State-of-the-art review for crop field boundaries delineation from satellite data

Introduction

The delineation of agricultural field boundaries from satellite imagery is crucial for precision agriculture, land management, policymaking and crop monitoring.

It can be a stand-alone solution or as the first stage for crop classification, advanced soil moisture analysis, yield forecast, the estimation of damages to crop yield due to natural (e.g. drought, floods) and more.

In the DaFab project, field delineation, which involves outlining agricultural parcels with polygons, offers a solution to several difficulties encountered in yield forecasting and crop classification tasks. This approach is more effective than traditional pixel-by-pixel classification maps because of the reasons:

Accuracy

Field boundary awareness enhances land cover classification (such as crop classification). It prevents the misclassification of crops that can happen within fields or at their edges when using pixel-based maps.

Computational resources

Filtering input data by agricultural field boundaries would significantly reduce the amount of data (e. g. high-resolution satellite imagery) needed to be downloaded and processed (without prior knowledge of field boundaries, non-agricultural land are processed as well as agricultural, which could otherwise be ignored).

Single fields typically contain dozens, hundreds, or even thousands of pixels. Reducing input data to one input vector per agricultural field instead of per pixel substantially reduce processing time and cost by several orders of magnitude.

Field delineation AI models and approaches

The European Union uses its Land Parcel Identification System (LPIS) to map agricultural field boundaries for land use monitoring and subsidy allocation. However, many other parts of the world lack such systems, relying on outdated maps that hinder effective agricultural management. The manual effort required to create and update LPIS data further emphasizes the critical need for automated, scalable ways to detect field boundaries using satellite data.

To identify and outline individual crop fields (a process called parcel delineation) automatically, researchers and practitioners employ a range of image segmentation strategies. These encompass methods like identifying boundaries through edge detection, grouping similar pixels via region segmentation, using graph-based algorithms, and integrating traditional segmentation techniques with modern machine learning (ML) and deep learning (DL) models.

Traditional methods for detecting edges rely on low-level local cues to identify changes in curvature, noise, or color, such as the Canny algorithm. However, shallow features aim at areas with evident changes primarily and do not consider contextual semantic information, resulting in a large amount of noise and blurred edges. In other words, commonly used edge detection methods only consider the changes between neighboring pixels, limiting the use of high-level image features. A major problem in edge detection is to obtain closed boundaries.

Region segmentation-based methods primarily aim to address the problem of non-closed boundaries in edge detection, e.g., using multi-resolution segmentation algorithms. Assumedly, adjacent elements inside a region have similar values. There are two basic methods for region segmentation, including merging and splitting.

The basic steps of region segmentation are:

1. obtain the initial segmentation result of the image;
2. merge similar fragments and divide dissimilar ones;
3. continue splitting until there are no remaining fragments.

However, region-based methods are dependent on well-segmented objects. If the segmentation quality of the image object is poor, it leads to reduced accuracy in the extraction of crop-field boundaries.

For instance, with poor parameterization, objects may stop growing before reaching the actual boundaries, creating sliver polygons and shifting the extracted boundaries inwards. Region based methods also tend to over-segment fields with high internal variability and under-segment small adjacent fields. Some of these adverse effects might be mitigated by purposefully over-segmenting images and deciding whether adjacent objects should be merged with machine learning (see Garcia-Pedrero et al., 2017, for instance; Fig. 2).

Field boundary delineation can also be approached as a graph-based problem. For example, the Graph-Based Growing Contours model (Wagner et al., 2020) offers a solution, though it still needs some supervision, mainly in terms of parameterization. This issue of parameterization, which often hinders the generalization of traditional segmentation techniques across diverse field types, regions, and environmental conditions, could be resolved by using pretrained Deep Learning (DL) models.

DL models

Understanding Field Delineation with AI

Effectively mapping agricultural fields from satellite using AI typically involves a three-step process: pre-processing (preparing satellite data), running the AI model, and post-processing the AI’s output (converting the detected field’s extent into vector polygons).

