**1. Introduction**

Korla fragrant pears are very popular among customers due to their thin skin, juicy, sweaty taste, and delicate flesh [1,2]. Nowadays, customers pay significant attention to both the organoleptic and physicochemical attributes of fruits with the improvement of their living standards. The organoleptic parameter, color of skin, is related to maturity of Korla fragrant pear. The sunward side of most mature Korla fragrant pears has blush which is also distinctive in all kinds of pears. However, only physicochemical parameters are used as quality evaluation attributes to grade Korla fragrant pears.

Several non-destructive studies have been carried out to evaluate the soluble solid content (SSC) of Korla fragrant pears [3,4]. Zhu et al. [5] used hyperspectral imaging and support vector regression to define this parameter. The correlation coefficient (RC) and the root mean square error (RMSEC) in their calibration set measured 0.986 and 0.186%, respectively. In their validation set the correlation coefficient (RV) and the root mean square error (RMSEV) measured 0.946 and 0.403%. Zhan et al. [6] quantitatively determined the

**Citation:** Liu, Y.; Wang, T.; Su, R.; Hu, C.; Chen, F.; Cheng, J. Quantitative Evaluation of Color, Firmness, and Soluble Solid Content of Korla Fragrant Pears via IRIV and LS-SVM. *Agriculture* **2021**, *11*, 731. https:// doi.org/10.3390/agriculture11080731

Academic Editor: Isabel Lara

Received: 21 June 2021 Accepted: 26 July 2021 Published: 31 July 2021

**Publisher's Note:** MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

**Copyright:** © 2021 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/).

SSC of Korla fragrant pears via least square support vector machine (LS-SVM) and partial least square regression (PLSR). The RV and RMSEV reported in this study measure 0.851 and 0.291%, respectively.

Other researchers investigated the firmness of Korla fragrant pears vie quantitative predictions. For instance, Sheng et al. [7] used near-Infrared (NIR) spectroscopy together with different variable selecting methods to construct a set of partial least square models to describe firmness. Yu et al. [8] predicted both the firmness and the SSC by developing a deep learning method based on Vis/NIR hyperspectral reflectance imaging. Their combination model of a series of stacked auto-encoders and a fully connected neural network achieved a reasonable prediction performance with RV and RMSEV values of 0.9434 and 1.81 N, respectively.

However, no investigation reported results on the simultaneous measurement of the organoleptic and physicochemical attributes on Korla fragrant pears. According to the requirements of the latest group standard on Korla fragrant pears [9], organoleptic and physicochemical attributes appear to have the same importance in the grade definition. The skin color of Korla fragrant pears changes from green to red-yellow, as the fruit ripes. The a\* value represents the color change from red to green in chromatic aberration data. Therefore, the organoleptic quality of the samples can be defined according to their a\* values. The firmness and the SSC are the most significant edible quality parameters in Korla fragrant pears, and they are directly related to consumers' satisfaction [10]. Thus, the three parameters, a\*, firmness, and SSC, must be carefully evaluated to determine the influence of the postharvest storage period on the fruit quality control process.

Both quality and safety parameters can be accurately evaluated via hyperspectral imaging [11–13], although the hyperspectral approach requires expensive equipment and complex data analysis Compared with other nondestructive testing methods [14]. However, in order to define simple predicting models and improve their prediction efficiency, a set of wavebands have to be selected. These wavebands can be related with several important chemical bonds, which can be used to discriminate the samples based on their quality and safety parameters. Successive projection algorithms (SPAs) [15–18], competitive adaptive reweighting sampling (CARS) [19,20], and uninformative variable elimination (UVE) [21,22] have been used by to choose such wavebands. Despite these selection methods are quite effective, they do not account for the combination effects among the wavebands. The iteratively retaining informative variables (IRIV) method ensures that each variable has the same probability to take part into the selection process and increases filtering speed by using a set of binary mixing filters [23,24].

To this date, the combination of IRIV and LS-SVM has not been investigated to quantitatively predict the quality parameters of Korla fragrant pears. In this work, IRIV-LS-SVM is used to (1) obtain the a\* value, the firmness, and the SSC of Korla fragrant pears, (2) analyze the spectral features of Korla fragrant pears in the 945–1670 nm wavelength range, (3) select the optimal wavebands related to the C-H, N-H, and O-H chemical bonds, and (4) construct a set of predicting models to define the quality parameters for Korla fragrant pears.

#### **2. Materials and Methods**

#### *2.1. Korla Fragrant Pears and Pretreatment*

Korla fragrant pears were collected from a plantation located near Tarim University (80◦30 –81◦58 E, 40◦22 –40◦57 N) from September 11th to September 15th 2019. A set of 200 samples with a uniform shape, a single fruit weight of 120 ± 10 g, and intact epidermis was selected. The side of each Korla fragrant pear, which was exposed to the sunlight, was labeled.

