FBIS-22M: Instance Segmentation for Field Boundaries

• FBIS-22M is a large-scale dataset for instance segmentation of agricultural field boundaries, containing over 22 million validated field instances from multi-resolution satellite imagery.
• It reformulates field delineation from semantic boundary detection to instance segmentation, providing closed masks that improve parcel precision across diverse European landscapes.
• Integrated with DelAny and DelAnyFlow systems, the benchmark demonstrates over 100% mAP improvements and enables efficient national-scale mapping in operational settings.

Searching arXiv for the specified papers and related field-boundary benchmarks. Field Boundary Instance Segmentation-22M (FBIS-22M) is a large-scale benchmark for agricultural field delineation from satellite imagery, introduced as the data foundation for the Delineate Anything and Delineate Anything Flow systems. It is described as the largest dataset currently available for agricultural field delineation, with 672,909 multi-resolution image patches and 22,926,427 validated field instances, spanning 0.25 m to 10 m imagery from multiple sensors and multiple European countries. Its central methodological premise is that field delineation should be treated as instance segmentation, with each field represented as a separate closed mask, rather than as a semantic boundary-detection problem over edge pixels (Lavreniuk et al., 17 Nov 2025).

1. Problem formulation and conceptual role

FBIS-22M was created in response to a persistent problem in agricultural remote sensing: semantic boundary-detection systems often produce incomplete contours, topologically invalid outputs, and merged adjacent fields. These errors are operationally consequential because many downstream tasks require distinct parcel identities rather than approximate cropland extents. The dataset therefore supports a reformulation of field delineation as instance segmentation, where each field is modeled as an individual object with a closed mask (Lavreniuk et al., 17 Nov 2025).

The earlier “Delineate Anything” paper states this reformulation explicitly and argues that instance-level supervision better matches the operational objective of identifying separate agricultural parcels. It also gives two illustrative contrasts between boundary-centric and instance-centric evaluation: a slight boundary offset can yield boundary IoU = 0.08 while the corresponding instance IoU = 0.98, whereas a partially missing boundary can yield boundary IoU = 0.93 but only instance IoU = 0.54 because adjacent fields become merged (Lavreniuk et al., 3 Apr 2025). These examples are used to show that boundary-only formulations can misrepresent practical delineation quality.

The significance of this shift is methodological as well as operational. Instance masks directly support parcel-level outputs, object-level evaluation, and later conversion to vector boundaries. This suggests that FBIS-22M is not only a larger dataset, but also a benchmark designed to realign supervision, inference, and evaluation with field-level use cases such as crop classification, yield estimation, subsidy auditing, cadastral support, and agricultural statistics (Lavreniuk et al., 17 Nov 2025).

2. Dataset composition, geography, and annotation

FBIS-22M combines scale, sensor diversity, and resolution diversity in a single corpus. The dataset spans Austria, France, Luxembourg, the Netherlands, Slovakia, Slovenia, Spain, Sweden, and Ukraine. Imagery is drawn from Sentinel-2, PlanetScope, Maxar, Pleiades, and other publicly available satellite sources. The listed spatial resolutions are 0.25 m, 0.3 m, 0.5 m, 1 m, 1.2 m, 2 m, 3 m, and 10 m (Lavreniuk et al., 17 Nov 2025).

Aspect	Reported value	Brief note
Total image patches	672,909	Multi-resolution
Total field instances	22,926,427	Validated instances
Training images	636,784	Reported split
Test images	36,125	Reported split
Resolution range	0.25 m–10 m	Eight listed resolutions
Geographic scope	9 countries	Europe-focused

For “most regions,” annotations come from official LPIS (Land Parcel Identification System) boundaries. Where LPIS was unavailable, especially in Ukraine, the labels were produced by manual annotation on high-resolution commercial satellite imagery. The authors also state that the dataset was manually cleaned by removing errors in field boundaries and inconsistencies, and that sampling was designed to preserve geographic and management diversity (Lavreniuk et al., 17 Nov 2025).

The paper reports 636,784 training images and 36,125 test images. A separate validation split is not explicitly reported. Likewise, the paper does not provide a formal inter-annotator agreement measure, annotation error rate, or a detailed quality-control protocol beyond the statements on manual cleaning and validation (Lavreniuk et al., 17 Nov 2025).

