Fields of the Planet: Field Boundary Mapping Beyond 10m

Abstract: Field-boundary maps support crop monitoring, irrigation planning, and yield estimation, but many smallholder parcels span only a few 10 m Sentinel-2 pixels. We introduce Fields of the Planet (FTP), a 3 m PlanetScope companion to Fields of The World (FTW) that pairs the same polygons, seasonal windows, and train/test splits with 133,168 co-registered PlanetScope patch-window targets across 24 countries. FTP evaluates field delineation as parcel recovery by vectorizing predictions before scoring panoptic quality (PQ), object F1, size-stratified PQ, and meter-scale matched-boundary error. Under matched architectures and training recipes, 3 m imagery raises PQ from 21.0 to 35.5, raises PQ on sub-0.5 ha fields from 5.8 to 15.7, and cuts matched-boundary error from 18.6 m to 7.4 m.

• The paper shows that deploying 3m PlanetScope imagery significantly enhances field-boundary segmentation, achieving nearly 70% higher panoptic quality than 10m Sentinel-2.
• It utilizes an object-centric evaluation with metrics such as polygon-level F1, panoptic quality, and boundary error to rigorously quantify segmentation improvements.
• The study provides actionable insights for agricultural monitoring by outlining how high-resolution data can accurately delineate smallholder plots and inform hybrid mapping strategies.

Fields of the Planet: Advancing Field Boundary Mapping with 3m Imagery

Introduction

Field-boundary segmentation underpins critical downstream tasks in agricultural monitoring, such as crop yield prediction, irrigation analysis, and global food security assessments. However, the global majority of farms—smallholder plots under 2 ha—pose a fundamental challenge for conventional remote sensing techniques, as their boundaries are typically unresolved in 10 m satellite imagery such as Sentinel-2. The "Fields of the Planet" (FTP) dataset (2607.04449) introduces a paradigm shift, pairing the extensive, multi-country Fields of The World (FTW) annotation protocol with 3 m PlanetScope imagery to systematically quantify the gains of higher spatial resolution for parcel delineation. This work evaluates the implications through rigorous, controlled benchmarking, advocates for object-centric evaluation protocols, and releases community resources to enable reproducible research.

Dataset Construction and Scope

FTP aligns with FTW's 1.6 million multi-continent field-boundary polygons, preserving identical patch/windows and train/validation/test splits. The critical innovation is the pairing of these polygons with co-registered 3 m PlanetScope scenes, yielding 133,168 patch-window targets in 24 countries. FTP implements robust patch-level quality control using Planet’s UDM2 masks, ensures seasonal alignment, and excludes low-quality acquisitions—resulting in coverage rates up to 99% in optimal regions, though some areas (notably Portugal, Finland, Netherlands) exhibit reduced retention due to cloud prevalence.

By keeping the label rasterization protocol consistent, FTP provides a uniquely controlled experimental design—isolating the effects of increased spatial resolution rather than dataset/label variations.

Resolution Limitations of 10 m Imagery

Critically, the paper documents the hard limit imposed by 10 m imagery on the recoverability of small fields, independent of model architecture or learning. Only approximately 7.5% of sub-0.5 ha parcels can be resolved to a distinct polygon after rasterizing to a 10 m grid, compared with 88% at 3 m. This result demonstrates that significant signal loss occurs before any downstream modeling.

Figure 1: PlanetScope, at 3 m resolution, resolves dense clusters of sub-0.5 ha parcels that are merged beyond recognition in Sentinel-2 10 m imagery.

Benchmarking Protocols and Model Evaluation

FTP employs a polygonal vector-matching evaluation, focusing on object F1, panoptic quality (PQ), and meter-scale boundary error—measures directly relevant for geospatial pipelines, in contrast to largely protocol-driven pixel IoU metrics. The baseline models are U-Nets with EfficientNet backbones, subjected to strictly matched training and inference regimes across imagery sources. Augmentation strategies—especially geometric and noise variants—proved particularly impactful for modeling on 3 m imagery.

