Infrared Spectroscopy Near Infrared overview

See also: Infrared Spectroscopy: Near-Infrared. Photo-
acoustic Spectroscopy.

Further Reading

Chalmers JM and Dent G (1997) Industrial Analysis with

Vibrational Spectroscopy. Cambridge: Royal Society of
Chemistry.

Chalmers JM and Grifﬁths PR (eds.) (2002) Handbook of

Vibrational Spectroscopy, vol. 2. Chichester: Wiley.

Coleman PB (ed.) (1993) Practical Sampling Techniques

for Infrared Analysis. Boca Raton, FL: CRC Press.

Colthup NB, Daly LH, and Wiberley SE (1990) Introduc-

tion to Infrared and Raman Spectroscopy, 3rd edn.
San Diego, CA: Academic Press.

Harrick NJ (1967) Internal Reﬂection Spectroscopy. New

York: Harrick Scientific Corporation.

Messerschmidt RG and Harthcock MA (eds.) (1988) In-

frared Microspectroscopy: Theory and Applications.
New York: Dekker.

Mirabella FM and Harrick NJ (1985) Internal Reﬂection

Spectroscopy: Review and Supplement. New York:
Harrick Scientific Corporation.

Siesler HW and Holland-Moritz K (1980) Infrared and

Raman Spectroscopy of Polymers, Practical Spectro-
scopy Series, vol. 4. New York: Dekker.

Willis HA, van der Maas JH, and Miller RGJ (eds.) (1987)

Laboratory Methods in Vibrational Spectroscopy, 3rd
edn. Chichester: Wiley.

Near-Infrared

C N G Scotter

, Cherington, UK

Introduction

Since the 1970s, near-infrared (NIR) spectroscopy
has been applied as a qualitative and quantitative
analytical tool in the industrial ﬁelds of agriculture,
food, chemical and textile production, and pharma-
ceuticals. The technique is capable of making rapid,
nondestructive multicomponent analyses of organic
materials in a complex background matrix. In situ-
ations where inorganic materials are dissolved in po-
lar solvents an indirect NIR measurement can be
obtained. If the concentration of an analyte exceeds
0.1% then it is likely that an acceptable analytical
error will be achieved. NIR spectroscopy is essen-
tially an applied instrumental technique not often
used in a pure research context, unlike the related
mid-infrared (mid-IR) spectroscopy, the spectra from
which can be interpreted to identify specific organic
bond pairs, e.g., CO, OH, NH. This article discusses
the physical theory underpinning the use of NIR,
NIR calibration methodology, and types of instru-
mentation. Examples are given of the application of
the technique in the industrial contexts noted above.

NIR Theory – Molecular Spectroscopy

Two main areas of physical theory are relevant to the
application and understanding of NIR: ﬁrst, the
chemical principles derived from quantum physics
and molecular spectroscopy, and, second, the physi-
cal principles used to relate the spectra to transmis-
sion or diffuse reﬂectance of electromagnetic energy.

Mid-IR and NIR spectra are generated via appro-

priate instrumentation by electromagnetic energy ab-
sorption between 400 and 12 500 nm. Subsets of the
wavelength range are employed by instruments with
detectors and energy sources appropriate for obtain-
ing spectra in their particular range. The mid-IR
range is about 2500–25 000 nm, the NIR range
about 1100–2500 nm, and the visible range about
400–800 nm. In the mid-IR range it is usual to refer to
wavenumber or frequency in cm

according to the

formula wave number (cm

)

¼ 10

/wavelength (nm).

The mid-IR region provides fundamental absorp-

tion data, whereby for a specific bond pair the energy
transfer from one level to another occurs at a unique
frequency. It is for this reason that the mid-IR region
has been used to analyze the bond pairs in organic
molecules. It is the 3600–1200 cm

region that

gives rise to the overtones in the NIR part of the
spectrum. The NIR region thus consists of overtones
and combination bands that result in broad and/or
overlapping absorption peaks. It is for this reason
that NIR spectroscopy employs more statistical,
mathematical approaches to the analysis of organic
constituents, compared with the approaches used in
mid-IR spectroscopy.

Thus, the two regions, although instrumentally

and chemometrically distinct, are theoretically parts
of the same vibrational and rotational spectroscopic
model. Figure 1 illustrates the basic models for
energy absorption by covalently bonded atoms of
dissimilar mass. A dipole moment (produced when
atoms of dissimilar mass are bonded) is a basic
condition for absorption to occur. The H–H bond,
for example, does not produce mid-IR or NIR
absorbance. When molecules are subjected to an

INFRARED SPECTROSCOPY

/ Near-Infrared

415

external energy source they acquire the potential for
energy changes and the amount of energy that can be
imparted from a source is dependent upon the
wavelength, according to the formula:

¼ hc%n

where E is the energy, h is Planck’s constant
(6.626

J s), c is the speed of light, and

%n is

the wave number in cm

When organic molecules absorb energy from an

external source, e.g., a tungsten or halogen (while
tight) lamp, some molecules will move from the
ground state

ðn ¼ 0Þ to the next energy level ðn ¼ 1Þ

(see Figure 1). The combined vibrational and rota-
tional energy changes involved in the energy level
change constitute the absorption process.

