Agricultural Parcel & Boundary Delineation
- APBD is the automated process of detecting and precisely delineating agricultural fields and boundaries from remote sensing imagery.
- Recent approaches integrate semantic and instance segmentation with multisensor data like Sentinel-2 and PlanetScope to achieve high geometric fidelity.
- Research addresses challenges such as label fidelity, functional versus cadastral distinctions, and adaptation of foundation models for robust performance.
Agricultural Parcel and Boundary Delineation (APBD) is the automated detection and precise delineation of agricultural field units and their boundaries from remote sensing imagery. In the recent literature, APBD is commonly organized into three hierarchical levels—Cropland Identification (CI), Boundary Delineation (BD), and Parcel Segmentation (PS)—with outputs ranging from binary cropland masks to closed parcel polygons and explicit boundary maps (Zheng et al., 20 Aug 2025). The topic spans cadastral, agronomic, and operational perspectives: some work distinguishes dynamic “functional field boundaries” from legal cadastral boundaries, while other work emphasizes parcel maps as analysis units for subsidy control, crop monitoring, land allocation, irrigation, fertilization, and greenhouse-gas policy (Zahid et al., 2024, Aung et al., 2020).
1. Problem formulations and boundary concepts
APBD is not a single task but a family of related formulations. The review literature separates CI, which distinguishes cropland from non-cropland, from BD, which targets parcel edges with high geometric fidelity, and from PS, which aims at instance-level segmentation and vectorization of individual fields (Zheng et al., 20 Aug 2025). This decomposition is also visible in task-specific papers: one early satellite formulation explicitly split farmland parcel delineation into “Segmentation of parcel boundaries” and “Segmentation of parcel areas,” using U-Net variants and reporting that the variant incorporating temporal information performed best on a France 2017 dataset (Aung et al., 2020).
A second distinction concerns what kind of boundary is being delineated. In European LPIS- or BRP-oriented settings, labels often derive from parcel registries and therefore reflect administratively maintained field units. By contrast, work on “functional field boundary delineation” defines the target as the operational edge of actively cultivated fields, which may shift with crop, management, and phenology and may diverge from static cadastral lines (Zahid et al., 2024). This distinction matters because model error can reflect either poor boundary recovery or disagreement between legal and functional definitions.
A third distinction concerns semantic versus instance formulations. Semantic APBD predicts classes such as background, field interior, and boundary, then derives instances through connected components, watershed, or vectorization. Instance-centric approaches treat each field as a distinct object and predict closed masks directly. The instance-segmentation literature argues that this reformulation addresses two recurring APBD failure modes: extreme sensitivity of boundary metrics to small misalignments and the merging of adjacent fields caused by broken contour continuity (Lavreniuk et al., 3 Apr 2025). A related transfer-learning study likewise framed field delineation as instance segmentation, but operationalized it through separate border and interior masks, polygonizing predicted interiors into individual fields for evaluation (Kerner et al., 2024).
2. Data sources, labels, and benchmark construction
APBD research is strongly conditioned by sensor resolution, temporal cadence, and label provenance. Sentinel-2 remains a dominant medium-resolution source because of its 10 m optical bands and seasonal coverage; FTW (“Fields of The World”) is a major benchmark built from bi-temporal Sentinel-2 Level-2A RGB+NIR imagery and more than 1.5 million manually validated field polygons spanning 24 countries (Muhawenayo et al., 28 Mar 2026). At higher resolution, FTP (“Fields of the Planet”) pairs the same FTW polygons, seasonal windows, and train/test splits with 133,168 co-registered PlanetScope targets at 3 m, explicitly quantifying the resolution gap for smallholders (Corley et al., 5 Jul 2026).
Large instance-level benchmarks have recently expanded the design space. FBIS-22M comprises 672,909 satellite image patches, 22,926,427 instance masks, and ground sample distances from 0.25 m to 10 m, integrating Sentinel-2, Planet, Maxar, Pleiades, and orthophotos across multiple European countries (Lavreniuk et al., 3 Apr 2025). In the SAM adaptation literature, ERAS (“ERAgriSeg”) adds a regional benchmark for Emilia-Romagna with 14,968 Sentinel-2 tiles and 3,742 high-resolution tiles, using farmer-declared parcel boundaries and consistent labels across modalities (Scribano et al., 19 Jun 2025).
Terraced APBD has motivated distinct benchmark construction. GTPBD introduces 47,537 high-resolution images, more than 200,000 manually annotated terraced parcels, and three-level labels—boundary, mask, and parcel—covering seven Chinese zones and multiple global terraced regions (Zhang et al., 19 Jul 2025). GTPBD-MM extends this setting to aligned optical imagery, structured text descriptions, and DEM, covering terraced regions across 25 countries and more than 900 km² while supporting Image-only, Image+Text, and Image+Text+DEM evaluation protocols (Zhang et al., 14 Apr 2026).