Figure 2: Successive introduction of PRUE+ augmentations delivers substantial polygon-level gains, positioning PlanetScope-based models ahead of all Sentinel-2 baselines.

Main Numerical Findings

The introduction of 3 m PlanetScope imagery drives strong, consistent improvements in every object-centric metric relative to 10 m Sentinel-2:

• Panoptic Quality (PQ): For the ten-country dense-label held-out split, PQ jumps from 21.0 (Sentinel-2 B3) to 35.5 (PlanetScope B3), a relative increase of nearly 70%.
• Sub-0.5 ha Parcel Recovery: For small fields, PQ surges from 5.8 (Sentinel-2) to 15.7 (PlanetScope).
• Boundary Localization: Mean symmetric boundary error decreases from 18.6 m to 7.4 m.
• Object F1 at IoU=0.5: Reaches 46.2 for PlanetScope B3, nearly matching the best PRUE Sentinel-2 B7 result, despite using a smaller model.
• Backbone Scaling: No meaningful improvement is observed moving from B3 to B7 with PlanetScope; the spatial information is already saturated at B3.

  Figure 3: Pixel-level metrics can overstate success—a high pixel IoU does not guarantee correct polygonal segmentation, especially when parcels fragment.

These results are robust across 22 of 23 evaluated regions, including diverse agro-ecological zones, and highlight the necessity of polygon-level, rather than pixel-based, reporting.

Upsampling and Output Grid Experiments

To disentangle the effects of a finer output grid from true sensing improvements, experiments with upsampled Sentinel-2 imagery reveal only partial recovery of object-level metrics, with boundary localization lagging far behind real 3 m input. This finding underscores that only actual high-resolution source data, not synthetic upscaling or post-hoc grid refinements, permit correct separation of adjacent fields and accurate geometry restoration.

Region- and Field-Scale Analysis

Across the macro-averaged global benchmark, PlanetScope delivers PQ gains of up to 14 points over the best Sentinel-2 models. Importantly, the impact is most pronounced for sub-0.5 ha parcels—high-resolution imagery does not merely smooth boundaries but enables distinction where Sentinel-2 merges or omits objects entirely. The only region with parity is Portugal, where field boundaries are primarily cadastral and do not correspond to image edges.

Figure 4: Across 23 regions, PlanetScope universally outperforms or matches Sentinel-2 in polygon PQ, most strongly in small-field strata.

Figure 5: In high-gain patches, PlanetScope models restore parcel separation and shape, resolving dense mosaics of smallholder plots far beyond Sentinel-2’s capabilities.

Implications and Future Directions

The adoption of 3 m PlanetScope imagery fundamentally expands the solution space for field-boundary segmentation, transforming previously unresolvable micro-parcel mosaics into actionable geospatial units. For practitioners, these findings argue for direct deployment of high-resolution commercial data in food security, rural economics, and agricultural monitoring pipelines, particularly in smallholder-dominated landscapes.

Theoretically, the results reinforce the importance of metric choice in assessing segmentation systems—polygon- and boundary-based criteria are essential when the application demands instance geometry, not mere semantic coverage.

The dataset’s structure imprints a path forward for multi-sensor, resolution-aware modeling: hybrid systems could exploit 3 m data where it exists and default to 10 m public data elsewhere, propagating improvements globally as commercial coverage expands. Moreover, with community tools and pipeline code released, the field is positioned to tackle additional open challenges: (i) country-scale deployment, (ii) multi-year and multi-band analysis, (iii) joint cross-sensor training, and (iv) improved label completeness in new regions.

Conclusion

"Fields of the Planet" (2607.04449) establishes that spatial resolution is a primary determinant of recoverable geometrical information in agricultural field-boundary segmentation. The controlled transition from 10 m to 3 m input results in order-of-magnitude improvements in smallholder parcel recovery and boundary localization, confirmed across diverse countries and consistent within rigorous polygonal evaluation frameworks. The work decisively supports the paradigm of high-resolution, object-centric mapping for operational geospatial AI, setting a new technical baseline for public benchmarks and future methodological advances.