The UNet architecture is a dominant force in image segmentation within deep learning, and many cutting-edge AI models for parcel delineation, such as ResUNet-a, U-TAE, ST-U-net, FAUNet and ftw, are built upon it. Transformer based architecture on the other hand is minority in field delineation, such as U-TAE.

Parcel delineation models can be further categorized based on their input and output. Inputs can be either single-scene (individual images) or multi-temporal (stacks of images over time), while outputs are generally segmentation masks (or instances) or direct polygons of individual field boundaries.

Single-scene models perform semantic segmentation on each image independently. Their predictions can be combined later to account for temporal changes. A key advantage here is that the model learns to be robust regardless of the specific time period.

In contrast, multi-temporal models apply semantic segmentation to a series of images taken over time. This allows the model to extract crucial spatio-temporal features, implicitly capturing how fields change over season. While these models tend to be larger and slower due to the richer input data (spatial, temporal, and spectral information), they offer a more robust understanding of crop field characteristics over time.

Output Formats and Post-Processing

Most deep learning models for field detection commonly produce three segmentation masks: one for the overall field extent, another for the field boundary/edge, and a third indicating the distance to the closest boundary. Examples include ResNet-a, BsiNet, SEANet, FAUNet (predicts only edge masks and extent masks; Fig. 3).

Getting usable field polygons from these masks requires sophisticated post-processing. This involves smoothing the raster masks, connecting broken boundaries, filling small gaps, and removing spurious detections to create topologically sound polygons.

However, an alternative and increasingly appealing approach is for models to directly predict polygons for individual crop fields, framing it as a graph-based task. This bypasses the need for complex post-processing steps and can reduce associated noise, as demonstrated by research Xia et al., 2024.

Full list of the resent state-of-the-art Deep Learning publications with simple summary presented in Table 1.

Method name	Code / License / weights	Data for training/validation	Summary
ResUNet-a 2020	code, CSIRO Open Source Software Licence Agreement (variation of the BSD / MIT License) No pretrained model weights	From JECAM sites. Sentinel-2 (blue, green, red, NIR) monthly composites. South Africa; Argentina, Australia, Canada, Russia, and Ukraine.	Single-scene architecture that proposed parcel delineation as a multi-task problem (extent, boundary, distance to the closest boundary of parcel).
FracTAL ResUNet 2021	code, CSIRO BSTD/MIT LICENSE, No pretrained model weights	Sentinel-2 images (blue, green, red, NIR channels). Australia, South Africa. Manually created dataset from mean-shift field detection algo.	In addition to the model, include a hierarchical watershed algorithm to post-process ResUNet model outputs and get defined instances of parcels.
U-TAE 2022	code, MIT license Have pretrained weights on Sentinel-2 for panoptic segmentation.	Sentinel-2 in France. PASTIS (Satellite Image Time Series of crop types dataset)	Multi-temporal model based on self-attention mechanism for the task of crop type segmentation.
BsiNet 2022	code, No license, Have pretrained weight on a Xinjiang GF-2.	China on GF-1/2 (1-2 m resolution). Bands - R, G, B.	Light-weighted model. The most computationally efficient among single-scene models. Additionally, the model has a flexible backbone structure that can be replaced by various state-of-the-art networks.
FAUNet 2023	code, No License, Have pretrained model weights	Demark dataset: EU Land Parcel Identification System (LPIS) + mosaic : Sentinel-2 dated 8 May 2016. Derived from Euro Data Cube Public Collections.	FAUNet (single-scene model) have a tendency to fall short in recall. One of the reasons could be that it only produces 2 output masks (edge and extent mask).
ST-U-net 2024	code , MIT license No pretrained model weights	Sentinel-2 France - for training, South Africa - tuning, and Kenya - validating.	Multi-temporal model based on simple architecture (U-Net).
Xia et al., 2024	code, No license No pretrained weights	High resolution (~1m) data from China drones. Dates 2020/2022.	Single-scene model that combines semantic edge and spatial structure graph frameworks and vectorizes the fusion of semantic edge and spatial structure graph outputs.
ftw 2024	code , MIT license Have pretrained model weights	Fields of The World dataset spans 24 countries on four continents (Europe, Africa, Asia, and South America) of multi-spectral Sentinel-2 satellite images.	Muti-temporal (two scenes from planting and harvesting seasons) U-Net with EfficientNet backbone model with output 3-class masks (field interiors, field boundaries, and background).
DelAny 2025	code , AGPL-3.0 license Have pretrained model weights.	FBIS-22M dataset of high-resolution satellite image patches (ranging from 0.25 m to 10 m – public data Sentinel-2, Planet, Maxar, Pleiades) and 22,926,427 instance masks of individual fields. Covering countries Austria, France, Luxembourg, the Netherlands, Slovakia, Slovenia, Spain, Sweden, and Ukraine.	The core of DelAny is the YOLOv11 instance segmentation model. Comparable performance of zero-shot experiments on geographic regions not included in the training set.