Label quality has itself become a research object. A continent-scale African dataset delineated 33,746 Planet NICFI sites from 2017 to 2023, producing 42,403 labeling assignments and 825,395 digitized polygons, together with built-in quality-control classes and a Bayesian risk metric for uncertainty (Estes et al., 2024). Reported quality was “moderately high (0.75) for measures of total field extent, but low regarding the number of individual fields delineated (0.33), and the position of field edges (0.05),” an explicit acknowledgment that 3–5 m imagery often cannot reliably resolve dense smallholder boundaries (Estes et al., 2024).
3. Methodological families
The APBD literature now spans traditional image analysis, conventional machine learning, semantic segmentation, instance segmentation, multimodal learning, and foundation-model adaptation. The broadest synthesis groups methods into three classes: traditional image processing, traditional machine learning, and deep learning-based approaches, with further subdivision of deep models into semantic segmentation-based, object detection-based, and Transformer-based methods (Zheng et al., 20 Aug 2025).
Earlier boundary-centric workflows relied on edge, region, and superpixel primitives. A UAV-oriented cadastral workflow combined gPb contour detection with SLIC superpixels and a QGIS plugin for semi-automatic boundary tracing, using shortest-path reasoning on a candidate network derived from image contours (Crommelinck et al., 2017). A cloud-deployable Sentinel-2 method combined SNIC superpixels with Canny edge detection on NDVI, using superpixel boundaries and edge masks as complementary cues for field polygonization in Google Earth Engine (Gayibov, 6 Feb 2025). These methods remained attractive where labels were scarce, but they were sensitive to weak gradients, shadows, and complex smallholder mosaics.
Label-efficient hybrid pipelines remained important in small-field contexts. A multi-stage method for sparse-ground-data environments used a pre-trained HED edge detector, polygon filtering rules, a “cut-point”/“min-cuts” procedure for under-segmentation, localized second-level edge detection, and a final random forest parcel classifier, specifically targeting faint boundaries, severe label scarcity, and interspersed non-field pockets (Marvaniya et al., 2020). This line of work treated APBD as a structured image-processing problem rather than end-to-end dense prediction.
Semantic segmentation then became the dominant paradigm. U-Net and its variants, DeepLab, FPN-style architectures, HRNet-like designs, and multi-task mask-edge-distance networks recur throughout the review literature (Zheng et al., 20 Aug 2025). PRUE formalizes this trend as a deployment-oriented recipe: an EfficientNet-B7 U-Net, log-cosh Dice loss, boundary class weighting with , and targeted augmentations such as channel shuffling and brightness/scale jitter, all tuned for large-scale FTW deployment (Muhawenayo et al., 28 Mar 2026). Spatio-temporal models extend the same family by using multi-date optical or SAR sequences. PTAViT3D and PTAViT3D-CA process Sentinel-1 and Sentinel-2 time series with a memory-efficient 3D Vision Transformer and cross-attention fusion, directly targeting cloud-contaminated imagery without pixel-level cloud masking (Diakogiannis et al., 2024).
Instance segmentation has reoriented APBD toward parcel recovery rather than pixel overlap. “Delineate Anything” treats each field as an instance and uses a YOLOv11-based segmentation model trained on FBIS-22M (Lavreniuk et al., 3 Apr 2025). DelAnyFlow adds a structured post-processing, merging, and vectorization sequence to produce “topologically consistent vector boundaries” at country scale, demonstrating a fully operational instance-first pipeline (Lavreniuk et al., 17 Nov 2025). Transfer-learning work also uses instance-wise evaluation even when the network predicts border and interior masks, because downstream use depends on closed parcel objects rather than only class-consistent pixels (Kerner et al., 2024).
Foundation-model adaptation has produced a separate methodological branch. Zero-shot SAM has shown proof-of-concept utility for smallholder delineation when labels are absent, especially when ensembling across checkpoints, tile sizes, dates, and edge-enhanced inputs (Tripathy et al., 2024). At the same time, fine-tuned SAM variants show that zero-shot performance is generally inadequate for APBD and that parameter-efficient adaptation is necessary. One study uses LoRA on the image encoder and decoder attention, with single-point prompts, to build a strong APBD baseline across AI4Boundaries and ERAS (Scribano et al., 19 Jun 2025). Another, fabSAM, combines a Deeplabv3+-based prompter with a fine-tuned SAM decoder for separate region and boundary heads (Xie et al., 21 Jan 2025).
