Evaluation and comparison of NIR multi-product calibration methods for Brix prediction

Abstract

Abstract Near-infrared (NIR) spectroscopy is a widespread technology for fruit and vegetable quality assessment. New fields of application of this technology, like mobile food analysis with handheld low-cost spectrometers, increase the demand for chemometric calibration models that are able to deal with multiple products and varieties thereof at once (so-called multi-product calibration models). While there are well studied methods for single-product calibration as partial least squares regression (PLSR), multi-product calibration is still challenging. Conventional approaches that work well for single-product calibration can lead to high errors for multi-product calibration. However, nonlinear methods as local regression and artificial neural networks were found to be suitable E. Micklander, K. Kjeldahl, M. Egebo, and L. Norgaard. Multi-product calibration models of near-infrared spectra of foods. Journal of Near Infrared Spectroscopy, 14:395–402, 2006. L. R. Lopez, T. Behrens, K. Schmidt, A. Stevens, J. A. M. Dematte, and T. Scholten. The spectrum-based learner: A new local approach for modeling soil vis-NIR spectra of complex datasets. Geoderma, 195–196:268–279, 2013. . Preliminary studies in multi-product calibration for quantitative analysis of food with near-infrared spectroscopy showed good results for memory-based learning (MBL) and a classification prediction hierarchy (CPH) M. C. Kopf and R. Gruna. Examination of multiproduct calibration approaches for quantitative analysis of food with near infrared spectroscopy. Bachelor's thesis, Karlsruhe Institute of Technology KIT, 2016. . In this study, three varieties of apples, pears and tomatoes with known sugar content (in ○Brix) are analysed with NIR hyperspectral imaging spectroscopy in the range from 900 nm to 2400 nm. Predictive performance of a linear PLSR model, two nonlinear models (CPH and MBL) and different pre-processing techniques are tested and evaluated. For error estimation, leave-one-product-out and leave-one-out cross-validation are used.