Farm Animal Metabolism and Nutrition

reduced by deliberately including the varia-
tion in the calibration; the difficult problem
lies in the nature of the samples themselves
and the variations that occur over a growing
season, between seasons, between growing
locations, between varieties, etc.

Closed and open populations

Closed populations and
calibration development
A closed population is one in which one
has all the samples of interest, e.g. 2000
lucerne samples from a pasture feeding
experiment. After grinding the samples, the
samples should be scanned in a random
order to avoid any pattern between the
samples and the order of scanning, i.e.
avoid scanning all the highest quality
samples on one day, at one temperature, and
the lowest quality samples on another day,
at a different temperature. The more any
differences due to personnel, room condi-
tions, instrument conditions, etc. are
randomly distributed across all the samples,
the better. The next step is to select the
samples for which the analyte of interest

will be measured by the standard technique,

i.e. NDF, ADF by extraction procedures (Van

Soest, 1994). It is these values which will be

used to determine the wavelengths needed

for the NIR-based determination of the

remaining samples. A smaller set of test

samples is also needed in order to test the

calibration after development. These are

only to determine if the calibration/equation

developed works properly on samples not

in the original calibration set.

The question then is how to find

samples in the set which represent all 2000

samples. Several methods have been used

over the years. First, N samples can be

randomly selected. Second, every nth

sample can be used. If this is done, it is

important to avoid a structured data set:

i.e. a set of 2000 samples consisting of 100

sets of 20 samples, each set in order of

increasing NDF content, and every 20th

sample taken for the calibration set. This

would result in only the samples with the

lowest NDF content being included in the

calibration. Under such circumstances, it is

unlikely that the calibration developed

would work on the test set or very many of

the remaining samples. The third method

Use of Near Infrared Reflectance Spectroscopy 191

Table 9.1.Some factors (as differences between samples) which can alter spectra and influence calibration
accuracy.

Factor Spectral effect Solution

Particle size Baseline shifts, multiplicative Mathematical pre-treatments of spectra
scatter

Temperature Spectral peak shifts Include temperature differences in
calibration

Sample packing Spectra do not represent sample Pack uniformly, scan a larger sample

Sample homogeneity Non-representative sample Grind more finely or scan a larger sample

Constituent out of range Inappropriate for calibration Add a sample with wider range of values

Different feed or samples Inappropriate for calibration Add samples to old or develop new
calibration

Spectrometer repaired or Peak shifts, absorbance changes, Redevelop calibration or use calibration
changes somehow multiplicative effects, etc. transfer

Different spectrometer used Peak shifts, absorbance changes, Redevelop calibration or use calibration
multiplicative effects, etc. transfer

Stray light in instrument Inaccurate determination Fix instrument or use constant lighting at
all times