Identifying animal species in NIR hyperspectral images of processed animal proteins (PAPs): Comparison of multivariate techniques

摘要

The use of PAPs in animal feed has several advantages over other feed ingredients, but requires rigorous and accurate control mechanisms that ensure the absence of ruminant meal. In order to differentiate between animal species while simultaneously offering the capacity to inspect PAPs in large volumes, a hyperspectral imaging (HSI) system operating in the NIR spectral range is proposed. This study investigates the sensitivity, specificity and other parameters with which HSI can discriminate between different animal species (ruminants, swine and poultry), making use of various classification methods. Diffuse reflectance spectra were acquired from 125 rendered meal samples in the 1000-1700 nm wavelength range; measured PAPs included particles of scale, hair, feather, blood, grease, skin, muscle and bone from both ruminant and non-ruminant animals, obtained in a rendering plant. Various classification methods were then applied to the dataset to determine the accuracy with which different animal species could be discriminated from each other. Support Vector Machine classification performed best in discriminating between animal species, with a sensitivity and specificity of around 90% and a Matthew's correlation coefficient of around 0.7 for non-ruminant species and higher than 0.95 for ruminant species. Other methods, such PLS-DA and Subspace Discriminant, also produced acceptable results and required less computational time. This study showed that spectral analysis of PAPs, based on diffuse reflectance spectroscopy, is a promising technique for differentiating between ruminant species and other terrestrial animal species. The technique may therefore offer accurate and fast analysis of large volumes of feed products, a necessary prerequisite for the lifting of the EU ban on non-ruminant processed animal proteins.