| Abstract: | The ability of five sample selection methods for local chemometrics and three variable selection algorithms were compared for the development and the transfer of whole soybeans protein and oil near infrared prediction models. Two new methods based on a similarity index considering Euclidian distance among Fourier coefficients were introduced and tested against more common approaches (locally weighted regression, LOCAL). Genetic algorithms were also challenged with the development of models based on particle swarm optimization (PSO). A modification to the original PSO model was introduced. Sample and variable selection methods, as well as their combinations, were tested in the transfer of models in intra‐ and inter‐brand situations using two Foss Infratecs and two Bruins OmegAnalyzerGs. For each brand, a master was designated and its models transferred onto the second unit of its network and the two units of the second brand. Calibration models were proven transferable from brand to brand with similar or better precisions than when all instruments were calibrated on their own calibration sets (relative predictive determinant (RPD) improving from 10.42 to 12.76 and 12.39 in intra‐brand standardization for Infratec network with local and variable selection methods respectively). These methods provided contrasted results depending on the instrument, the parameter, and the variability of interest. Copyright © 2010 John Wiley & Sons, Ltd. |
