contained several forages, treated with
different levels of manure and collected at
four or five cuttings over a 2-year period.
Using only 120 calibrations (based on
spectral diversity) and 24 test samples,
excellent calibrations for NDF and CP were
obtained, resulting in a reduction of almost
95% in the number of chemical determina-
tions.
Two other considerations should be
taken into account when considering NIR
for assaying research samples. The first
relates to the calibration nature of an NIR
spectrometer. An NIR calibration is based
on the reference method used to provide
the assay values needed for building the
calibration. As such, NIR determinations
are often stated to be, at best, no better than
the calibration standards (assays) used to
develop the calibration. While there is
debate about this, many feel this to be
generally true. Thus, if one has the choice
between NIR determinations and accurate
chemically determined values, the latter
may be preferable. On the other hand, it is
highly likely that if several operators are
involved the NIR results may well be more
consistent than will be the reference
method. This is because there are fewer
steps involved in using an existing calibra-
tion than in performing many chemical
procedures, and different individuals
always do things slightly differently, intro-
ducing errors and biases. The second
consideration deals with error distributions
in NIR calibrations. It has been reported
(Buxton and Mertens, 1991) that the errors
in NIR calibrations are not always
randomly distributed across samples. Such
an effect is demonstrated in Fig. 9.5 (hypo-
thetical). If the samples in sets 1 and 2
were matched sets whose true values were
the same, the conclusion that there is a
difference between the sets (i.e. varieties
or treatments) would be incorrect.
Examination of the statistics alone (R^2 of
0.971, mean value of 5.5 for all samples)
would give no indication of a problem.
Even by examination of the plotted data
(Fig. 9.5) a problem might not be indicated,
unless the samples were labelled in the
plot as to source (A or B). Then, as shown,
one would notice all the As were predicted
high and all the Bs low. In the real world,
one needs to check some test samples to
see that this is not occurring.
The point of this is not to show that
NIR is useless and should not be used, only
that like all procedures, one should not
simply assume that NIR solves all one’s
problems at the press of a button. Finally, it
should be noted that work to improve NIR
calibrations has done a lot to show how
poor the wet chemical procedures, against
which NIR is rated, can be. Many NIR
calibrations have been improved simply by
cleaning up or changing the procedures
used to obtain the calibration values.Chemometrics and NIRAs previously discussed, the area of
statistics used to develop NIR calibrations
has come to be known as chemometrics and
is a vital part of using NIR spectroscopy.
While a subject unto itself, at least a brief
discussion is needed to understand what
NIR is all about and what is involved in its
use. For a more in-depth discussion, the
reader is advised to check one or more of the
following references or software packages:
Sharaf et al. (1986); Massart et al. (1988);
PLSplus (1992); Shenk and Westerhaus
(1993); or Unscrambler (1994). There is also
a web site at http://newton.foodsci.kvl.
dk/chdb_asc.htmlwhich contains a search-
able database containing >1200 references
on chemometrics.
The basic objective of calibration
development is to find a relationship
between spectral information and the
analyte of interest. In the beginning, much
of the NIR data were obtained from filter
instruments, and thus consisted of a dozen
or few dozen wavelengths at most. In addi-
tion, the computers available were limited
in speed and memory which limited the
algorithms which could be used practically.
As a result, most of the early work used
stepwise multiple linear regression (MLR)
techniques. Today, a wide range of tools
and statistical procedures, which serve
three basic functions (data exploration,Use of Near Infrared Reflectance Spectroscopy 193