Modelos de calibração multivariada por NIRS para a predição de características de qualidade da carne bovina

Abstract

Near infrared reflectance spectroscopy (NIRS) has been successfully applied in the quantitative determination of the main constituents of beef but it has been presenting inconsistent results in determining characteristics relating to tenderness. In addition, the various aspects related to data processing (mathematical pre-treatments, spectral bands, sample presentation, regression method), should be constantly evaluated, since they affect the prediction cap acity of NIRS. In this context, the present study was developed to determine which spectral data-processing methods make it possible, using the PLS regression method, to obtain robust calibration models that determine the chemical composition and tenderness characteristics of beef. The accuracy of the models was determined by external validation, which has been little used in previously published studies. To develop the calibration models, three spectra were collected from each sample of the Longissimus dorsi muscle of 25 mixed-breed castrated dairy calves, divided into five treatments (five repetitions in each) based on supplying diets containing millet and including babassu mesocarp bran at proportions of 0; 12; 24; 36 and 48% in the dry matter of the total diet, comprising 75 spectra. For the external validation set, samples were used from five mixedbreed castrated dairy calves fed on a diet based on maize and soybean, totalling 15 spectra. To determine the chemical composition (fat content, protein, ash content and moisture) and the tenderness properties (water holding capacity – WHC -, total and soluble collagen, shear force, FMI and pH), 135 calibration models were developed with mathematical pre-treatments available on VISION software, version 3.1, using PLS regression, from which 37 (27.41% of the total) presented coefficients of determination considered good or excellent in their predictive capacity. The pre-treatment with “first derivatives” made it possible to develop more robust models for the chemical composition properties, except for RMF, in which “Savitzky-Golay” and “second derivatives” were more efficient, obtaining R 2 and RPD values above those available in the literature. For determining the tenderness properties in beef, the models develope d with “first and second derivatives” pre-treatments, in isolation or with “Savitzky -Golay” or “multiplicative scatter correction” smoothing methods, presented the highest values of RPD, demonstrating that themselves are efficient chemometric tools for obtaining robust calibration models. Models were obtained with limited predictive capacity only in the determination of total fats and total collagen quantification. This was probably due to the low variability presented in the samples used a nd to the low sensitivity of NIRS for total collagen. It was concluded that NIRS can be used to replace conventional methods, being a fast and precise technique, as well as allowing simultaneous analysis of beef quality characteristics.