R2CV values for WBSF in the current study were greater than those obtained by De Marchi (2013), who reported R2CV values of 0.08 and 0.13. Standard error of cross validation values were within in the same region as the present study (SECV 6.51 and 6.98). Most interestingly, De Marchi (2013) showed the application of a mathematical pre-treatment upon spectra (multiplicative scatter correction and first derivative) had a negative effect upon predictive accuracy, reducing it from 0.13 for untreated spectra down to 0.08. This pre-treatment methodology gave a result that corresponded with the majority of meat quality traits modelled in the present study. Ripoll et al. (2008) achieved higher coefficients of determination for WBSF prediction (R2P 0.743). Though, this value was achieved by collecting spectra on homogenized rather than intact sample, which previously has been shown to increase predictive capabilities (Weeranantanaphan, Downey, Allen, & Sun, 2011).
A homogenization step would not be suitable for implementation into a commercial setting, as it is both destructive of sample and time consuming. In the present study the aim was to achieve an acceptable calibration on intact muscle, in order to be compatible with future online or at-line installation. With industry increasingly moving towards in-pack ageing, this may be a useful point at which to record spectra, predict meat quality at-line and manage the resultant product for tenderness. While WBSF is a robust and well publicised instrumental measurement of tenderness in meat, analysis by means of a trained sensory panel is still the gold standard measurement and supersedes any WBSF result of a sample. Nevertheless, shear force analysis does offer some advantages as it is a more rapid method that offers less subjectivity than sensory assessors; even for those that have been highly trained this can still be an issue (Prieto, Roehe, Lavín, Batten, & Andrés, 2009; Wariss, 2010).
Sensory tenderness predictions resulting from spectra collected offline in the laboratory at 49h post mortem (R2CV 0.13) were similar to that acquired by Prieto et al. (2009) (R2CV 0.16; SECV 0.6). However, the present study had a higher margin of standard error (RMSECV 8.6). The difference in cross-validation standard error values observed between the two studies may be explained by the contrast in standard deviation values for the tenderness trait. The present study has a SD of 10.47 for tenderness, while Prieto et al., (2009) had a lower SD of 0.7. Yancey et al., (2010) predicted sensory tenderness values with a coefficient of determination of cross validation value of 0.4 and 0.7 (2nd derivative), with predictive error values of 0.71 and 0.57. Though, the small sample size used in that study may be a hindering factor when applying this chemometric model on a larger sample set. In the present study, spectra collected at the 49 h time-point were recorded on LTL samples that were sliced and bloomed 1 h, indicating that oxygenation and exposure to light, or perhaps colour stability may have a role to play in optimal prediction of this sensory trait (Mancini, 2009). The 49 h model that gave the highest R2 was constructed using the full spectral wavelength, 350 – 2500 nm, demonstrating that both visible and NIR wavelength regions may have a role to play in the prediction of tenderness.
When 49 h spectral measurements for the 5 most tender samples are averaged and plotted with the average spectra for the 5 toughest samples, the tougher samples have higher absorption between 550 nm – 1450 nm as shown in Figure 2. Prediction results for tenderness achieved by Ripoll et al. (2008) were more accurate than those obtained in this study with R2CV value of 0.981 obtained. The result achieved in their study may be attributed to the use of animals with differing maturity rates and varying fatness and conformation grades, leading to greater variation within their dataset and in turn higher coefficients of determination. Moreover, in their study NIR spectral measurements and sensory panel analysis were carried out on the same day (7 d PM), which possibly contributed to the higher R2 values attained. While this is a good experimental design (collecting spectra and performing lab analysis on the same day PM), collecting spectra offline at 7 d PM may not be suitable for implementation online within the factory, especially for the aim of the present study. In the current study the tougher sensory textural attributes chewiness, stringiness and difficulty to swallow were all highly positively correlated with each other (P < 0.001). These traits were all predicted in a similar fashion with a low degree of accuracy (R2CV 0.1 – 0.22; RMSECV 7.1 – 9.3). Best fitting models for all three sensory traits were built using spectra from the visible wavelength region, 450 – 779 nm, and collected 2 d PM (48 h chewiness and stringiness; 49 h difficulty to swallow). Both day 1 and day 2 PM best fitting models for the prediction of difficulty to swallow trait were constructed using spectra collected on meat left to bloom 1 h (25 and 49 h), indicating that colour stability and the oxygenation of meat may play a role in prediction of this trait.