Scene complexity is explicitly heterogeneous. Figure 1 groups images from fewer than 10 to more than 300 masks per image, indicating substantial variation in field density and fragmentation. This variability is presented as important for learning to separate adjacent parcels in both sparse and dense agricultural landscapes (Lavreniuk et al., 17 Nov 2025).

3. Label representation and position among field-boundary datasets

FBIS-22M is explicitly framed as an instance segmentation dataset. It contains instance masks of individual fields, and the associated DelAny model is trained to predict instance masks and bounding boxes. The paper repeatedly refers to “instance masks,” “field-level instance,” and “closed-field masks.” It also notes that field boundaries can be recovered from these masks via contour extraction (Lavreniuk et al., 17 Nov 2025).

The dataset description does not explicitly state that FBIS-22M is distributed with polygon labels, boundary maps, or signed distance maps. Because part of the source annotation came from LPIS boundaries, polygon geometry likely existed upstream, but the benchmark as described is organized around mask-based supervision rather than polygon-native learning targets. Accordingly, FBIS-22M directly supports instance masks, and only indirectly supports field boundaries through downstream mask-to-contour conversion (Lavreniuk et al., 17 Nov 2025).

The benchmark is positioned against prior agricultural delineation datasets by scale and annotation volume. In the comparison summarized in the DelAnyFlow paper, FBIS-22M is larger than previously listed datasets such as Farm Parcel, PASTIS, PASTIS-HD, AI4SmallFarms, AI4Boundaries, and Fields of the World (Lavreniuk et al., 17 Nov 2025).

Dataset	Reported scale	Representation context
AI4Boundaries	55K images, 2.5M instances	Prior large benchmark
Fields of the World	70K images, 1.63M instances	Global benchmark
FBIS-22M	673K images, 22.9M instances	Instance masks

The paper repeatedly states that FBIS-22M is more than twelve times larger than AI4Boundaries in annotated field instances and is the first field-boundary benchmark to include high-resolution commercial satellite imagery (Lavreniuk et al., 17 Nov 2025). By contrast, the FTW benchmark emphasizes geographic breadth—24 countries, 1.63M field polygons, and both semantic and instance labels—but at much smaller raw annotation volume (Kerner et al., 2024).

The dataset also has acknowledged structural bias. Section 6.3 states that FBIS-22M is dominated by European agricultural landscapes, which may bias models toward temperate, high-input systems. The authors specifically identify the underrepresentation of sub-Saharan Africa and Southeast Asia as a limitation (Lavreniuk et al., 17 Nov 2025).

4. DelAny and DelAnyFlow as the primary systems built on FBIS-22M

The principal model trained on FBIS-22M is Delineate Anything (DelAny), described as a resolution-agnostic instance segmentation architecture based on YOLOv11. In the DelAnyFlow paper, “resolution-agnostic” means that a single trained model is applied zero-shot across RGB imagery from 0.25 m to 10 m sources without inference-time fine-tuning. Operationally, images from different sensors and resolutions are tiled into a common 512 × 512 RGB format, and the model learns across the multi-resolution training distribution rather than via per-resolution architectural changes (Lavreniuk et al., 17 Nov 2025).

The full DelAny model uses the complete depth of YOLOv11, with a 1.5× width multiplier and a maximum of 512 channels, for 379 layers and 62 million parameters. The lightweight DelAny-S uses half the depth and one-quarter of the width with up to 1024 channels, for 203 layers and 2.9 million parameters. Both are initialized with pre-trained COCO weights and output instance masks and bounding boxes on 512 × 512 RGB tiles (Lavreniuk et al., 17 Nov 2025).

The reported training setup is 30 epochs, batch size 320, 40 images per GPU across 8 NVIDIA H100 GPUs, and learning rate $2 \times 10^{-5}$. Data augmentation includes horizontal flipping, vertical flipping, color jittering, mixup, copy-paste augmentation, and mosaic augmentation during the first 20 epochs (Lavreniuk et al., 17 Nov 2025). The paper states that the standard YOLO loss is used, combining bounding box regression, objectness, classification, and task-alignment terms, but does not print the full loss equation.

DelAnyFlow extends DelAny into a production pipeline for raster-to-vector field mapping. Its conceptual workflow begins with overlapping RGB images from any satellite, followed by tiling, instance segmentation, post-processing, tile merging, and vectorization. Two spatial masks are then applied: a primary quality mask to exclude missing or corrupted pixels and non-agricultural areas such as water and urban areas, and a more conservative context mask that expands exclusion zones near gaps and linear features such as roads and forest edges (Lavreniuk et al., 17 Nov 2025).