Figure 1 illustrates the simplest theoretical treat-

ment for the energy absorption behavior of covalent
bonds. Hooke’s law curve corresponds to bonds
vibrating in simple harmonic motion. In practice,
covalent bond vibrations are anharmonic and follow
the potential energies depicted in the Morse curve,
also illustrated in Figure 1. It is this phenomenon of
anharmonicity that allows the overtones or harmonics
to occur in the NIR region. Thus, the frequencies of
the overtones are slightly less than whole number

multiples of the fundamental frequency. Greater
amounts of energy are required to displace mole-
cules to higher overtones and the numbers of mole-
cules displaced is in inverse proportion to the
overtone number. Transitions occur most frequently
from the ground state (n

¼ 0, see Figure 1) to the ﬁrst

level, and are increasingly infrequent for the higher
levels. Consequently, higher overtone absorptions are
much weaker than the fundamental.

Combination bands are also encountered in NIR

spectra. These result from energy transitions invol-
ving two or more different vibrational modes for the
same functional group. For example, CH

groups

have two symmetrical vibrational modes, a symmet-
ric stretching mode, and a corresponding in-plane
bending, or scissoring. The combination band result-
ing from these two vibrational modes is seen at
2320 nm. A ﬁnal factor that has been identiﬁed as
complicating NIR spectra even further is that of
Fermi resonance. In this condition, unalike sets of
bonds, e.g., the O–C–O bending mode and the sym-
metrical C–O stretching mode, occur at approxi-
mately the same frequency and a repulsion occurs.
One type of bond energy level is lowered and the
other type gains energy. In the case of carbon dioxide
this results in bands at 1388 and 1286 cm

instead

of one strong and one weak line at 1340 and
1330 cm

, respectively. Although relatively obvious

in the mid-IR region, Fermi resonance phenomena
are usually hidden beneath the broad overlapping
peaks in the NIR.

Although NIR absorbance peaks are generally

broad, spectral assignment is possible. A number of
authors have discussed bond assignment in the NIR
region and tables and lists of assignments of partic-
ular bonds and vibrational modes have been pub-
lished.

Transmission and Diffuse Reﬂectance

When NIR radiation is passed through a non-
scattering organic medium, the intensity of absorp-
tion can be described in terms of transmittance, T:

¼ I=I

where I is the intensity of the emergent radiation and
I

is the incident energy. For absorption spectra, I can

be expressed in terms of the Beer–Lambert relation-
ship:

log

ðI

=IÞ ¼ log

ð1=TÞ ¼ kcl ¼ A

where A is the absorbance, k is the molar ab-
sorptivity, c is the concentration of the absorbing
molecular species, and l is the pathlength of energy

Energy

Interatomic distance

Harmonic

potential

Anharmonic potential

(Morse curve)

The second overtone
from the energy transition
between

0 and

The first overtone from
the energy transition
between

0 and

The fundamental vibration
from the energy transition
between

0 and

Figure 1

Hooke’s law and Morse curves as functions of energy

and interatomic distance.

416

INFRARED SPECTROSCOPY

/ Near-Infrared

through the sample. For a ﬁxed pathlength the
absorbance is thus directly proportional to the con-
centration of the absorbing species.

In the case of samples that produce scatter in

transmission or diffuse reﬂectance spectra, a number
of factors corrupt the linearity of the Beer–Lambert
absorbance concentration relationship. Sample scat-
tering of radiation results in an alteration of the pro-
portion of absorbed and reﬂected radiation so that
pathlength becomes another unknown in the Beer–
Lambert relationship. Particle size, particle shape,
crystalline form, bulk, density, and the nature of the
pore space (ﬁlled with air, water, or oil) are all
variables that dictate the effective pathlength of the
radiation. Sample surfaces also reﬂect specular
energy that has not interacted with molecular struc-
tures. This form of energy has an overall effect on
spectra, contributing primarily to the curvilinearity
of the spectral baseline.

A number of attempts have been made to describe

both theoretically and practically diffuse reﬂectance
and scattering functions to enable a linear relation-
ship to be established between absorbance (A), ex-
pressed as log

(1/R) where R is the reﬂectance, and

molecular concentration. Perhaps the commonest re-
lationship encountered is that ascribed to Kubelka
and Munk, who established nine assumptions and 16
variables. These can be simpliﬁed to the Kubelka–
Munk function, namely:

ð1 R

where R

is the reﬂectance for a sample of ef-

fectively inﬁnite depth, K is a constant proportional
to the kc term in the Beer–Lambert relationship, S is
a scattering coefﬁcient usually taken as d

where d

is the diameter of the particles in a densely packed
scattering medium, k is the molar absorptivity, and c
is the concentration of the absorbing analyte.

NIR Calibration Development

The most important ﬁrst step in the practical appli-
cation of NIR is appropriate sample selection. To
take advantage of the potential rapidity and accuracy
of analysis that NIR spectroscopy can offer, the NIR
instrument must be calibrated against an acceptable,
accurate, and reproducible laboratory method for the
analyte or analytes of interest. The selection of a
sample set for NIR scanning and laboratory analysis
should be, as far as practically possible, represen-
tative of the material that the calibration will en-
counter when it is used to predict the targeted analyte
or analytes. Another critical aspect of NIR analysis

is appropriate and repeatable sample presentation.
Instrument manufacturers provide a plethora of
sampling devices that attempt to optimize presentation
criteria for specific sample types.

