Field Notes

The Machine Checks Its Homework

For thirty years, cleaning a combine's bad yield readings was a job for free federal software. A new Deere filing moves it inside the machine, into the control loop.

Clay-diorama editorial illustration of a small tractor and field-measurement markers.

What changes in this design

A reader might assume that a farming machine simply uses whatever readings it has gathered. That picture does not say whether the readings behind a result are identified or whether the result is tied to a single place.

Here, readings are collected together, the individual readings used are identified, and a result associated with one particular place is generated. A farming-related action signal is then generated; what happens after that is not stated.

Knowing this helps a reader follow how the readings relate to a result for one place and to the later signal. It does not state what the machine does with that signal.

Twelve seconds between cutting and weighing

There's a gap on a combine nobody can engineer away. Grain is cut at the header, then travels up to a load cell behind a plate at the top of the clean grain elevator, and that trip takes about twelve seconds. The monitor stamps the weight with wherever the machine is when the grain arrives - twelve seconds and some distance down the row from where it grew.

Assume nine seconds instead of fourteen and the yield map shifts visibly; researchers showed exactly that comparison in 2007. It isn't a rounding issue, I think. Readings get attributed to ground they never came from, and then the machine averages them, which smears the error wider still.

A free federal tool that reads Deere's files

I keep coming back to how old this problem is. The impact-plate sensor arrived in the early 1990s and hasn't fundamentally changed. Properly calibrated it's accurate to roughly 1 to 3 percent averaged across a field, but pass to pass most studies find 0 to 10 percent error. Buildup on the plate, chain tension, paddle wear, the machine pitching on a slope.

The Agricultural Research Service published a tool called Yield Editor in 2007 to clean this up: twelve filters, from grain flow delay to velocity limits to manual deletion. It's public domain, and it imports Greenstar text format, which means free federal software has been reading Deere's harvest files for nearly twenty years. The paper's summary of the literature is blunt. Researchers have reported that 10 to 50 percent of the observations in a given field contain significant errors and should be removed.

What Deere filed in January

Which brings me to US20260223763A1, published on August 6 2026, filed by Deere & Company that January, one named inventor, earliest priority November 2022. The application's own framing of the problem is candid, I think: it can be difficult to obtain instantaneous sensed values that are meaningful.

The claimed system aggregates samples, records where they came from, rebuilds a corrected value for a smaller patch of ground out of a subset of them, and hands that value to something that generates a machine action signal. That last part is the story. Yield Editor cleans a map afterward, on a desktop, for a person to read. This does it on board and steers with the result.

It's a continuation. The parent granted in late February 2026, and this one was filed about four weeks earlier - the family kept open rather than closed out. The parent also carries a classification the newer publication doesn't: autonomous vehicle control.

Two arguments against needing it

The first objection is academic. Seven weeks before that November 2022 priority date, two researchers posted an algorithm rejecting the whole filtering premise. Existing approaches, they wrote, rely on heuristic rules that delete problematic data. Their alternative throws nothing away. It treats each observation as an overlapping, oddly shaped patch and spreads harvested mass across a tessellation of the field, then smooths. Keep the bad samples, apportion them. The opposite instinct.

The second is duller. An extension agronomist reports that an uncalibrated monitor can be off by as much as 100 percent, that calibrating once at the start of harvest still leaves 7 to 10 percent error late in the season, and that calibrating properly across several flow rates brings errors to 0.5 percent or less. The manual fix already beats the problem by an order of magnitude. Some of what's being automated here is, I think, operator discipline.

And the payoff is thinner than the engineering suggests. Federal agricultural economists found that GPS soil and yield mapping raises operating profit on an average-size corn farm by almost 3 percent. Variable-rate technology: 1.1 percent.

The part worth watching

Deere has said it wants 10 percent of revenue recurring by 2030, with autonomy carrying much of it. This family was prosecuted through a bad stretch for iron - combine sales down 40.7 percent in one September against the year before, more than 2,000 production and salaried workers cut over about a year. A patent that turns noisy samples into a control signal sits on the subscription side of that ledger.

Four weeks before the application published, the FTC and five state attorneys general settled a monopolization case requiring Deere, for ten years, to give farmers and independent repairers the same repair resources its dealers get: fault codes, component pairing, restarting a machine after an emissions shutdown. The industry's 2014 pledge, which Deere signed, says farmers own the information generated on their operations.

So the raw samples belong to the farmer by pledge, and the repair tools by consent decree. The corrected value computed from those samples, on the farmer's own machine, is the thing under patent. Nobody has answered who owns that; I can't find an answer in this record.

Sources

Reporting for this piece drew on the published filing and the references below.

The record, briefly. US20260223763A1, “Agricultural Machine Control Based on Corrected or Localized Data Sample Values” The abstract describes aggregating sensor samples, identifying the samples used, generating a localized value, and generating an agricultural action signal.

More angles on this

When a Cart Needs to Know Which Vehicle It Is MeetingAnother reading from the same G05D classification family, systems for controlling or regulating non-electric variables.
The Leveling Bar Has a Very Firm Opinion About LevelAnother reading from the same A01B classification family, soil working in agriculture or forestry.
CPC A01B: Soil working in agriculture or forestry; parts, details, or accessories of agricultural machines or implements, in generalThe official classification family this record was filed into, published in the USPTO Cooperative Patent Classification scheme.
CPC A01D: Harvesting; mowingThe official classification family this record was filed into, published in the USPTO Cooperative Patent Classification scheme.
Ruts Have Filed a Formal ComplaintAnother reading from the neighbouring G05B classification family, control or regulating systems in general.

Read around this one:Public patent record USPTO published figure Patent issue-date index