- The paper introduces DelAnyFlow, a resolution-agnostic instance segmentation pipeline that achieves over 100% mAP improvements and 400x faster inference than SAM2.
- It employs adaptive tiling, structured post-processing, and vectorization to generate complete and topologically consistent field boundary layers at national scales on consumer-grade hardware.
- The approach outperforms existing methods in detecting smallholder and fragmented fields, supporting enhanced operational applications in agricultural monitoring and policy governance.
Delineate Anything Flow: An Expert Synthesis of Country-Level Field Boundary Detection
Context and Motivation
The accurate delineation of agricultural field boundaries from satellite imagery underpins a suite of geospatial intelligence activities, including land management, crop monitoring, and statistics for both operational and policy contexts. However, legacy pixel-wise and semantic approaches are structurally limited, frequently yielding incomplete field boundaries, merged parcels, and suboptimal generalization across heterogeneous agro-ecologies. The need for scalable, high-confidence, and topologically correct parcel layers is especially acute given the political-economic stakes for subsidy auditing, food security assessments, and governance, notably in areas lacking cadastral data.
Methodology: DelAnyFlow Pipeline
The Delineate Anything Flow (DelAnyFlow) methodology constitutes a resolution-agnostic, scalable field boundary extraction framework leveraging instance segmentation via deep learning. The backbone is the Delineate Anything (DelAny) model, built upon YOLOv11, trained on FBIS-22M—by far the most extensive agricultural boundary segmentation dataset assembled (>22M instances, 672K images, resolutions spanning 0.25-10m).
The pipeline consists of:
- Adaptive tiling and data preparation for multi-sensor imagery.
- DelAny instance segmentation for direct field-level mask output.
- Structured post-processing, including morphological operations and mask cleaning.
- Cross-tile instance merging for topological consistency.
- Final vectorization, yielding analysis-ready, validated field boundary shapefiles suitable for operational applications.
FBIS-22M's heterogeneity in resolution, sensor type, geographic coverage, and field size/density enhances the generalization capability of DelAny, allowing for robust inference on previously unseen territories and smallholder settings.
Empirical Evaluation and Results
DelAny, when benchmarked against state-of-the-art models such as SAM2 and MultiTLF, achieves:
| Model |
[email protected] |
[email protected]:0.95 |
Inference Latency (ms) |
| DelAny |
0.720 |
0.477 |
25.0 |
| DelAny-S |
0.632 |
0.383 |
16.8 |
| SAM2 |
0.382 |
0.235 |
>10,000 |
| MultiTLF |
0.257 |
0.110 |
55.8 |
DelAny delivers over 100% higher mAP and is 400x faster than SAM2 under identical inference conditions. The zero-shot generalization is strong, with high accuracy across geographies (including Africa, Southeast Asia, Americas) not present in the FBIS-22M training corpus.
Scalability and Operational Deployment
DelAnyFlow generated a complete vector field boundary layer for all of Ukraine (603,000 km²) in less than six hours on consumer-grade hardware. For fields ≥0.25 hectares, DelAnyFlow detects 3.75M parcels at 5m resolution and 5.15M parcels at 2.5m—compared to Sinergise's 2.66M and NASA Harvest's 1.69M counts in the same region—demonstrating superior smallholder field detection and completeness.
Increasing dataset size and diversity yields consistent performance gains; model accuracy does not saturate even with the largest public dataset, emphasizing the importance of coverage over scale alone.
Comparative Assessment
SAM2 and MultiTLF show limitations in small, fragmented, or spectrally ambiguous fields, exhibiting field merging and spurious boundary errors. Operational products from Sinergise and NASA Harvest demonstrate lower detection rates, especially for small and medium-sized fields, and less geometric fidelity compared to DelAny-derived results.
Use of higher-resolution imagery (Planet, Maxar) improves detection rates and boundary sharpness for small parcels, but even with interpolated Sentinel-2 (down to 2.5m), DelAny maintains near-parity, suggesting diminishing returns and practical trade-offs in computational burden for extreme resolutions.
Methodological Advances and Claims
- Instance segmentation formulation: Replacing semantic segmentation with instance-level mask prediction resolves critical shortcomings in field merging and boundary misalignment. Instance IoU as an evaluation metric is robust to minor spatial errors but penalizes topological errors fundamentally relevant for operational parcel mapping.
- Resolution-agnostic model architecture and dataset: DelAny's YOLOv11 backbone and FBIS-22M's multi-resolution design underpin robust generalization and enable zero-shot deployment.
- Analysis-ready output: Integration of post-processing and vectorization ensures compatibility with downstream routines (crop classification, cadastral mapping, subsidy audit).
Implications
Practical
- Rapid national-scale mapping: On modest hardware, DelAnyFlow supports countrywide deployments, enabling temporal field monitoring in near real-time (hours).
- Support for countries lacking cadastral data: Outputs bridge critical gaps in land parcel layers for regions with incomplete or outdated official registries.
- Operational utility: High completeness and topological fidelity in smallholder and fragmented systems facilitate crop type mapping, conflict damage assessment, yield estimation, and subsidy governance.
Theoretical
- Instance segmentation advances operational geospatial intelligence: Reformulating agricultural mapping as an object detection task offers a pathway for similar advances in related domains (e.g., environmental monitoring, urban parcel mapping).
- Benchmarking for Earth observation: The release of FBIS-22M and open-source models enables new standards for validation and reproducibility in agricultural ML.
Limitations and Future Directions
- Global generalization biases: While zero-shot transfer is demonstrably robust, European continental dominance in the training set may induce subtle biases; extending annotation in smallholder-dominated geographies remains essential.
- Multi-modal and temporal data: Fusion with SAR and multi-temporal composites is recommended to mitigate cloud and phenology effects.
- Uncertainty quantification: Inclusion of confidence scores or posterior intervals would bolster downstream decision-making.
- Operational scaling: Cloud-native deployment and standardized APIs will be crucial for adoption by governmental and international entities.
Conclusion
DelAnyFlow establishes a technically robust, resolution-agnostic workflow for large-scale, topologically consistent field boundary delineation, outperforming contemporary baselines in both accuracy and efficiency. Its empirical successes across smallholder, fragmented, and industrial agricultural systems lay a foundation for scalable national and global mapping programs. The model, codebase, and outputs are open-sourced, facilitating reproducibility, benchmarking, and integration into agricultural monitoring initiatives such as GEOGLAM and FAO programs. Expansion of annotated datasets, modality fusion, and uncertainty metrics are clear avenues for future research.