PRUE: A Practical Recipe for Field Boundary Segmentation at Scale

Gedeon Muhawenayo; Caleb Robinson; Subash Khanal; Zhanpei Fang; Isaac Corley; Alexander Wollam; Tianyi Gao; Leonard Strnad

cs.CVMar 28, 2026

PRUE: A Practical Recipe for Field Boundary Segmentation at Scale

Authors: Gedeon Muhawenayo, Caleb Robinson, Subash Khanal, Zhanpei Fang, Isaac Corley, Alexander Wollam, Tianyi Gao, Leonard Strnad, +5 more

Organizations: Arizona State University · Microsoft AI for Good · Washington University in St. Louis · Oregon State University · Wherobots · Clark University

Abstract

Large-scale maps of field boundaries are essential for agricultural monitoring tasks. Existing deep learning approaches for satellite-based field mapping are sensitive to illumination, spatial scale, and changes in geographic location. We conduct the first systematic evaluation of segmentation and geospatial foundation models (GFMs) for global field boundary delineation using the Fields of The World (FTW) benchmark. We evaluate 18 models under unified experimental settings, showing that a U-Net semantic segmentation model outperforms instance-based and GFM alternatives on a suite of performance and deployment metrics. We propose a new segmentation approach that combines a U-Net backbone, composite loss functions, and targeted data augmentations to enhance performance and robustness under real-world conditions. Our model achieves a 76% IoU and 47% object-F1 on FTW, an increase of 6% and 9% over the previous baseline. Our approach provides a practical framework for reliable, scalable, and reproducible field boundary delineation across model design, training, and inference. We release all models and model-derived field boundary datasets for five countries.

Figures & tables

Figure 2 : Example visualization of predictions over Illinois, USA (top, MGRS tile 16TDL) and Mato Grosso, Brazil (bottom, MGRS tile 20LRQ) with the FTW Baseline, Terramind and PRUE. The FTW Baseline and Terramind models show strong sensitivity to scene characteristics, producing discontinuous and noisy field boundaries. The PRUE model achieves stable boundaries across regions and imaging conditions.

Figure 3 : Spatial consistency. Four overlapping crops from each image corner are independently segmented. The consistency mask shows pixel-level agreement, with yellow indicating unanimity across all four predictions, and purple disagreement. This metric quantifies grid artifact resistance for large-scale field delineation.

Semantic segmentation baselines
Model	Backbone	Pixel-level			Object-level					#Params	Throughput
Model	Backbone	IoU $\uparrow$	Prec $\uparrow$	Recall $\uparrow$	Prec $\uparrow$	Recall $\uparrow$	F1 $\uparrow$	AP 0.5:0.95 $\uparrow$	AP 0.5 $\uparrow$	(M) $\downarrow$	(km 2 /s) $\uparrow$
FTW-Baseline	U-Net+EfficientNet-B3	0.70	0.90	0.72	0.40	0.37	0.38	0.22	0.39	13.2	623.28
DECODE	FracTALResUNet	0.71	0.83	0.83	0.27	0.17	0.21	0.09	0.17	64.8	113.47
Instance & panoptic segmentation
Mask2Former	Swin-S (fine-tuned, 8ch)	0.68	0.88	0.75	0.62	0.30	0.39	0.28	0.44	68.8	26.66

Table 1 : Performance comparison across model families on FTW test set (excluding presence-only countries). Semantic baselines: Post-processed with connected components, using presence-only label masking in training [ 34 ] . Instance/panoptic models: Fine-tuned for 8-channel input and FTW-specific classes but without presence-only masking due to architecture constraints. GFM models: Frozen encoders with our trained convolutional decoder. Total parameter count includes both the frozen encoder (28M–300M) and our learnable components, which consist of a frames-fusion module and a convolutional decoder (an additional 30M–110M parameters, depending on the encoder’s output dimensionality). Bold indicates best performance, underline indicates second-best. Throughput measured on a V100-32GB GPU with batch size 64. * Galileo did not fit into VRAM at any batch size.

