Variable selection by double competitive adaptive reweighted sampling for calibration transfer of near infrared spectra

Abstract A new variable selection method called double competitive adaptive reweighted sampling (Double CARS) was proposed for calibration transfer of near infrared spectra. In Double CARS, many variable subsets are generated by the calibration model of primary spectra. Then in each variable subset of primary spectra, many variable subsets of secondary spectra are also generated by the transfer set of secondary spectra. After that, the root mean square error of validation (RMSEV) of each paired variable subsets of primary and secondary spectra is computed and the paired ones with minimal RMSEV are selected as the best variable subsets. Three batches of spectra including corn, tablet and wheat datasets were tested by Double CARS. The results showed that in contrast to CARS solely selecting variables by primary spectra, the Double CARS can use the information of both primary and secondary spectra to select variables with lower errors. Moreover, the Double CARS can also execute variable selection for the paired spectra with different numbers of variables.