ftw-prue-efnet-b7-ccby

A semantic segmentation model for agricultural field boundary delineation from Sentinel-2 satellite imagery. A semantic segmentation model for agricultural field boundary delineation from Sentinel-2 satellite imagery. Part of the Fields of the World (FTW) initiative, this model uses a UNet architecture with EfficientNet-B7 encoder to identify field interiors, field boundaries, and background from bi-temporal Sentinel-2 inputs. The model is trained on the FTW dataset — a large-scale benchmark dataset designed to advance machine learning models for instance segmentation of agricultural field boundaries. The dataset supports the need for accurate and scalable field boundary data, which is essential for global agricultural monitoring, land use assessments, and environmental studies.

Training Methodology

The model uses the PRUE (Pre-trained, Regularized, UNet Ensemble) training methodology with an EfficientNet-B7 encoder from the segmentation_models_pytorch library. This CC-BY version is trained using the prue_efnet7.yaml config with non-commercial licensed countries removed from the training dataset, making it suitable for commercial use.

Key Features

• Global coverage — Trained on field annotations from 24 countries across 4 continents
• Bi-temporal input — Uses two Sentinel-2 windows (planting and harvesting seasons) to capture seasonal crop variation
• 3-class output — Background (0), field interior (1), field boundary (2)
• 10m resolution — Operates on Sentinel-2 Level-2A imagery (R, G, B, NIR bands)
• CC-BY licensed — Free for commercial and non-commercial use with attribution

Dataset

The model is trained on the Fields of the World dataset , which includes:

• 24 countries across Africa, Asia, Europe, and South America
• 1.6 million field annotations from government and open-source datasets
• Sentinel-2 L2A imagery with bi-temporal (window A / window B) pairs
• 3-class semantic segmentation labels — background, field interior, field boundary

Source

• Repository: fieldsoftheworld/ftw-baselines
• Checkpoint: v3.1 release (CC-BY)
• Dataset: Fields of the World
• Paper: Kerner et al., "Fields of the World: A Machine Learning Benchmark Dataset for Global Agricultural Field Boundary Delineation"
• Code License: MIT
• Checkpoint License: CC-BY 4.0 International