Multivariate calibration of non-destructive spectral sensors with a particular focus on food applications: Validation issues and guidelines

Abstract

Multivariate calibration methods have enabled the use of non-destructive spectral sensors in a wide range of applications but carry a risk of overfitting to the available training samples. For this reason, the prediction of unseen samples plays a vital role both in tuning the prediction algorithm and in assessing its performance, two activities that need to be carefully distinguished. Methods employed include data-splitting, cross-validation, and the use of genuinely independent sets of data. These approaches are described and some common issues with them are identified. The focus is on food applications but the methods discussed are widely used in other areas.