 countries with field sizes ranging from <1 ha to >100 ha. The model outputs per-pixel probabilities for field interiors, boundaries, and background.
• Post-Processing: The ftw-tools package provides a command-line interface for inference. Post-processing involves thresholding softmax outputs at 0.5, applying morphological opening to remove small artifacts, and converting rasters to polygons via connected components. Optional filtering using land-cover masks (e.g., ESA WorldCover) removes non-agricultural areas like forests and water.
• Downstream Applications:
  • Crop Classification: The paper integrates FTW polygons with MOSAIKS (Random Convolutional Features). For each polygon, 256 random 3×3 filters are applied to Sentinel-2 pixels to generate 256-dimensional embeddings. A logistic regression classifier is trained on these embeddings using USDA Cropland Data Layer (CDL) labels.
  • Forest Loss Attribution: The system spatially joins FTW-derived polygons with the Hansen Global Forest Change (GFC) dataset (30m resolution, 2000–2023) to compute zonal statistics (mean, min, max loss year) per field, identifying parcels with recent deforestation relevant to EUDR compliance.
  • Country-Scale Inference: Pre-computed predictions for five countries (Japan, Mexico, Rwanda, South Africa, Switzerland) covering 4.76M km² are stored in cloud-optimized Zarr archives. These use sliding window inference (256×256 pixels with 25% overlap) and Gaussian-weighted averaging to suppress edge artifacts.

Key Contributions

1. FTW Ecosystem: A comprehensive resource comprising a benchmark of 1.6 million field polygons across 24 countries, pre-trained segmentation models, and open-source inference tools (ftw-tools).
2. Reproducible Workflows: Two Python notebooks demonstrating end-to-end processes: (1) local-scale field boundary extraction coupled with crop classification and forest loss attribution, and (2) country-scale inference and change detection using cloud-native data.
3. Few-Shot Capability: Demonstration that MOSAIKS embeddings derived from FTW polygons enable effective crop classification with limited labeled data.
4. Cloud-Native Access: Provision of pre-computed, country-scale predictions in Zarr archives, enabling lazy loading and spatial queries without full dataset downloads.

Results

• Crop Classification: On MGRS tile 14TPN (Iowa, USA), the system achieved macro F1 scores between 0.65 and 0.75 for crop type classification (corn, soybeans, grass/pasture, alfalfa) using limited labels. Training with only 10% of labeled fields yielded scores within 5% of full-data training. PCA of embeddings revealed separable clusters by crop type, with high accuracy for dominant classes (F1 > 0.8) and expected confusion among minority classes.
• Field Size Variation: Pre-computed predictions revealed median field areas ranging from 0.06 ha in Rwanda (reflecting smallholder systems) to 0.28 ha in Switzerland (reflecting mechanized systems).
• Change Detection: The system successfully detected year-over-year changes (2023 vs. 2024) by differencing field probability maps, flagging pixels exceeding the 99th percentile of absolute difference as potential agricultural expansion or abandonment.

Significance
The paper positions the FTW ecosystem as a foundational tool for scalable agricultural monitoring. It demonstrates that segmentation models trained on diverse global data can generalize across different geographies and climate zones. By integrating extracted field boundaries with downstream tasks like crop classification and forest loss monitoring, the ecosystem supports critical applications in food security and regulatory compliance (e.g., EUDR). The availability of pre-computed, cloud-optimized data allows researchers and practitioners to perform temporal analysis and spatial queries efficiently without the computational burden of re-running inference locally.