Instance segmentation for field delineation

Existing solutions and challenges

Land Parcel Identification System

LPIS has been done in many countries of EU

Solution?

U-net

Existing solutions (results)

Some results are already done for Sentinel-2 data (example for the Ukraine and Brazil)

Automatic field delineation: New release | by EO Research | Sentinel Hub Blog | Medium

Issues:

• The model detects boundaries through per-pixel segmentation, resulting in fields that appear visually delineated. However, in reality, they remain a single polygon in the vector file because the per-pixel approach does not ensure that the boundaries are fully enclosed.

FracTAL ResUNet

MultiTLF (Unet-based)

Segment Anything

How to apply?

Segment Anything

Existing solutions (results)

Some results are already done for Sentinel-2 data (example for the UK)

GitHub - Spiruel/UKFields

Issues:

• Contours have jagged edges due to raster vectorization.
• Some contours are cropped, most likely due to slicing into squares for model inference (red vertical line).
• For some fields, contours are missing or have an unusual shape.

Existing datasets

Field Delineation (France, 2017): Sentinel-2 224x224. The dataset consists of 1572 training samples, 198 validation samples, and 196 test samples.

Field Delineation - SustainBench (sustainlab-group.github.io)

Aung HL, Uzkent B, Burke M, Lobell D, Ermon S. Farm Parcel Delineation Using Spatio-temporal Convolutional Networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops 2020 (pp. 76-77).

Field Delineation (7 countries, 2019) consists of

Orthophoto imagery resampled at 1 m resolution, 512 by 512 pixels�Sentinel-2 monthly composites (March to August 2019), 256 by 256 pixels

5319 files for training, 1140 for validation, 1139 for testing

2.5 M parcels

GitHub - waldnerf/ai4boundaries

d'Andrimont, Raphaël, et al. "AI4Boundaries: an open AI-ready dataset to map field boundaries with Sentinel-2 and aerial photography." Earth System Science Data 15.1 (2023): 317-329.

Existing field boundaries

Agri-food Data Portal (Data on national and European agriculture and common agricultural policy (CAP), provided by the European Commission's agricultural and rural development department)

​

EuroCrops - is a dataset collection combining all publicly available self-declared crop reporting datasets from countries of the European Union.

Border Region Austria - Slovakia around Bratislava (© EuroCrops)

Challenges

Planet data for France

Dataset cleaning

Dataset cleaning

Proposed methodology:

Delineate Anything

Our solution

• A novel task formulation of field boundary detection as an instance segmentation problem.
• A new, large-scale, multi-resolution satellite imagery dataset.
• A resolution-agnostic model that significantly outperforms current SOTA for field boundary detection, with superior inference speed and strong zero-shot generalization.

Semantic vs Instance Segmentation�

Intersection over Union (IoU)

A novel task formulation

Field boundary masks (semantic segmentation)

Individual field masks (instance segmentation)

A new, large-scale, multi-resolution FBIS-22M dataset

• Existing Datasets: utilized existing Sentinel-2 and orthophoto datasets for various countries
• Sentinel-2:

France: 2022 field boundaries

Slovakia: 2020-2024 field boundaries

• Planet:

France & Slovakia: 2023 field boundaries

• Planet, Pleiades, Maxar:

Ukraine (10 target areas): 2022/2023 manually created & cleaned field boundaries

• Resolutions: 0.25m, 0.3m, 0.5m, 1m, 1.2m, 2m, 3m, 10m
• Note: Existing field boundary data requires manual cleaning

A new, large-scale, multi-resolution FBIS-22M dataset

Austria

France

Luxembourg

Netherlands

Slovakia

Slovenia

Spain

Sweden

Ukraine

Train/Test Distribution in Ukraine

W - train sites

W - test sites

Dataset coverage

A new, large-scale, multi-resolution FBIS-22M dataset

Image & Field Boundaries Samples

Maxar 0.3m

Pleiades 0.3m

Ortho 1m

Planet 3m

Sentinel-2 10m

Our solution

Normalize satellite images using 1 and 99 percentiles based on this study �How To Normalize Satellite Images for Deep Learning | by EO Research | Sentinel Hub Blog | Medium

​

Values for S2 calculated based on the AI4boundaries dataset

percentiles = {

'B2': {min: 94.5, max: 2266.0},

'B3': {min: 192.0, max: 2502.0},

'B4': {min: 132.0, max: 2818.0},

'B8': {min: 141.0, max: 5640.0}

};

​

​

​

Exclude the test squares (AI4boundaries dataset) from the new satellite data for 2022 for France.

Linear normalization

• c, d as min, max
• c, d as 1st and 99th percentile
• c, d as 1st and 99th percentile with the range bounded between 0 and 1

Dynamic World normalization scheme

• 30th / 70th percentile with log transform
• 30th / 50th percentile with log transform
• 30th / 95th percentile without log transform
• 20th / 95th percentile without log transform

Histogram equalization

Our Final Dataset

Resolution-Agnostic Model

AI models

Segment Anything (SAM)

SAM 2

1 stage detector vs 2 stages

You Only Look Once (YOLOv11)

Results

† : Models retrained on our FBIS-22M dataset for fair comparison. Latency (ms) represents the total time required to generate field boundaries.

Quantitative comparisons on the FBIS-22M

Qualitative results on the FBIS-22M test set

Qualitative results of zero-shot

Delineate Anything is applied to geographic regions with different climates, terrains, and agricultural practices, highlighting its field boundary delineation capabilities outside the training data.

Inference pipeline

Inference pipeline

​

​

​

Delineate

Compose

Instances

Input: 512x512 RGB

Output: masks & confidences

…

0.75

0.34

0.59

For each mask:

Original

Erode

Select largest

Dilate

• Compose largest first.
• Filter by land cover.

Delineate Anything

Image Tiling

Input: Multiple satellite images

Output: overlapping tiles

Inference pipeline

Merge

Vectorize

Before:

After:

Input: Final instance map

Output: Geopackge

• Polygonize
• Remove holes
• Area-based filtering

Result

• Find intersecting instances.
• Merge if overlap > threshold.

Input: Tiles with instances

Country level

• mAP @0.5 = 0.72
• mAP @0.5:0.95 = 0.477
• Latency < 25 ms
• Total time = 15 hours
• 2M+ of fields detected

Explore Our Results

Project page: https://lavreniuk.github.io/Delineate-Anything

Thank you!