Using artificial neural network in predicting the key fruit quality of loquat

Abstract The formation and regulation of loquat fruit quality have always been an important research field to improve fruit quality, commodities, and market value. Fruit size, soluble solids content, and titratable acid content represent the most important quality factors in loquat. Mineral nutrients in abundance or deficiency are among the most important key factor that affect fruit quality. In the present study, we use artificial neural network (ANN) to explore the effects of mineral nutrients in soil and leaves on the key fruit quality of loquat. The results show that the ANN model with the structure of 12–12–1 can predict the single fruit weight with the highest accuracy (R 2 = .91), the ANN model with the structure of 10–11–1 can predict the soluble solid content with the highest accuracy (R 2 = .91), and the ANN model with the structure of 9–10–1 can predict the titratable acid content with the highest accuracy (R 2 = .95). Meanwhile, we also conduct sensitivity analysis to analyze the relative contribution of mineral nutrients in soils and leaves to determine of the key fruit quality. In terms of relative contribution, Ca, Fe, and Mg content in soils and Zn, K, and Ca content in leaves contribute relatively largely to a single fruit weight, Mn, Fe, and Mg content in soils and the N content in leaves contribute relatively largely to the soluble solid content, and the P, Ca, N, Mg, and Fe in leaves contribute relatively largely to the titratable acid content of loquat. The established artificial neural network prediction models can improve the quality of loquat fruit by optimizing the content of mineral elements in soils and leaves.

its dual-purpose as medicine and food (Liu et al., 2016). In recent years, loquat cultivation has become more commercial, mainly because of the price advantage of loquat (Liu et al., 2016). During the hot summer, fresh fruit is scarce in the market, so that loquat may be recognized as the most promising fruit (Chalak et al., 2014).
Fruit quality (flavor, nutrition, taste, etc.) directly affects grading, the commodity value of fruit, fruit size represents one of the most important quality factors in loquat (Agustí et al., 2003). The soluble solids content and titratable acid content are the main factors affecting fruit quality, and the balance between them is important to the taste and quality of fruit . Adequate mineral nutrition is an important factor for normal growth and development of fruit trees, that ensures high-quality fruit with maximum yield (Amiri et al., 2008;Kader, 2008). The leaf is a fundamental source for supplementation of nutrition and mineral elements for fruit, and tree growth therefore directly affect fruit formation and quality (Iglesias et al., 2007;Morgan et al., 2005). The mineral element content in leaves is directly affected by the mineral element content in the soil. Therefore, to study the correlation between mineral element contents in leaves and soil, and its effect on fruit quality, this study greatly improves yield, fertilizer efficiency, and balanced fertilization.
ANN is an operational model similar to the neural network of the human brain for information processing (Saffari et al., 2009). It is mainly composed of three parts: the input layer, hidden layer, and output layer (Tracey et al., 2010). Many neurons are connected between different layers. The structure and function of each neuron are not complicated, and the strength of the connection between each neuron has a certain plasticity (Besalatpour et al., 2013). It has strong a self-learning ability. It can process existing data internally through different functions, adjust weights for different environments, optimize network systems behavior, analyze hidden relationships between data, and thus reveal unknown data, which is of special significance for prediction. Its powerful computing power allows it to process more data, save more time, and solve many very complex nonlinear problems. The origin and variety of loquat can be analyzed qualitatively by combining near infrared and artificial neural network model (Fu et al., 2007), and the extraction rate of ursolic acid from loquat leaves was optimized based on neural network model (Sun et al., 2010). Meanwhile, many previous studies (Gholipoor & Nadali, 2019;Niedbała, 2019;Torkashvand et al., 2017Torkashvand et al., , 2019 indicated that the artificial neural network model is a very effective and reliable forecasting tool, which has been widely used in agriculture and has a very high prediction accuracy.
The present research was aimed to detect the reliability of ANN models to predict the single fruit weight, soluble solids content, and titratable acid content of loquat and to explore the effects of mineral elements in soil and leaves on the key fruit quality of loquat, so as to provide theoretical basis for nutritional diagnosis and scientific fertilization of loquat in practical production.

| Experimental site and samples collection
The experimental study was performed in loquat orchards, in Wuzhong district, Suzhou, Jiangsu Province of China, which is located at 31°02'-31°06'N, 120°20'-120°25'E ( Figure 1). We selected 150 plots of "Baiyu" loquat cultivar at the same growth stage regarding age and growing conditions.
In each plot, 20-40 fresh fruits were randomly collected from 8 trees with the same maturity time and size. At the same time, two leaves were taken from all four directions, east, south, west, and north of the tree, and a total of 40-60 leaves were collected.

