developing a calibration, which can some-
times result in improved calibrations. Also,
some chemometric procedures, such as
Neural Networks, still require considerable
computer time even on the fastest personal
computer, and, by reducing the wave-
lengths to be examined, the process can be
sped up considerably (McClure et al.,
1992).
Data exploration also includes deter-
mining which samples are to provide the
analyte calibration and test values. As
previously discussed, although random
selection can be used, it is also possible
and best to select samples on the basis of
spectral similarities and differences. Using
a spectrally diverse set of samples offers
the best chance of developing a calibration
which will be applicable to other samples.
The techniques used can be visual (cluster
analysis based on spectral similarity) or
numerical (regressing one spectrum against
the other). The objective is the same: to find
groups of samples that are similar and pick
one to represent each group. Unfortunately,
there is no magical way to know how many
samples will be needed. Most of the
techniques used normalize the data before
starting in such a way that only relative
differences are determined. Thus, the result
might be the 50 most diverse out of 100 or
1000 or 10,000 samples. Whether 50 will be
enough or too many for any of these three
sets depends on how similar the samples
are, something which is very difficult to
determine absolutely a priori.Regression analysis
Once the samples needed for calibration
development and testing have been deter-
mined, and the analyte values determined
by the reference method, the process
moves to finding the relationship between
the spectral data and analyte. Many proce-
dures have been tried and are being used to
various degrees. While an in-depth discus-
sion would require a book unto itself (see
earlier references), a short discussion is
necessary.
One of the problems with MLR is the
possibility of over-fitting the data. In such a
case, relationships are found between theUse of Near Infrared Reflectance Spectroscopy 195Absorbance500 1000 1500 2000
Wavelength (nm)Fig. 9.6.Demonstration of the value of examining spectra before calibration development.