Cleaning and Reading Yield Maps Properly

Most raw yield maps are wrong in ways you can measure. That isn’t a complaint about the combine. A yield monitor is a mass-flow sensor bolted into a clean grain elevator, guessing how much grain came off a patch of ground ten or fifteen seconds after the header cut it. It does that thousands of times per field, and every one of those guesses carries some error. If you load that file straight into a variable-rate seed or nitrogen plan, the errors go into the plan too.

This piece is about removing those errors before anyone draws a management zone. Done properly, yield map analysis is the cheapest precision agronomy data you will ever own, because you’ve already paid for the combine and the harvest. Done badly, it’s a pretty colour picture that sends extra seed to the headlands and puts less nitrogen on your best ground.

What’s actually wrong with a raw map

Open almost any untouched harvest file in John Deere Operations Center, Climate FieldView or CNH’s FieldOps and you’ll see the same faults. Learn to recognise them, because each one needs a different fix.

Flow delay. Grain cut at the header takes time to reach the sensor. On most modern combines that’s 10 to 16 seconds. At 6 km/h the combine covers about 1.7 m every second, so a 12-second delay puts each yield reading roughly 20 m behind where the crop actually grew. When the delay setting is wrong, you get a sawtooth pattern: alternate passes look shifted up and down the field, and a wet hollow shows up as two half-hollows offset from each other.

Fill and empty. At the start of a run the elevator is filling, so the first few seconds read low. At the end the combine lifts the header while the elevator is still emptying, which leaves a short tail of logged points with grain but no real cut area behind them. Those points read very high or very low depending on how the monitor handles header position.

Headlands. Headlands get compacted, overlapped, sprayed twice, shaded by hedges and cut at odd speeds with partial headers. Some of the yield loss there is real. Most of what the map shows is a mix of real loss and junk from slowing down and turning. Either way, headland data shouldn’t drive in-field zones.

Swath width errors. When the operator cuts a part-width pass on a finishing run or next to a laid patch and doesn’t adjust the width, the monitor divides the full grain flow by a 9 m strip that was really 4 m wide. Yield on that pass falls by half.

Calibration drift and mismatch. Monitors drift with crop, moisture and a worn elevator. When two combines worked the field, you’ll often see stripes: one machine’s passes reading 9.4 t/ha next to the other’s 8.1 t/ha on identical ground.

GPS and speed glitches. Points stacked on top of each other when the combine stopped, points that jump 30 m sideways for one second, and readings logged at 0.3 km/h while unblocking a concave.

A worked example: one wheat field, one harvest

Take an 18.4 ha field of winter wheat in Lincolnshire, cut with a single combine and a 9.1 m header in August 2025. The raw file had 41,260 points. The monitor’s field total was 152.3 t, which works out at 8.28 t/ha. The weighbridge tickets from the grain store added up to 168.9 t at the same moisture basis, or 9.18 t/ha.

So before any filtering, the monitor is under-reading by about 10%. That matters less than people expect for zoning, because you’re looking at relative differences across the field. It matters a lot if you use the map to cost a variable-rate return per hectare.

The cleaning was done in USDA’s free Yield Editor and then checked in QGIS (also free). The same steps exist in paid platforms: FieldView has automatic cleaning on its paid tier, Operations Center lets you edit calibration and delay after the fact, and the Ag Leader SMS and Trimble Farmer Pro desktop tools have full filter sets. Here’s the log of what came out:

Filter                         Setting               Points removed   % of raw
------------------------------------------------------------------------------
Flow delay                     shifted 12 s → 14 s   0 (repositioned)   0.0
Start-of-pass delay            4 s                   1,612              3.9
End-of-pass delay              3 s                   1,208              2.9
Header up / zero flow          -                     2,044              5.0
Speed                          <2.5 or >9.0 km/h     1,391              3.4
Swath width                    <6.0 m                  877              2.1
Position jump                  >4 m off pass line      214              0.5
Headland buffer                18 m from boundary    5,960             14.4
Yield outliers                 mean ± 3 SD             692              1.7
------------------------------------------------------------------------------
Total removed                                       13,998             33.9
Points retained                                     27,262             66.1

A third of the data gone sounds brutal. It’s normal. Published cleaning studies regularly strip 20–40% of points, and the retained data is the part worth trusting.

