When people hear "we monitor farms from space," they picture the hard part as the space part — the satellites, the spectral bands, the physics of light bouncing off a leaf. That part is genuinely hard, but it is solved. The part nobody warns you about is deceptively small: where, exactly, does the field end? Draw that line wrong and every number you compute afterwards is quietly, invisibly wrong.
A satellite doesn't see fields. It sees a grid.
The European Space Agency's Sentinel-2 satellites already deliver free, calibrated imagery of nearly every field on Earth every few days. Their most-used bands land at 10-metre resolution — the satellite carves the world into a grid of 10 m × 10 m squares and reports one number per square, per band. It has no concept of "field," "hedge," or "road." Just pixels. Turning that grid back into your field is our job, not the satellite's.
The naïve approach is to draw a rectangle — a bounding box — around roughly where the field is, and average every pixel inside it. It's fast, it's easy, and it's wrong in a specific, damaging way: a rectangle around a real field almost always swallows things that are not the field.
Real fields are not rectangles
They curve along rivers, taper into corners, and sit next to roads, ditches, tree lines, and a neighbour's completely different crop. A bounding box averages all of that in — so the "crop health" you report is really a blend of the crop, the tractor track, and someone else's barley.
The bounding box (left, dashed red) pulls in a road and a neighbour's field. The true polygon (right) is traced to the field and then shrunk inward, so only clean interior pixels count.
Why a contaminated average is worse than no average
Imagine a healthy wheat field with an NDVI (a standard greenness/vigour index) around 0.80. Now imagine the bounding box around it also captures a gravel access road, a strip of bare soil, and the edge of a fallow field next door — all with NDVI near 0.15–0.25. The number you compute isn't 0.80 and it isn't 0.20. It's some muddy middle value that describes no real place on the farm.
That's the trap. A missing number is obviously missing — you know not to trust it. A contaminated number looks perfectly reasonable. It sits in a plausible range, it trends smoothly, and it will happily tell a farmer their crop is "slightly stressed" when it's actually thriving, because you diluted it with a road.
At a glance: bounding box vs. true field polygon
| Aspect | Bounding box easy way | True field polygon right way |
|---|---|---|
| What it captures | The field plus roads, ditches, hedges, neighbouring crops | Only the pixels that are actually the field |
| Effect on the average | Diluted, misleading, hard to detect as wrong | Represents the real crop condition |
| Non-rectangular fields | Worst case — more of the box is not the field | Handled correctly by design |
| Cost to build | Trivial | The real engineering work |
Even with a perfect polygon, the edge fights back
Say we do it properly and trace the exact outline of the field. We're still not done, because the boundary itself is fuzzy at 10 metres. A pixel sitting half on the crop and half on the road is a mixed pixel — its single reported value is a physical blend of both surfaces. No amount of careful outlining un-mixes it; the smearing is baked into the resolution. Research on Sentinel-2 boundary delineation shows exactly this: parcel edges get smeared across mixed pixels, and small misalignments between images can even masquerade as change along the edges (see peer-reviewed research on Sentinel-2 field-boundary delineation).
The practical fix is to treat the boundary as a zone, not a line — typically by shrinking the polygon inward (a small negative buffer, or "erosion") so the analysis uses only pixels safely in the field's interior. You trade a sliver of coverage for a big gain in trust. On a large field that costs almost nothing. On a small one it's a real tension — which is the next hard problem in the queue.
How SatField handles it
Our pipeline treats the boundary as a first-class problem, not an afterthought — four steps, in order:
Real geometry in from the start
We begin from the field's real outline — a boundary the grower supplies, whether drawn at sign-up, uploaded as a file, or referenced from an official parcel record — never a rectangle guessed from a single dropped pin.
Keep the interior clean
We buffer the polygon inward so mixed edge pixels never enter the average, leaving only confident, in-field pixels.
Remove cloud, shadow, water, bare soil
These pixels are dropped before anything is averaged — so a stray cloud can't quietly pull the number down.
An honest field number
The health figure a farmer receives is computed over genuine, in-field, valid crop pixels — and nothing else.
None of this is glamorous. There's no clever satellite in it. But it's the difference between a number a farmer can act on and a number that merely looks like one — and in a business built on trust, that's the whole game.