Retained instances are sorted from largest to smallest area and undergo a morphological refinement sequence:

1. Erosion
2. Connected-component analysis retaining only the largest contiguous region
3. Dilation

After that, cross-tile duplicates are handled through cross-tile field unification using empirically chosen overlap criteria: Intersection over Union, Intersection area, Intersection / Novel, and Intersection / Original. The exact thresholds are not disclosed. Final outputs are mosaicked into a single raster, assigned unique identifiers, filtered for small artifacts, converted to vector polygons, and subjected to topological validation and sliver removal to produce non-overlapping, geospatially accurate field boundaries (Lavreniuk et al., 17 Nov 2025).

5. Benchmark results, scaling effects, and operational demonstrations

On the FBIS-22M test set, the DelAnyFlow paper reports the following headline results for instance segmentation quality and latency on 512 × 512 patches: MultiTL+ at $[email protected] = 0.257$ and $[email protected]{:}0.95 = 0.110$ with 55.8 ms latency; SAM2 at 0.382 and 0.235; DelAny-S at 0.632 and 0.383 with 16.8 ms latency; and DelAny at 0.720 and 0.477 with 25.0 ms latency (Lavreniuk et al., 17 Nov 2025).

The evaluation metrics are explicitly given as:

$mAP = \frac{1}{n} \sum_{k=1}^{n} AP_k$

$$Precision = \frac{TP}{TP + FP}, \qquad Recall = \frac{TP}{TP + FN}$$

$$AveragePrecision (AP) = \sum_{k=1}^{n-1} [r(k) - r(k-1)] \times \max(p(k), p(k-1))$$

$IoU = \frac{|P \cap G|}{|P \cup G|}$

where $P$ is the predicted mask and $G$ is the ground-truth mask (Lavreniuk et al., 17 Nov 2025).

The paper interprets DelAny’s performance relative to SAM2 as 88.5% higher mAP@0.5 and 103% higher [email protected]:0.95, while the abstract summarizes the result as “over 100% higher mAP and 400× faster inference than SAM2” (Lavreniuk et al., 17 Nov 2025). DelAny-S also exceeds SAM2 by 65.5% and 63% on the two mAP metrics.

Ablation evidence links these gains directly to FBIS-22M’s scale and diversity. Training on AI4Boundaries (45K images) yields 0.358 [email protected] and 0.211 [email protected]:0.95, whereas a same-sized FBIS-22M subset (45K images) yields 0.597 and 0.335. Larger FBIS subsets further improve to 0.678 / 0.429 at 150K images, and 0.720 / 0.477 on the full 636K-image training set (Lavreniuk et al., 17 Nov 2025). The paper presents this as evidence that resolution diversity, sensor diversity, and broader geographic variation matter in addition to raw sample count.

Qualitative zero-shot experiments evaluate the trained model without fine-tuning in Brazil, Cambodia, New Zealand, Rwanda, the United States, Vietnam, and South Africa. The paper reports strong generalization across smallholder farms, large industrial fields, and irregular mixed-use mosaics, but does not provide quantitative mAP values for those regions (Lavreniuk et al., 17 Nov 2025).

The paper also includes a cross-resolution study on a 100 km² block in Lvivska Oblast. The reported minimum reliable thresholds and detected-field counts are: Sentinel-2 10 m at 0.5 ha with 444 detected fields; Sentinel-2 5 m at 0.5 ha with 632; Sentinel-2 2.5 m at 0.3 ha with 781; Planet 3 m at 0.3 ha with 1038; Maxar 2 m at 0.1 ha with 1120; and pan-sharpened Maxar 0.5 m at 0.05 ha with 1516 polygons. Manual delineation on the same area produced 9468 fields, of which 8188 were smaller than 0.25 ha and 9060 were smaller than 0.5 ha (Lavreniuk et al., 17 Nov 2025). The paper concludes that higher resolution improves delineation of small fields, including when finer inputs are obtained by interpolation of Sentinel-2, but it also notes that very high resolution can promote unnecessary merging of adjacent small fields into larger groups.

