**1. Introduction**

Korla fragrant pears are popular with consumers worldwide due to their beautiful skin color, sweet and crisp flesh, and rich fragrance. Consumers pay close attention to appearance quality and edible quality when they purchase Korla fragrant pears. A special characteristic of pears is that various stress factors induce the formation of stone cells, composed of large amounts of lignin and cellulose, which negatively affects the edibility and quality of the fruits [1]. There is a large number of soluble cells in some Korla fragrant pears of inferior quality, such as 'green top fruit' and rough skin fruit [2,3]. The taste is more delicate if there is lower stone cell content (SCC) in fresh Korla fragrant pears. Up to now, research on stone cells has mostly been concerned with their content [4] and structural characterization [5–7] in different germplasms of pears. Moreover, traditional methods of evaluation of the quality of edible fruit are destructive and require long processing times.

Visible and near-infrared (Vis/NIR) spectroscopy is fast, safe, and contactless technology that has been used to evaluate some qualities of fruits, including soluble solids content (SSC) [8–12], firmness [13–16], titratable acidity [17], dry matter [18], and polyphenols to amino ratio [19]. Intelligent evaluating research about SCC has not been conducted. In addition, samples are needed to change detecting places manually, and the illuminating angles cannot be changed automatically when the Vis/NIR spectroscopy systems are used.

While Vis/NIR spectrometry usually yields thousands of response signals per detection point, not all wavelengths are associated with corresponding chemical bonds as a

quality parameter of one kind of agro-product. Furthermore, the spectra at certain wavelengths do not make significant contributions to the evaluation models. Therefore, it is essential to extract characteristic wavelengths to establish intelligent models that are robust, require less computing time, and are highly predictive. Successive projective algorithm (SPA) [20,21] and uninformative variable elimination (UVE) [22,23] are two filter-type [24] algorithms that have been widely used to extract characteristic wavelengths. The SPA and UVE were evaluated using the results of correlation coefficient (R), root mean square error (RMSE), multiple linear regression (MLR), and partial least square regression (PLSR) as bases to estimate the combinations of the wavelengths. MLR [25,26] and PLSR [27,28] have different principles of establishing evaluating models. It is unreasonable to use different evaluating models to estimate the combinations of characteristic wavelengths.

A support vector machine (SVM) is a generalized linear classifier that categorizes data into two classes based on supervised learning. It has several advantages, including stability, sparsity, and simplicity. Previous applications of this algorithm mainly included qualitative analyses [29–32]. Recently, support vector regression (SVR) has been applied in quantitative evaluations as part of the development of SVM theories [33–35].

Taking all these factors into account, this study was conducted with three aims: (1) to set up an intelligent spectra acquisition system for Korla fragrant pears in which the optical subsystem has an adjustable irradiation angle and samples can be rotated at a specific angle; (2) to choose characteristic wavelengths to simplify the detection of SCC in Korla fragrant pears by SPA and UVE; and (3) to establish an SVR evaluating model after optimizing penalty factor C and kernel function parameter gamma (g).
