Refining transfer set in calibration transfer of near infrared spectra by backward refinement of samples

Abstract

The transfer set in calibration transfer can be further refined using a backward refinement (BR) method. Our BR includes many loops, and in each the sample can be removed which reduces the root mean square error of validation (RMSEV) most effectively. This procedure can be repeated many times so that many subsets of samples can be obtained. Then the sample subset with a small value of RMSEV can be chosen as the best sample subset. Two batches of data including corn and wheat datasets are used to test this method. The results show that compared with the ordinary Kennard–Stone method, BR can largely reduce errors when it is applied in calibration transfer methods, including: canonical correlation analysis combined with informative components extraction, direct standardization, piecewise direct standardization PDS and spectral space transformation.