Partial least squares analysis of near-infrared hyperspectral images for beef tenderness prediction

Abstract

Tenderness is a primary determinant of consumer satisfaction of beef steaks. The objective of this study was to implement and test near-infrared (NIR) hyperspectral imaging to forecast 14-day aged, cooked beef tenderness from the hyperspectral images of fresh ribeye steaks (n = 319) acquired at 3–5 day post-mortem. A pushbroom hyperspectral imaging system (wavelength range: 900–1700 nm) with a diffuse-flood lighting system was developed. After imaging, steaks were vacuum-packaged and aged until 14 days postmortem. After aging, the samples were cooked and slice shear force (SSF) values were collected as a tenderness reference. After reflectance calibration, a Region-of-Interest (ROI) of 150 × 300 pixels at the center of longissimus muscle was selected. Partial least squares regression (PLSR) was carried out on each ROI image to reduce the dimension along the spectral axis. Gray-level textural co-occurrence matrix analysis with two quantization levels (64 and 256) was conducted on the PLSR bands to extract second-order statistical textural features. These features were then used in a canonical discriminant model to predict three beef tenderness categories, namely tender (SSF ≤ 205.80 N), intermediate (205.80 N < SSF < 254.80 N), and tough (SSF ≥ 254.80 N). The model with a quantization level of 256 performed better than the one with a quantization level of 64. This model correctly classified 242 out of 314 samples with an overall accuracy of 77.0%. Fat, protein, and water absorption bands were identified between 900 and 1700 nm. Our results show that NIR hyperspectral imaging holds promise as an instrument for forecasting beef tenderness.