What is actually inside a grain result

What the instrument sees

Near infrared analysis measures light. Protein, starch, oil and water each absorb light at their own characteristic wavelengths, because of the way their molecular bonds respond to that energy.

Send light through a sample of whole grain, measure what comes out the other side, and the absorption pattern carries information about what the sample is made of. Measuring through the sample is called transmittance. It is the mode used for whole grain, because light that passes through the kernels samples the inside of them as well as the outside.

On Infratec™ and Infratec™ NOVA that means a scanning monochromator and a silicon detector, with an optical bandwidth of 7 nanometres. The cell the grain sits in has a variable path length, automatically controlled between 6 and 33 millimetres. The instrument sets that distance itself once the sample type is selected, because the right distance for light to travel differs by commodity.

What the detector records at this stage is a curve.

Turning a curve into a number

Turning that absorption pattern into a protein figure takes a calibration. A calibration is a model built by analysing real samples using laboratory reference chemistry, then matching each reference result to the absorption pattern the same sample produced.

The laboratory work still happens. It happens during data collection, season after season, which is what allows the model to be developed and kept current. At intake, each measurement then takes under a minute.

Two modelling techniques are used. Partial Least Squares finds a linear relationship between the spectrum and the constituent. Artificial Neural Networks allow the relationship to be non-linear, which matters because the effect of temperature, moisture and kernel structure on the spectrum does not scale in a straight line. The Infratec database holds calibrations of both kinds, and the global grain models are ANN.

The database behind those models holds over 50,000 cross checked samples, built on 30 years of harvest data. The samples come from different growing regions, varieties, seasons and conditions, which is what allows one calibration to work on grain the model has never seen. Our global ANN models are in use across the major grain producing regions.

The sample in front of the light

A calibration can only work on what it is shown, and one scan of one cell-full is a narrow view of the sample. The sampling that happens inside the instrument matters as much as the sampling that happens before it.

The instrument therefore measures the sample in portions, a cell-full at a time, and averages those readings into the figure that appears on the display. Ten of these sub-samples is the standard.

How much that matters depends on the sample. A uniform one gives much the same answer from any portion of it. A variable one does not, and averaging across portions is what keeps the result representative of the whole.

Dynamic sub-sampling takes this a step further. The instrument runs a statistical analysis of the sub-sample data in real time, during the analysis itself, and judges whether the sample is common and homogeneous enough to stop early. When it is, the measurement comes down to as little as 40 seconds. When it is not, sub-sampling continues until the figure is sound.

The instrument makes that call itself, judging how much sampling the sample needs and carrying it out.

How good is good

A calibration is tested on samples held back from the set used to build it, an independent validation set the model has never seen, and the result is published.

The figure to look for is the standard error of prediction, or SEP. It is the typical disagreement between the instrument and the reference laboratory, quoted as constituent per cent weight for weight. Alongside it sits the correlation between predicted and reference values.

The wheat and barley model is a worked case. We publish the figures for WB003034, the combined wheat and barley ANN model, which is built on 30,092 protein reference samples and 10,572 moisture samples, and was optimised using more than 50 independent validation sets from 14 countries.

Validated against an independent set of 11,822 protein samples across all grain types, it returns an SEP of 0.27 with a correlation of 0.99. On barley alone, 3,963 samples, the figures are 0.30 and 0.97. Moisture across 4,600 samples comes in at 0.24 and 0.99.

One further figure in the same paper carries more than any of those. Where a smaller validation set was analysed by a single reference laboratory, 67 wheat samples from the 2007 harvest, protein came back with an SEP of 0.16 against 0.27 across the full set.

Both sets ran on the same instrument. The smaller one simply had less variation between reference laboratories folded into it. Some of what reads as instrument error in any published accuracy number is laboratory error, which matters the next time a result is disputed against a lab report.

These figures come from our white paper on the global ring test study, and they are reported without bias correction. Performance on any given site will differ. Reference method, reference laboratory, harvest year, grain type and weather all move it.

Random error and bias are different problems

An SEP figure describes scatter. Readings fall on both sides of the true value, and across a run of samples that error averages out. Bias sits in one direction and stays there, so it does not average out, and an accuracy figure on its own does not say whether any is present.

Finding bias means comparing against a reference laboratory over a series of samples. One disagreement is scatter. A consistent offset in the same direction across a whole run is something else, and the two call for different responses.

Correcting it is the delicate part. An instrument can be adjusted to agree with a reference laboratory, but that assumes the laboratory is the one in the right. Our position in the ring test reporting is that adjustments should be applied only once the reference laboratory has been shown to align with other approved laboratories. The ring test exists partly to answer that question, and the answer is sometimes that the reference method needs the adjustment.

Keeping a model honest after it ships

A calibration built once and left alone would soon be describing a crop that no longer exists. Varieties and agronomy change, and seasons arrive that sit outside anything in the original data.

So the database keeps growing. Reference samples continue to be collected season after season, and new models are downloaded to instruments already in service.

The model is also checked against the rest of the industry every year. We run an annual ring test for the World Grain Network, in which around ten samples each of wheat, barley and rapeseed are cleaned, homogenised, divided and sent out to participating reference laboratories. It has been a proficiency testing scheme compliant with ISO/IEC 17043 since 2022, the global calibrations are maintained in accordance with ISO 12099:2018 and EN 15948:2020, and the 2026 round ran this spring.

Over seventeen years of annual ring tests, more than 100 organisations from 34 countries have taken part. The 2023 round covered 23 countries, the United Kingdom among them.

In that round, on samples from the 2022 harvest, the reproducibility standard deviation on protein was 0.17 for the reference methods and 0.11 for the global ANN model. On moisture it was 0.18 for the reference methods and 0.06 for the model. On those samples the instrument agreed with itself across laboratories more closely than the laboratory methods agreed with each other.

In Britain the same discipline operates nationally, through the UK Grain Testing Network. UKGTN is run by a steering committee drawn from the industry itself, with farmers, merchants, maltsters, millers, scientists and researchers on it.

Its work bears directly on anyone taking grain in. Before harvest, members' NIR instruments are standardised to the network sub-master, so the season starts with every site on the same footing. During harvest the committee monitors calibration accuracy against the reference method. It also runs ring checks on wheat, barley, oats and oilseed rape, along with erucic acid in oilseed rape and mycotoxin in cereals, and operates a TASCC accredited laboratory of its own.

That is the layer most often missing from a discussion about accuracy. A calibration validated globally still has to hold at a particular weighbridge in a particular British harvest, and the national network is how that gets checked.

Transferability is the related problem. Two analysers at two sites should return the same figure on the same sample, which is handled by manufacturing instruments to be identical from the factory.

Temperature is handled separately, by Infratec's patented stabilising technique. Protein is predictable to within 0.1 per cent at temperatures anywhere from minus 4 to plus 40 degrees Celsius, which covers an intake shed in February and a trailer standing in August sun.

Where results settle payment, the method behind them has standing of its own. Our wheat and barley ANN model is approved as a European standard for the simultaneous prediction of protein and moisture in whole grain, described in EN 15948:2020, the European method for determining moisture and protein in whole kernels by near infrared spectroscopy. Approval of the instrument itself for trade use is granted country by country.

What the number rests on

The instrument is the part you can see. It measures light and returns a figure in under a minute. Whether that figure is right is settled elsewhere.

What settles it sits behind the instrument, in the number of reference samples standing behind the parameter, the harvests and growing regions those samples span, the way the model has been validated, and whether the laboratories it is aligned to agree with each other. That work has no finishing point. It carries on through every season, which is the only reason a figure taken in a difficult harvest can be read the same way as one taken in a straightforward one.