Sample and feature augmentation strategies for calibration updating

Calibration updating—transfer and/or maintenance—has historically been implemented using a simple but effective technique: in addition to primary samples, include a small number of secondary samples and weight them. It would be beneficial if these classical weighting techniques could be enhanced. Moreover, it would be ideal if we could only use secondary spectra without reference measurements. In this paper, we examine multiple calibration updating scenarios involving unlabeled and labeled secondary spectra. First, we propose three new updating approaches involving sample augmentation whereby unlabeled secondary spectra are used to construct an “undesirable” subspace. This subspace is then used to steer the model vector away from a spectroscopically undesirable solution. Second, we propose two new feature augmentation approaches using labeled secondary samples. These three approaches involves the sum of two model vectors, a dedicated primary model vector plus a perturbation vector, that can accommodate new secondary samples. We rigorously vet the proposed approaches across two near‐infrared (NIR) data sets and across multiple data splits. Out of all of the approaches examined, one feature augmentation approach provides improved results compared with existing approaches, and one sample augmentation approach utilizing only unlabeled secondary spectra appears promising. This paper extends the traditional approach to calibration updating when a standardization set or transfer samples are not available. One extension sample augments (vertically concatenates) unlabeled secondary samples while another approach feature augments (horizontally concatenates) labeled secondary samples. We compare these approaches to baseline calibration updating methods.