Table 1: Summary of the state-of-the-art field delineation DL models analyzed

CDSE : pipelines on Sentinel-2 satellite data

The Copernicus Data Space Ecosystem (CDSE) provides two distinct pipelines, leveraging Sentinel-2 satellite imagery and CDSE APIs (openEO, Sentinel Hub), for automated field boundary delineation. These comprehensive pipelines manage everything from data loading and pre-processing to deep learning inference and the final conversion into vector polygons for individual crop fields.

One pipeline, openEO Delineate, is open-source (GNU GPL) and offers a ready-to-use solution with included model weights. The other, from the Niva project (Sentinel Hub), is designed for large-scale operations, efficiently downloading data and parallelizing pipeline execution across Sentinel-2’s tiled data. The Niva project has two versions: V1 (2020) with MIT-licensed code (archived and no AI model pretrained weights), overview, webinar, results and V2 (2023) with available overview and results.

Geo Foundation models

Geospatial foundation models (GFM) are large-scale deep learning models pre-trained on massive and diverse geospatial datasets. Unlike traditional models trained for a specific task (e.g., crop classification, field delineation), GFMs learn generalized representations of Earth observation data (like satellite imagery, aerial photos, topographic maps, and even temporal sequences). This allows them to be adapted or “fine-tuned” for a wide range of downstream tasks with significantly less task-specific labeled data and computation resources.

Examples of the promising recent state-of-the-art GFM to be pretrained for the geo specific task – DOFA, Prithvi (already tested for the downstream task of crop classification), TerraMind.

Sentinel-2 field delineation datasets

Developing effective deep learning (DL) models for tasks like field delineation, whether by training from scratch or fine-tuning pre-trained models, critically depends on the availability of robust training, validation, and testing datasets.

Several prominent global satellite programs, including Landsat, Sentinel, and Gaofen, provide publicly accessible data with global coverage and consistent acquisition schedules. Among these, Sentinel-2 (part of the European Union’s Copernicus Sentinels program) stands out as the optical imagery satellite offering the highest resolution publicly available data.

The analysis of open-source field boundary datasets based on Sentinel-2 (Table 2) reveals diverse multi-country coverage and varying timeframes. A significant challenge for many of these open datasets (e.g., EuroCrops, LPIS data, fiboa) is their lack of accompanying Sentinel-2 imagery. To utilize these datasets for DL model training, an additional step of acquiring relevant Sentinel-2 data for the specified crop fields is necessary. Furthermore, the Sentinel-2 data provided within datasets that do include it typically consists only of the 10-meter resolution bands (Red, Green, Blue, Near-Infrared), with Sen4AgriNet being a notable exception. The most common Sentinel-2 product provided is Level-2A.