| Sample treatment
The leaf samples were washed with 0.1% detergent solution for 30 s, rinsed with tap water, followed by ddH 2 O three times, and suspended to dry place in the nylon net bag. Samples were oven-dried at 105°C for 30 min and then dried at 75°C for constant weight and lapping, and these were then screened by a 100-mesh sieve and stored in plastic bottles for determination of mineral elements in leaf and fruit (Xu et al., 2012). The soil samples are left to dry naturally in a ventilated area until it obtained a constant weight and then screened by a 100-mesh sieve used to determine mineral element contents.

| Sample digestion
0.200 g leaf sample and 5 ml of HNO 3 were added to glassware for digestion, and 0.200 g soil sample, 1.5 ml HNO 3 , and 4.5 ml HCl were added to glassware for digestion, respectively. After digestion and cooling, dilute the liquid mixture to 100 ml with ultrapure water for mineral elements determination of P, K, Ca, Mg, Fe, Mn, Cu, and Zn in leaf and soil samples (Huang et al., 2018). 0.100 g leaf sample, 5 ml of H 2 SO 4, and a few drops of H 2 O 2 were added to glassware for digestion, and 0.100 g soil sample, 5 ml of H 2 SO 4, and 1.000 g mixed catalyst [m (K 2 SO 4 ):m (CuSO 4 ) =10:1] were added to glassware for digestion, respectively. After digestion and cooling, dilute the liquid mixture to 100 ml with ultrapure water for mineral element determination of N in leaf and soil samples (Ying-Li et al., 2006).

| Mineral element and fruit quality determination
The mineral elements of P, K, Ca, Mg, Fe, Mn, Cu, and Zn in soil and leaf samples were determined by inductively coupled plasma-optical emission spectroscopy (ICP-OES) (American Agilent Corporation) (L. Li et al., 2018). The N element in leaf and soil samples was determined by the flow analyzer (Shenzhen Yicheng Science and Technology Co., Ltd) (Ying- Li et al., 2006).
The single fruit weight was determined by electronic analytical balance (Mettler Toledo instrument Co., Ltd., precision 0.0001 g), the content of soluble solids was determined by pal-1 portable digital display sugar meter (Atago Ajon Company of Japan), and titratable acid content was determined by indicator titration (Jiangkang, 2007).
Some statistical properties of the data are presented in Table 1.

| Statistical analysis
We used SPSS 22 (Cary, NC., USA) to calculate the simple correlation coefficient between single fruit weight, soluble solids content, titratable acid content, and mineral elements in leaves and soils. We used mineral element values in leaves and soils as input layer (independent variables), single fruit weight, soluble solid content, and titratable acid content of loquat as output layer (dependent variables) ( Figure 2). To make more effective use of the neural network, we normalize the two-layer data. In this study, we used 70% of the data randomly to train the ANN model, 15% of the data were used to verify the ANN model, and the remaining 15% of the data were used to test the ANN model. The following equation was used to normalize the data (Shabani et al., 2017): In the equation, T is the original data, T n is the normalized output or input values, and T min and T max are the minimum and maximum values of the related variable. We used MATLAB software (2015

version) to apply and test various training functions, as shown in
Tables 3-5. The equations for the transfer function are shown below.
•Tangent-sigmoid function: •Log-sigmoid function: We tested different training functions, transfer functions, and hidden layers to find the final model and evaluated the performance of the ANN models by determining the coefficient (R 2 ), root mean square error (RMSE), relative standard error (RSE), mean absolute error (MAE), and mean absolute percentage error (MAPE). The equations were as follow.
In the equations, n is number of data, and the bar is the mean of a variable, and M i and P i represent the measured value and the predicted value, respectively. (1) Structure of multilayer perceptron artificial neural networks (ANNs) to predict single fruit weight, soluble solids content, and titratable acid content of loquat

| Simple correlation analysis and input variables selection
The simple correlation analysis results between single fruit weight, soluble solid content, titratable acid content, and mineral elements in leaves and soils are shown in Figure   structure is 12-12-1 (Table 3) The results indicate this ANN model is reliable and effective in predicting the single fruit weight of loquat. and RMSE (0.0.08116) and the highest R 2 value (.91), and the neural network structure is 10-11-1 (Table 4)

