Texturized vegetable proteins surface characterization with NIR hyperspectral imaging

Over the last years, an increasing interest in vegetable sources of proteins has been observed from the consumers and industry. This is demonstrated by the increasing investments in the plant-based sector, which reached 3.1 billion dollars in 2020. The driving forces of such ascending trend could be summarized basically in i) the nutritional concerns linked to the awareness of the side effects of high animal products consumption in the industrialized countries; ii) the need for environmentally-friendly and sustainable food productions, and iii) the ethical aspects linked to the animal welfare. In this context, special attention deserves the plant-based meat analogues (PBMA), a category of products that accurately imitate the structure and the sensory properties of meat. PBMA mainly consist of texturized vegetable proteins (TVP) produced by raw proteins historically extracted from wheat gluten, soybean, and pea, although new sources and different extraction/separation processes are available nowadays. Considering the increasing usage of TVP for a wide spectrum of applications in the industrial context, a critical aspect regards developing suitable and convenient methods for quality control. Quality assessment of TVP is of primary importance as they represent the main ingredient in PBMA formulation after water and strongly affect the nutritional and textural properties of the final product. Proximate composition (proteins, carbohydrates, lipids, ashes) is the first mandatory aspect to be monitored. However, macronutrients analysis using standard methods is a demanding task that requires time, polluting solvents, and resources. Generally, around 24 hours are needed to determine the proximate composition by the official methods of analysis. On the opposite, near infrared spectroscopy (NIRS) has been proven to be an efficient technique for rapid and non-destructive food quality control. Besides, when NIRS is coupled with imaging techniques, a complex and comprehensive spatial and spectral information of the product under study could be achieved. Hyperspectral Imaging (HSI) was previously applied to meat products with the aim of quality control, detection of defects, classification of the origin and the identification of chemical constituents. In the light of these applications, it has been recently hypothesized that HSI can be potentially used to evaluate the quality of TVP and PBMA. In this framework, this work is aimed at studying the feasibility of NIR-HSI for the analysis of TVP chemical composition. Four different TVP have been produced in duplicate by a low-moisture extrusion process, combining different protein sources as follows: F1, dry-fractionated pea protein and oat protein (70:30 w/w); F2, pea protein isolates and oat protein (70:30 w/w); F3, dry-fractionated pea protein, pea protein isolate, and oat protein (35:35:30 w/w/w); F4, soy isolates and oat protein (70:30 w/w). TVP were produced by the KETSE 20/40 twin-screw extruder (Brabender GmbH, Duisburg, Germany). TVP were analysed for moisture (% w/w), total protein content (% w/w), total fat (% w/w) and ashes (% w/w). Carbohydrates were determined as 100 - (proteins + fat + ashes + moisture). All the analyses were carried out in duplicate on the two batches of TVP produced. NIR hyperspectral images were collected in reflectance mode using a spectrometer (Headwall photonics model 1002A-00371) working in the wavelength range of 1009-1694 nm with a spectral resolution of 4.85 nm (a total of 142 bands were recorded for each spectrum). The sample was illuminated with diffuse white light in an angle of 45° concerning the sample. The spectrometer was adapted to a line mapping configuration with a line of 320 pixels. The spatial resolution was 30 µm. This hyperspectral camera was kindly provided by FOSS (FOSS A/S, Denmark). Twenty TVP replicates per formulation (4) and extrusion process (2) were analyzed (n = 160). TVP were randomly positioned on the plate for image acquisition. A total of seven images (24 samples per image) were collected. Images were exported and processed in Matlab environment using PLS_toolbox and HYPER-Tools. Data were explored using PCA and then subjected to regression analysis by PLS1 algorithm. The figures of merit of the regressions in calibration and cross-validation showed excellent performance of the developed models, with values of R2 always higher than 0.90 and low values of RMSE. Prediction of external test set confirms these results. In figure 1, the prediction obtained for the protein content is reported.