Category	Ablation	Performance		Input order		Brightness		Scale		Agree.
Category	Ablation	Object F1 $\uparrow$	Pixel IoU $\uparrow$	F1 $\|\Delta\|\downarrow$	IoU $\|\Delta\|\downarrow$	F1 $\|\Delta\|\downarrow$	IoU $\|\Delta\|\downarrow$	F1 $\|\Delta\|\downarrow$	IoU $\|\Delta\|\downarrow$	Avg $\uparrow$
	FTW-Baseline	0.39 $\pm$ 0.08	0.68 $\pm$ 0.08	0.07	0.11	0.04	0.05	0.15	0.12	0.93
Data augs	Brightness+Resize	0.38 $\pm$ 0.08	0.66 $\pm$ 0.09	0.06	0.10	0.02	0.03	0.00	0.01	0.95
Data augs	Channel shuffle	0.39 $\pm$ 0.07	0.68 $\pm$ 0.09	0.00	0.00	0.04	0.05	0.17	0.14	0.94
Class weights	$\omega=0.75$	0.42 $\pm$ 0.06	0.74 $\pm$ 0.07	0.08	0.11	0.07	0.07	0.29	0.15	0.95
Encoder	EfficientNet-B7	0.42 $\pm$ 0.07	0.71 $\pm$ 0.08	0.07	0.09	0.03	0.04	0.20	0.13	0.94

Table 2 : Ablation results for controlled experiments on FTW test set (excluding presence-only countries) in which each row varies a single design choice (data augmentations, class weighting, encoder, loss function, or architecture). The Combination rows report the best-performing joint configurations for the FCSiam and U-Net models. Bold indicates best performance, underline indicates second-best.

Country	Area processed (million km 2 )	Cost ($)	Year	Fields (M)	Median area (ha)
Rwanda	0.02	3.23	2023	0.18	0.05
Rwanda	0.02	3.23	2024	0.26	0.06
Switzerland	0.09	5.11	2023	0.36	0.32
Switzerland	0.09	5.11	2024	0.36	0.28
Japan	0.65	7.59	2023	1.55	0.20
Japan	0.65	7.59	2024	1.61	0.19

Table 3 : Agricultural field statistics derived using PRUE . The table reports total land area, distributed inference cost, field counts, and median field area in hectares for five diverse countries.

Figure 4 : PCA visualization of frozen GFM encoder features for a representative subset of image examples. The top 3 principal components of patch embeddings are displayed as RGB channels. Clay ( $8\times 8$ patch sizes) and Galileo ( $4\times 4$ patch size) capture finer spatial structure and more distinct field boundaries compared to models using $16\times 16$ patches, demonstrating how tokenization granularity affects feature quality for segmentation tasks. See Table 4 for quantitative performance.

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Model	Backbone	Pixel-level			Object-level
Model	Backbone	IoU $\uparrow$	Prec $\uparrow$	Recall $\uparrow$	Prec $\uparrow$	Recall $\uparrow$	F1 $\uparrow$	AP 0.5:0.95 $\uparrow$	AP 0.5 $\uparrow$
Clay	ViT-Large	0.56	0.88	0.60	0.22	0.16	0.18	0.07	0.17
Galileo	ViT-Base	0.53	0.83	0.59	0.11	0.19	0.13	0.08	0.18
DINOv3	ViT-Large	0.47	0.89	0.50	0.25	0.09	0.12	0.03	0.08
Prithvi 2.0	ViT-Large	0.44	0.84	0.48	0.20	0.06	0.10	0.02	0.06
TerraMind	ViT-Base	0.44	0.85	0.47	0.19	0.07	0.10	0.02	0.06

Appendix

Table 4 : GFM linear probing results using a lightweight decoder (1

\times

1 convolution + bilinear upsampling), sorted by object-level F1. Clay (ViT-Large, 8

\times

8 patches) and Galileo (ViT-Base, 4

\times

4 patches) outperform other GFMs that use coarser 16

\times

16 patch sizes, due to the finer patch resolutions as well as techniques intentionally designed to handle missing spectral bands.

