PCA was executed on both the MSC- and SNV-preprocessed
FT-IR spectral data to extract relevant information. However
the data might be classified according to moisture content or
particle size, rather than their chemical composition characteristics,
because the IR data patterns are determined by not only
absorption characteristics but also scattering characteristics.
These spectral variations were removed by oven drying all of
the samples and pretreating the spectra with MSC- and SNVpreprocessing
methods. Neither the MSC- nor the SNVpreprocessed
FT-IR data were able to quantify the level of
adulteration by means of data clustering using PCA. The data
were scattered over a vast range, and no distinct separations
among the data groups were found from the PCA score plot
(data not shown). PCA is a popular primary technique for data
clustering. It is not, however, optimized for class separability.
The loading plots of the first three PCs are depicted in Figure 6.
The PC1 loading accounts for 96.1% of the total sample
variance, producing a pattern similar to the raw FT-IR spectra
(Figure 2a). The loadings of PC1 and PC2 show the highest
intensity at ∼980 cm−1
. Sankaran et al. also found a similar peak
for starch in the spectral range of 950−1000 cm−1
.
27 In
addition, Danwille observed a sharp peak at 975 cm−1 arising
from an −OCH3 bond vibration related to cellulose.15