Recent Advances in the Modeling and Predicting Quality Parameters of Fruits and Vegetables during Postharvest Storage: A Review

ABSTRACT

Artificial neural network (ANN), genetic algorithm (GA), fuzzy logic (FL), and adaptive neurofuzzy inference system (ANFIS) have been applied in every aspect of food science in the recent years. These models are useful tools for fruit and vegetable monitoring; grading and classification; modeling the respiration rate; predicting and modeling quality properties; modeling of microbial growth; and forecasting chemical, physical, and sensorial characteristics during processing and postharvest storage. These models hold an enormous deal of promise for modeling difficult task;s in practice control and simulation and in the use of machine perception including machine vision system and electronic nose for fruit and vegetable quality control. In addition, these models were used for different fruit and vegetable storage process modeling, for detecting chilling injury, to detect defects, for controlling various drying process, and for improving climate control. The present study reviews the efficiency and applications of ANN, GA, FL, and ANFIS models to predict and control the quality parameters of various fruits and vegetables during postharvest storage.

KEYWORDS:

• Adaptive neurofuzzy inference system
• artificial neural network
• fuzzy logic
• genetic algorithm

Introduction

Postharvest storage is one of the several methods that follow after harvesting period in an agricultural food-producing process. One way of guaranteeing food security is through the enhancement of postharvest storage procedures. Under storage, postharvest products produce heat, moisture, CO₂, and ethylene gases. The physicochemical changes of fruits and vegetables following postharvest treatments are a highly nonlinear one involving respiration rate and quality loss.

Artificial neural network (ANN), genetic algorithm (GA), fuzzy logic (FL) and adaptive neurofuzzy inference system (ANFIS) are analytical alternative to usual modeling methods, which are often limited by strict assumptions of linearity, normality, homogeneity, and variable independence (Bahramparvar et al., 2014; Ramzi et al., 2015; Salehi and Razavi, 2016b). These models are useful tools for fruit and vegetable monitoring; grading and classification; modeling the respiration rate; predicting and modeling quality properties; modeling of microbial growth; and predicting physical, chemical, and sensorial characteristics during processing and postharvest storage (Table 1). ANN, GA, FL and ANFIS models were used for detecting chilling injury in red delicious apple (ElMasry et al., 2009); for potato storage process modeling (Abdulquadri Oluwo et al., 2013); for improving the climate control for stored potato (Gottschalk et al., 2003); for selection of best wavelength features for decay detection in citrus fruit (Lorente et al., 2013); to sense defects such as bruises, russet, bitter pit, puncture and leaf roller in apples (Kavdır and Guyer, 2004); for respiration rate modeling of guava (Wang et al., 2009); for predicting moisture content of grain (Liu et al., 2007); for controlling drying process of olive stones (Javadikia et al., 2011); for controlling in microwave-based Chinese herbs drying systems (Lu et al., 2006); for grain drying (Mansor et al., 2010); and for developing a dynamic model for the rotary drying plant (Areed et al., 2012). May et al. (2011) and Wali et al. (2013) used FL to decrease the operational times and cooling energy generation of the chilling system. The FL controller tracked the reactor desired temperature precisely with minimal overshoot and a fast warmup phase. Trouble in the form of changing flow rate in the process input was well rejected by the FL controller.

Table 1. ANN model parameters to modeling quality properties of fruits and vegetables.