Category	Ablation	Performance		Input order		Brightness		Scale		Agree.
Category	Ablation	Object F1 $\uparrow$	Pixel IoU $\uparrow$	F1 $\|\Delta\|\downarrow$	IoU $\|\Delta\|\downarrow$	F1 $\|\Delta\|\downarrow$	IoU $\|\Delta\|\downarrow$	F1 $\|\Delta\|\downarrow$	IoU $\|\Delta\|\downarrow$	Avg $\uparrow$
	FTW-v1	0.39 $\pm$ 0.08	0.68 $\pm$ 0.08	0.07	0.11	0.04	0.05	0.17	0.08	0.93
Data augs	Brightness	0.39 $\pm$ 0.08	0.68 $\pm$ 0.08	0.07	0.09	0.00	0.00	0.14	0.05	0.93
Data augs	Resize	0.38 $\pm$ 0.08	0.67 $\pm$ 0.09	0.07	0.10	0.04	0.05	0.00	0.01	0.94
Data augs	Brightness+Resize	0.38 $\pm$ 0.08	0.66 $\pm$ 0.09	0.06	0.10	0.03	0.03	0.00	0.02	0.95
Data augs	Channel shuffle	0.39 $\pm$ 0.07	0.68 $\pm$ 0.09	0.00	0.00	0.05	0.05	0.18	0.09	0.94

Appendix

Table 5 : Ablation results for controlled experiments on the FTW test set (excluding presence-only countries) in which each row varies a single design choice (data augmentations, class weighting, encoder, loss function, or architecture). Bold indicates best performance, underline indicates second-best.

Figure 5 : Pareto front between Object F1 and Throughput for all Table 1 models, including PRUE-EF (B3/B5) throughput. PRUE-EF3 is a Pareto improvement over the FTW baseline. The PRUE family forms a substantially stronger Pareto frontier than any prior model.

Figure 6 : Consistency dependence on overlap size. Consistency scores as a function of overlap window size for the FTW baseline (left) and PRUE (right). Larger overlaps measure intrinsic translation sensitivity, while smaller overlaps additionally capture context dependency. Shown: top 5 and bottom 5 countries by mean consistency for each model. PRUE achieves higher consistency across all countries and overlap sizes, with particularly strong improvements in challenging regions. Hard countries (Kenya, India) show low consistency even at large overlaps, while well-represented regions (Finland, Netherlands) have high consistency, suggesting consistency metrics can identify out-of-distribution samples at inference time.

Figure 7 : Country-level consistency vs. object F1. Each point represents one FTW country. Dashed lines show linear fits for the FTW baseline (orange, R²=0.48) and PRUE-EF7 (blue, R²=0.30). Consistency is weakly correlated with performance, suggesting it may serve as a partial out-of-distribution signal but not a reliable sole predictor of model accuracy.

Country	Pixel IoU					Object F1
Country	FTW	M2F	Clay	Clay-FT	PRUE	FTW	M2F	Clay	Clay-FT	PRUE
Austria	0.71	0.74	0.69	0.76	0.78	0.41	0.38	0.39	0.47	0.50
Belgium	0.75	0.78	0.73	0.79	0.82	0.57	0.57	0.57	0.62	0.66
Cambodia	0.40	0.19	0.27	0.40	0.66	0.19	0.10	0.09	0.19	0.36
Corsica	0.45	0.52	0.47	0.51	0.51	0.18	0.24	0.18	0.22	0.24
Croatia	0.67	0.70	0.64	0.71	0.77	0.28	0.34	0.25	0.33	0.45

Appendix

Table 6 : Per-country performance comparison for the top-performing models of each architecture family: FTW baseline (semantic), Mask2Former with Swin-S (instance/panoptic), Clay (frozen GFM), Clay-FT (finetuned GFM), and PRUE (our optimized semantic model). Bold indicates best performance per country. PRUE achieves the highest or second-highest pixel IoU scores across nearly all countries. All models evaluated using the protocol in Section 3.4 on presence/absence labeled countries only.

