Near-infrared hyperspectral imaging for beef tenderness prediction

Beef tenderness is an important quality attribute for consumer satisfaction. The traditional methods to assess meat quality features are subjective and/or non-accurate, time consuming and destructive. Hyperspectral imaging can provide information with high spatial and spectral resolution, and has been successfully applied to predict chemical, microbiological, technological and sensory attributes of meat. The purpose of this study was to develop a prediction model for beef tenderness, selecting the most important wavelengths to optimize the calibration models, thereby allowing to obtain further tenderness distribution maps. One hundred eighty-eight samples (2.5 cm thick) were removed from left half carcass of 94 Nellore cattle (94 Longissimus dorsi and 94 Biceps femoris muscles). The steaks were first scanned by hyperspectral reflectance imaging system SisuCHEMA SWIR (Specim Spectral Image Ltd., Oulu, Finland) in the spectral range of 1,000-2,500 nm (resolution of 10 nm – total of 256 wavelengths), and following the Warner Bratzler Shear Force (WBSF) values, to determine tenderness, were measured for each steak. The region of interests (ROI) was selected by positioning a box in the center of muscle, and then the spectra of every pixel within the ROI were extracted and averaged at each wavelength to achieve a mean value representing the spectrum of the ROI. Two approaches, Martens’ uncertainty test (MT) and regression coefficients (RC), were used to select optimal wavelengths for model optimization. The dataset was split into calibration-set (n = 125; mean WBSF = 65.9±22.31 N) and a test-set (n = 63; mean WBSF = 67.6±20.67 N). Spectral data within full wavelength range or at the selected wavelengths (MT or RC – total of 64 and 73 wavelengths selected, respectively) were processed using partial least squares regression (PLSR). All wavelength selection and multivariate calibration and prediction were completed with software Unscrambler X 10.3 (CAMO, Oslo, Norway) and Matlab R2013a (The Mathworks Inc., Natick, MA, USA). The predictive models were evaluated in calibration and prediction processes by correlation coefficient of calibration (rc) and prediction (rp), the root mean square error of calibration (RMSEC) and prediction (RMSEP) as well as the residual predictive deviation (RPD). The calibration and prediction models established based on the full wavelengths to predict tenderness had high performance (rc = 0.91, RMSE = 8.81, rp = 0.77, RMSEP = 12.46 and RPD = 1.79). Prediction modeling based on reduced wavelengths had a little higher results using RC (rc = 0.91, RMSE = 8.89, rp = 0.81, RMSEP = 11.81 and RPD = 1.92) and MT (rc = 0.88, RMSE = 10.23, rp = 0.81, RMSEP = 11.50 and RPD = 1.98), respectively, when compared to full model. Both, RC and MT methods were similar in performance; however, wavelength reduction is an important step to improve data processing speed and modeling efficiency. This result indicates that NIR hyperspectral imaging could be used for rapid and non-destructive prediction of WBSF values with reasonable efficiency. Also it would enable further development of WBSF distribution maps of beef steaks and still enhance the understanding of its heterogeneity within beef samples.