CALIBRATION METHODS FOR NIRS INSTRUMENTS - A THEORETICAL EVALUATION AND COMPARISONS BY DATA SPLITTING AND SIMULATIONS

The properties of the recently proposed calibration method called rest ricted principal component regression (RPCR) were evaluated and compar ed with partial least-squares regression (PLSR) and two types of princ ipal component regression (PCR1, selected according to the size of the eigenvalues, and PCR2, selected according to the t-value). RPCR can b e considered a compromise between PCR and PLSR, since the first compon ent of RPCR is equivalent to the first component of PLSR, while the re st can be regarded as principal components on a space orthogonal to th e first. The methods showed almost the same properties when the irrele vant components had small eigenvalues. The prediction error of RPCR se lected according to the size of the eigenvalues was intermediate to th ose of PCR1 and PLSR when the number of components was low, while RPCR and PCR1 nearly coincided when the number of components exceeded the number of relevant ones. The prediction error minimum was about the sa me for RPCR, PCR1, and PLSR, but the minimum of PLSR was obtained when a lower number of components were included in the calibration model.