Algorithm 1 Planting Season Heuristic Algorithm

Algorithm 2 Harvest Season Heuristic Algorithm

Algorithm 3 Greedy Scene Selection Algorithm

Figure 8 : Example visuals over each country in our large-scale inference set (Japan, Mexico, Rwanda, South Africa, and Switzerland). For each region of interest, we show: (1) the PRUE field boundary predictions from 2023; (2) the 2023 predictions with changes detected in 2024 highlighted in bright green; (3) Planet monthly basemaps from 2023 with the vectorized change mask outlined in red; and (4) Planet monthly basemaps from 2024 with the vectorized change mask outlined in red. The basemaps shown for each pair are from the same month in consecutive years.

Explore similar work

Delineate Anything v2: A Global Foundation Model for Field Delineation
Jul 21, 2026Mykola Lavreniuk, Nataliia Kussul, Andrii Shelestov +4Delineation Wireless Foundation Models
European Space Agency, Frascati, Italy · Space Research Institute NASU-SSAU, Kyiv, Ukraine · University of Maryland, College Park, Maryland, USA +1
The first global agricultural field boundary map at 10m resolution
May 11, 2026Caleb Robinson, Gedeon Muhawenayo, Subash Khanal +9Agricultural Monitoring Precision Agriculture
Microsoft AI for Good Research Lab · Arizona State University · Washington University in St. Louis +3
Farmland Extent and Visible Boundary Mapping from 1 m NAIP Imagery Using Residual U-Net and Text-Prompted SAM 3 Refinement
Jul 24, 2026Mohammadreza Narimani, Vikram Anand, Parastoo FarajpoorAgricultural Monitoring Precision Agriculture
Department of Biological and Agricultural Engineering, University of California, Davis, Davis, CA, 95616, USA · Del Norte High School, San Diego, CA, USA

Jul 21, 2026, cs.CVJ/K move · Enter open · S save

Delineate Anything v2: A Global Foundation Model for Field Delineation

Mykola Lavreniuk, Nataliia Kussul, Andrii Shelestov, Yevhenii Salii+3

European Space Agency, Frascati, Italy · Space Research Institute NASU-SSAU, Kyiv, Ukraine · University of Maryland, College Park, Maryland, USA+1

Delineation Wireless Foundation Models

Open paper AskPDF arXiv

Accurate agricultural field boundary delineation at large scale is a foundational task for food security, supply chain transparency, and carbon accounting. While vision foundation models like SAM show remarkable zero-shot capabilities, they frequently fail in geospatial domains due to topological complexity, cropland texturing patterns, and a lack of physical scale awareness. In this work, we introduce Delineate Anything v2, a globally scalable foundation model designed specifically for wide-area field boundary mapping. We construct FBIS-73M, a 73-million-instance multi-resolution dataset spanning 61 countries. To address the pervasive issue of multi-field administrative parcel merging, we introduce a resolution-specific data curation pipeline that leverages topological image-space adaptation to homogenize merged parcels and strengthen weak physical boundaries. Furthermore, we establish a novel, manually curated evaluation benchmark covering 100 countries to assess independent zero-shot generalization. Our results show that Delineate Anything v2 surpasses the current state-of-the-art, including the Delineate Anything framework, by 0.284 mAP@0.5 (+103.3% relative gain), while maintaining execution speeds suitable for rapid national- and global-scale deployment, as demonstrated by nationwide mapping of Ukraine (603,000 km^2) in 5.4 hours on a consumer-grade workstation. Code, pre-trained weights, the FBIS-73M dataset, and ready-to-use national-scale vector boundary products are publicly available at https://lavreniuk.github.io/Delineate-Anything/.

Abstract

Explore similar work

Delineate Anything v2: A Global Foundation Model for Field Delineation

The first global agricultural field boundary map at 10m resolution

Farmland Extent and Visible Boundary Mapping from 1 m NAIP Imagery Using Residual U-Net and Text-Prompted SAM 3 Refinement