Near-Infrared Spectroscopy

A near-infrared instrument records how strongly a sample absorbs light across a range of near-infrared wavelengths. Because different chemical bonds absorb at different, partially overlapping wavelengths, the resulting spectrum is a composite signal that reflects the sample's overall composition rather than a clean, single-constituent readout. Extracting a usable prediction from that composite spectrum requires a calibration model — typically built with multivariate statistical regression — that has learned, from a training set of samples analysed by both near-infrared and a primary reference method, which spectral features correspond to which constituent value.

Because the calibration model is fitted to reference values, near-infrared spectroscopy inherits every limitation of those reference methods and adds its own on top. A moisture calibration is only as good as the oven-drying results it was trained against; a protein calibration is only as good as the Kjeldahl or Dumas combustion results used to build it; an oil calibration is only as good as the solvent-extraction results behind it. Near-infrared spectroscopy can never be more accurate than the primary methods it is calibrated to reproduce — its role is to approximate those results faster and non-destructively, not to exceed them.

This dependence is ongoing, not a one-time event. Calibrations are monitored over time and re-standardised using fresh reference-sample checks, because the relationship between spectrum and constituent can drift as instruments age and as the population of samples the instrument encounters shifts.

• A calibration is built for a specific commodity, and often for a specific cultivar set, growing region, or season; applying it to a different commodity is unreliable, and applying it to grain from an unrepresented cultivar, region, or season can introduce a systematic bias that the instrument does not disclose.
• The model predicts reliably only within the range of constituent values and sample types represented in its training set. A sample well outside that range is extrapolated rather than interpolated, and the resulting number can look entirely plausible while being wrong.
• Outlier detection can flag spectra that fall outside the training population, but only if it is built into the calibration and actively used; a model without outlier detection, or an operator who ignores its warnings, loses this safeguard entirely.

The spectrum an instrument records depends on more than composition: particle size, grinding, packing density, optical path length, and temperature all shift the absorbance pattern for the same underlying sample. A calibration built on whole-grain samples is not interchangeable with one built on ground samples, even for the same commodity and the same constituent, because the physical presentation changes the spectrum the model was trained on.

Moving a calibration from the instrument it was built on to a different unit is not automatic either; instrument-to-instrument variation in optics and electronics means calibrations must be formally standardised across units before a model developed on one instrument can be trusted on another.

For intact fruit measured non-destructively, near-infrared light penetrates only a short distance into the tissue, so the reading reflects composition near the surface on the side actually measured — not an average of the whole fruit. A non-destructive fruit reading is therefore a local inference about one part of one piece of produce, not a whole-unit measurement, and this matters most for soluble solids content, where sugar distribution within a fruit is often uneven.