A data-driven near infrared calibration process including near infrared spectral thumbprints

Abstract

Near infrared spectra are highly correlated, complex and noisy, and potentially have many more predictor variables than are required to estimate a parsimonious calibration equation. It is difficult to appreciate the implication of pre-processing choices that are made during calibration, especially in connection with the relationship between the transformed data and the reference values. Graphical methods can be used to understand these relationships better and decisions made during the calibration process can be based on the data alone. In this paper, new graphical tools are introduced to help the researcher better understand these complex relationships in the data. When combined with the proposed algorithm to explore spectra in relation to calibration, these tools enable a parsimonious calibration model to be formed. The results from two different (diesel and wheat) near infrared spectra show that it is possible to form successful calibration equations based on the proposed algorithm, which includes the two new graphical tools. There is a high level of correlation between the results of the different transformations considered, suggesting that in terms of parsimony, developing a calibration using the raw spectra could provide the most judicious outcome.