Farm Animal Metabolism and Nutrition

regression analysis and calibration valida-
tion and testing) are available for analysing
NIR data.

Data exploration
Once a set of NIR spectra is obtained, one
needs to examine the spectra and see if
there are any apparent problems, as
demonstrated in Fig. 9.6. As can be seen,
one of the five spectra differs in several
ways from the others: first, the peaks are
different, especially in the 400–600 nm
region, and second the baseline is shifted
upwards. Such baseline shifts are often due
to particle size differences between the
samples but, in this case the samples were
scanned on different spectrometers. Since
NIR spectra of a given material tend to be
very similar, such spectral differences
indicate potential problems. The last thing
about all the spectra is the discontinuity

visible at 1100 nm, due to the use of two

detectors: one for 400–1098 nm, and the

second for 1100–2498 nm. By examining

spectra before performing a calibration, bad

spectra can be found and corrective actions

taken, i.e. rerun the sample, do not use the

region around 1100 nm, etc. Although it is

highly likely that such problems will be

discovered during the calibration process

(i.e. a single bad sample will be tagged as

either a spectral or compositional outlier),

they would influence the results and

require redevelopment of the calibration

after removal or replacement of the bad

spectra.

Data exploration can also include

examination of plots showing the correla-

tions at each wavelength between the

spectra and the analyte of interest. Such

‘Correlation Plots’ can be used to select

regions of the spectra to use or avoid when

194 J.B. Reeves III

120

100

80

60

40

20

0

Protein (g kg

^1

)

1 2 3 4 5 6 7 8 9 10

Sample number

Set 1

Set 2

True values

Fig. 9.5.Demonstration of effect of non-random partitioning of NIR errors between two truly identical
sample sets.