Fruit or vegetable	Model inputs	Predicted parameters	Number of hidden layers	Number of neurons in hidden layer	Activation function	Main results	References
Apple	The reflectance features at the selected optimal wavelengths	Firmness	1	3	Sigmoid and linear	Their experimental results confirmed that a spectral imaging system linked with ANN can successfully distinguish between chilling-injured apples and normal apples, in addition to sense firmness changes.	(ElMasry et al., 2009)
Apricot	Osmotic solution temperature and concentration, and dehydration time	Weight and water loss and solid gain	1	14	Hyperbolic tangent	Sensitivity analysis results by optimum ANN showed the osmotic solution temperature was the most sensitive factor for controlling the weight and water loss and solid gain.	(Salehi et al., 2015)
Avocado	Coated and uncoated avocado fruit and temperature	Quality parameters of avocado during storage	2	5 + 5	Hyperbolic tangent	The ANN models developed could be successfully used for predicting respiration rate, weight loss, firmness, and total color difference.	(Maftoonazad et al., 2011)
Button mushroom	Calcium ascorbate concentration and shelf-life time	Weight loss, firmness, TSS, pH, L*, a*, b*, chroma, Hue angle, ΔE, browning index, vitamin C, total phenol and polyphenol oxidase activity	1	8	Sigmoid	Results of sensitivity analysis by optimum neural network (2-8-14) defined shelf-life time as the most effective factor in predicting button mushroom attributes during postharvest storage.	(Sayyari et al., 2015)
Button mushroom	Air temperature, flow rate and drying time	Weight loss	1	16	Hyperbolic tangent	The modeling results showed the GA-ANN can well predict the weight loss in button mushroom drying by hot air (R = 0.999).	(Salehi et al., 2017)
Grain	Grain temperature, initial moisture content, and air temperature	Moisture content	1	6	Sigmoid	The ANN optimized using GA performed well and the accuracy of the predicted values was excellent.	(Liu et al., 2007)
Guava	Temperature, CO2 and O2 concentrations	Respiration rate	1	9	Logsig	Their results showed that the ANN approach is an high precise method, and it can be used for predicting and modeling the respiration rate of fruits.	(Wang et al., 2009)
Kiwifruit	Temperature and concentration of osmotic solution, and time	Color changes, water loss and solid gain	1	16	Hyperbolic tangent	The ANN could potentially be applied to estimation color changes and mass transfer kinetics of dehydrated kiwifruit.	(Fathi et al., 2011)
Peach	Wavelengths	Normal and cold damaged peaches	1	3	Hyperbolic tangent	The prediction of quality parameters of peach by using ANN model achieved well, with the r for prediction set of 0.698–0.903.	(Pan et al., 2016)
Pomegranate	Methyl jasmonate and storage time	Chilling injury index, weight loss, ion leakage, pH, ethylene, anthocyanins, respiration, total antioxidant activity and polyphenols	1	14	Sigmoid	The total agreement between GA-ANN predictions and experimental data was considerable (r = 0.87).	(Sayyari et al., 2017)
Potato	Ambient air temperature and relative humidity and inlet air temperature	Potato temperature	1	3	Log-sigmoid and	The suggested model was useful in simulation processes involving intelligent controllers.	(Abdulquadri Oluwo et al., 2013)
Tomato	Drought stress and chilling stress effects	Chlorophyll a, chlorophyll b, total phenol, relative water content, root electrolyte lekage, F0, Fm and proline	1	7	Hyperbolic tangent	According to sensitivity analysis results by optimum neural network, severity of PEG-induced drought stress was effective factor in predicting chilling resistance and growth parameters of tomato seedlings.	(Ghanbari et al., 2018)

As a computer-aided possibility, these models display the most excellent performance at various conditions by comparing with the conventional mathematical regression models that were created on the theories of chemical and enzyme kinetics. So, the present study summarizes the efficiency and applications of ANN, GA, FL, and ANFIS models to predict and model quality parameters of various fruits and vegetables during postharvest storage.

Artificial Neural Networks

ANN is an analytical tool based on the structure of biological neurons, and it has been widely used in numerous fields, including postharvest quality controlling of fruits and vegetables and food industry. Contrasting other analytical methods, where prior knowledge of relationships among process parameters is required, ANN draws on previously gathered information and uses this when analyzing new data input. It is particularly useful in managing uncertainties and nonlinear data relationships. In the area of fruit and vegetable quality control, ANN has been successfully applied to predict the quality of agricultural raw and stored material. ANN has proven to be very successful in identification, grading, and classification of fruits and vegetables grown or stored in various conditions, where noncoherence or nonlinearity often exists (Fan et al., 2013; Ghazanfari et al., 1996; Jayas et al., 2000; Kim et al., 2000; Lorente et al., 2013; Sayyari et al., 2017; Zhang et al., 2014). For example, Sreekanth et al. (1998) predicted psychometric parameters using different ANN models. Kaminski et al. (1998) used an ANN for data smoothing and for predicting material moisture content and temperature.

Estimation of shelf-life and quality of fruits and vegetables with an electronic nose system is a very active area of research. The ANN technologies for pattern detection may make electronic nose applications more powerful for fruit quality analysis (Garcıa-Gonzalez and Aparicio, 2003; Huang et al., 2007).

In another study, neural network models were used to predict the shelf-life of greenhouse lettuce by Lin and Block (2009). Using two-stage ANN models, an R² of 0.61 could be achieved for predicting the remaining shelf-life. This study indicated that neural network modeling has potential for cold chain quality control and shelf-life prediction.