| DISCUSS IONS
The results of this study showed that there were complex correlations between the mineral element content (the variable in the input layer) and loquat per fruit weight, soluble solids content, and titratable acid content (the variable in the output layer) (Agusti et al., 2010;Huang et al., 2018Huang et al., , 2019Yu et al., 2018), respectively. Therefore, it is very important to establish a model to predict the key fruit quality of loquat and reveal its internal relation to improve the quality of loquat. However, traditional mathematical methods and modeling techniques cannot accurately quantify the complex internal relations mentioned above. At present, this difficulty can be solved by another ANN modeling technique (Rohani et al., 2011), which can capture the association between some data objects and find their internal laws by self-learning some data objects. In recent years, the ANN has been tested in predicting the accuracy of fruit quality, including the soluble solid content in pineapple (Chia et al., 2012) and Korla fragrant pears (Lan et al., 2020); aroma in strawberry, lemon, cherry, and melon (Adak & Yumusak, 2016); titratable acid, soluble solid, and vitamin C in Nanfeng mandarin fruit (Xudong et al., 2009).
The results indicate that the ANN model is a very useful and reliable predictive tool, and the high prediction accuracy is very satisfactory.
To explore the most effective prediction model, we compare the It is usually used in multilayer networks trained by back-propagation algorithm, mainly because this function easy to differentiate and is differentiable. It has excellent mathematical properties, and many numerical optimization algorithms can be directly used to obtain the optimal solution. LM is a network training function that can provide F I G U R E 7 Sensitivity analysis of the governing traits on the single fruit weight (a), soluble solids content (b), and titratable acid content (c) of loquat numerical solutions that minimize nonlinearity. By modifying parameters during execution, this algorithm can realize the advantages of both gradient descent method and Gauss-Newton algorithm and improve the shortcomings of both algorithms (Burney et al., 2005;Kumalasari et al., 2018). It is usually the fastest back-propagation algorithm in the network toolbox, and although it is memory-intensive in computation, it is still the most widely used algorithm. The LM algorithm has a second-order convergence speed and fewer iterations, which can greatly improve the convergence speed and stability of the algorithm (Haykin & Network, 2004).
The size, soluble solids content, and titratable acid content of fruit play important roles in the external quality and edible flavor of fruit (Melke & Fetene, 2014), and the mineral elements, especially in soil and leaves, have important influences on fruit growth, fruit formation, and fruit quality control (Aular et al., 2017;Nestby et al., 2005). Soil is an important factor for material and energy exchange in the ecosystem, soil nutrients elements are closely related to the growth, development of fruit trees yields, and fruit quality and leaves are the main organs that plants use for photosynthesis, production of nutrients, and directly supply mineral nutrition for tree growth and fruit (Velemis et al., 1999).
The present study showed that the Ca, Fe, and Mg content in soils and the Zn, K, and Ca content in leaves have a relatively large influence on the single fruit weight of loquat. In this regard, Ca affects pollination and fruit formation in fruit trees (Freitas & Mitcham, 2012). Torkashvand (Torkashvand et al., 2020) showed a significant correlation between Ca concentration in leaves and fruit yield. Egilla (Egilla et al., 2005) assumed that the increase of potassium would lead to the increase of photosynthesis, and the yield and dry matter content will increase accordingly, and a lack of Fe content leads to a significant reduction in the fresh weight and number of fruit per tree (Álvarez-Fernández et al., 2003). The soluble solid content is mainly affected by the Mn, Fe, Mg, and N content in soils and the leaf N content, and the titratable acid content is mainly affected by the mineral nutrients in the leaves, and the P, Ca, N, Mg, and Fe content in the leaves had the most significant impact on it.
The Mn, Mg, and Fe content play so important roles in plant chlorophyll synthesis and photosynthesis. Mn participates in water photolysis and electron transfer in photosynthesis. In the absence of Mn, chloroplasts can only produce a small amount of oxygen, and photosynthetic phosphorylation is weakened, and the synthesis of sugar and cellulose is also reduced (Lu, 2003). Mg absorbed by plants from the soil can activate ribulose diphosphate carboxylase and promote CO 2 assimilation, which is beneficial for the synthesis, accumulation of sugars and starch in fruits (Lu, 2003). The content of Fe is the best activator of sucrose phosphate synthase, and Fe deficiency in plants can reduce sucrose synthesis in the body. The P content is an essential component of plant nucleus and various plasma membranes, which can promote cell division, improve stress resistance and adaptability of fruit trees, and increase fruit yield; however, if the content of P in fruit is too high, the respiration is enhanced, and the consumption of carbohydrates and energy is large, which has adverse effects on fruit development (Huang et al., 2018).

| CON CLUS ION
In this study, we use the LM training function and Log-sigmoid transfer functions to build three effective and reliable artificial neural networks that can predict the key fruit quality of loquat by the mineral content in leaves and soils. Among them, when the structure of the single fruit weight prediction model is 12-12-1, it reaches the highest accuracy (R 2 = .91); when the structure of soluble solid content prediction model is 10-11-1, it has the highest accuracy (R 2 = .91); when the structure of titratable acid content prediction model is 9-10-1, it achieves the highest accuracy (R 2 = .95). We conduct sensitivity test analysis, and the results show that the content of Ca, Fe, and Mg in soils and K, Zn, and Ca

CO N FLI C T O F I NTE R E S T
The authors declare that they have no conflict of interest.

E TH I C A L A PPROVA L
This article does not contain any studies with animal or human subject.