LAUSR

Abstract This paper proposes a method based on Near Infrared Hyperspectral Imaging for discriminating between pork, poultry and fish species in processed animal protein meals. First, an investigation was conducted into the possible importance of incorporating into the discrimination models anomalous (or singular) pixels as probable discriminant pixels for each species. Subsequently, partial least squares discriminant analysis (PLS-DA) spectral and textural models were constructed. The former reflected the spectral information (spectral trace), and the latter the spatial (textural trace) information based on different groups of features. Finally, the spectral and textural information was integrated using classification trees, to ascertain whether the combined use of such information represented an improvement in accuracy in the effort to discriminate between species. The method was applied to a set of 40 pork, 40 poultry and 40 fish meals analysed in the 1000–1700 nm range. Models were then tested using an external validation set comprising 45 samples (15 pork, 15 poultry and 15 fish meals). The results demonstrated that combining spectral and appearance characteristics in a single classification tree generated better classification results for the samples used in the study (92% correct) than when using the PLSDA spectral model (83% correct).