Genetic Algorithm

The performance of an ANN model depends strongly upon its structure. Determination of the optimal number of neurons in the hidden layer is usually performed by trial and error method. GA optimization technique can be used to overcome this inherent limitation of ANN (Figure 1). GAs are search techniques for an optimal value, mimicking the mechanism of biological evolution (Ramzi et al., 2015; Salehi et al., 2015; Sayyari et al., 2017). Liu et al. (2007) optimized the ANN topology for predicting the moisture content of grain during drying process using GA. The GA was used for selecting the suitable network architecture in determining the optimal number of nodes in the hidden layer of the neural network. The number of neurons in the hidden layer was optimized for six backpropagation neurons and 10 radial basis function neurons using GA. Modeling test on the moisture content prediction of grain drying process showed that the structural modular neural network optimized using GA performed well and the accuracy of the predicted values is excellent.

Figure 1. Genetic algorithm–artificial neural network model for prediction of calcium ascorbate effects on the button mushroom.

Fuzzy Logic

Fuzzy sets provide mathematical methods that can characterize the uncertainty of human expression. The FL approach can model complex nonlinear behavior of some systems under study (Davidson and Sun, 1998; Johansen and Babuska, 2003; Lee and Kwon, 2007; Moghaddam et al., 2011).

FL was applied for sensory evaluation by panel test as well as the food process and sensory-related quality control. The corresponding responses to sensory attributes, such as appearance, taste, and firmness, were transformed to the fuzzy sets and then manipulated according to fuzzy mathematics (Moghaddam et al., 2011). Lu et al. (2006) applied an FL controller in microwave-based Chinese herbs drying equipment. Good results were obtained from matlab software simulations in the FL Toolbox. May et al. (2011) developed two FL controllers to decrease operating times and cooling energy generation for air-conditioning purposes of some buildings. Simulation results showed promising results in achieving optimal operations of the chilling system. Mansor et al. (2010) designed and applied the FL control method for grain drying. Simulation results obtained proved to be good in comparison with those obtained in literature in the areas of settling time and steady-state error. Gómez-Melendez and López-Lambraño (2011) developed an FL greenhouse fertigation control system based on a field-programmable gate array. Their results from simulation and experiments showed the simplicity of the design, the viability of its implementation, and the low cost that the use of a fuzzy logic controller and  field programmable gate array system represents.

Adaptive Neurofuzzy Inference System (ANFIS)

Fuzzy inference systems (FISs) and ANNs are model-free numerical estimators. They share the ability to develop the predictive capability of a system working in imprecise, uncertain, and noisy environments. FIS and ANN may be combined into an integrated system described as ANFIS; the combined system then has the benefits of both ANN (learning and optimization abilities and connectionist organization) and FIS (humanlike if-then rules, and ease of incorporating expert knowledge accessible in linguistic terms). The ANFIS model has the unique benefit that no clear relationship between the input and output variables needs to exist before the model is applied since the relationship is recognized through a self-learning procedure (Bahram-Parvar et al., 2017; Becker and Karri, 2010; Ramzi et al., 2015; Salehi and Razavi, 2016a; Sayyari et al., 2017).

Several efforts have been made to apply intelligent control to nonlinear procedures. Takagi et al. (1990) developed a technique for tuning the fuzzy control rules automatically, using neural networks. In this method, two networks were used where one of the networks classified the present control performance while the other simulated control performances against combinations of fuzzy labels in the control rules. Karr and Gentry (1993) developed an adaptive fuzzy controller that altered membership functions (MFs) optimally using GA. It was applied to the pH control of a solution. Morimoto et al. (1997) jointed some intelligent methods for optimization of the storage process. ANN was used to identify the relationship between the relative humidity and ventilation and GA was used to determine the MFs and control rules efficiently during storage. GA due to its iterative nature affects the controller response. Morimoto Kang and Hashimoto (1999) used two decision systems consisting of both GA and ANN to recognize and optimize the storage process. The ANN identified the fruit responses as affected by the relative humidity and the GA selected the optimal values of the MFs and control rules. In both the cases, the controller adjusted only the storage relative humidity using on–off control of the dampers and temperature was not controlled. Kiralakis and Tsourveloudis (2005) compared FL and ANFIS controller for monitoring of olive stones drying. They concluded that in terms of stability and set-point tracking the ANFIS performed better than the FL controller, but the fuzzy did better at higher initial moisture content.

FG Areed et al. (2012) developed a dynamic model for the rotary drying plant and an ANFIS controller for the drying process and compared it with an FL and proportional integral derivative (PID) controllers. Modeling results proved that the ANFIS controller yielded the best dynamic performance followed by the FL controller, in terms of rising time, settling time, maximum overshoot, and steady-state error.

