sample left out once) with the sample left
out used as the validation set. The results
are used to estimate the prediction error
and to provide guidance in selecting the
final number of factors to use. This proce-
dure is felt to provide the best way to find
the optimal equation and avoid over-fitting
to the calibration data, although again an
independent test set should also be used.
An obvious question which arises is,
what happens if everything looks great,
until the equation is used on the test set?
Unfortunately, there is no easy answer to
this question. One of the problems, which
has become more apparent as NIR use has
expanded beyond the laboratory, is that
developing and maintaining NIR calibra-
tions, while economically beneficial, is not
a simple thing to learn. It takes a lot of
experience and knowledge to know both
what to do and what to look for. At a recent
meeting, it was stated that it was felt that a
year working in a chemometrics laboratory
was needed to really know how to develop
calibrations.
Part of the problem arises from the
black box nature of NIR; take the spectra,
apply the statistical procedure and see
what happens. Unlike the case where we
know that proteins absorb in the UV at
280 nm, data on NIR absorption bands are
scattered throughout the literature and are
very fragmented. While there are sections
in various books (e.g. Murray and Williams,
1987), there are no books dedicated to the
subject, as is true in the MIDIR (Colthup et
al., 1990 and many others). Only recently
have efforts turned to the better under-
standing of NIR spectra in an effort to
better understand calibrations.
Black box versus understandable NIR
While black box calibrations may be
adequate when dealing with closed popula-
tions, it can be quite a different matter with
open populations. For example, take the
case of acetone–water mixtures; as the
amount of acetone decreases, there are peak
shifts particularly at 2390 and 2412 nm. A
simple linear regression (percentage
acetone = a + b wavelength position)
yielded an R^2 of 0.994 using the position of
the peak at 2390 nm and 0.984 using that at
2414 nm, based on eyeball peak positions.
Therefore, an accurate calibration could be
developed based on peak position alone.
This might be fine, unless the peak posi-
tions shifted for reasons other than water
content, e.g. temperature changes. A
similar effect has been used to determine
salt concentrations in water. Salt does not
have any absorption bands in the NIR, but
does cause changes in the spectra of water,
which can be used to determine the
amount of salt present (Hirschfeld, 1985).
However, it is important to understand that
other salts (KCl, KNO 3 , etc.) can easily
cause the same effects, and that one is not
determining salt concentration directly.
Therefore, the calibration would not detect
a substitution based on the spectral
information alone. While this may be a
simple case, similar things have occurred
with calibrations, and the more one under-
stands about the process and spectra, the
less likely they are to occur.
For this reason, more and more effort
is going into understanding the spectra and
the basis for calibrations. Such knowledge
can also explain why something does not
work in order to eliminate the waste of
trying something in the first place.Calibration transferThe final issue to be discussed dealing
with calibrations is calibration transfer. As
discussed above, almost anything which
differs between the samples used to
develop a calibration and the samples
upon which it is to be used can cause
errors, and this includes differences in the
instrument at the times the samples were
scanned, and applying a calibration to
samples scanned on different machines.
This represents a big problem when, for
example, one wishes to develop a network
of instruments to determine forage quality.
One certainly does not want to scan all the
samples on each instrument, and go
through the calibration procedure for each,
and then to have to repeat the process
periodically due to repairs or even wearUse of Near Infrared Reflectance Spectroscopy 197