1

Fig. 1. Average uncorrected NIR spectra of filled-viable, empty and petrified seeds of L. sibirica (A), and difference spectrum of non-viable and viable seeds (B).

Table 1. A statistical summary of discriminant models computed to classify viable, empty and petrified seeds of L. sibirica into three and two classes.
Class A R2Xp R2X0 R2Y Q2 cv
Three 2 + 13 0.267 0.733 0.842 0.820
Two 1 + 12 0.041 0.959 0.937 0.931
A = number of significant components to build the model (the first and second values represent the number of components for predictive and orthogonal variations, respectively); R2Xp = the predictive spectral variation; R2Xo = Y-orthogonal variation (spectral variation uncorrelated to class discrimination) for the OPLS-DA models; R2Y = the modelled class variation; and Q2cv = the predictive power of a model according to cross validation.
2

Fig. 2. Score plot for the first (tp[1]) and second (tp[2]) predictive components (A) showing clear clustering patterns of filled-viable (green star), empty (blue box) and petrified (brown triangle) seeds, and loading plots for the first (B) and second (C) predictive components showing absorption bands accounted for class discrimination.

Table 2. Predicted class membership of viable, empty and petrified seeds in the test sets by three-class OPLS-DA modelling of single seed NIR spectra.
Seed lot
fractions
No. of test 1 2 3 No class 1 & 2 1 & 3 2 & 3 1 & 2 & 3 Classification
accuracy (%)
Viable (1) 45 44 0 0 1 0 0 0 0 98
Empty (2) 45 0 37 5 1 0 0 2 0 82
Petrified (3) 45 0 1 39 4 0 0 1 0 87
3

Fig. 3. Score plot for the first predictive (tp[1]) versus orthogonal (to[1]) components showing symmetrical separation of viable (green stars) and non-viable (blue dots) seeds (A); loading plot for the first predictive component (P1[p]) showing absorption bands correlating to seed classes (B); and loading plot for the first orthogonal component (P1[o]) showing absorption bands that are not correlated to class discrimination (C).

4

Fig. 4. Predicted class membership of non-viable (empty and petrified) and viable seeds in the test set by two-class OPLS-DA modelling of NIR spectral region (A); and Plot of Variable Influence on Projection, VIP, (B) showing spectral regions that influenced the discrimination of viable from non-viable seeds. The horizontal dotted line in panel A is the class limit (Ypredicted > 0.5) for assigning the test sets into viable class otherwise non-viable, and that in panel B is the cut-off limit (1.0) for discriminating relevant and irrelevant predictors.