Recent work also pushes beyond single-image delineation. SEED-SR performs “super-resolution” in a segmentation-aware latent space, conditioning on multi-source low-resolution time stacks and an older high-resolution reference to generate 0.5 m segmentation maps at a scale factor for smallholder APBD (Agarwal et al., 18 Nov 2025). Time2Agri addresses label efficiency through agriculture-specific self-supervised temporal pretext tasks and reports that Future-Frame Prediction transfers best to field-boundary delineation on FTW India (Gupta et al., 6 Jul 2025). Terraced-scene research adds multimodal fusion of image, text, and DEM through ETTerra, explicitly separating semantic disambiguation from terrain-guided boundary reinforcement (Zhang et al., 14 Apr 2026).
4. Evaluation protocols and representative results
APBD evaluation uses both raster and object criteria. The most common pixel-wise metrics are IoU and F1:
Boundary-specific work additionally uses ODS and OIS for edge detection, while object-centric studies use AP or PQ. FTP evaluates parcel recovery after vectorization and reports panoptic quality, object F1, size-stratified PQ, and meter-scale matched-boundary error, emphasizing that polygon quality rather than only raster overlap determines utility (Corley et al., 5 Jul 2026). Terraced benchmarks further introduce GOC, GUC, and GTC to quantify over-classification, under-classification, and total object-level error (Zhang et al., 19 Jul 2025, Zhang et al., 14 Apr 2026).
Several quantitative results have become reference points. PRUE reports “76% IoU and 47% object-F1 on FTW,” improving by 6% IoU and 9% object-F1 over the previous baseline under unified experimental settings (Muhawenayo et al., 28 Mar 2026). DelAny reports $0.720$ mAP@0.5 and $0.477$ mAP@[0.5:0.95] on FBIS-22M, with gains of in [email protected] and in mAP@[0.5:0.95] over SAM2 (Lavreniuk et al., 3 Apr 2025). FTP shows why these object metrics matter: moving from 10 m Sentinel-2 to 3 m PlanetScope raises PQ from $21.0$ to $35.5$, raises PQ on sub-0.5 ha fields from $5.8$ to 0, and reduces matched-boundary error from 1 m to 2 m (Corley et al., 5 Jul 2026).
Smallholder and sparse-label studies report a different performance profile. The multi-stage HED-based small-field pipeline reaches an F-Score of 3 in large-field areas and 4 in small-field areas, with the latter explicitly described as “reasonable accuracy” under sparse ground data (Marvaniya et al., 2020). Zero-shot SAM in Bihar, using 2 m SkySat imagery and multi-date ensembling, correctly identifies 5 of fields with delineation accuracy of approximately 6 mean IoU and F1 around 7 (Tripathy et al., 2024). fabSAM reports composite mIoU values of 8 on AI4Boundaries and 9 on AI4SmallFarms, clearly outperforming both zero-shot SAM and Deeplabv3+ under the study’s metric definition (Xie et al., 21 Jan 2025).
Multi-date and temporal-pretraining results reinforce the importance of seasonal information. A Netherlands/Pakistan study reports that three-date NDVI stacks with SE-ResNeXt-50 achieve approximately 0 mean IoU on Sentinel-2 and approximately 1 on PlanetScope in the Netherlands, with combined training across geographies also yielding approximately 2 mean IoU (Zahid et al., 2024). Time2Agri reports 3 IoU on FTW India for field boundary delineation when using Future-Frame Prediction pretraining, outperforming MAE, DoFA, and a supervised ViT-S baseline in that setting (Gupta et al., 6 Jul 2025).
Terraced APBD highlights the gap between pixel and parcel quality. On GTPBD, Mask2Former achieves IoU 4 and F1 5 for semantic segmentation, while REAUNet-Sober leads edge detection with ODS 6, OIS 7, and AP 8 (Zhang et al., 19 Jul 2025). On GTPBD-MM, ETTerra reaches Recall 9, F1 $0.720$0, OA $0.720$1, mIoU $0.720$2, OIS $0.720$3, ODS $0.720$4, and GTC $0.720$5, illustrating how multimodal evaluation extends beyond standard semantic metrics (Zhang et al., 14 Apr 2026).
5. Operational workflows and application domains
APBD is operationally important because parcel boundaries are not merely cartographic outputs. The climate-policy literature links parcel delineation to land allocation, irrigation, fertilization, greenhouse-gas accounting, and agricultural insurance compensation after extreme weather damage (Aung et al., 2020). The broader review literature adds crop-type mapping, yield modeling, pest and disease surveillance, irrigation scheduling, and CAP-style compliance auditing as recurrent downstream applications (Zheng et al., 20 Aug 2025).
