and probably the best requires scanning all
the samples first, and then choosing
samples on the basis of their spectral
differences.
Since the spectra are supposed to be
determined by the composition, choosing
the 100 most spectrally different samples
can be a good place to start. This of course
supposes that 100 samples can represent
the remaining 1900 samples. Depending on
the diversity of the samples, this may or
may not be true. For example, if one had a
sample set of 2000 lucerne hays from one
growing season and one farm, it would be
much more likely that 100 samples could
represent the entire group than if one had
2000 samples consisting of 100 different
feedstuffs each collected over 5 years from
four different locations. The diversity in
such a sample set might require many more
(or even all) samples to be tested. The latter
was found to be true for a set of 325
chemically/physically treated samples, in
which no two samples were the same feed-
stuff treated in the same way. The samples
produced were so diverse that even using
all 325 samples did not produce a satis-
factory NIR calibration (Reeves, 1994b).
Unfortunately, only by testing the calibra-
tion can one determine if enough samples
were used. If the results are unsatisfactory,
more samples can be added to the calibra-
tion and the process repeated.
Unfortunately, it is also true that only
experience and knowledge of one’s
samples can be used for guidance.
Open populations
While in a closed population, one has all
the samples there will ever be at the start of
the process, in an open population, this is
never true. For example, one of the
successes of NIR has been its use to deter-
mine the nutritive value of forages. By
using a van carrying a NIR spectrometer,
and the means to collect and grind hay
samples on the spot, NIRS has even been
used at hay auctions. In such cases, it is
obvious that one can never have all the
possible samples. This is the case for many
of the most useful applications of NIR.
The problems arising from open
populations give rise to the biggest mis-
understanding of NIR and subsequent dis-
illusionment and that is the belief that if
one buys an NIR spectrometer or finds
someone with one, then one never needs to
perform a wet chemical analysis again. On
the contrary, as new samples arise, one
must periodically run a few samples by
both the standard and the NIR methods
and compare the results to determine if the
calibration is performing properly. If not,
then one can add some of the new samples
to the calibration, and perhaps remove
some of the old ones (i.e. a variety or feed
no longer grown), and redevelop the
calibration. This is what the private feed
testing companies utilizing NIR do.
However, even if one out of every ten
samples were to be checked, this would
still reduce the need for the standard
analysis by 90%.
One point about results performed by
commercial feed testing laboratories using
NIRS should be understood: such results
are likely to be accurate only if the firm has
developed equations which cover your
geographic area, your type of samples and
the season of interest.Research samples
The questions revolving around open and
closed populations probably have a great
impact on the usefulness of NIRS for
research samples. For example, for a
farmer, the ability to get a rapid analysis of
his feedstuff, a matter of minutes at a hay
auction, may outweigh the need for
extreme accuracy, and thus the use of a
general hay calibration may be more than
acceptable. In the research environment,
the problem is that accuracy is generally
paramount, and the samples may represent
rather unique population sets containing
rather small numbers of sample. For such
sample sets, it is unlikely that NIRS will be
of great benefit.
On the other hand, if either a large
sample set is available or will be accumu-
lated over time as a result of repetitive
efforts, developing an NIR calibration may
well be worth the time. As an example,
we recently had 1800 samples which192 J.B. Reeves III