Construction of a Chlorophyll Content Prediction Model for Predicting Chlorophyll Content in the Pericarp of Korla Fragrant Pears during the Storage Period

Abstract

A chlorophyll content prediction model for predicting chlorophyll content in the pericarp of Korla fragrant pears was constructed based on harvest maturity and storage time. This model predicts chlorophyll content in the pericarp of fragrant pears after storage by using the error backpropagation neural network (BPNN), generalized regression neural network (GRNN) and adaptive neural fuzzy inference system (ANFIS). The results demonstrate that chlorophyll content in the pericarp of fragrant pears decreased gradually as the harvest time lengthened. The chlorophyll content in the pericarp of fragrant pears with different maturity levels at harvest decreased continuously with the increase in storage time. According to a comparison of the prediction performances of the BPNN and ANFIS models, it was discovered that the trained GRNN and ANFIS models could predict chlorophyll content in the pericarp of fragrant pears. The ANFIS model showed the best prediction performances when the input membership functions were gasuss2mf (RMSE = 0.006; R² = 0.993), dsigmf (RMSE = 0.007; R² = 0.992) and psigmf (RMSE = 0.007; R² = 0.992). The findings of this study can serve as references for determining the delivery quality and timing of Korla fragrant pears.

1. Introduction

Korla fragrant pears, a pear species native to Xinjiang, China, are highly appreciated by consumers due to their thin skin, rich juice and sweet and crispy taste [1,2,3]. After harvest, storing in a refrigeration house at ice temperature and then marketing at a proper time are the main ways to prolong the supply stage and increase the economic value of Korla fragrant pears [4]. Practitioners use the apparent quality of fragrant pears as a benchmark for determining the delivery quality and timing. Chlorophyll is an essential pigment substance in fragrant pears, and the amount contained in the pericarp affects the appearance and color of pears while also indicating physiological metabolism changes and nutrition conditions [5,6]. The content of chlorophyll in the pericarp is often used as a physical indicator to study the post-ripening characteristics and aging process of fruits [7]. It directly impacts the apparent quality and commodity value of fragrant pears. Therefore, studying the variation laws of chlorophyll content in the pericarp of fragrant pears during storage and developing an accurate prediction model for chlorophyll content in the pericarp can provide references for controlling quality at delivery and determining the time for delivery of fragrant pears. These considerations are crucial for promoting the industrial development of fragrant pears.

Currently, associated studies focus on the effects of storage conditions on chlorophyll content in the pericarp of pears [8,9]. Nevertheless, fruit quality after harvest is primarily determined by maturity at harvest [10]. Therefore, some scholars have begun to pay close attention to the influences of maturity at harvest on chlorophyll content in fruits. Lan et al. discovered that maturity at harvest impacted the chlorophyll content in Korla fragrant pears during storage [11]. Niu et al. discovered a significant correlation between harvest maturity and chlorophyll content in the pericarp of Korla fragrant pears during storage [12]. Notably, fragrant pears in storage mature gradually and lose quality progressively as their storage time increases. Wang et al. demonstrated that even after long-term storage with a CaCl₂ or carboxymethyl chitosan coating, the quality of the Korla fragrant pear was still unsatisfactory [13]. Jia et al. were unable to maintain the apparent colors of fragrant pears using 1-methylcyclopropene processing [14]. As a result, all storage methods must take into account the storage time constraints caused by the post-ripening characteristics of fragrant pears. In a nutshell, maturity at harvest and storage time are important factors influencing chlorophyll content in the pericarp of fragrant pears. The dynamic model is widely applied to predict the quality indices of food during storage [15,16]. However, it has poor adaptation. Most models and methods are applicable only to the same type of food. Artificial neural networks have strong adaptation and fault tolerance and therefore may bring prediction results closer to objective reality [17]. Hence, neural networks can offset the shortcomings of the traditional dynamic model. In recent years, error backpropagation neural network (BPNN), generalized regression neural network (GRNN) and adaptive neural fuzzy inference system (ANFIS) have been widely applied to predict the quality of fragrant pears due to their strong universality and high prediction accuracy. For instance, they have been used to predict hardness, soluble solid content and vitamin C content [18,19,20]. To date, no research has been published on the prediction of chlorophyll content in the pericarp of fragrant pears during storage.

