This paper proposes a new technique for calibration transfer, which combines the Successive Projections Algorithm (SPA) for robust variable selection with the subsampling and model aggregation technique known as subagging. The proposed technique is aimed at building Multiple Linear Regression (MLR) models that are robust with respect to differences in the instrumental response of two spectrometers (primary and secondary). For this purpose, a small set of transfer samples with spectra acquired at the secondary instrument is employed to guide the variable selection procedure. The efficiency of the proposed technique is demonstrated in a case study concerning the FT-IR determination of specific mass and two distillation temperatures (T10%, T90%) for gasoline samples and the NIR determination of moisture in corn samples. In terms of the root-mean-square error of prediction at the secondary spectrometer, the MLR models obtained according to the SPA-subagging approach provided better results in comparison with Partial Least Squares employing Piecewise Direct Standardization. In particular, the use of subagging resulted in a more systematic reduction in the prediction error with the progressive inclusion of transfer samples.
multivariate calibration; calibration transfer; variable selection; successive projections algorithm; subagging
