There are two issues that can clarify the results obtained in this
feasibility study. First of all, despite the fact that SECV values are
not high, compared with other studies cited in literature, most of
the prediction models obtained in this work for minor constituents are not appropriate for quantitative analytical purposes, as their
coefficients of determination are small. NIR prediction models with
r2 values between 0.30 and 0.69 could be used to separate samples
with higher and lower analytical values; whilst r2 values higher
than 0.70 indicate a good predictive ability (Shenk & Westerhaus,
1996). Low coefficients of determination could be caused by the
similarity between standard deviations (SD) and SEL values for
some constituents that can be observed in Table 1, as the maximum
value of r2 can be estimated according to the following
expression (Shenk & Westerhaus, 1996):