Coefficients of calibration for chewiness obtained in the current study (R2Cal 0.38) are lower than in other studies (Liu et al. 2003) (Rødbotten et al., 2000), who reported R2Cal values of 0.58 and 0.38, respectively. R2CV values are not presented in the studies mentioned so cannot be compared. It is notable that the models presented by Liu et al. (2003) were constructed using just 24 carcasses from the similar Angus and Hereford breeds; with the possibility that this animal group used for calibration is too homogenous for prediction of chewiness in a larger scale. Prediction of crumbly sensory texture gave the highest coefficient of determination of all traits analysed in the present study. In the frame of cross validation, the 48 h spectra model outperforms the 25 h model in nearly every aspect with higher R2CV (0.41; 0.34) and lower RMSECV values (9.4; 10.5). This suggests that VisNIRS may indeed be a useful technology for the prediction of this important sensory trait. When 25 h spectral measurements for the five average lowest scoring and the five average highest scoring crumbly samples are plotted together, the lower scoring samples (indicating toughness) have higher absorption between 600 nm – 2500 nm, with prominent differences in absorption visible between 1400 nm – 2500 nm.
Best fitting models for crumbly texture (Table 3) were built using the full spectral wavelength range (350 – 2500 nm) and without the use of any mathematical pre-treatment as they were not found to improve accuracy of prediction. The ability to predict instrumental and sensory textural information of carcasses at an early PM stage on-line is desirable for the meat industry, as it allows for the formation of differentiated product lines of various grades of meat with higher quality and high added value (Balage, da Luz e Silva, Gomide, Bonin, & Figueira, 2015). In turn, improved grading of meat based on quality may lead to a fairer cost to the consumer and additionally a fairer payment to the producer. As mentioned previously, methods currently employed for the measurement of meat quality do not allow for industry wide quality monitoring or the recording of these traits for breeding objectives. The ability to implement on-line systems to predict meat quality early PM within the factory could be a major advantage to breeding companies for selection of animals for breeding based on meat quality phenotypes (Cecchinato, de Marchi, Penasa, Albera, & Bittante, 2011; De Marchi, 2013). To the best of the author’s knowledge there are no other reports in the literature relating to the use of near infrared spectroscopy for the prediction of chewiness, stringiness, difficulty to swallow and crumbly texture traits in beef. Sensory panel juiciness scores predicted via VisNIRS had the lowest R2 value of all traits predicted in the current study, with the highest R2CV achieved being 0.09 (Table 3), indicating that VisNIRS spectroscopy may not be suitable for the prediction of this trait.
The best fitting PLSR models for juiciness shown in Table 3 were constructed using spectra collected off-line (49 h PM), when the muscle had been sliced and bloomed for 1 h. Both 24 h and 49 h models were shown to have the relatively low RMSECV value of 6.8. This result for the prediction of juiciness by way of VisNIRS is in agreement with that of Prieto et al. (2009), who also achieved a low R2CV value (0.13). Studies carried out by Liu et al., (2003) and Ripoll et al., (2008) however were able to achieve more accurate predictions of juiciness by way of VisNIRS/NIRS with R2 calibration values between 0.5 – 0.54 and standard error values of 0.18 – 0.64 reported. Clearly, juiciness is a more complex trait than previously considered and aspects of the physical structure may impact on the release of juice in the mouth which was not well-predicted here. Rødbotten, Nilsen, & Hildrum (2000) were unable to obtain any predictive model for sensory juiciness scores in their study even after multiplicative scatter correction mathematical pre-treatment was applied to the spectral data, again underlining that prediction of juiciness scores is a complex affair and a difficult attribute to consistently predict with accuracy. Prediction of beef LTL sensory flavour In Table 2, beef flavour and beef after affect are shown to be highly positively correlated (R2 0.72, P<0.001), however best fitting predictive models for both traits were calibrated using different spectral wavelength ranges, different PM times, and different mathematical pre-treatments (Table 3). The best fitting model for beef flavour prediction provided an R2CV value of 0.13 and an error value of 6.4, and was constructed using the full spectral wavelength and no spectral pre-treatment. Beef after-effect was predicted with an R2CV of 0.2, and a lower standard error (in comparison to beef flavour) of 5.4.
