*3.3. Accurate Prediction of Muscle Texture Profiles Based on the Full Spectral Range*

Since the texture profiles of the four muscle regions were different, we tried to determine whether it is possible to predict the muscle texture profiles. We used all reflectance values of the skin HSI data to predict the corresponding muscle texture indicators with different ML methods. For one texture indicator of one region, we only retained the prediction with the highest *r*<sup>P</sup> for the downstream analysis (Supplementary Table S7). Predicting the chewiness of the four muscle regions had the highest *r*<sup>P</sup> (from 0.9555 to 0.9836). The overall prediction accuracies of the gumminess (*r*<sup>P</sup> from 0.9234 to 0.9863) and cohesiveness (*r*<sup>P</sup> from 0.8952 to 0.9224) of the four muscle regions ranked second and third, respectively. The overall prediction accuracies of the hardness and adhesiveness indicators were also higher than 0.88 for all four regions. Among all the best predictions, the prediction accuracy of dorsal springiness was the lowest, with only 0.5612.

The BP-ANN, LS-SVM, and PLSR models were the best three methods to predict the muscle TPA parameters (Tables 1–4). The BP-ANN method had the best calibration results mainly for gumminess, chewiness, cohesiveness, hardness, and adhesiveness, including dorsal gumminess (0.9863), pectoral gumminess (0.9620), dorsal chewiness (0.9673), pectoral chewiness (0.9555), abdominal chewiness (0.9690), gluteal chewiness (0.9836), dorsal cohesiveness (0.9224), pectoral cohesiveness (0.9306), abdominal hardness (0.9401), and gluteal adhesiveness (0.9303). LS-SVM had the highest prediction accuracy for adhesiveness, including dorsal adhesiveness (0.9206) and abdominal adhesiveness (0.9206). PLSR had the optimal prediction effect mainly for gumminess, cohesiveness, and hardness, including abdominal chewiness (0.9318), gluteal chewiness (0.9836), abdominal gumminess (0.9318), gluteal gumminess (0.9234), and pectoral hardness (0.9033).


**Table 1.** Predicting the dorsal muscle texture profiles using the skin HSI with the reflectance values in the full-wavelength range.

**Note:** *r*C: coefficients of determination for calibration. *r*P: coefficients of determination for prediction. RMSEC: rootmean-square error for calibration. RMSEP: root-mean-square error for prediction. PLSR: partial least-square regression. LS-SVM: least-square support vector machines. BP-ANN: backpropagation artificial neural network.

**Table 2.** Predicting the pectoral muscle texture profiles using the skin HSI with the reflectance values in the full-wavelength range.


**Table 3.** Predicting the abdominal muscle texture profiles using the skin HSI with the reflectance values in the full-wavelength range.


**Table 4.** Predicting the gluteal muscle texture profiles using the skin HSI with the reflectance values in the full-wavelength range.


*3.4. Accurate Prediction of the Muscle Texture Profiles Based on the Optimum Wavelengths*

Equivalent calibration results were obtained based on the optimal wavelengths compared with full wavelengths. This is because the optimal wavelengths carry the most

important information relevant to the determination. Some peaks and valleys (positive and negative relationships with the TPA parameters) were selected at certain wavelengths, and the selection of optimal wavelengths was successfully conducted for the eight texture parameters using the RC method. We selected 60 to 114 optimal wavelengths of the dorsal, pectoral, abdominal, and gluteal skin regions to predict the muscle textures, respectively (Figure 3).

For one texture indicator of one region, we only retained the prediction with the highest *r*<sup>P</sup> for the downstream analysis (Supplementary Table S8). In general, the prediction accuracies based on the optimal wavelengths were equal to those based on the full-wavelength range (Tables 5–8). Using the values in the range of the full wavelength, nine predictions had accuracies lower than 0.85. The accuracies of these regions were still lower than 0.85 using the optimal wavelengths. The remaining regions had accuracies over 0.85 using either the values of the full wavelength or the ones of the optimal wavelength. Moreover, the absolute prediction differences between the full wavelength values and the optimal wavelength values ranged from 5.93% to 15.20%, showing that the wavelength selection could make the reduced models more stable and robust.

**Table 5.** Predicting the dorsal muscle texture profiles using the reflectance values in the optimal wavelength range.


**Note:** No. of WLs: number of wavelengths.

**Table 6.** Predicting the pectoral muscle texture profiles using the reflectance values in the optimal wavelength range.


**Table 7.** Predicting the abdominal muscle texture profiles using the reflectance values in the optimal wavelength range.


**Figure 3.** Selection of optimal wavelengths in the dorsal muscle. Regression coefficients method for (**a**) gumminess, (**b**) springiness, (**c**) cohesiveness, (**d**) resilience, (**e**) hardness, (**f**) brittleness, (**g**) adhesiveness, and (**h**) chewiness.


**Table 8.** Predicting the gluteal muscle texture profiles using the reflectance values in the optimal wavelength range.

Compared with the prediction of TPA values based on the full-wavelength range, the *r*<sup>P</sup> values of four TPA parameters (cohesiveness, hardness, springiness, and resilience) in the dorsal region were enhanced using the optimal wavelength, with the increase ranging from 0.0009 to 0.1014. The remaining four values slightly decreased. In the pectoral muscle, the *r*<sup>P</sup> values of five TPA parameters, including adhesiveness, were increased by 0.0475. In the abdominal muscle region, three indicators (gumminess, springiness, and resilience) had improved *r*<sup>P</sup> values. However, in the gluteal muscle region, only the accuracies of the gumminess raised by 0.0105. These results indicate that for, the former three regions, the prediction using the optimal wavelengths would be better than using the full-range wavelengths.

The BP-ANN, LS-SVM, and PLSR models were also the best three methods to predict the muscle TPA parameters. The BP-ANN method had the best calibration tool mainly for chewiness, gumminess, hardness, and adhesiveness, including dorsal chewiness (0.9469), pectoral chewiness (0.9552), abdominal chewiness (0.9776), gluteal chewiness (0.9768), dorsal gumminess (0.9847), pectoral gumminess (0.9574), pectoral hardness (0.9033), dorsal adhesiveness (0.9191), and gluteal adhesiveness (0.9304). LS-SVM had the highest prediction accuracy for gumminess and adhesiveness, including gluteal gumminess (0.9339) and pectoral adhesiveness (0.9370). PLSR had the optimal prediction effect mainly for gumminess, cohesiveness, and hardness, including abdominal gumminess (0.9517), dorsal cohesiveness (0.9367), pectoral cohesiveness (0.9056), abdominal cohesiveness (0.9070), dorsal hardness (0.9298), and abdominal hardness (0.9392).
