Dataset of near-infrared hyperspectral images and 2D point-based spectra of beef patties with plant-based proteins added.

Spectra in the near-infrared (NIR) range of the electromagnetic spectrum were non-destructively recorded from beef patties with added plant-based protein. The data were recorded with two instruments measuring in the NIR range. First, with a point-based hand-held microNIR spectrometer covering the 908.1 - 1676.2 nm spectral range and, second, with a hyperspectral camera (HSI) recording data in the 900 - 1700 nm range. The dataset is composed of spectral data from a total of 240 beef patties fabricated with Spanish Ternera de Navarra beef that were enriched with pea (PP, nPP = 60), soybean (SP, nSP = 60) and chickpea (CP, nCP = 60) proteins. Additionally, 60 samples of pure beef were also considered as control samples. The plant-based proteins were incorporated at different percentages: 1%, 2%, 3%, 4%, 5% and 6% w/w with a total of 10 replicates per content and percentage. No additional ingredients apart from beef meat and protein were included. Partial-least squares discriminant analysis (PLS-DA) and linear discriminant analysis (LDA) by protein type (PP, SP, CP) and content added (≥3% or <3%) yielded accurate results. Briefly, excellent results were achieved for the classification according to the type of protein added with >90% of correctly classified samples in Test using the microNIR. Besides, >94% of the samples were correctly identified when using the legal content (3%) as threshold with the HSI camera. Also, hierarchical modelling was carried out to tackle double classification tasks, first, by type of protein and, second, by the amount of protein added in a unified approach. In this scenario, the microNIR device reached higher classification rates (>94% in Test) than the HSI camera (>83% in Test). The dataset is organized into folders and subfolders containing raw and pre-processed data. Concisely, it includes a CSV file with 1200 rows in which the normalized spectra of all the replicates measured with the microNIR device are gathered. Besides, it includes a HSI folder that comprises, for instance, the white and black references taken every measurement day, the segmented masks of each patty, the 3D cubes, the unfolded 2D matrices, and the mean spectra of each patty. The potential reuse of this dataset encompasses training machine learning algorithms able to classify or predict the added protein, to perform calibration transfer among hyperspectral sensors and hand-held devices or even to train image segmentation methods. Consequently, the meat industry could move towards identifying potential fraudulent practices and; therefore, the digitalization of quality control measures.