Operational pipelines therefore include more than neural inference. DelAnyFlow exemplifies a production-grade sequence: native-resolution inference, masking to valid agricultural areas, morphological cleaning, cross-tile merging, raster mosaicking, contour extraction, and topology enforcement to obtain “topologically consistent vector boundaries” (Lavreniuk et al., 17 Nov 2025). Using Sentinel-2 data for 2024, the pipeline generated a complete field boundary layer for Ukraine, covering $0.720$6, in under six hours on a single workstation, and delineated $0.720$7 million fields at 5 m and $0.720$8 million at 2.5 m, compared with $0.720$9 million detected by Sinergise Solutions and $0.477$0 million by NASA Harvest (Lavreniuk et al., 17 Nov 2025).
PRUE shows a semantic-segmentation route to similar scale. The released PRUE maps cover Japan, Mexico, Rwanda, South Africa, and Switzerland over 2023 and 2024, with more than $0.477$1 million km² processed (Muhawenayo et al., 28 Mar 2026). The same study reports throughput of approximately $0.477$2 for the 67.1M-parameter EfficientNet-B7 model, together with blockwise vectorization into fiboa GeoParquet for large-scale analytics (Muhawenayo et al., 28 Mar 2026). This suggests that operational APBD now depends as much on stitching, vectorization, overlap handling, and metadata serialization as on the choice of backbone.
Temporal and multimodal workflows also alter operational assumptions. PTAViT3D reduces the need for manual cloud-free curation by learning directly from raw cloud-affected Sentinel-2 time series and by fusing Sentinel-1 where required (Diakogiannis et al., 2024). SEED-SR changes the revisit–resolution trade-off by using weekly low-resolution stacks and an older high-resolution reference to generate high-resolution segmentation maps for in-season boundary updates (Agarwal et al., 18 Nov 2025). Terraced-scene systems further suggest that production APBD in mountainous regions may require not just imagery, but terrain geometry and scene-level semantic priors (Zhang et al., 14 Apr 2026).
6. Limitations, misconceptions, and emerging directions
A recurring misconception is that pixel accuracy alone captures APBD quality. Multiple recent studies argue the opposite. Instance-segmentation work shows that boundary IoU can remain deceptively high when adjacent fields are merged, while FTP demonstrates a “representation ceiling” at 10 m: when FTW polygons are rasterized and re-vectorized at 10 m, only about $0.477$3 of sub-0.5 ha parcels remain separable, versus about $0.477$4 at 3 m (Lavreniuk et al., 3 Apr 2025, Corley et al., 5 Jul 2026). This suggests that smallholder APBD is constrained not only by model design but by the recoverable geometry of the sensor grid itself.
A second misconception is that foundation models solve APBD without adaptation. Zero-shot SAM can be useful as a label-scarce first-pass delineator, and its Bihar study is explicit about the proof-of-concept value of multi-date ensembling (Tripathy et al., 2024). However, dedicated adaptation studies report very poor zero-shot mAP on AI4Boundaries and ERAS and show that LoRA-based or prompt-guided fine-tuning is necessary for strong APBD performance (Scribano et al., 19 Jun 2025, Xie et al., 21 Jan 2025). The controversy is therefore not whether foundation models matter, but whether promptable segmentation transfers without agricultural specialization; the current evidence indicates that it usually does not.
A third limitation concerns label fidelity. The African multi-year label effort makes this explicit: total extent quality can be acceptable while edge placement remains highly uncertain, and this uncertainty can be quantified through Qscore, Rscore, and Bayesian risk (Estes et al., 2024). Functional-versus-cadastral mismatch introduces an additional source of label noise in registry-based datasets (Zahid et al., 2024). This suggests that uncertainty-aware losses, quality-weighted sampling, and risk-stratified evaluation are likely to become more central.
Current future directions are comparatively consistent across the literature. Review and benchmark papers repeatedly call for multi-sensor fusion, stronger multimodal foundation models, domain adaptation, topology-aware or vectorization-oriented extraction, and longer temporal modeling (Zheng et al., 20 Aug 2025, Zhang et al., 14 Apr 2026). Smallholder-focused work adds all-weather SAR fusion and frequent monitoring, while SEED-SR argues for directly optimizing segmentation-relevant latent structure rather than perceptual super-resolution (Agarwal et al., 18 Nov 2025). Transfer-learning studies emphasize bridge-region adaptation for label-poor geographies, especially in under-resourced regions where local training data remain sparse (Kerner et al., 2024). Taken together, these developments indicate that APBD is shifting from isolated segmentation experiments toward integrated geospatial systems in which sensing geometry, temporal coverage, annotation uncertainty, instance topology, and deployment-scale vector processing are treated as first-class design variables.