In this study, the variation laws of chlorophyll content in the pericarp of fragrant pears with different maturity levels at harvest during the storage period are discussed. BPNN, GRNN and ANFIS models for predicting chlorophyll content in the pericarp of fragrant pears were constructed. According to a comprehensive comparison of the BPNN, GRNN and ANFIS models for predicting chlorophyll content in the pericarp of fragrant pears, the optimal model was determined and verified.

2. Materials and Methods

2.1. Sample Preparation

Fragrant pear samples were collected from 5 September 2020 to 15 October 2020 in the fragrant pear garden in the Alaer City 12th Regiment, the First Division of the Xinjiang Production Construction Corps, using the Korla Fragrant Pear Association standards [21]. Samples were collected every five days to obtain fragrant pears with different degrees of maturity. The harvest periods of the fragrant pears were defined as H1, H2, H3, H4, H5, H6, H7, H8 and H9. The maturity at harvest of the fragrant pears was defined as either 1, 2, 3, 4, 5, 6, 7, 8 or 9, and these values were used as input for the neural network model. During the harvest period, as chlorophyll in the pericarp degrades continuously, the apparent color of fragrant pears changes from green to yellow-green and finally to yellow, indicating that fragrant pears are completely ripe (Figure 1). At each sampling, fragrant pears (45 in total) devoid of distortion, damage, diseases and insect damage were chosen. Among these, nine fragrant pears were selected randomly to measure the chlorophyll content in the pericarp, and the mean values were recorded. Picking of fragrant pears was conducted 9 times, and a total of 405 fragrant pears were needed for the whole experiment. On the day of harvest, fragrant pears of the same maturity were transported to the refrigeration house and stored at −1 ± 1 °C with a relative humidity of 90 ± 5%, oxygen concentration of 5–8%, carbon dioxide concentration of less than 2.5% and ethylene concentration of 30–90 µL/L.

2.2. Measurement of Chlorophyll Content in the Pericarp

The measurement of chlorophyll content in the pericarp using the formulas for calculating chlorophyll content in the pericarp were carried out following the method of Gao et al. [22,23,24]. Three fragrant pears were chosen randomly. The pericarp was collected and cut into pieces. Later, 0.3 g of the pericarp was placed in a mortar with quartz sand, calcium carbonate and 2 mL of 95% ethyl alcohol. The mixture was ground into a homogenate, and then 10 mL of 95% ethyl alcohol was added and the mixture was ground continuously until the pericarp tissues were completely whitened. The whole mixture was kept static for five minutes. Next, the homogenate was filtered into a 25 mL brown flask. The mortar, rod and residues were rinsed several times with 95% ethyl alcohol, and the residual homogenate was poured into the funnel for filtration. Later, some ethyl alcohol was collected with a dropper, and all of the chlorophyll on the filter paper was rinsed into the flask. Finally, a constant volume of 25 mL of ethyl alcohol was added, and the mixture was vibrated uniformly. The chlorophyll extract was poured into a cuvette with a diameter of 1 cm. The 95% ethyl alcohol was used as the blank control. Absorbance values at 665 nm and 649 nm were tested using a spectrophotometer. The test was repeated three times, and the mean values were calculated. The formulas for calculating chlorophyll content in the pericarp were as follows:

$$\mathrm{Concentration}\mathrm{of}\mathrm{chlorophyll}\mathrm{a}:C_a=13.95A_{665}-6.88A_{649}$$

(1)

$$\mathrm{Concentration}\mathrm{of}\mathrm{chlorophyll}\mathrm{b}:C_b=24.96A_{649}-7.32A_{665}$$

(2)

$$\mathrm{Total}\mathrm{concentration}\mathrm{of}\mathrm{chlorophyll}:C=C_a+C_b$$

(3)

$$\mathrm{Chlorophyll}\mathrm{content}:M=CVn/1000m$$

(4)

where A₆₆₅ indicates absorbance at 665 nm, A₆₄₉ indicates absorbance at 649 nm, C_a is the concentration of chlorophyll a (mg/L), C_b is the concentration of chlorophyll b (mg/L), C is the total concentration of chlorophyll (mg/L), M is the chlorophyll content (mg/g), V is the volume of the extract (mL), n is the dilution ratio and m is pericarp mass (mg).