One approach to establishing a calibration sample

set is to scan a large number of samples that repre-
sent the potential population to be predicted and to
submit the spectra to an analysis that establishes
which combination of samples shows the largest
spectral variability overall. This smaller subset of
samples can then be analyzed by the laboratory
method. In practice, it will seldom be possible to
establish calibrations that will optimally predict all
future samples. It is therefore helpful to include new
samples in the calibration that are spectrally different
from the original calibration set. The standard error
of calibration (SEC) is used as an indication of the
efﬁcacy of a calibration.

It is equally important that a calibration is

tested against a set of samples different from
those of the calibration set. They should ideally
be separately collected but they must embody the
chemical and physical features encompassed by
the calibration set. It is the standard error of predic-
tion or performance (SEP), which deﬁnes the success
of a calibration. It is important, however, to relate
the SEP to the overall range of the analyte in ques-
tion.

The factors listed below are commonly responsible

for increasing the prediction error:

faulty laboratory analysis,

sample heterogeneity,

lack of repeatability in the packing and presenta-
tion of the sample,

instrumental malfunction.

The standard deviation for the residuals is due to

differences between the laboratory analytical values
and the NIR predicted values for samples in the cali-
bration set (SEC) or in a test set (SEP).

Mathematical Data Treatment

Although the theoretical models noted above attempt
to satisfy well-established physical principles, in prac-
tice most NIR workers start the calibration equation
development process with the orthodox log

1/R or

log

1/T spectra. Although calibrations have been

generated using the ‘draw’ log reciprocal data, it has
been found in most cases that some form or forms of
mathematical data treatment prior to regression anal-
ysis improves a calibration substantially.

Mathematical approaches that have been em-

ployed to obviate instrumental noise and other

INFRARED SPECTROSCOPY

/ Near-Infrared

417

physical effects in favor of chemical features in NIR
spectra include:

Smoothing functions

The simplest smoothing func-

tion employed is to average the absorbance (log 1/R
or log 1/T) values over a segment of the spectrum.
Many other smoothing algorithms are available to
smooth noisy spectra and reduce the numbers of
wavelengths in spectra. It should be noted, however,
that a high signal-to-noise ratio is characteristic of
NIR spectra and smoothing algorithms are primarily
used as an initial data reduction step.

Derivative functions

Derivative functions (strictly

termed differencing) provide a plot of the rate of
change of the slope in the spectrum. The average
slope is deﬁned over a spectral segment that is fol-
lowed by the omission of several wavelengths
(a gap). The gap may be zero. The process is repeated
along the spectrum to its ﬁnal wavelength. This ﬁnite
difference technique is usually combined with a
smoothing algorithm. The size of the gap and
segment may be varied. This technique has the
effect of removing parallel baseline shift due to particle
size differences in the samples, but it retains slope
differences between spectra, and thus retains the
basic peak height molecular species concentration
relationship. Derivatized spectra contain lobes and
valleys and inﬂected peaks and are therefore difﬁcult
to interpret. Up to four orders of derivative have
been employed with NIR spectra. Trial and error and
experience with gap and segment sizes provide
optimal solutions for calibrations and predictions.

Multiplicative scatter correction (MSC)

The model

is based partly upon theoretical considerations of
radiation scatter and partly upon empirical observa-
tions of spectra.

If x

; x

; y; x

are the absorbance values of a sam-

ple and

; %x

; y; %x

are the average absorbance values

for the set of samples. The regression model used is:

¼ a þ b %x

þ e

The coefﬁcients a and b are estimated by least
squares over the wavelength range and the scatter-
corrected spectrum is obtained by subtracting a and
dividing by b, i.e.,

knew

¼ ðx

kold

aÞ=b

e is a term accounting for ‘error’ caused by, for
example, random measurement noise. MSC also
requires a linearity transform to ensure a linear
relationship between absorbance values at wave-
lengths absorbing for the analyte of interest.

Standard normal variate (SNV) and detrending

For

a given set of similar samples the reﬂectance spectra
will vary in their slope and curvilinearity according
to the particle size distribution between samples,
differences in crystallinity, and differences in the
chemically nonspecific scatter at the sample surfaces.

The SNV function for a 700 wavelength spectrum

may be expressed as:

SNV

¼ y

ðy

%yÞ

1=2

ðn 1Þ

¼ 1; y; 700

where SNV

ðn ¼ 1; y; 700Þ are the individual SNV

terms for each of the 700 wavelengths, y

are the

700 log 1/R values, and

%y is the mean of the 700 log 1/R

values.

A dominant feature of NIR diffuse reﬂectance

spectra is the increase in absorbance values from
1100 to 2500 nm. This trend is curvilinear for densely
packed samples. A second-degree polynomial func-
tion has proved to be an adequate model to linearize
the spectral baseline.

The list of NIR spectral transformations is not ex-

haustive but it exempliﬁes the physical features that
are perceived to interfere with the chemical absorb-
ance features and the approaches used to overcome
the perceived interference in diffuse reﬂectance NIR
spectroscopy.

