Abstract
Large calibration matrices are usually needed to cover all the variability of forage samples. The present work applies experimental design techniques to reduce the number of samples necessary to calibrate quality properties of pellets of forage. Near infrared spectra were recorded from the raw material which consisted of dehydrated alfalfa milled, pressed and packed in cylindrical pellets, 5 mm in diameter. Partial least squares calibration models for moisture and crude protein, built up from 771 samples, were used as reference. Full factorial, central composite and Box–Behnken designs, using the principal components of the raw spectra as design factors were tested for sample reduction. The effect of increasing the number of replicates for each design point was also studied. Comparisons between models have been made in terms of prediction accuracy, but also looking at the model stability by means of Martens' uncertainty test. Results obtained indicate that the prediction accuracy is similar in all assayed designs and similar to that of the reference model, although the stability of the models obtained using a reduced sample set is lower. Increasing the number of replicates of the design points increases stability.
© 2004 NIR Publications
PDF Article
More Like This
Cited By
You do not have subscription access to this journal. Cited by links are available to subscribers only. You may subscribe either as an Optica member, or as an authorized user of your institution.
Contact your librarian or system administrator
or
Login to access Optica Member Subscription