2.3. Modeling Methods

In this study, three modeling methods were used: BPNN, GRNN and ANFIS. The network input was the fragrant pears’ harvest maturity and storage time, and the network output was chlorophyll content in the pericarp. According to the statistical experimental data, a total of 45 groups of datasets were gained, from which 70% of the experimental data were randomly chosen for a training set and the remaining 30% were chosen for a testing set. On this basis, a prediction model for chlorophyll content in the pericarp of fragrant pears with different maturity levels at harvest during the storage period was developed, and the optimal prediction model was selected. During the following year, the practical performances of the optimal prediction model were verified.

2.3.1. BPNN Model

BPNN is a feedforward network that uses the error backpropagation algorithm, and thus far, it is the network that has been most frequently used. It is mainly applied for pattern recognition, data mining, system identification and automatic control, among other things. In fact, it calculates the minimum total error function of the network and corrects the weight coefficient along the negative gradient direction of the error function by using the learning rule of the steepest descent method [25].

2.3.2. GRNN Model

GRNN is an important variation of the radial basis function neural network. Because of its strong nonlinear mapping ability, flexible network structure, high fault tolerance and robustness, GRNN is useful for solving nonlinear problems [26]. GRNN outperforms the radial basis function neural network in terms of approximation capability and learning rate. The network finally converges at an optimized regression surface with more sample accumulation, and the prediction effect is relatively good when the sample data are relatively small.

2.3.3. ANFIS Model

ANFIS is a new fuzzy inference system structure that forms an organic combination of fuzzy logics and neural networks. It adjusts the premise and conclusion parameters using a combination of the backpropagation algorithm and the least square method [27]. It can automatically generate ‘If-Then’ rules and gradually deploy the appropriate membership function to meet the desired fuzzy inference input/output relation [28].

2.3.4. Determination of the Optimal Prediction Model

To select the optimal prediction model, the root-mean-square error (RMSE) and coefficient of determination of the linear regression straight (R²) were used to evaluate the prediction performances of the models [18]. The formula for RMSE was determined by Niu et al. [12] and is shown in Equation (5).

$$\mathrm{RMSE}=\sqrt{{\displaystyle \sum _i^N\frac{{\left(K_E-K_P\right)}^2}{N}}}$$

(5)

where N denotes the total data size, K_P represents the predicted value of the model output and K_E denotes the measured value.

The lower the RMSE value and the higher the R² value, the higher the prediction accuracy of the models.

3. Results and Analysis

3.1. Variation Laws of Chlorophyll Content in The Pericarp of Fragrant Pears during the Storage Period

The variation laws of chlorophyll content in the pericarp of fragrant pears with different maturity levels at harvest during the storage period are shown in Figure 2. As shown in Figure 2, with the increase in maturity at harvest, the chlorophyll content in the pericarp of fragrant pears decreased gradually during H1–H10. The chlorophyll content in the pericarp of fragrant pears with different maturity levels at harvest declined continuously over time during storage. Subsequently, there were significant differences in chlorophyll content in the pericarp of high-maturity and low-maturity pears during the early storage process. The chlorophyll content of Korla fragrant pear fruits during H1–H5 decreased considerably faster than that during H6–H9. In the middle and late periods of the storage process, there were no significant differences in the decreasing rate of chlorophyll content in the pericarp between high-maturity and low-maturity pears. Low-maturity pears had higher chlorophyll content in the pericarp than high-maturity pears throughout the storage process. When Korla fragrant pears grow on trees, chlorophyll synthesis strengthens continuously in the first stage, from mitotic time to cell elongation. After reaching the peak, the chlorophyll content decreases quickly. At this point, fragrant pears enter their maturation period, which lasts until the fruits are completely ripe. Chlorophyll content gradually decreases after reaching the maturity stage, but carotenoid remains in colorful aging tissues and the mature pericarp after the disintegration of chlorophyll [29]. In appearance, fruit colors gradually change from green to yellow-green, with localized areas of dark red. With the extension of the storage time, the ability of the fruit to resist various stresses and scavenge oxygen free radicals gradually weakens, resulting in the degradation of pericarp tissue cell contents [30]. In addition, chloroplasts, which are important sources of cellular energy, require additional energy to function properly [31]. However, the cells themselves cannot provide additional energy, leading to chlorophyll metabolism enhancement and chloroplast degradation to maintain basic cellular life activities [32]. Consequently, chlorophyll in the pericarp continues to degrade, and pericarp colors shift from green to yellow.