None of the above approaches optimizes the rela-

tionship between NIR absorbances and analyte for a
range of sample types. Derivative transformations
have been found to be generally useful when stepwise
multiple linear regression (SMLR) techniques are
used. When multidimensional statistics are em-
ployed, e.g., partial least-squares (PLS), principal
component regression (PCR), or neural nets, it has
been observed in some cases that the untransformed
log 1/R data can perform just as well in correlation
coefﬁcient and error terms as in any kind of trans-
formation. It is considered that in some cases physi-
cal manifestations of the sample contained in the
spectra provide valid and useful discriminant data.

Fourier and Wavelet Transforms

In the past, the large amount of spectral data gene-
rated by NIR instruments challenged the ability of
computers to provide computations within a reason-
able time frame. Mathematical techniques, therefore,
which offered a reduction of the raw data, but with
minimum loss of information, were often employed.
The decomposition of 1000 spectral data points to
100 Fourier transform coefﬁcients provides a great
saving in computational time with very little loss of

418

INFRARED SPECTROSCOPY

/ Near-Infrared

spectral information. Large memory computers have
meant, however, that NIR workers can derive func-
tional calibrations or discriminant functions using all
the spectral data. With the advent of spectral imaging
datasets, and very large sample databases, the need
has reemerged for efﬁcient data reduction tech-
niques. One of the most recent, and promising tech-
niques is wavelet transformation. This technique
employs various wave functions that can be com-
bined to represent a spectrum. Wavelet transforms
have not been used widely so far, but there is hope
that they may provide a powerful way of data re-
duction without loss.

Developing the Regression Equation

Until recently, it was most common to use SMLR to
develop a calibration model. A condition for the use
of the technique is that the relationship between ab-
sorbance expressed as log

1/R or log

1/T and the

laboratory measurement should be very close to linear.
The specular reﬂectance and particle size effects seen
in spectra discussed below also affect the optimal
reduction of the SEC and SEP.

When the laboratory value is plotted against the

NIR predicted value for the calibration sample set it
may well be noted that some points lie well away
from the computed regression line. This will, of
course, reduce the correlation between laboratory
and NIR data and increase the SEC or SEP. These
samples may be outliers. The statistic h

describes the

‘leverage’ or effect of an individual sample upon a
regression. If a particular value of h

is exceeded this

may be used to determine an outlier sample. Evalu-
ation criteria for selecting outliers, however, are
somewhat subjective so there is a requirement for
expertise in multivariate methods to make outlier
selection effective.

Intercorrelation (or multicollinearity) phenomena

also reduce the effectiveness of a calibration. First,
there can be intercorrelation between the constitu-
ents in the samples. For example, water and fat in
some meats are negatively correlated and it is pos-
sible to obtain the same calibration wavelengths for
both components although the two components do
not absorb in the same band.

Particle size differences between samples cause all

calibration coefﬁcients to change because of the
baseline shift of the spectra. Furthermore, adjacent
wavelengths are also highly correlated, so it is
difﬁcult for the SMLR algorithm to ‘focus’ on a single
combination of wavelengths that will optimize the
terms in the calibration. There are also correlations
between areas in the spectrum that represent over-
tones of the same covalent bond.

It is often suggested that slope and bias corrections

should be made to computed regression models, i.e.,
an adjustment be made to the regression line so that
it can be extrapolated through the zero laboratory/
NIR prediction values (bias or intercept). This pro-
cedure will also adjust the slope term in the equation.
The most recent view of this technique is that it
should be used with utmost caution, in that if such
adjustments are apparently required to improve the
statistical results then it probably indicates that the
original model requires improvement or that labora-
tory or NIR measurements are at fault.

Multidimensional Regression Techniques

There are a number of multidimensional, multi-
variate techniques that are capable of overcoming
many of the effects that corrupt calibrations, without
necessarily using data pretransformations. These
techniques also involve wavelength data compres-
sion, resulting in a decrease of computational time.
Equations are also derived from functions of the
whole spectrum. PLS and PCR are the most com-
monly used of these techniques. Both technique
derive linear functions from the wavelength data
(latent variables in PLS, principal components in PCR).
Typically, a few of the functions in relationship to the
sample number will provide sufﬁcient data to gene-
rate coefﬁcients for good calibrations equations.

In PCR, the components are derived by extracting

the maximum amount of residual variation from the
spectral data. Each principle component is ortho-
gonal to the next (they have 0 correlation). In this
way the wavelength intercorrelation problem noted
above is alleviated. PLS functions are derived by in-
terrelating the laboratory value variability and the
NIR spectral variability and latent variables have low
intercorrelation. Although a number of writers point
to relative advantages for these techniques, in prac-
tice, used properly, they provide similar solutions.
Fourier transform techniques are also used for cali-
bration. Once again the wavelengths are compressed
into a relatively few (Fourier) coefﬁcients. In com-
mon with PLS and PCR, the adjacent Fourier func-
tions (coefﬁcients) have a low intercorrelation.

The three techniques of multidimensional calibra-

tion noted above also generate functions that are
valuable for the chemical and physical interpretation
of spectra.

