NIR Spectroscopy Glossary

ANKOM filter bag

Standard methods: AOAC 973.18, 2002.04, ISO 13906. Reproducibility ±0.5–1.0%. The reference for NIR fiber calibrations in forage, silage, and feed. Replaces older Goering-Van Soest manual procedure.

Bias

Mean prediction error

Bias = mean(NIR predicted − reference). A constant bias across the prediction range is the easiest calibration problem to fix — apply a slope/bias correction and you regain accuracy without rebuilding. A bias that grows with concentration means non-linearity, which slope/bias cannot fix.

Rule of thumb: Constant bias → slope/bias adjust. Bias trend → recalibrate.

Calibration transfer

Standard methods: Piecewise Direct Standardization (PDS), Direct Standardization (DS), slope/bias correction. Requires a transfer set of 10–20 diverse samples scanned on both instruments. Critical when expanding from one NIR to a fleet, or replacing an aging unit.

Dispersive NIR

NIR using a grating or prism to spatially separate wavelengths. Often diode-array detection — fast and rugged.

Dumas combustion

Modern alternative to Kjeldahl for protein. Combusts sample in oxygen, measures total nitrogen by thermal conductivity.

FT-NIR

Fourier Transform Near-Infrared

FT-NIR uses a moving mirror to encode wavelength information in time, then computes the spectrum via Fourier transform. Multiplex advantage (Fellgett) gives high SNR; wavelength-stable laser reference (Connes) gives sub-wavenumber accuracy. Preferred for multi-product labs and method development.

Karl Fischer

Reference method for moisture in low-water powders and oils. Volumetric or coulometric titration with iodine reagent.

Kjeldahl

Standard methods: AOAC 991.20 (milk), 928.08 (meat), 988.05 (cereals). Reproducibility ±0.10–0.20% protein. Takes 2–4 hours per sample. The reference NIR protein calibrations are built against. Multiply nitrogen by 6.25 (or 5.7 for grain) to get crude protein.

Leverage

Related to but distinct from Mahalanobis distance. High-leverage samples in the calibration set disproportionately influence the model — sometimes useful (extending the range), sometimes harmful (pulling the model toward an artifact). Always inspect leverage outliers before removing them.

Mahalanobis distance

Each new sample's spectrum is compared to the calibration set's mean and covariance. A high Mahalanobis distance means the sample doesn't look like anything the calibration was trained on — the prediction may be unreliable. Most chemometrics packages report this as a leverage flag or an "X-residual" warning.

Mean centering

Default first step in most chemometrics packages. Without mean centering, the first PLS factor is dominated by the average spectrum, wasting a factor on no useful information. Always applied unless you have a specific reason not to.

MSC

Multiplicative Scatter Correction

MSC fits each spectrum to the mean spectrum (slope and intercept) and removes the multiplicative + additive components. Works well when you have a representative mean spectrum, but the mean can be skewed if outliers are in the calibration set — SNV is more robust in those cases.

PCA

Principal Component Analysis

PCA is unsupervised — it doesn't use reference values. The first PC captures the most spectral variance, the second the next-most, etc. Score plots (t1 vs t2) are the standard view: clusters indicate sample groups, isolated points indicate outliers. Loadings reveal which wavelengths drive each PC.

PCR

Principal Component Regression

PCR runs PCA on the spectra to get scores, then regresses scores against reference values. PLS is generally preferred because it considers both spectra and reference values when forming factors, leading to lower-rank models with similar accuracy.

PLS

Partial Least Squares Regression

PLS solves the problem that NIR spectra have hundreds of variables but you typically have far fewer samples. Each factor explains some spectral variance and some reference variance simultaneously. Factor count must be tuned: too few = underfit, too many = overfit. Standard tools: Camo Unscrambler, Bruker OPUS, Foss WinISI, Thermo TQ Analyst.

