analysis. Basically, it can be said that if
there is a determination presently being
used in the analysis of animal feedstuffs,
then someone has probably investigated the
use of NIR to carry out the determination.
Sample populations and effect on calibrations
The tall fescue results in Table 9.3 (Blosser
et al., 1988) demonstrate the result of some
of these effects. Samples were collected
over three consecutive growing years from
the same fields, ground with the same
grinder, scanned in the same cells, on the
same instrument under similar conditions,
with chemistries performed by the same
technician using the same equipment and
chemicals. Thus, many of the variations
which affect a calibration developed or
used over a period of years did not exist.
Yet as can easily be seen, the equations
developed using samples collected in one
year, often would not work on samples
collected in a different year. One possible
explanation is that if samples from one
year all contain the same percentage of
some constituent, such as hemicellulose,
then the calibration equation does not need
to include wavelengths to account for
variations in hemicellulose concentration.
The result is that if that equation is applied
to samples from a year where the hemi-
cellulose content varies, the values for fibre
will probably be incorrect. It is similar
events which make samples from one year
different from those of another, thus
causing many of the calibration problems.
Calibration improvement and redevelopmentWhen an existing calibration fails to
perform satisfactorily on new samples, one
can often correct the problem by including
some of the new samples in the calibration
set, followed by redevelopment of the
calibration. Typically, only a few of the
new samples are needed. An example
using two closed population sets may be
seen in Table 9.4 (Reeves et al., 1991).
These samples were residues of hays from
two nylon bag in situ digestion studies.
While the original hays were the same for
both studies, the animals used and their
diets were different. As can be seen, the
determination of constituents in samples
in study 2 from equations developed using
samples from study 1 often resulted in
reasonable correlations, but extremely
large biases. To correct this, six samples
for each feed (9%) from study 2 were
added to the study 1 samples and new
equations developed, with the results
shown.
It is also important to avoid
continually adding new and unique
samples to a calibration set. This produces
not only a very large and therefore
difficult set to work with, but also a set
which may determine all samples equally,
but few samples particularly well.
Therefore, the development and main-
tenance of calibrations is a continuous job,
the reward of which is the reduced time,
waste and costs of performing wet
chemical analyses.Use of Near Infrared Reflectance Spectroscopy 201Table 9.3.Analysis of ADF in tall fescue by NIRS using equations prepared from samples of various growing
years (Blosser et al., 1988).
VAL year Mean SD CAL year NVAL PMean PSD SEA Bias r^2
1976 34.6 2.07 1976 21 34.2 2.07 0.72 0.40 0.88
1976 34.6 2.35 1977 64 35.5 2.66 1.66 0.89 0.61
1976 34.6 2.35 1978 64 33.4 1.94 1.90 1.25 0.34
1978 37.0 3.30 1978 20 37.1 3.63 0.63 0.14 0.96
1978 36.9 3.63 1976 61 33.9 2.81 1.63 3.00 0.82
1978 36.9 3.63 1977 61 36.1 3.91 1.08 0.81 0.92VAL = validation set, CAL = calibration set, N= no. of samples, P = predicted, SD = standard deviation, SEA
= standard error of analysis, Bias = mean Pmean.