Source / license / year	Sentinel-2 (S2) images	Labels / annotations	Coverage (ToI, AoI)	Dataset size
AI4 boundaries MIT license 2022	Sentinel-2 level-2A monthly composites. Size 256*256 pixels. Bands - R, G, B, NIR, NDVI.	Sourced from GSAA data. Pre-processed raster format (vector label, boundary mask, distance mask, and field enumeration)	Austria, Catalonia, France, Luxembourg, the Netherlands, Slovenia, and Sweden 2019 year (March to August).	7831 tiles (image/mask). ~ 31 GB for masks, ~58 GB for images
EuroCrops CC-BY-SA-4.0 license 2023	None for now (Sentinel-2 data is not provided)	sourced from LPIS data (vector format). Includes Hierarchical Crop and Agriculture Taxonomy (HCAT) that harmonizes all declared crops in EU.	Austria, Belgium, Germany, Denmark, Estonia, Spain, France, Croatia, Lithuania, Latvia, Netherlands, Portugal, Romania, Sweden, Slovenia and Slovakia	~8.6 GB
Sen4AgriNet MIT license 2022	S2 12 bands, L1C images less than 10% cloud coverage. 366x366 images for 10m bands, 183x183 for 20m bands and 61x61 for 60m bands	sourced from Land Parcel Identification System (LPIS) to corresponding pixel based crop type maps	2016-2020, multi-country (France, Catalonia)	225,000 patches
AgriSen-COG CC BY 4.0 2023	S2 Level-2A: 366×366 less than 30% cloud. Bands - red, green, blue, and near-infrared. For crop classification tasks on parcel level and crop mapping on pixel-based level (in Zarr format)	sourced from original LPIS data (Geopackage and Parquet format) and the rasterized corresponding tiles - COGs (cloud-optimized Geotiff).	Austria (2019, 2020), Belgium (2019, 2020), Catalonia (2019, 2020), Denmark (2019, 2020), Netherlands (2019, 2020))	~28 GB (41,100 patches)
AI4SmallFarms CC BY 4.0 2023	Multi-temporal composites (S2) images to ensure cloud-free data. Bands - blue, green, red, NIR. patches 256 × 256	No information	Vietnam and Cambodia	62 tiles having a size of approximately 5×5 km
Fields of The World CC BY 4.0 2024	Multi-temporal (two scenes from planting/mid-season and harvesting/off-season) multi-spectral Sentinel-2 Level 2A. Bands - blue, green, red, NIR. patches 256 × 256	from government databases, published literature, and other websites with qualified quality assessment	24 countries on four continents (Europe, Africa, Asia, and South America) with varied ToI	70,000 samples
FBIS-22M CC BY-NC-SA 4.0 2025	dataset of high-resolution satellite RGB image patches (ranging from 0.25 m to 10 m – public data Sentinel-2, Planet, Maxar, Pleiades) for field boundary instance segmentation	LPIS boundaries were utilized for most regions, while high-resolution commercial satellite imagery was manually annotated for regions where LPIS data was unavailable, such as Ukraine. Additionally, the dataset was manually cleaned	Covering countries Austria, France, Luxembourg, the Netherlands, Slovakia, Slovenia, Spain, Sweden, and Ukraine.	672,909 high-resolution satellite image patches
Registered polygons of parcels updated annually	None	LPIS data	Netherlands, France (2010-now), Belgium, Denmark, Luxembourg

Table 2: Open-source field boundaries datasets

Conclusion, challenges and what’s next

As DaFab have explored, the landscape for tackling field delineation tasks offers a wealth of solution approaches and datasets. However, several critical considerations shape the effectiveness of these solutions:

• Data Challenges: The core need for high-quality, diverse, and meticulously annotated training data remains paramount.
• Field Variability: The inherent heterogeneity of fields worldwide – in size, shape, crop type, and agricultural practices – poses a significant challenge for models to generalize universally.
• Environmental Factors: Cloud cover frequently impedes optical satellite imagery, necessitating multi-temporal data fusion or integration with SAR (Synthetic Aperture Radar) data.
• Resolution Limitations: Lower resolution imagery, such as Sentinel-2’s 10m bands, can struggle with small or irregular fields, while higher-resolution alternatives, though more accurate, incur greater costs.
• Topological Accuracy: Ensuring that delineated boundaries form clean, closed polygons without overlaps or gaps often requires advanced post-processing.

DaFab ongoing research focuses on leveraging and adapting cutting-edge solutions to overcome these very challenges. If you are interested in more details, please get in touch.

P.S. Want to see DaFab work in action? Check out the code for DaFab current AI model for field delineation here: https://github.com/DaFab-AI-eu/NIVA-model

• Tech
• Agriculture
• Events