Development of a non-targeted approach using three handheld spectrometers combined with ensemble classifiers for authentication of bovine milk

Abstract

In this study, three Vis/NIR handheld spectrometers coupled with four machine learning methods were developed for the analysis of the authentic bovine milk samples and adulterated samples with hydrogen peroxide and sodium hypochlorite. On this matter, ensemble bagged tree (EBT), random under sampling-boosted ensemble (RUS-BE), random subspace discriminant ensemble (RSDE), and random subspace ensemble k-nearest neighbor (RSE-kNN) were proposed and compared. The experiments were performed in two different scenarios. In the first scenario, pure milk and adulterated samples with one adulterant were analyzed separately with three handheld spectrometers. By comparing the results of ensemble classifiers for three spectrometers, RSE-KNN represented the best results for the accuracy and Youden index (96%, and 93% respectively). Interestingly, the results of RSE-kNN were much better than kNN as base learner. In the second scenario, the mixtures of two adulterants were analyzed with three handheld spectrometers. Among them, one presented the best results, with the predictive accuracy and Youden index of 100% for EBT, RSDE and RSE-kNN.