Delineate Anything Flow: Fast, Country-Level Field Boundary Detection from Any Source

1. Problem Statement

Accurate delineation of agricultural field boundaries from satellite imagery is critical for land management, crop monitoring, yield estimation, and agricultural statistics. However, existing automated methods face three major challenges:

• Topological Inconsistency: Traditional semantic segmentation approaches often produce geometrically invalid outputs, such as intersecting boundaries, overlapping polygons, or merged adjacent fields (under-segmentation).
• Metric Limitations: Standard evaluation metrics like boundary Intersection over Union (IoU) are overly sensitive to minor pixel misalignments, penalizing models for errors that do not significantly impact operational utility, while failing to adequately penalize the merging of distinct fields.
• Scalability and Cost: State-of-the-art foundation models (e.g., SAM, SAM2) offer zero-shot capabilities but suffer from prohibitive computational costs (slow inference) and systematic errors (over-merging or over-segmentation) when applied to national scales. Furthermore, a lack of large-scale, multi-resolution, instance-annotated datasets hinders the training of robust models.

2. Methodology: DelAnyFlow

The authors propose DelAnyFlow, a resolution-agnostic, instance-level segmentation framework designed for large-scale field boundary extraction. The methodology consists of five key stages:

A. Task Reformulation

The authors reframe field boundary delineation from a semantic segmentation task to an instance segmentation task.

• Rationale: Instance segmentation treats each field as a distinct object, inherently preventing the merging of adjacent fields.
• Metric Shift: They prioritize Instance IoU over Boundary IoU. Instance IoU is robust to minor boundary shifts and correctly penalizes the merging of distinct fields, aligning better with operational requirements.

B. The DelAny Model

• Architecture: Based on YOLOv11, a high-speed real-time object detection and instance segmentation backbone.
• Training: Trained on the new FBIS-22M dataset (see below).
• Input: Accepts optical imagery (RGB or bands closest to RGB) at any spatial resolution (0.25m to 10m).
• Operation: Operates in a zero-shot setting, requiring no fine-tuning for new geographies or sensors.

C. The FBIS-22M Dataset

To address the data scarcity gap, the authors created the Field Boundary Instance Segmentation–22M (FBIS-22M) dataset:

• Scale: 672,909 multi-resolution image patches and 22.9 million annotated field instances.
• Diversity: Covers resolutions from 0.25m to 10m (including Maxar, Planet, Sentinel-2, Pleiades) and spans multiple countries (Ukraine, EU nations, etc.).
• Quality: Derived primarily from official Land Parcel Identification Systems (LPIS) and manually annotated high-resolution imagery, ensuring strict focus on crop-based agriculture.

D. Structured Post-Processing Pipeline

To convert raw model outputs into operational vector data, DelAnyFlow employs a rigorous pipeline:

1. Adaptive Tiling: Handles variable input resolutions and overlaps.
2. Morphological Refinement: Uses erosion, connected-component analysis, and dilation to remove artifacts (e.g., spurious bridges merging fields) and ensure topological validity.
3. Cross-Tile Unification: Merges overlapping instances from adjacent tiles based on IoU and area criteria to ensure seamless national-scale coverage.
4. Vectorization & Simplification: Converts raster masks to vector polygons, applying the Ramer–Douglas–Peucker algorithm to simplify vertices while preserving geometric fidelity.

3. Key Contributions

1. DelAnyFlow Framework: A scalable, end-to-end pipeline for generating topologically consistent, vector-ready field boundaries from multi-resolution satellite imagery.
2. FBIS-22M Dataset: The largest agricultural field boundary dataset to date (22.9M instances), enabling robust training and zero-shot generalization.
3. Performance Breakthrough: The DelAny model achieves state-of-the-art performance, outperforming foundation models like SAM2 by 88.5% in mAP@0.5 and 103% in mAP@0.5:0.95, while being 400× faster in inference.
4. National-Scale Demonstration: Successfully generated a complete field boundary layer for Ukraine (603,000 km²) in under 6 hours on a single workstation, detecting 3.75 million fields (5m resolution) and 5.15 million fields (2.5m resolution).

4. Experimental Results

The study evaluated the method across six dimensions:

• Benchmarking: DelAny significantly outperformed SAM, SAM2, MultiTLF, and existing global products (Sinergise, NASA Harvest, Fields of the World) in IoU, small object recall, and boundary discrepancy metrics (avg GOC, GUC, GTC).
• Zero-Shot Generalization: The model demonstrated strong performance in unseen regions (Brazil, Cambodia, Rwanda, USA, etc.) without fine-tuning, handling diverse field shapes and agricultural practices.
• Resolution Impact: Higher resolution (0.25m–1m) significantly improved the detection of small fields (<1 ha). However, even interpolated Sentinel-2 (2.5m) performed competitively, detecting far more small fields than native 10m data.
• Temporal Stability: Performance remained robust across different acquisition dates (early vs. mid-season) and years, with IoU variance <0.005.
• Preprocessing: Simple RGB composites performed best; complex temporal encoding strategies did not yield significant improvements over standard spectral inputs.
• Comparison with Global Products: DelAny detected 3.75M fields in Ukraine compared to 2.66M by Sinergise and 1.69M by NASA Harvest. It showed superior accuracy in detecting smallholder plots (1–10 ha) and producing geometrically realistic boundaries.

5. Significance and Impact

• Operational Scalability: The ability to process a country-sized area in hours on consumer-grade hardware makes this approach viable for national agricultural monitoring agencies, particularly in regions lacking digital cadastral data.
• Policy and Food Security: The generated boundary layers support critical applications such as crop type classification, yield estimation, conflict damage assessment, and subsidy auditing (e.g., EU CAP).
• Open Science: The authors have released the DelAny model, source code, FBIS-22M dataset subset, and national-scale vector outputs publicly, fostering reproducibility and integration into global initiatives like GEOGLAM and FAO programs.
• Paradigm Shift: The work demonstrates that reformulating the problem as instance segmentation, coupled with large-scale diverse training data, is superior to semantic segmentation or foundation model adaptation for agricultural boundary delineation.

In conclusion, DelAnyFlow provides a cost-effective, high-accuracy, and scalable solution for global agricultural intelligence, bridging the gap between advanced deep learning and operational land management needs.