3.2. Prediction Model for Chlorophyll Content in the Pericarp of Fragrant Pears

3.2.1. Prediction of Chlorophyll Content in the Pericarp Based on the BPNN Model

The maturity at harvest and storage time of fragrant pears were chosen as network inputs, while chlorophyll content in the pericarp of fragrant pears was used as the network output. Forty-five groups of datasets were obtained from the statistical experimental data. Among these, 70% of the test data were chosen randomly to serve as the training set. The cross-validation method was used in the training, and four rounds of training were carried out. The target RMSE was set to 0.0004, and the number of iterations was set to 1000. Since the initial weight of the BPNN was chosen randomly, each group of BPNNs was trained 100 times before storing the network with the best prediction effect. The BPNN parameters were introduced as follows: number of neurons in the input layer = 2; number of neurons in the hidden layer = 12; number of neurons in the output layer = 1; learning rate = 0.1. The remaining 30% of the data were used as a test set and input into the trained model, thus obtaining prediction values.

According to the linear fitting of the measured values and predicted values, it was found that the R² and RMSE of the BPNN model for predicting chlorophyll content in the pericarp of fragrant pears were 0.63 and 0.06, respectively. The fitting effect is shown in Figure 3. The measured values were poorly correlated with the predicted values, and the prediction error was relatively high. The trained BPNN model could not accurately predict chlorophyll content in the pericarp.

3.2.2. Prediction of Chlorophyll Content in the Pericarp Based on the GRNN Model

The network inputs were the maturity of the fragrant pears at harvest and their storage time, while the network output was chlorophyll content in the pericarp. The statistical experimental data yielded 45 dataset groups. Among these, 70% of the test data were chosen randomly to serve as the training set. The smooth factor (σ), the most important parameter of the GRNN model, was then determined using the cross-validation searching algorithm. Finally, the optimal value for predicting chlorophyll content in the pericarp was σ = 0.26. Under this circumstance, GRNN demonstrated the best prediction performance, with the smallest error. The remaining 30% of the data were used as the test set and input into the trained model, thus obtaining prediction values. The R² and RMSE of the GRNN model for predicting chlorophyll content in the pericarp of fragrant pears were 0.978 and 0.01, respectively, based on the linear fitting of the measured and predicted values. Figure 4 depicts the fitting effect. Thus, the trained GRNN model could effectively predict chlorophyll content in the pericarp.

3.2.3. Prediction of Chlorophyll Content in the Pericarp Based on the ANFIS Model

The maturity at harvest and storage time of fragrant pears were chosen as network inputs, while chlorophyll content in the pericarp of fragrant pears was used as the network output. Among the test data, 70% were chosen randomly to serve as the training set, while the remaining 30% of the test data were selected as the test set. In the model training stage, the initial ANFIS model was generated using the meshing method. Fuzzification of input data was performed by using eight types of input membership functions, including trimf, trapmf, gbellmf, gaussmf, gasuss2mf, pimf, dsigmf and psigmf. The fault tolerance was set to 0, and the number of iterations was set to 100.

After the training stage, the ANFIS model was tested on an independent dataset. The test set data were input into the trained ANFIS model, yielding the prediction values. For the model training and prediction stages, the correlation between the predicted values and the measured values is shown in Figure 5. The RMSE and R² of the measured values and predicted values for the training and test sets are listed in

Table 1. Root-mean-square error (RMSE) and determination coefficient (R²) for the prediction of chlorophyll content in the pericarp by the adaptive neural fuzzy inference system (ANFIS) model with different input membership functions (MFs) in the training and prediction stages.

