On the selection of samples for multivariate regression analysis: application to near-infrared (NIR) calibration models for the prediction of pulp yield in Eucalyptus nitens

The effects of using reduced calibration sets on the development of near-infrared (NIR) calibration models for the prediction of kraft pulp yield in Eucalyptus nitens (Dean & Maiden) Maiden trees were explored. Three selection techniques based on NIR spectral data (CADEX (computer-aided design of experiments), DUPLEX, and SELECT algorithms) and one selection method based on a measured property (RANKING algorithm) were used for analysis and compared against a model using all data. The effect of using calibration sets of different sizes was also evaluated. All sample-selection methods resulted in models of similar performance compared with the model fitted using all samples. For calibration purposes, RANKING selection resulted in models with the lowest errors of cross-validation, followed by the DUPLEX, CADEX, and SELECT methods. In terms of validation, the SELECT and CADEX methods resulted in lower errors of prediction compared with the DUPLEX and RANKING algorithms. In general, cross-validation and prediction errors decreased as the number of calibration samples increased. These results show that it is possible to obtain adequate NIR calibration models with a reduced number of samples allowing the remaining samples to be used for model validation and that sample selection based on NIR spectral data alone is as successful as selection based on a measured property.