Harnessing AI for Crop Boundaries to Scale Field Mapping

Join us for a webinar on October 2 where we will showcase how Land & Carbon Lab and WRI are harnessing the power of AI to create an open-source toolkit for accurate crop field boundary detection, starting with an AI model that automatically delineates boundaries in South America. This work is being carried out in partnership with Arizona State University and is supported by the Walmart Foundation.

2. AI Toolkit for Crop Field
Mapping
WEBINAR
October 2, 2025

3. Agenda
Agenda3
▪AI modelling approach
▪Creating a tailored model for South America
▪Release timeline
▪Panel discussion
▪Q&A

4. AI modeling approach

5. 5
Baseline: Fields of The World (FTW)
1
Kerner et al. (2025). Fields of The World: A Machine Learning Benchmark Dataset For Global Agricultural Field Boundary
Segmentation. AAAI 2025.https://fieldsofthe.world/
●FTW
1
is a deep learning modeland training
datasetfor field boundary segmentation
Baseline training data

6. 6
Baseline: Fields of The World (FTW)
1
Kerner et al. (2025). Fields of The World: A Machine Learning Benchmark Dataset For Global Agricultural Field Boundary
Segmentation. AAAI 2025.https://fieldsofthe.world/
●Segmentation model analyzes two scenesfrom the
plantingand harvestingmonths of a season
Baseline training data

7. 7
Baseline: Fields of The World (FTW)
1
Kerner et al. (2025). Fields of The World: A Machine Learning Benchmark Dataset For Global Agricultural Field Boundary
Segmentation. AAAI 2025.
2
Oldoni et al. (2020).LEM+ dataset: for agricultural remote sensing applications.
●FTW provides 70,000 sampleswith1.4
million fieldsfrom sources around the world.
1
●FTW dataset covers 24 countries
○Only ~1.5% of training data from South America
(Bahia, Brazil
2
)
Baseline training data

8. 8
Baseline: Fields of The World (FTW)
1
Kerner et al. (2025). Fields of The World: A Machine Learning Benchmark Dataset For Global Agricultural Field Boundary
Segmentation. AAAI 2025.
Baseline FTW model performs particularly
wellin large-scale agricultural contexts where
fields are regularly-shaped.
Segmentation Model
Salta, Argentina

9. 9
Baseline: Fields of The World (FTW)
1
Kerner et al. (2025). Fields of The World: A Machine Learning Benchmark Dataset For Global Agricultural Field Boundary
Segmentation. AAAI 2025.
Segmentation Model
Baseline FTW model strugglesin many
regions of South America where fields are
partially cleared, irregularly shaped, or
growing different crops than model was
trained on (e.g., soy-pasture rotation systems).
Chaco ecoregion

10. 10
New training data
We added new training data
through expert hand annotation.

11. 11
We added new training data
through expert hand annotation.
We targeted:
●diverse agricultural contexts
●areas to “challenge” the model
●frontiers of nature loss
New training data

12. 12
>37,000
fields

13. 13
from 17
ecoregions

14. 14
AI Toolkit for Field Boundaries
FTW dataset
1
has >70k examples
across 24 countriesand four
continents, butonly ~1200 in
South America(Bahia, Brazil
2
)
1
Kerner et al. (2025). Fields of The World: A Machine Learning Benchmark Dataset For Global Agricultural Field Boundary
Segmentation. AAAI 2025.
2
Oldoni et al. (2020).LEM+ dataset: for agricultural remote sensing applications.
Manually annotated >900 new examples
from 17 ecoregions in South America
Partner data(e.g., Mundo Agropecuario)
in South America
Pampas
Paraguay
Trainedsemantic
segmentation models
on combined dataset

15. 15
AI Toolkit cloud-based inference pipeline
Sentinel-2
Data
Sentinel-2 Raw Fields Processed FieldsTiles
Visual
Quality
Assessment
Configure location
(Sentinel-2 tile ID)
and planting +
harvest windows
(crop calendar)
Sentinel-2 input
mosaics
Segmentation
model predictions
(raster)
Field boundary
polygons
(vector)

16. 16
AI Toolkit outputs and quality assessment (QA)
AI-generated boundariesin Mato Grosso, Brazil
Annotators visually assess the quality of the AI-
generated boundaries
1. Visual QA
2. End-user feedback
End-users assess utility of field boundary products
in supply chain transparency applications
Early Access Programcoming soon!
Formal unbiased accuracy assessment of field
boundary maps in select regions
3. Quantitative validation

17. 17
A Generalized
Segmentation
Model (SAM) vs
AI Toolkit for
Crop Field
Mapping

18. 18
What we’re releasing and when
Segmentation Models
Models with commercial usage licensing and non-
commercial licensing that can:
-can be fine tuned
-applied to new geographies
Training Data
Over 40,000 field boundaries, >1000 training samples
-fully open-source, commercial usage
-organized by features: crop type, geography, country
for model adaptation
Field Boundaries
Field boundaries in target geographies
-Several states in Brazil, Argentina, and Paraguay

19. 19
What we’re releasing and when
Segmentation Models
-Models to be released in January/February 2026
Training Data
-Training data to be released January/February 2026
Field Boundaries
-Several states in Brazil, Argentina, and Paraguay
boundaries in January 2026, accessible NOW through
the early access program

20. Thank you
20