The samples were sprayed with a special fruit cleaning agent (Almawin, Germany), soaked in water for about 30 s, and then rinsed with distilled water twice. The cleaned pears were dried at room temperature (20 ◦C), and then stored in a preservation box at

4 ◦C. The samples were placed on the desk at room temperature for 30 min to eliminate the influence of the temperature change before the hyperspectral image data acquisition.

#### *2.2. Hyperspectral Imaging System and Diffuse Reflectance Spectrum Data Acquisition*

The hyperspectral imaging system used in this study is shown in Figure 1. It consists of a push-broom scanning system composed of a spectrograph (N17E, Spectral Imaging Ltd., Oulu, Finland), an enhanced near-infrared hyperspectral camera (Xeva-1.7-320, Xenics Infrared Solutions, Leuven, Belgium), four halogen light sources with a maximum power of 150 W each, a stepper-motor-driving stage, a dark box, and a computer.

**Figure 1.** Schematic diagram of the hyperspectral imaging system.

Before data acquisition, the system was preheated for about 30 min to ensure its temperature stability. A sample with a uniform shape and a moderate weight was placed on the stage to adjust the calibration parameters of the instrument. In order to ensure the hyperspectral image integrity of the samples, the distance between the bottom of the spectrograph and the stage was set to 310 mm and maintained fixed. Moreover, the focal length was adjusted by rotating the aperture until sharp reflection peaks appeared. The moving speed of the platform and the exposure time of the camera were set to 18 mm/s and 20 ms, respectively.

A strip-shaped standard reflecting whiteboard was placed under the spectrograph to obtain the white and the black references separately by opening and closing the lens cover. The long axis of a Korla fragrant pear sample was positioned along the moving direction of the stage to ensure a uniform irradiation. The sunward side of each Korla fragrant pear was placed upside in order to reduce the influences on spectral data and measured quality parameter values of the sunward side and nightside. The sample hyperspectral image data were corrected by taking into account the black and white references to eliminate the influence of the light source intensity differences and the camera dark current noise, as described in Formula (1).

$$\mathbf{I} = (\mathbf{I}\_{\rm O} - \mathbf{I}\_{\rm b}) / (\mathbf{I}\_{\rm W} - \mathbf{I}\_{\rm b}) \tag{1}$$

Here, Io corresponds to the original hyperspectral image data, Ib to the black reference data collected when the lens cover is closed, Iw refers to the white reference image data of the strip-shaped standard reflecting whiteboard when the lens cover is open.

The spectral data of the region of interest (ROI) were extracted by using the ENVI 5.1 software (Exelis Visual Information Solutions, Boulder, Colorado, USA). The shape of the ROI was rectangular, and its center was located near the intersection between the long axis and the equator of the pear. The corresponding pixel numbers of each ROI were 90 along the long axis and 70 along the equator.

#### *2.3. Measurement of the Sensory and Physicochemical Parameters*

The sensory and physicochemical parameters of the Korla pears were measured after the hyperspectral image data acquisition. The sensory parameter a\* was obtained by employing a precision chromatic aberration meter (HP-C220, Shenzhen HanPu Testing Instrument Co., Ltd., Shenzhen, China). Each measurement consisted of an average of five points randomly selected on the ROI surface.

The firmness was obtained by averaging the values collected at five different locations of the pears. They were set at 12 mm center distance between two adjacent ROIs and were measured by a firmness tester (GY-4 with a probe diameter of 7.9 mm, Top instrument). The SSC was measured by using a digital refractometer (PAL-1, ATAGO, Tokyo, Japan). Before the measurement, the refractometer was calibrated with distilled water. Three small pieces of pulp of about 5 g each were cut out from the ROI. Their liquid content was dropped into a sample tank by manual extrusion. The average value of the solid content of the three pulp samples was taken as the measurement value.

#### *2.4. Spectral Preprocessing*

The standard normal variable transformation (SNV) is a normalization, which is sometimes employed in near infrared spectroscopy [25,26]. This preprocessing algorithm can center and scale each spectrum. Multiple scatter correction (MSC) is used to compensate for the non-uniform scattering effects in spectral data, when heterogenous sample sizes, irregular distributions, and other physical effects are present [27]. Whereas the Savitsky-Golay (S-G) algorithm can be used to improve smoothness of spectral curves. The different preprocessing effects obtained with MSC, SNV, MSC-SG, and SNV-SG were compared to evaluate the characteristics of the PLSR models.