The model for after-effect was calibrated using spectra from the near-NIR range, which is far narrower and in contrast with the full wavelength range which gave most accurate determinations for beef flavour. Due to the high significant correlation and similarities between beef flavour and beef after affect, it is somewhat surprising that one trait is predicted more accurately than the other. Low predictions for beef flavour were also reported by Prieto et al., (2009) (R2CV 0.13 – 0.14; SECV 0.41 – 0.46) and Byrne, Downey, Troy, & Buckley, (1998) (R2CV 0.24; SECV 0.39). It is also worth noting that beef after-effect was the only sensory trait in this study where a spectral pre-treatment (SNV) model provided the best fitting model for both day 1 and 2. In all other models within the present study, a spectra pre-treatment has only improved the accuracy of models for one PM day. Fatty mouthfeel and fatty after-effect are two phenotypic traits examined that are both comparable and highly positively correlated (Table 2). Best fitting models for both traits (Table 3) also share a number of similarities in their calibration, such as similar time-point (49 h PM), wavelength range utilized (780-1099 nm), and RMSECV value (3.2), in contrast to the aforementioned beef flavour and after-effect calibrations. R2CV predictions for fatty after-effect were marginally more accurate of the two at 28 % (compared to 23 % for mouthfeel), with this model calibrated with the use of Savitzky-Golay smoothing.
This was the only model within the present study where SG smoothing increased the predictive capabilities of a model, in comparison to untreated spectra. Interestingly, it was observed that R2 calibration models for both traits related to fattiness had higher predictive values in models constructed with Day 1 spectra (R2Cal 0.46-0.53), indicating that some information for these traits is captured in NIR spectra recorded on the day of quartering. In Figure 4, 25 h spectral measurements for the average 5 lowest scoring and the average 5 highest scoring fatty mouthfeel samples are plotted together. The 5 higher scoring samples have higher absorption between 580 nm – 1300 nm. However, this changes between 1380 nm and 2500 nm where the lower scoring fatty mouthfeel samples have increased absorption. Peaks are visible at 980 nm, 1100 nm and 1400 nm. The peak visible at 980 nm is most likely due to water (Liu et al., 2003), while the peaks visible at 1100 nm and 1400 nm are most likely related to C – H molecular bonds of fatty acids (Williams & Norris, 2001). High positive correlation was observed between metallic flavour and metallic after effect (Table 2), as expected as they are complementary phenotypic traits. Best fitting predictive models for both traits were found to be very similar, with both models calibrated using spectra from the same time-point (24 h PM), the same PLS terms utilized (10 terms), and the same wavelength region (780-1099 nm). The models discussed also had a similar R2CV prediction of 0.17 and 0.19, respectively. While predictions for metallic traits mentioned are low, it is noteworthy that when models are built using so many different variables that the best fitting models found are so similar, indicating that a significant amount of information for these traits may be present in the near-NIR region of spectra captured immediately after carcass quartering. To the best of the author’s knowledge no other reports exist in the literature relating to the use of visible and near infrared spectroscopy for the prediction of the sensory traits beef after-effect, fatty mouthfeel, fatty after-effect, metallic flavour and metallic after-effect.
This essay has been submitted by a student. This is not an example of the work written by our professional essay writers. You can order our professional work here.