Principle components regression, for instance,

generates ‘loadings’ for each principal component
(PC) that may be plotted against wavelength, to
illustrate the importance of specific wavelength
regions to the regression analysis. The peaks and
troughs of the plot correspond to high positive or

INFRARED SPECTROSCOPY

/ Near-Infrared

419

low negative loadings on the ordinate axis and are
indicative of the degree of inﬂuence that the
wavelength regions, encompassed by the peaks or
troughs, have on the ﬁnal model. Since the loadings
are calculated from spectral variation in the whole
sample set, the loadings against wavelength plot re-
ﬂects spectral features in the dataset. Figure 2 shows
the loadings against wavelength plot for the second
PC derived from a set of orange juice samples. In this
plot, the peak at 2100 nm is homologous with a peak
in the orange juice spectra at 2100 nm (see Figure 3),
which in turn is a characteristic peak in the spectrum
of sucrose, and sucrose is a major varying constituent
of the orange juice samples from which the PC was
derived. Figure 3 illustrates an orange juice spectrum.
In Figure 4, the ﬁrst PC loadings are plotted against
wavelength for the orange juice reﬂectance data
referred to above. In this case, the plot shows a weak
‘smoothed’ orange juice spectrum (cf. Figure 3). This

represents a spectral pattern that is generated by
specular (mirror-like, no interaction with molecular
bonds) reﬂection and particle size differences be-
tween samples. Figure 5 shows a series of leaf tea
reﬂectance spectra where the spectra are very similar
in shape but the spectra are spaced up and down the
ordinate axis. This is a typical spectral response to
samples with particle size differences.

Fourier analysis offers the process of self-

deconvolution that is capable of resolving under-
lying peaks from broad overlapping bands. Fourier
self-deconvolution has the advantage over derivati-
zed spectra that there is no shift or averaging effect
and the self-deconvoluted spectrum is directly com-
parable with the log/reciprocal wavelength scale.

Two other techniques of calibration equation com-

putation have recently been used by NIR workers,
locally weighted regression (LWR) and artiﬁcial
neural nets.

0.80

0.60

0.40

0.20

0.00

Loadings

−

0.20

−

0.40

−

0.60

1100

1300

1500

1700

Wavelength (nm)

1900

2100

2300

2500

Figure 2

Second principal component loadings versus wavelength plot of a set of 25 components used for canonical variates

analysis of orange juice NIR reﬂectance spectra.

8.36

5.42

2.49

−

0.45

−

3.38

−

6.32

1100 1300 1500 1700

Wavelength (nm)

1900 2100 2300 2500

log 1/

Figure 3

log 1/R spectrum of orange juice dried onto a glass

ﬁber disk.

1100

1300

1500

1700

Wavelength (nm)

1900

2100

2300

2500

0.90

0.72

0.54

0.37

0.19

0.01

Loadings

Figure 4

First principal component loadings versus wavelength

plot for a set of orange juice reﬂectance spectra.

420

INFRARED SPECTROSCOPY

/ Near-Infrared

LWR has been recommended to counteract the

common inherent nonlinearity existing between labora-
tory values and the spectral absorbances. Essentially,
segments of the NIR versus laboratory value regre-
ssion computed are used to give a much better ap-
proximation to linearity than a regression line model
for the whole dataset.

Neural Networks

Neural networks have been used, most effectively, as
a way of establishing robust NIR calibrations. It is
the calibration method of choice for a whole cereal
grain, transmission instrument, which is being used
worldwide for many of the analyses required for esti-
mating grain quality. An artiﬁcial neural network
program performs many iterations in order to estab-
lish the optimum solution required from large, com-
plex datasets.

Discriminant Analysis

NIR instruments may not only be calibrated for
quantitative analysis, but they may also be ‘trained’
for qualitative purposes. This process is usually
termed discriminant analysis. The criteria noted
previously for establishing quantitative calibrations
with minimum prediction error are equally applica-
ble to discriminant calibration sets.

Spectral matching algorithms are also employed

using multidimensional mathematics. Such tech-
niques are often used in quality assurance programs
where a spectrum of a known material is compared
with incoming raw materials to check its purity or
authenticity. For example, samples of orange juice
concentrate were collected to represent the types
available from two countries. In addition, samples
that were known to be adulterated were obtained
from one of the countries. All other samples were

characterized as authentic. The question asked was
‘Can NIR spectra be used to discriminate between
the three categories of juice concentrate?’ That is,
country A, country B (authentic), and country B
(adulterated).

The technique chosen to perform the analysis was

canonical variates analysis. Briefly, the method ab-
stracts functions from combinations of PCs. It has
been found with large sample sets (

D100 or more)

that up to 25 PCs give optimal predictive ability. It is
unwieldy to select the predictively important analytes
from 25 PC dimensions. For this reason the
PC-derived canonical varieties are computed wherein
the dimensions available are one less than the number
of ascribed characteristics or groups. In the example
noted above, therefore, the two CVs describe the 25
PC dimensions, since there are three groups in the
analysis. In this example, a calibration discriminant
analysis was calculated using 71 samples, and 23
prediction samples were used to test the calibration.
Figure 6 illustrates the two-dimensional CVs’ cali-
bration and prediction.