Orange

The advantages of machine vision systems (MVSs) in the food industry area are to accomplish tasks related to visual quality control and/or remote sensing to replace human inspectors in an adverse environment, reduce inspection errors and/or increase efficiency. ANN is one of the best tools for the pattern recognition of MVS purposes (Guyer and Yang, 2000; Huang et al., 2007; Kondo et al., 2000). Kanali et al. (1998) sorted eggplants and oranges through 3D-shape recognition. Primary image features of eggplants and oranges were acquired through an MVS chiefly consisting of neural retina and data transfer and conversion units. Image features were compressed with a charge-simulation method and then used as input vectors for an ANN. Overall classification rate was 74–94% depending upon the extent of the sample shape difference.

MVS-based quality evaluation of orange using ANN was studied by Kondo et al. (2000). They predicted the sugar and acid content of orange from images acquired. Red, Green and Blue (RGB) values and shape features were extracted from the images, and then employed as input vectors for an ANN trained with a Kalman filter learning model. While the image features did not show a high correlation with sugar content or pH, the study reported the feasibility of using this technique to evaluate the quality of orange fruits.

Cherry

Guyer and Yang (2000) used the GA-ANN and spectral imaging (680–1280 nm) for defect detection on cherries. An improved near-infrared (NIR) range vidicon black and white camera (400–2,000 nm) was used to acquire cherry images. Gray values from 16 wavelengths at each pixel were used as input vectors for an enhanced genetic ANN to classify cherries. The average prediction accuracy was 73% for identification and quantification of all types of cherry defects. The total approach of combining spectral information and GA-ANN classification methods with imaging capabilities has the promising potential for improving both the accuracy and the efficiency of automated quality detection of vegetables and fruits.

Pomegranate

Pomegranate is a typical fruit from many subtropical and tropical countries. Pomegranate arils contain high concentration of organic acids, sugars, polysaccharides, vitamins, essential minerals, and polyphenol and have antioxidant properties (Al-Maiman and Ahmad, 2002).

ANFIS and genetic algorithm–artificial neural network (GA-ANN) models (Figure 2) were used to predict the effect of storage time (0, 14, 28, 42, 56, 70, and 84 days) and methyl jasmonate concentration (0, 0.01, and 0.1 mM) on physiological changes and quality parameters of pomegranate fruits during storage by Sayyari et al. (2017). Their results showed that GA-ANN predictions agreed with experimental data and the GA-ANN with 14 neurons in one hidden layer can predict physiological changes and quality parameters of pomegranate (chilling injury index, weight loss, ion leakage, pH, ethylene, anthocyanins, respiration, polyphenols, and total antioxidant activity) with correlation coefficients equal to 0.87. The ANFIS model with three Gaussian type MFs for input variables (methyl jasmonate and time), 9 rules and linear for output gives the best fitting with the experimental data that predict quality parameters with a correlation coefficient equal to 0.92. However, the ANFIS model performs was better than GA-ANN model and this method was suggested to relevant postharvest storage projects with acceptable results.

Figure 2. Artificial neural network model for prediction of quality parameters of pomegranate fruit.

Apple

Different ANN methods were investigated for bruise prediction in apple, peach, and pear (Barreiro et al., 1997; Kavdır and Guyer, 2004). An ANN with the textural features extracted from spatial distribution of color/gray levels was used by Kavdır and Guyer (2004) to detect defects (bitter pit, leaf roller, puncture, russet, and bruises) in apple. In another study, the ANN model was used by ElMasry et al. (2009) to examine the ability of hyperspectral imaging (HI) and ANN methods for the finding of chilling injury in red apples. Their experimental results confirmed that a spectral imaging system linked with ANN can successfully distinguish between chilling-injured apples and normal apples, in addition to sense firmness changes. They reported that classification accuracy of above 90% was obtained with the application of selected five optimal wavelengths.

Peach

Peaches in cold storage may develop chill damage, as symptomized by deteriorated texture and lack of juice. To study peach quality, Pan et al. (2016) established a HI system to detect cold injury, and an ANN model was developed for which eight optimal wavelengths were selected. Between normal and chilling damaged peaches, significant differences in peach quality parameters (firmness, soluble solid content, extractable juice, chlorophyll content, and titratable acidity) and the spectral response to correlating selected wavelengths were observed. Evidencing this relationship, the correlation coefficients (r) between quality parameters and the respective spectral response of eight selected wavelengths were −0.59 to −0.70, 0.39 to 0.55, 0.51 to 0.75, and 0.57 to 0.77. With optimal representative wavelengths as inputs for the ANN model, the overall classification accuracy of chilling damage was 96% for all cold-stored peaches. The ANN prediction models for quality parameters performed well, with r = 0.698 to r = 0.903.

