Multivariate calibration and chemometrics for near infrared spectroscopy: which method?

The four most important regression methods are evaluated on very large data sets: Multiple Linear Regression (MLR), Partial Least Squares (PLS), Artif icial Neural Network (ANN) and a new concept called "LOCAL" (PLS with selec tion of a calibration sample subset of the closest neighbours for each samp le to predict). The Standard Errors of Prediction (SEPs) are statistically tested and the results show that the regression methods are almost equal an d that the data matrices are more important than the fitting methods themse lves. The types of pre-treatments (Multiplicative Scatter Correction, Detre nd, Standard Normal Variate, derivative etc.) of the spectra are too numero us to be able to test all the combinations. For each test, the pre-treatmen t found as the best with the PLS method is fixed for the other ones. The se cond part of the paper emphasises the importance of the number of samples. If any agricultural commodity, and probably any kind of product measured by an NIR instrument, can be considered as a mixture of several constituents, the databases built by collecting actual samples bringing new information can reach hundreds, if not thousands, of samples.