Farm Animal Metabolism and Nutrition

those things which commonly cause prob-
lems is shown in Table 9.1.
While any of these factors can vary
from one sample to the next, the problem
arises when they vary between the samples
used to develop a calibration, and new
samples on which the calibration is used.
Ultimately, the objective in developing a
calibration is to find a set of samples whose
spectra/composition are representative of a
larger set of samples, i.e. one does not want
any systematic differences. Thus, one does
not want to develop a calibration using
samples ground to one mesh size or
scanned at one temperature, and then use
that calibration to determine the values for
samples ground to a different mesh (or even
ground on a different type of grinder) or
scanned at a different temperature. While it
may not be possible to take into account the
repair of an instrument, one can avoid the
problems caused by the other factors by
scanning the calibration samples over a
range of temperatures or scanning each
sample at several different temperatures,
and so forth. Thus, to a large degree,
variations likely to influence spectra can be

built into the calibration set. The calibra-

tion process will then avoid those wave-

lengths where, for example, variations due

to temperature have too great an influence.

Also, spectral pre-treatments, such as

multiplicative scatter correction (Isaksson

and Kolwalski, 1993), can be used to

eliminate some effects, such as those due to

particle size differences. The real problems

occur with differences in spectrometers and

samples. Spectrometer to spectrometer

variation is difficult and expensive to build

into a calibration because it requires

running the same samples on many

different instruments and using all the

spectra obtained to develop the calibration.

Even then, another spectrometer may vary

too greatly from those used for the calibra-

tion for it to work properly. Therefore,

generally, the problem is handled outside

the calibration development process in

what is called ‘Calibration Transfer’.

Sample differences lie at the heart of

calibration development, and revolve

around ‘Sample Populations’. In developing

calibrations, many effects (temperature

variations, etc.) can be eliminated or

190 J.B. Reeves III

Absorbance

Data points

6 5 4 3 2 1

E

F

D

C

A B

Fig. 9.4.Hypothetical effects of peak shift (B), baseline shifts (C and D) and a second component (E and F)
on an NIR spectrum (A).