*3.4. SCC Evaluation Based on PSO-SVR*

When the selected 57 wavelengths or 83 wavelengths were used as inputs, the optimal values of C were 2.5133 and 6.2561, respectively, as shown in Figure 5a,b. This indicated that the error tolerability produced by the first wavelength group was stricter than that produced by the second group and might result in evaluation model overfitting. The optimal values of g were 0.1128 and 0.2643 for the first and second group, respectively, indicating that the support vector number of the first wavelength group was less than that of the second. The optimal combination of C and g was obtained in the 36th iteration for MCUVE where the fitness was 0.01394%, and for SPA in the 100th iteration where the fitness was 0.01404%.

**Figure 5.** Optimization of SVR parameters. (**a**) SPA; (**b**) MCUVE.

The evaluation of SCC based on SPA-POS-SVR is shown in Figure 6a. The correlation coefficients of the Cs and Vs were 0.949 and 0.928, respectively, and the RMSEs

of both sets were 0.0253% and 0.0297%, respectively. The evaluation ability in the SCC range of 0.2000–0.4000% and 0.6000–0.7000% was better than that in the SCC range of 0.4000–0.7000%, with samples of Cs and Vs evenly distributed in the interval with smaller deviations. This showed that the robustness of SPA-PSO-SVR was poor and that the adaptability of the global evaluation was low.

**Figure 6.** Scatter plot of the calibration set (×) and verification set (o) of stone cell content. (**a**) SPA; (**b**) MCUVE.

The evaluation results of SCC based on MCUVE-POS-SVR are shown in Figure 6b. The correlation coefficients of the Cs and Vs were 0.966 and 0.951, respectively, and the RMSEs of both sets were 0.0209% and 0.0239%, respectively. The evaluation ability in the SCC range of 0.2000–0.7000% was better. Moreover, several scattering points of the Cs and Vs had relatively large deviations in the SCC ranges of 0.3000–0.4000% and 0.6000–0.7000%. Overall, the evaluation accuracy and robustness of MCUVE-PSO-SVR were better than those of SPA-PSO-SVR. The MCUVE-PSO-SVR model of SCC can be applied in online systems or portable equipment for evaluating the qualities of Korla fragrant pear.

The parameters of our spectroscopy system were set according to Korla fragrant pear. The SCC evaluating model and parameters of spectroscopy system would not be suitable for other kinds of pear, such as 'Yali' pear, 'Dangsha' pear, and so on. The minimum SCC of all samples was 0.240% in Cs and 0.315% in Vs according to Table 1. The largest evaluating errors were 8.711% and 10.845% in Cs and Vs. The evaluating precision met the requirements of application. Therefore, the minimum limit of SCC that could be detected was 0.240% by the MCUVE-PSO-SVR model.