One problem that is evident from this analysis is

that samples that lie outside the deﬁned circles (95%
of the samples sample points should lie within the
group circles providing each group has the same vari-
ance), may (1) belong to one of the deﬁned groups or
(2) belong to a population not deﬁned in the cali-
bration. At present no satisfactory method exists to
make an objective decision about such ‘outlier’ sam-
ples. One approach to help with this problem is to
employ cluster analysis techniques. In this case sam-
ple spectra are allotted to specific groups in a hier-
archical fashion, beginning with a division of the
sample set into two groups, then four, and so on.
This type of analysis requires decisions to be made
about appropriate distance and similarity functions,
which are required to perform the computations.
There are many such functions and only experience
of their use leads to valid analysis and samples that
may be truly outside the group populations will at
some level of the analysis be included in a group.

A Nonregression-Based Technique

Comparison analysis using restructured near-infrared
and constituent data (CARNAC) can be used to
provide quantitative or discriminant analysis for
large databases. In the current version of the tech-
nique, CARNAC D, wavelet transforms are used to
reduce the spectral information, and then the spec-
trum of the sample requiring a particular analyte
measurement or discriminant comparison (trans-
formed in the same way) is compared with the ‘cali-
bration’ database. The sample giving the closest

1100

1300

1500

1700

Wavelength (nm)

1900

2100

2300

2500

1.05

0.86

0.67

0.48

0.29

0.10

log 1/

Figure 5

log 1/R spectra of tea leaf samples showing the

‘stacking’ of similarity shaped spectra along the ordinate axis.

INFRARED SPECTROSCOPY

/ Near-Infrared

421

match gives the solution to the required analyte or
discriminant identiﬁcation.

Types of NIR Instrumentation

Until the late 1980s there were only two basic types of
NIR instrument, those employing discrete wavelength
ﬁlters and wavelength scanning instruments using
holographic gratings. Both types employ high-intensity
white light sources and one or two types of detector
depending upon the wavelength range being used.
For the NIR region (1100–2500 nm) a lead sulﬁde
detector provides appropriate sensitivity and in the
higher-energy visible/far visible region from 400 to
1100 nm less sensitive silicon detectors are used.

Detector geometry is important to optimize the

collection of diffuse reﬂectance energy in favor of
specular reﬂectance energy that contains no chemical
absorbance

information.

Two

approaches

are

commonly used. One uses an integrating sphere that
provides a double-beamlike instrument, whereby the
arrangement of the detectors within the reﬂective
sphere enables a reference and a sample beam to be
employed. The reference spectrum is subtracted from

each sample spectrum obviating the inclusion of
purely instrumental contributions to the sample
spectrum. Single-beam instruments, using a detector
array set to 45

1 to the sample surface, must have a

reference material to scan, the spectrum of which is
subtracted from the sample scan. The full-wave-
length scanning instruments are more expensive than
ﬁlter instruments and are in general used for cali-
bration development and applied research.

In the late 1980s and into the new millennium in-

strumental approaches to the NIR region have in-
creased substantially and some of these new
approaches are noted briefly below.

Calibration Transfer

NIR instruments of all kinds differ in there optical
and operational properties. Instruments of the
same kind and make have small differences, but
these produce differences in spectra that are sufﬁcient
to erode the accuracy of a calibration if it is trans-
formed from one instrument to another. There has
been, and there continues to be, an interest in ways of
transforming calibrations, such that, when they are

−

First canonical variate

Second canonical variate

−

Country B

(adulterated)

Country B

(authentic)

Country A

Figure 6

Canonical variates calibration and prediction sample plot of NIR diffuse reﬂectance spectra characterized as:

m, D, country A;

K, Jcountry B (authentic); and ’, &, country B (adulterated). Solid symbols, calibration samples; hollow symbols, prediction samples.

422

INFRARED SPECTROSCOPY

/ Near-Infrared

transferred from one instrument to another they re-
tain their predictive accuracy in the new instrument.

Transmission through Solids

Using the higher-energy visible region, instruments
have been made to provide transmission spectra of
solid samples that have been satisfactorily calibrated
for kernel hardness of gain, meat components, and
tenderness of peas, for example.

Photodiode Array Instruments

Photodiode array (PDA) type of instrument is solid
state (no moving parts) and because of the electronic
switching of the diodes providing the NIR illumina-
tion, it is very rapid scanning, compared with other
instrument types. Such a construction lends itself to
miniaturization, and the result is physically robust,
enabling it to be used in many situations where other
instrument types could not be used, including situ-
ations where a hand-held analyzer is required. At
present the wavelength range is limited between 500
and 1100 nm (other instruments can provide wave-
lengths up to 2400 nm). The spectra from a PDA
instrument also tend to be noisier than spectra from
other instruments. Currently, this type of NIR ana-
lyzer is being used as an integral measurement system
in a cereal combine harvester. It is also being used
experimentally to estimate quality parameters of
whole fruit on the tree. The uses for this NIR tech-
nology are increasing rapidly.

NIR Image Analysis

Developments are currently taking place in using the
visible and NIR region for video imaging purposes.
Experiments have been carried out to image meat sur-
faces and the textural appearance of bread crumbs.

A leading pharmaceutical manufacturer has devel-

oped an NIR-based image analysis technique to
image the ingredients of capsules and tablets to check
there even distribution in a formulated product.

