The stack
One county, one season.
Irrigated circles and dry range in southwest Kansas: 160 km² read in May, July and September, by optical and radar.
Sentinel-2 · Sentinel-1
Industry · Agriculture
Know what is growing, field by field.
Crop type, acreage and stress from a season of satellite passes. AgriScope compares each field with fields that grew like it, and a field with weak evidence goes to a person.
- AgriScope
- TerraScope
- Atlas
37.90° N · 100.92° W · 160 km²
A season of fields, read in seven passes.
Radar
SeasonOptical
Field 01 · 45 ha
Contains modified Copernicus Sentinel data (2024).
Real
Every layer is public satellite data. Fields, groups and stress are computed by a script in this site's repository.
Illustrative
The groups are unsupervised clusters of season curves, not crop labels. The county is a public example, not a client.
What runs here.
Each product below works on agriculture today. Atlas is the engine they share.
What it has shown.
Numbers from real runs, each with the caveat it needs. Nothing here is a claim the run did not make.
- 0.809
overall accuracy, seven crop classes, kappa 0.755
Five-fold spatial cross-validation on 136 photo-interpreted fields, for a national crop programme.
- 0.562
median field IoU from zero-shot segmentation, 6.5 times the baseline
Segmentation on vegetation imagery against a published baseline at 0.086.
- 0.901
boundary recall from graph segmentation
The classical method, kept beside the learned one.
How it works.
Open any part for the method, the numbers and where to read more.
Where parcel lines are old or missing, fields are drawn from the imagery itself. A segmentation model prompted on vegetation imagery found field shapes with a median IoU of 0.562, and half the fields cleared 0.5.
A classical graph segmentation recovered 90% of field boundaries. The two disagree in useful places, and the disagreement goes to review.
A season of optical and radar features per field, classified into seven crop types. Validation holds out whole areas, not random pixels, so the score is not flattered by neighbours.
Overall accuracy
0.809Kappa
0.755Labels
136 photo-interpreted fields
Agriculture
On screen.
Product screens from the cleared bundle. Names and accounts are painted out.
Why it matters.
Public figures on the size of the problem.
Saskatchewan insured 37.3 million acres in 2024 to 2025, carrying CA$13.3 billion in liability, with no live satellite view of the field.
Source: industry research (Saskatchewan Crop Insurance annual report)A record CA$2.6 billion drought payout followed the 2021 season.
Source: industry research (Government of Saskatchewan)
Agriculture · Next step
See it on your own ground.
Send an area and a question. I will say what the sensors can and cannot answer before anyone builds anything.