Development and comparison of regression models for the determination of quality parameters in margarine spread samples using NIR spectroscopy

Abstract

Fat and dry material contents (connected to moisture) are one of the most important parameters in the quality control of butter, margarine and margarine spreads (dairy spreads). More than a hundred margarine samples were used to model their fat and dry material content based on Fourier transform-near infrared (FT-NIR) spectroscopy in transmission and reflectance modes for the quality control of margarine. We also carried out a systematic comparison of various modeling techniques such as PLS regression, principal component regression (PCR) and support vector machines (SVM). Moreover, three types of cross-validation, three types of variable selection and the effect of different spectral types (transmission and reflectance) were also compared with factorial ANOVA tests. We examined the effect of the applied datasets (calibration, test samples, and both sets) based on the original predicted values. Sum of ranking differences (SRD), a novel comparison tool, was applied for the task. We showed that the SRD values can be used as a promising and useful performance parameter for the ranking and evaluation of numerous regression models. Four datasets with 42–42 transmission and 34–34 reflectance models were used for the evaluations. Finally, we have found the best models in each case based on their SRD values. The properly validated SVM models proved to be the best for all of the four used datasets. Although the method comparison is data set dependent, the suggested methodology is applicable generally and unambiguously. These final models can be used for fast and easy quality control of margarine samples instead of the time-consuming original analytical techniques.