Fixing the flow delay made the biggest visual change, and it removed nothing. Yield Editor lets you slide the delay in one-second steps while watching adjacent passes. At 12 s, the passes into and out of a known wet strip didn’t line up. At 14 s the strip came together as one clean band. Pick the delay where the field looks least stripy along the direction of travel. That’s the whole method.

After filtering, the field mean rose from 8.28 to 8.61 t/ha, because most of the junk was low readings. Rescaling to the weighbridge total (multiply every point by 168.9 ÷ 152.3 = 1.109) gave a calibrated cleaned mean of 9.55 t/ha for the in-field area. That’s higher than the weighbridge whole-field figure of 9.18 because the headlands, which you’ve just excluded, really were poorer.

Two combines, one field

If a contractor’s combine helped out, don’t average the two. Split the points by machine (most files carry a machine or session ID) and work out each machine’s mean on the passes where they cut similar ground. In one case, machine A read 9.40 t/ha and machine B read 8.10 t/ha on interleaved passes, a 16% gap. Scale machine B up by 9.40 ÷ 8.10 = 1.16, then rescale the whole field to the weighbridge total. The stripes disappear. If you skip this, a zoning algorithm will draw a high zone and a low zone in exact alternating 9 m bands, and the variable-rate drill will faithfully follow them.

Going multi-year: where yield map analysis earns its keep

One year of yield tells you about one season. A dry June, a slug-hit autumn or a lodged patch can dominate it. Zones worth acting on come from stacking years, and you can only stack them once each year is cleaned.

You can’t average tonnes across crops: 4.2 t/ha of oilseed rape and 9.5 t/ha of wheat aren’t comparable. So convert every cleaned year to relative yield, meaning each point as a percentage of that year’s field mean. Then lay a grid over the field (20 m squares suit a 24 m tramline system) and take the average relative yield in each cell for each year.

Here’s the same 18.4 ha field over five harvests:

For each grid cell, calculate two numbers: the mean relative yield across all five years, and the coefficient of variation (standard deviation divided by mean), which tells you how consistent that cell is. Then classify:

(The remaining slivers between 10% and 15% CV were merged into whichever neighbouring zone they bordered.)

The unstable zone is the interesting one. On this field it sat across a band of heavier clay that yielded well in the dry 2022 season and badly in wet 2024. Feeding it a flat “low zone” rate would be wrong in half the years. It’s a drainage question, not a seed rate question.

The stable low zone overlapped almost exactly with an EM conductivity scan done by SOYL in 2020, showing lighter, stonier ground. That’s the cross-check you want. When cleaned multi-year yield and an independent soil layer agree, you can act with some confidence. If you’re also pulling satellite NDVI or drone imagery into the picture, our guide to satellite, drone and variable-rate agronomy covers how those layers fit alongside harvest data.

Using AI tools without trusting them blindly

General AI assistants like ChatGPT, Claude or Copilot can genuinely speed this up for someone who isn’t a GIS specialist. Ask one to write a QGIS Python script that converts each year’s cleaned points to relative yield and grids them at 20 m, and you’ll usually get something that runs on the second or third attempt. Ask it what a coefficient of variation means for your field and it will explain it plainly.

Don’t ask it to choose your filter settings. It has no idea your header is 9.1 m, that the 2023 contractor’s combine reads low, or that the far corner was cut in the rain at 1 a.m. Those judgements are yours, made while you look at the map. Platform “auto-clean” buttons have the same blind spot. They’re a sensible first pass, and you should still check the flow delay by eye.

What it costs, roughly

Software can be free: Yield Editor and QGIS cost nothing. Budget your own time at about 45 minutes per field-year once you’ve done a few, so five years across 20 fields is around 75 hours. That’s a winter’s worth of wet afternoons, or a job you pay an agronomist or a precision service to do. If you’re buying it in, ask whether the price includes rescaling to weighbridge totals and per-combine correction. If it doesn’t, you’re paying for a tidier version of the same errors.

The return comes when you vary rates. Take seed: on that 18.4 ha field, moving the 2.9 ha stable low zone from 350 to 400 seeds/m² and the 5.1 ha stable high zone down to 300 seeds/m² roughly nets out on seed cost. The gain is in establishment where you actually need it, measured on the next cleaned yield map and not on the one that came straight off the combine.

Before you do any of that, check whether this year’s file has a calibration date logged. If it’s blank, the monitor has probably been running on last season’s numbers since the spring beans.