The best-known operational demonstration is a national-scale deployment over Ukraine. Using Sentinel-2 imagery for the 2024 season, DelAnyFlow generated field-boundary layers for 603,000 km² using interpolated 5 m and 2.5 m Sentinel-2 composites. On a desktop equipped with an AMD Ryzen 9 9900X 12-Core CPU, NVIDIA GeForce RTX 5070 Ti 16 GB GPU, and 64 GB RAM, the 5 m national product was generated in 5.4 hours, excluding data acquisition time; the abstract summarizes this as under six hours on a single workstation (Lavreniuk et al., 17 Nov 2025).

The reported field counts for Ukraine are 3,755,804 fields at 5 m and 5,151,789 fields at 2.5 m, compared with 2,659,421 from Sinergise Solutions and 1,687,310 from NASA Harvest 2023, under a threshold of fields larger than 0.25 hectares (Lavreniuk et al., 17 Nov 2025). The paper states that DelAnyFlow substantially improves representation of fragmented and smallholder-like systems, particularly in the 0.25–1 ha range, while also noting that the 2.5 m map is less practical because of longer delineation time and over-detection of very large fields. The 5 m product is therefore presented as the preferred operational compromise.

6. Limitations, related paradigms, and resource availability

The DelAnyFlow paper is explicit about several limitations of FBIS-22M and systems trained on it. First, the corpus is Europe-heavy, which creates potential domain bias toward temperate, high-input systems. Second, the main experiments rely on single-date optical imagery, so cloud contamination and phenological variation may induce false boundaries. Third, DelAnyFlow can produce overly large polygons, particularly where spectral boundaries between neighboring fields are weak. Fourth, the method is stated to be operationally useful but does not meet EU LPIS accuracy requirements, which depend on orthophotos and stricter standards. Finally, several implementation details are omitted, including exact merge thresholds, masking thresholds, polygon simplification parameters, and topology-repair rules (Lavreniuk et al., 17 Nov 2025).

A common misconception is that the instance-segmentation formulation embodied by FBIS-22M is the only technically credible approach to field delineation. Related work does not support such a universal conclusion. The FTW benchmark provides 1.63M field polygons across 24 countries, with semantic masks, instance masks, and explicit boundary labels, and was designed around blocked country-wise splits and transfer evaluation (Kerner et al., 2024). The FTW ecosystem tutorial describes a practical pipeline based on U-Net with an ImageNet-pretrained EfficientNet-B3 encoder, bi-temporal Sentinel-2 inputs, thresholding at 0.5, morphological opening, connected components, and polygonization to GeoPackage (.gpkg) outputs (Corley et al., 8 Feb 2026).

This alternative paradigm is further strengthened by PRUE, which evaluates 18 models on FTW and reports that a U-Net semantic segmentation model outperforms instance-based and geospatial foundation model alternatives on that benchmark, achieving 76% IoU and 47% object F1 (Muhawenayo et al., 28 Mar 2026). A plausible implication is that the relative merits of direct instance segmentation and semantic-boundary pipelines depend strongly on input modality, label regime, field morphology, and deployment constraints.

Transfer-learning work also highlights a complementary concern that FBIS-22M does not fully resolve: geographic label asymmetry. The multi-region transfer-learning paper studies France, South Africa, and Kenya, uses temporal composites and semantic border/interior masks, and shows that a supervised bridge region can substantially improve delineation in a target region with no target training labels (Kerner et al., 2024). This suggests that large in-domain supervision and cross-region adaptation are distinct research axes rather than substitutes.

At the boundary-refinement level, generic instance-segmentation research such as SharpContour is relevant because it addresses long straight edges and sharp corners via iterative contour evolution and an Instance-aware Point Classifier, although it is not a field-delineation paper and was not evaluated on FBIS-22M (Zhu et al., 2022). For agricultural parcels with polygon-like geometry, such refinement strategies are methodologically adjacent to mask-to-boundary conversion.

On openness, the project provides a landing page, a public code repository, pretrained weights, a Ukraine 2024 national-scale demo, and a subset of FBIS-22M for research use. The stated resources are:

• Project page: https://lavreniuk.github.io/Delineate-Anything/
• Code, training scripts, weights, DelAnyFlow pipeline: https://github.com/Lavreniuk/Delineate-Anything
• Ukraine 2024 field boundaries demo: https://delineate-anything.projects.earthengine.app/view/ua2024fields
• Subset of FBIS-22M: https://huggingface.co/datasets/MykolaL/FBIS-22M

The wording is specific: the paper states that a subset of FBIS-22M is available for research purposes; it does not state that the entire 22.9M-instance dataset is fully public (Lavreniuk et al., 17 Nov 2025).