R²

Coefficient of Determination

R² = 0.95 sounds good, but on a narrow protein range (8.5–10.5%), even a poor model can hit R² = 0.80. R² is misleading without RMSEP and RPD context. Always report all three together.

Rule of thumb: Don't cite R² without RMSEP and RPD on the same data.

Reflectance

Diffuse reflectance is what the term means in NIR — light penetrates 1–3 mm into the sample, interacts with multiple molecules, and the back-scattered light carries the chemistry. Used for ground feed, flour, milk powder, kibble, ground meat, anything that scatters strongly.

RMSEC

Root Mean Square Error of Calibration

RMSEC is computed using the calibration set. Because the model has seen those samples, RMSEC always underestimates real-world prediction error. Use RMSECV (cross-validation) or RMSEP (independent test set) for any decision about whether the calibration is good enough.

Rule of thumb: A calibration with low RMSEC but high RMSECV is overfit.

RMSECV

Root Mean Square Error of Cross-Validation

RMSECV approximates how the model will perform on samples it hasn't seen, without requiring a separate test set. Plotted against PLS factor count, RMSECV reveals the optimal number of factors — the elbow before overfitting begins.

Rule of thumb: Stop adding PLS factors when RMSECV stops decreasing.

RMSEP

Root Mean Square Error of Prediction

RMSEP is the gold standard performance estimate. You hold back 20–30% of samples from calibration development; use this set only once after the model is finalized. RMSEP close to RMSECV means the cross-validation gave a fair estimate. RMSEP much higher than RMSECV signals overfitting or sample population drift.

Rule of thumb: RMSEP > RMSECV by a meaningful margin = overfit model.

RPD

Ratio of Performance to Deviation

RPD compares the model's prediction error to the natural variation in the reference data. RPD = 1 means the model is no better than predicting the mean every time. RPD = 3 means real quantitative work for QC. RPD > 5 is excellent. The Williams scale (Williams 2014) is the common interpretation in food/feed.

Rule of thumb: RPD ≥ 3 acceptable for QC; ≥ 5 is excellent; < 2 is poor.

Savitzky-Golay derivative

First derivative removes constant baseline offsets; second derivative removes linear baseline drift and enhances peak resolution. Common settings: window 11–21 points, polynomial order 2–3, derivative order 1 or 2. Tradeoff: derivatives amplify noise — keep window large enough to suppress it.

SEC

Standard Error of Calibration

Similar relationship to RMSEC as SEP has to RMSEP: SEC² + bias² = RMSEC². On the calibration set, bias is typically zero by construction, so SEC ≈ RMSEC.

SEP

Standard Error of Prediction

SEP measures random error after subtracting any constant offset. SEP² + bias² = RMSEP². If SEP is small but RMSEP is large, the model has bias drift — fixable by a slope/bias correction without rebuilding the calibration.

Slope

A slope < 1 means NIR underpredicts at high values and overpredicts at low values (or vice versa) — the calibration "compresses" the range. Slope correction is the second half of the standard slope/bias adjustment used to extend calibration life without a full rebuild.

SNV

Standard Normal Variate

For each spectrum, subtract its mean and divide by its standard deviation. Each scan is treated independently, so SNV is robust to outlier samples in the calibration set. Standard preprocessing for ground/powder samples where particle size variation causes baseline shifts.

Rule of thumb: Default for most ground-sample applications.

Soxhlet

AOAC reference method for crude fat. Solvent extraction with petroleum ether or hexane over 4–6 hours.

Transflectance

Used when transmission would need too short a pathlength to be practical (very absorbing liquids). The reflector at the back of a 0.1–1 mm sample cup folds the optical path, giving 0.2–2 mm effective pathlength.

Transmission

Requires the sample to be optically thin enough to transmit light. Pathlength is precisely controlled (0.1–10 mm depending on matrix). Standard for raw milk in defined-pathlength cells, juice, beer, and whole-grain analyzers like the Infratec.