COMPARISON OF MULTIVARIATE CALIBRATION AND DISCRIMINANT-ANALYSIS IN EVALUATING NIR SPECTROSCOPY FOR DETERMINATION OF MEAT TENDERNESS

Often the primary goal of analytical measurement tasks is not to find good estimates of continuous reference values but rather to determine whether a sample belongs to one of a number of categories or subgroups . In this paper the potential of different statistical techniques in t he classification of raw beef samples in tenderness subgroups was stud ied. The reference values were based on sensory analysis of beef tende rness of 90 samples from bovine M. longissimus dorsi muscles. The samp le set was divided into three categories--very tough, intermediate, an d very tender--according to degree of tenderness. A training set of sa mples was used to find the relationship between category and near-infr ared (NIR) spectroscopic measurements. The study indicates that classi cal discriminant analysis has advantages in comparison to multivariate calibration methods [i.e., principal component repression (PCR)], in this application. One reason for this observation seems to be that PCR underestimates high measurement values and overestimates Low values. In this way most samples are assigned to the intermediate group of sam ples, causing a small number of erroneous classifications for the inte rmediate subgroup, but a large number of errors for the two extreme gr oups. With the use of PCR the number of correct classifications in the extreme subgroups was as love as 23%, while the use of discriminate a nalysis increased this number to almost 60%. The number of classificat ions in correct or neighbor subgroup for the two extreme subgroups was equal to 97%. A ''bias-correction'' was also attempted for PCR, and t his gave results comparable to the best results obtained by discrimina nt analysis methods. Test sets used NTR analysis of fresh, raw beef sa mples with different processing. While this spectroscopic approach had previously been shown to be useful with frozen products, it appears u nsuitable at this time for fresh beef. However, its marginal analytica l utility proved useful in evaluating the two classification approache s employed in this study.