Diffuse reﬂected NIR radiation from the surface of

the sample is collected by a series of imaging optics
and passes through an NIR tunable ﬁlter prior to for-
ming an image on an IR focal plane array detector. The
ﬁlter is continuously stepped through a predetermined
spectral interval and an image at each wavelength is
stored. The resulting data hypercube is analyzed by
proprietary principal component-based software.

Industrial Uses of NIR Instrumentation

NIR spectroscopy is an application-driven technique
and perhaps one of the greatest motive inﬂuences in
the development of the many kinds of NIR instrument

has been their application to at-, on-, or in-line anal-
ysis for factory processes. Once a calibration has
been developed, the NIR technique does not require
chemicals and is capable of rapid, multicomponent,
noninvasive analysis. Two features of NIR instru-
mental development are most signiﬁcant in this con-
text. The ﬁrst is the increasing use of optical ﬁbers,
which enables the instrument to be multiplexed and/
or remote from the line. Optical ﬁbers are also used
within instruments to optimize the light path geometry.
The second feature is the development of nonmoving
part or solid-state spectrometers that are light,
rugged, long-lasting, small, and up to 10 times less
expensive than their equivalents with moving parts.
As has been noted, this type of instrument can also be
developed to be handheld.

Industrial Applications of NIR
Technology

Applications in Agriculture and the Animal Feed
Industries

In the agriculture and animal feed industries NIR
instruments, and in particular discrete ﬁlter instru-
ments, have found the widest application in the
analysis of grain and ﬂour. Protein, moisture, hard-
ness, and baking quality are all examples of analysis
that have been performed by NIR instrumentation,
both off- and online. Several countries, including the
USA and Canada, have adopted NIR protein analysis
as national standards to gauge the level of payment
for wheat. The protein method was approved by the
AACC (American Association of Cereal Chemists) in
1983.

In the feed and forage industries NIR methods

have been used extensively. NIR factory-based quality
control schemes have been implemented which
enhance the economy with which feed components
are used, and the quality and consistency of the ﬁnal
product. Moisture, crude protein, and acid detergent
ﬁber analyses of feeds are being steered through the
AOAC (Association of Ofﬁcial Analytical Chemists)
validation process. NIR is also used widely in the US
tobacco industry for measuring nicotine and other
process parameters both on- and offline.

Applications in the Food and Beverage Industry

It is perhaps the food and beverage industry that
presents the most difﬁcult challenge to both off- and
online NIR analysis. The materials are organically
complex, and often need to be analyzed in the pres-
ence of a high percentage of moisture, the broad,
highly absorbing bands of which, in the NIR region,
tend to obscure the lesser absorbing constituents.

INFRARED SPECTROSCOPY

/ Near-Infrared

423

Table 1

Matrix of food products and components and characteristics analyzed offline and online by NIR

Food product

Acids,

AlcoholAmino acids

Ash Caffeine

Capsaicin

CaseinCellulose

Chlorophyll

Color Dietary fiber

Egg Fat/oilFungal spores

-Glucan

GlutenHardness

Heat treatment

Hot water extract

Howard mold count

Insect

Iodine value

Lactose

Methionine

Moistur

Nicotin

Nitrogen (total)

NylonOriginal gravity

Pasteurization

Polypropylene

Polythene

Proportions in a mixture

Protei

Salt Sedimentation volume

Soya bean flour

StarchStarch damage

Sucrose

SulfurSugarsTenderomete

Total solids

Water absorption

Barley

Beer

Beet sugar

Biscuit

Bread

Breakfast cereal

Cake mixes

Cheese

Chocolate

Cocoa

Coffee

Dough

Edible oils

Fish

Flour

Hops

Ice cream

Liquors

Maize

Malt

Margarine

Meat

Milk products

Nondairy creamer

Oats

Oilseed

Packaging laminate

Pasta

Peas

Peppers

Potato

Pulses

Rice

Salad cream

Sausages

Spices

Tea

Tobacco

Toffee

Tomatoes

Triticale (rye)

Wheat

Wine

Yogurt

−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
×
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
×
−

−
×
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
×
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
×
−

×
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
×
−
−

−
−
−
−
−
−
−
−
−
−
−
−
−
−
×
−
−
−
−
−
−
−
×
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−

−
−
−
−
−
−
−
−
−
−
×
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
×
−
−
−
−
−
−
−

−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
×
−
−
−
−
−
−
−
−
−
−
−
−
−
−

−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
×
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−

−
−
−
−
−
−
−
−
−
−
−
−
−
−
×
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−

−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
×
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−

−
−
−
−
−
−
−
−
−
−
−
−
−
−
×
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−

−
−
−
−
−
×
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−

−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
×
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−

−
−
−
×
×
−
×
ο
×
×
−
×
−
−
−
−
×
−
−
−
×
×
ο
×
−
×
−
×
−
−
−
×
−
×
×
−
−
−
×
−
−
−
−
×

×
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
×
−
−
−
−
−
−
−
−
−
−
×
−
−

×
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
×
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−

−
−
−
−
−
−
−
−
−
−
−
−
−
−
ο
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−

−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
×
−
−

−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
×
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−

×
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−

−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
×
−
−
−
−

−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
×
−
−

−
−
−
−
−
−
−
−
−
−
−
−
×
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−

−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
×
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−

−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
×
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−