Avocado

ANN and HI methods were used to model quality changes in avocados during storage at different temperatures by Maftoonazad et al. (2011). ANN models were used in two ways to develop models for predicting quality parameters during storage. The optimal configuration of the neural network model was obtained by varying the different model parameters. Results showed the ANN models to be accurate and versatile, and they predicted the quality changes in avocado fruits better than the usual regression models; furthermore, the storage time–temperature-based ANN models were better than the hyperspectra-based ANN models.

Button Mushroom

ANN was used to model the effect of calcium ascorbate on extending button mushroom shelf-life by Sayyari et al. (2015). In order to predict the calcium ascorbate effects on button mushroom shelf-life, multilayer perceptron neural network with two inputs (calcium ascorbate concentration and shelf-life time) and 14 outputs (weight loss, firmness, total soluble solid (TSS), pH, L*, a*, b*, chroma, Hue angle, ΔE, browning index, vitamin C, total phenol, and polyphenol oxidase activity) was used by researcher. Their results showed that an ANN with eight neurons in a hidden layer and using sigmoid function and levenberg–marquardt optimization technique and 40%-20%-40% data for training/testing/validating process can well predict the effect of calcium ascorbate on button mushroom shelf-life with correlation coefficient equal to 0.91.Sensitivity analysis results showed that the shelf-life time was the most sensitive factor for prediction of  button mushroom attributes during postharvest storage.

Tomato

Low temperatures lead to numerous physiological disturbances in the cells of chilling-sensitive plants and result in chilling injuries and death of tropical and subtropical plants (Fagundes et al., 2014; Ghanbari et al., 2018). ANN modeling was used for predicting chilling resistance of tomato seedlings following imposing drought stress pretreatment with application 0%, 10%, and 20% polyethylene glycol (PEG) by Ghanbari et al. (2018). In order to predict the chilling effects on tomato seedling attributes, multilayer perception neural network with two inputs (drought stress and chilling stress effects) and eight outputs (chlorophyll a, chlorophyll b, total phenol, relative water content, root electrolyte lekage, F₀, F_m, and proline) was used. Their results showed that a network with seven neurons in a hidden layer and using hyperbolic tangent function and Levenberg–Marquardt optimization technique and 40%-20%-40% data for training/testing/validating process can well predict drought stress effects on chilling resistance of tomato seedlings with high correlation coefficient (r = 0.92). According to sensitivity analysis results by an optimum neural network, the severity of PEG-induced drought stress was an effective factor in predicting chilling resistance and growth parameters of tomato seedlings. Hahn et al. (2004) analyzed visible and NIR spectra with an ANN for finding of fungal rots in tomato fruits. They reported that 96% of the polluted tomato were correctly detected.

Potato

Abdulquadri Oluwo et al. (2013) used ANN optimization model for a potato storage system. The optimum model had a mean-squared error value of 0.831 and a coefficient of determination (R²) value of 0.735. The ANN was based on the min-max method of normalization, and the network provided a better representation of the storage process. The suggested model was useful in simulation processes involving intelligent controllers. In another study, Gottschalk et al. (2003) improved the climate control for potato during storage using an FL controller supported by GA. The GA was used to fit some parameters to the criteria to minimize the total storing cost.

Conclusion

The postharvest storage procedure is a highly nonlinear one involving mass and heat transfer. ANN, GA, FL, and ANFIS are analytical alternatives to usual modeling methods, which are frequently limited by strict assumptions of normality, linearity, homogeneity, and variable independence. These models do not require the previous knowledge of the relationship between the input and output variables because they can discover the relationship through successive training. Moreover, these models can predict several output variables at the same time, which is difficult in general regression methods. The ANN, GA, FL, and ANFIS models were used for detecting chilling injury; for storage process modeling; for improving the climate control for stored products; for selection of best wavelength features for slowdown detection in fruits and vegetables; to detect defects such as bruises, russet, bitter pit, puncture, and leaf roller; for respiration rate modeling; for predicting moisture content; for controlling drying process; and for detecting chilling injury. In addition, these models are suitable for the classification of fruit and vegetable quality for different research objectives and are suitable to be considered for evaluating the cold injury. In summary, the ANFIS model performs better than other models, and this method can be applied to relevant postharvest storage projects with satisfactory results.