Research map / Detection and segmentation / Open-set and few-shot detection
Fields of The World: A Machine Learning Benchmark Dataset for Global Agricultural Field Boundary Segmentation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces Fields of The World (FTW), a 70k-sample Sentinel-2 benchmark for agricultural field instance segmentation across 24 countries, with U-Net baselines showing that pre-training on diverse FTW countries improves zero-shot and fine-tuned transfer.
Problem
Automatic extraction of crop field boundaries from satellite imagery lacks geographically diverse, labeled datasets and generalizable models, as existing benchmarks are concentrated in a few European countries.
Approach
A globally diverse benchmark of Sentinel-2 two-window (multi-temporal) chips with instance, 2-class and 3-class masks harmonized to the fiboa specification, plus baselines showing that models pre-trained on diverse countries transfer better to held-out regions.
Key results
- Models pre-trained on FTW minus target country outperform AI4Boundaries-pretrained models on India (zero-shot pixel IoU 0.57 vs 0.16) and Cambodia+Vietnam (zero-shot pixel IoU 0.43 vs 0.16).
- 3-class masks with ignored presence-only background give higher object recall (e.g., Slovenia 0.20 vs 0.07 for 3-class vs 2-class ignoring bg setting) and were adopted as the default.
- Best baseline U-Net with EfficientNet-b3 reaches pixel IoU 0.83 in France and South Africa using 2-class masks and 0.59 in Slovenia.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Pre-train on FTW excluding the target country rather than a smaller geographic subset | U-net + EfficientNet-b3 transfer learning · India test | Pixel IoU: 0.51 / 0.50 -> 0.57 / 0.55; 0.16 / 0.54 -> 0.57 / 0.55 | Helped |
| Pre-train on FTW excluding the target countries rather than a smaller geographic subset | U-net + EfficientNet-b3 transfer learning · Cambodia + Vietnam test | Pixel IoU: 0.04 / 0.48 -> 0.43 / 0.55; 0.16 / 0.51 -> 0.43 / 0.55 | Helped |
| Use 3-class masks rather than 2-class masks (ablation) | U-net + EfficientNet-b3 · Slovenia, France, South Africa | Object recall: SVN 0.06 -> 0.07; FRA 0.29 -> 0.45; ZAF 0.24 -> 0.35 | Helped |
| Ignore presence-only background pixels during training with 3-class masks (ablation) | U-net + EfficientNet-b3 · Slovenia, France, South Africa | Object recall: SVN 0.07 -> 0.20; FRA 0.45 -> 0.58; ZAF 0.35 -> 0.55 | Mixed |
| Remove the NIR channel (ablation) | U-net + EfficientNet-b3 with stacked Windows A and B · Slovenia, France, South Africa | Pixel IoU: 0.58 -> 0.58; 0.79 -> 0.78; 0.80 -> 0.79 | Hurt |
| Stack rather than average Windows A and B (ablation) | U-net + EfficientNet-b3 · Slovenia, France, South Africa | Pixel IoU: 0.54 -> 0.58; 0.77 -> 0.79; 0.78 -> 0.80 | Helped |
| Add Window B to Window A (ablation) | U-net + EfficientNet-b3 · Slovenia, France, South Africa | Pixel IoU: 0.55 -> 0.58; 0.77 -> 0.79; 0.78 -> 0.80 | Helped |
| Add Window A to Window B (ablation) | U-net + EfficientNet-b3 · Slovenia, France, South Africa | Pixel IoU: 0.52 -> 0.58; 0.78 -> 0.79; 0.79 -> 0.80 | Helped |
| Use U-net rather than DeepLabv3+ (ablation) | Semantic segmentation models with EfficientNet-b3 · Slovenia, France, South Africa | Pixel IoU: 0.50 -> 0.59; 0.77 -> 0.79; 0.78 -> 0.80 | Helped |
| Use a different architecture or backbone (ablation) | Field boundary segmentation models · Slovenia, France, South Africa | Pixel IoU | Mixed |
Surprising findings
- FTW-pretrained models without any fine-tuning perform similar to or better than fully fine-tuned models pre-trained on AI4Boundaries countries.
- Object recall is low for 2-class masks (e.g., 0.06 in Slovenia), so 3-class masks substantially improve instance delineation despite lower object precision.
Limitations
- Some candidate datasets (GloCAB, Zambia, Romania, Brazil CAR) were excluded due to label quality or licensing issues, limiting coverage.
- Only limited modeling configurations were explored and no hyperparameter tuning was performed.
- Ethiopia deployment assessment is qualitative only, with no quantitative metrics.
Experimental setup
Datasets: FTW (24 country sources including EuroCrops-derived data, India, Kenya, Brazil, Rwanda, Cambodia, Vietnam) AI4Boundaries AI4SmallFarms PASTIS PASTIS-R Sentinel-2 via Microsoft Planetary Computer
Benchmarks: FTW per-country test sets India Cambodia+Vietnam
Metrics: pixel IoU pixel precision pixel recall object precision object recall
Baselines: U-Net DeepLabv3+ AI4Boundaries-pretrained models France-pretrained models Netherlands-pretrained models
Builds on: U-Net DeepLabv3+ EuroCrops fiboa
Closest papers by meaning
- Can Large Multimodal Models Understand Agricultural Scenes? Benchmarking with AgroMind (NeurIPS 2025 · similarity 0.89)
- GTPBD: A Fine-Grained Global Terraced Parcel and Boundary Dataset (NeurIPS 2025 · similarity 0.87)
- Combining Deep Learning and Street View Imagery to Map Smallholder Crop Types (AAAI 2024 · similarity 0.87)
- HarvestNet: A Dataset for Detecting Smallholder Farming Activity Using Harvest Piles and Remote Sensing (AAAI 2024 · similarity 0.86)
- Exact: Exploring Space-Time Perceptive Clues for Weakly Supervised Satellite Image Time Series Semantic Segmentation (CVPR 2025 · similarity 0.85)
- Towards Open-Vocabulary Remote Sensing Image Semantic Segmentation (AAAI 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.