×
×
−
ο
×
−
−
×
−
×
×
×
−
×
×
×
−
−
×
×
×
×
ο
×
×
×
−
×
−
−
−
×
−
×
×
×
−
ο
×
−
−
×
−
−

−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
ο
−
−
−
−
−
−

−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
×
−
−
−
−
−
−
−

−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
×
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−

−
ο
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−

−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
×
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−

−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
×
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−

−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
×
−
−
−
−
−
−
−
−

×
−
−
−
×
−
−
×
−
×
−
×
−
×
ο
−
−
−
−
×
−
×
×
×
×
×
−
×
×
−
−
×
×
×
×
−
−
−
−
−
×
×
−
×

−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
×
×
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−

−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
×
−
−

−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
×
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−

−
−
−
−
×
−
−
−
−
×
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
×
×
−
−
−
−
−
−
−
−
−
−
−
−

−
−
−
−
−
−
−
−
−
−
−
−
−
−
×
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−

−
−
ο
×
−
−
×
−
×
−
−
×
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
×
−
−
−
×
−

−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
×
−
−
−
−
−
−
−
−
−
−
−

−
−
ο
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−

−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
×
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−

−
−
−
−
−
−
−
−
−
−
×
−
−
−
−
−
−
×
−
−
−
−
×
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−

−
−
−
−
−
−
−
−
−
−
−
−
−
−
×
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−

−

, data not known;

, offline analysis;

, online analysis.

In spite of the challenging nature of NIR analysis

of food, applications of the technique are most com-
mon and diverse in this industry. Table 1 lists NIR
food applications both off- and online. Table 2 lists
those applications where NIR has been used to anal-
yze for broad, relatively indeﬁnable quality charac-
teristics in food.

Whole Fruit and Vegetable Analysis

Following on from the success with whole cereal
grain analysis using NIR transmission technology,
many workers have been using a variety of instru-
ment types to obtain spectra from whole fruit and
vegetables, nondestructively. Some success has been
gained, for example, in measuring dry matter and
sugar concentration in this way.

Chemicals and Textiles

Quantitative analysis by NIR of petrochemicals dates
back to the 1930s. Since then a range of NIR cali-
brations have been developed, for example, for octane
number, methyl group analysis, and methanol content
in petroleum. NIR has been used to monitor water,
detergent solids, and glycerol in shampoo, and to
analyze moisture and lubricant levels on polymer
ﬁlms. Process NIR spectrometers have been used to
monitor naphtha composition and NIR instrumenta-
tion has been used to monitor ethylene polymerization.

For over a decade NIR quality control has been

used in the textile and ﬁber industries. Perhaps the
widest use of NIR has been in the cotton industry.
Cotton blending, mercerization, and ﬁber maturity
measurement have been used offline for rapid proc-
ess control. The wool industry has also employed
NIR to measure the residual grease after scouring.
Calibrations have been developed for measuring
moisture and heat set temperature in nylon yarn.
Online NIR analysis is being developed for the quality
control of the dyeing procedure for carpet yam and
for measuring yarn diameter.

Pharmaceuticals and Medical Applications

Discriminant analysis of NIR spectra was used in
1986 to assay the level of lincomycin in a pharma-
ceutical formulation. This was the ﬁrst NIR analysis
to be accepted by this US Food and Drug Adminis-
tration (FDA). NIR is now used to measure the sali-
cylic acid content of aspirin and pharmaceutical
companies use discriminant NIR procedures to check
incoming raw materials for drug production. Sample
identiﬁcation can be achieved using at-line NIR ﬁber
optic systems.

Antibiotic production in fermentations can now be

monitored by NIR. A highly signiﬁcant development
in the medical ﬁeld is the use of a handheld non-
invasive NIR-based ﬁlter instrument for the moni-
toring of blood glucose levels for diabetics. Research
is currently being undertaken to explore the feasibil-
ity of measuring human body temperature using
NIR. Another potential medical application being
examined is the measurement of oxygen levels in
brain blood using a ﬁber optic probe placed on the
skull. In the medical laboratory context, user-friendly
fecal fat analysis is performed using NIR.

Other Industrial Applications of NIR

An NIR method has been developed for measuring the
hardwood content of bleached hardwood and the
lignin content of unbleached hardwood pulp. Work
has also been carried out using NIR reﬂectance spectra
of hardwoods to discriminate between the different
species. Using a Fourier transform NIR instrument
carbonate measurements have been made on soil sam-
ples. NIR analysis of forest humus samples has
provided satisfactory calibrations for microbial basal
respiration, based on the organic polymer content of
the humans. NIR analysis of the lake sediments for pH
has been used to construct the lake water history.

The examples noted above are only a small frac-

tion of the industrial applications of NIR, and the
advent of solid-state instrumentation, will further
widen the way in which NIR measurements can be
made in an industrial context.

See also: Chemometrics and Statistics: Statistical
Techniques; Multivariate Classiﬁcation Techniques; Multi-
variate Calibration Techniques. Food and Nutritional
Analysis: Overview. Fourier Transform Techniques.
Fuels: Oil-Based. Infrared Spectroscopy: Overview.
Pharmaceutical Analysis: Drug Purity Determination.
Process Analysis: Overview. Proteins: Foods. Quality
Assurance: Quality Control. Textiles: Natural; Synthetic.