*3.2. Spectral Characteristics and Different Preprocessing Methods*

Spectra in the range of 498–1020 nm were considered effective, owing to the large amount of noise at both ends of the original spectrum. The effective original spectral curves of Korla fragrant pears are shown in Figure 2a. There were two reflective valleys near 680 nm and 980 nm and two reflective peaks near 550 nm and 750 nm. The spectra near 750 nm and 980 nm were related to carbohydrate content [42], and O-H [43] in the flesh of fruits. SCC had a negative correlation with carbohydrate content; therefore, the spectra at these wavelengths were indirectly related to stone cells. Spectra near 550 nm and 680 nm were related to anthocyanins and chlorophyll in the sample epidermis, respectively [44]. Korla fragrant pears with high levels of stone cells usually have green skin; therefore, SCC also had some relationship with spectra near 550 nm and 680 nm.

**Figure 2.** Reflective spectral curves. (**a**) Raw spectrum; (**b**) spectrum after SNV pretreating.

Using preprocessing algorithms could improve the evaluation accuracy compared with not using them. The principal component numbers of PLSR models were all 10 after different preprocessing algorithms. The optimal combination of frame size and fitting order was (7, 5) where R of the calibration set and validation set grew the largest, according to Table 2. The evaluation results of PLSR models based on different spectral preprocessing algorithms are shown in Table 3. Evaluation models based on MSC and SNV had higher Rs and lower RMSEs, while those based on S-G(7, 5) had lower Rs and higher RMSEs, in Cs and Vs. The robustness of the PLSR model based on SNV was better than that of MSC according to the different values of Rs between Cs and Vs. The addition of S-G(7, 5) did not improve the ability of evaluation because the combination of two-point smoothing and S-G(7, 5) eliminated some effective spectral information. The model established on the basis of the SNV preprocessing algorithm achieved the best results, with R and RMSE of 0.9189 and 0.0277% in the Cs, and 0.8935 and 0.0315% in the Vs. Spectral curves based on SNV are shown in Figure 2b.


**Table 2.** Correlation coefficients of Cs and Vs with different S-G parameters.


RC: the correlation coefficient of the calibration set; RMSEC: root mean square error of the calibration set; RV: the correlation coefficient of the validation set; RMSEV: root mean square error of the validation set; S-G(7,5): Savitzky–Golay filter with a frame size of 7 and fitting order of 5.

SNV 10 0.8999 0.0308 0.8641 0.0356