Membership Function	Training Stage		Prediction Stage
	RMSE	R2	RMSE	R2
trimf	0.001	0.999	0.018	0.989
trapmf	0.002	0.999	0.008	0.989
gbellmf	0.001	0.999	0.010	0.980
gaussmf	0.001	0.999	0.015	0.959
gasuss2mf	0.001	0.999	0.006	0.993
pimf	0.002	0.999	0.022	0.905
dsigmf	0.002	0.999	0.007	0.992
psigmf	0.002	0.999	0.007	0.992

. Clearly, the R² of the measured values and predicted values in the prediction stage was higher than 0.905. This proves that the trained ANFIS model could effectively predict chlorophyll content in the pericarp of fragrant pears during storage. Moreover, it was found from the analysis that the prediction accuracy of the model was RMSE = 0.006 and R² = 0.993 when gasuss2mf was used as the input membership function; RMSE = 0.007 and R² = 0.992 when dsigmf was used as the input membership function; and RMSE = 0.007 and R² = 0.992 when psigmf was used as the input membership function. The prediction effects of these three membership functions were similar, and they were better than those of other membership functions. Hence, gasuss2mf, dsigmf and psigmf were the most suitable membership functions of ANFIS for predicting chlorophyll content in the pericarp.

3.2.4. Determining the Optimal Prediction Model

Predicting chlorophyll content in the pericarp of fragrant pears during the storage period based on a neural network found that the R² of the BPNN and GRNN was 0.63 and 0.978, respectively. Under eight input membership functions, the R² of the ANFIS model was higher than 0.905. As a result, as long as the maturity at harvest and storage time are known, both the trained GRNN and ANFIS models can predict chlorophyll content in the pericarp of fragrant pears during the storage period. Moreover, the prediction performances of the GRNN and ANFIS models were compared. We suggest using the ANFIS model with gasuss2mf, dsigmf or psigmf to predict chlorophyll content in the pericarp of fragrant pears, as this showed the best prediction performances.

3.2.5. Model Verification

In 2021, a refrigeration storage experiment was carried out on fragrant pears to verify the practical performance of the optimal prediction model. Fragrant pears at H1, H3, H5, H7 and H9 (maturity levels at harvest were defined as 1, 3, 5, 7 and 9) were chosen for refrigeration storage for 0, 60 and 120 days. At the end of storage, the chlorophyll content in the pericarp of fragrant pears was evaluated. Linear fitting of the values measured in the experiment and the predicted values of the ANFIS model with gasuss2mf, dsigmf and psigmf was carried out. The results are shown in Figure 6.

According to the verification experiment, the ANFIS model, which used gasuss2mf (RMSE = 0.009; R² = 0.989), dsigmf (RMSE = 0.009; R² = 0.988) and psigmf (RMSE = 0.009; R² = 0.988) as the input membership functions, achieved a relatively high prediction accuracy of chlorophyll content in the pericarp. This demonstrates that entering maturity at harvest and storage time into the trained optimal prediction model can enable the accurate prediction of chlorophyll content in the pericarp of fragrant pears during the storage period.

In this paper, the neural network modeling method was used to predict the chlorophyll content in the pericarp of Korla fragrant pears during storage, and the optimal prediction model for chlorophyll content in the pericarp was constructed. The ANFIS model showed the best prediction performance when the input membership functions were gasuss2mf, dsigmf and psigmf. The refrigeration storage experiment was repeated the following year to verify the practical performance of the optimal prediction model. Korla fragrant pears from the same pear orchard were chosen as the research objects. The results demonstrate that the optimal prediction model can predict chlorophyll content in the pericarp of fragrant pears from the same pear orchard during storage. However, due to differences in environmental climate and cultivation conditions between different orchards, the growth and post-ripening characteristics of fragrant pears will be affected. Therefore, the applicability of the prediction model to other fragrant pear orchards still needs to be further verified. Nevertheless, this study demonstrated that maturity at harvest and storage time are important factors influencing fruit quality during storage, and the effect of post-ripening characteristics on storage quality was revealed. The findings of this study can provide methodological guidance for the accurate prediction of storage quality indices for other fruits and serve as references for determining the delivery quality and timing of other fruits.

4. Conclusions

This study concludes that the chlorophyll content in the pericarp of fragrant pears decreased over time during the harvest period. The chlorophyll content in the pericarp of fragrant pears with different maturity levels at harvest decreased continuously during storage. The GRNN and ANFIS models could predict chlorophyll content in the pericarp of fragrant pears. The ANFIS model performed best when the input membership functions were gasuss2mf (RMSE = 0.006; R² = 0.993), dsigmf (RMSE = 0.007; R² = 0.992) and psigmf (RMSE = 0.007; R² = 0.992).