Application of a multilayer perceptron artificial neural network for identification of peach cultivars based on physical characteristics

Experimental site and plant materials

This research was conducted during 2019 using nine peach cultivars: Dixon, Early Grande, Flordaprince, Flordastar, Flordaglo, Florda 834, TropicSnow, Desertred, and Swelling (Fig. 1).The peach trees, which were eight years old, were grown in sandy soil (pH of 7.7–7.8) in a private commercial orchard near the Pico Company, El-Behera Governorate, Egypt. The peach trees were budded in Nemaguard rootstock spaced at 3.5 × 5 m apart. All tree cultivars were similarly pruned, trained, and thinned; thus, color descriptions accurately reflect the relative colors of the investigated peach fruits. The orchard was irrigated by drip irrigation system. Technical and horticulture processes such as fertilization, disease and pest control were carried out regularly based on standard (El-Etreby, 1996; Ikinci & Bolat, 2018). Four trees (as uniform as possible) were chosen from each peach cultivar to collect fruit samples.

Fruits were harvested when fully mature according to commercial practice and immediately transported to the laboratory located at the Department of Food Science and Technology, Faculty of Agriculture, Alexandria University, Alexandria, Egypt. All fruits were harvested at the ripe stage and samples with defects (sunburn, cracks, bruises, and/or cuts in the husk) were discarded.

Measurements of fresh peach fruit properties

Peaches were neither washed nor brushed before being measured. For each cultivar, 20 random fruits were used to define the flesh firmness of each of the two sides using an Effegi pressure tester with a 5/16-inch plunger (Effegi, 48011 Alfonsine, Italy); flesh firmness was expressed as Lb/inch². Fruit diameters, polar and cross, were measured in mm immediately after harvesting using a digital caliper. According to Commission Regulation (EC) no. 1861/2004, 56 mm is the minimum diameter for a fruit to be considered in the “extra” category, which can be subdivided into different categories as follows: AAAA (diameter > 90 mm), AAA (90 mm > diameter ≥ 80 mm), AA (80 mm > diameter ≥ 73 mm), A (73 mm > diameter ≥ 67 mm), B (67 mm > diameter ≥ 61 mm), C (61 mm > diameter ≥ 56 mm), and D (i.e., not in the “extra” category; 56 mm > diameter ≥ 51 mm) (Mirás-Avalos et al., 2017).

As the measurement of skin color is nondestructive, the parameter was measured prior to the analysis of flesh firmness using a simple digital imaging method (Yam & Papadakis, 2004). Specifically, a fresh peach was placed in a white container inside a lighted box, which was illuminated by two 26 W fluorescent lamps (lumen = 1250 ± 20%). The lamps (13 cm in length) were situated 45 cm above the peach sample. A high-resolution digital camera (Canon XUS105, 12 megapixels, 4 × digital zoom) was used to capture images, which were stored on an Acer T6500 laptop (4.0 GB RAM, 320 GB hard disk). The digital camera was located vertically over the background at a distance of 45 cm. The angle between the camera lens and the lighting source axis was approximately 90°. The camera was fixed on top of the lighting box (Fig. 2), which was constructed from white plastic.

Twenty peach images were captured for each cultivar, and peach color was quantified in the images using the Histogram window in Photoshop (Ado, 2002). The mean color values of lightness (L), a, and b were determined, and these nonstandard color values were converted to L^∗, a^∗, and b^∗, respectively, using the following formulas (Yam & Papadakis, 2004):

(1)${\displaystyle L^{\ast }=\frac{\text{Lightness}}{255}\times 100}$ (2)${\displaystyle a^{\ast }=\frac{240a}{255}-120}$ (3)${\displaystyle b^{\ast }=\frac{240b}{255}-120}$

Given that the L^∗, a^∗, and b^∗ coordinate axis defines the three dimensional CIE color space (Kortei et al., 2015) in which is the parameters represent lightness, redness, and yellowness, respectively, positive values of b^∗ indicate yellow color; thus, peaches with high b^∗ values are considered high quality.

Artificial neural network

An ANN can be used as a data processing tool that simulates the learning technique of a biological neural network. ANNs originate from the human nervous system, which comprises a hugely parallel interconnection of nodes that initiate various perceptual and recognition tasks in a small amount of time (Halagundegowda & Singh, 2018). Multilayer feed-forward neural networks contain more than one layer of artificial nodes, which allow unidirectional forward connections of inputs and outputs; these networks are also known as multilayer perceptrons (MLPs). MLPs comprise regular input signals, an output neural layer, and a number of hidden layers, with different nodes found between the input and output layers.

In multilayer feed-forward networks, information is transmitted from the input layer to the output layer, as is the case in the human brain where signals move in one direction. Feed-forward networks use any Boolean function and are guaranteed to reach stability provided that the number of hidden nodes is sufficiently large. Indeed, MLPs can be considered as special cases of non-linear regression models. In economics, finance, and agriculture, not all relationships are direct; hidden layers form indirect relationships between input and output variables. Due to the absence of information on the parameters in hidden layers, ANNs are often termed “black boxes”. The number of hidden layers and the number of nodes in each hidden layer relates to the ability of the network to approximate more complex functions. However, networks with complex structures do not necessarily perform better. The block diagram of proposed methodology illustrating the procedure for identification and classification of peach varieties is shown in Fig. 3.

In the present study, a MLP neural network was fitted to the data using SPSS 19.0. The number of hidden nodes was changed; thus, different neural network model structures were tested before the final architecture of the model was confirmed. Ultimately, a neural network model with six input nodes and five hidden nodes performed better than the other competing models in terms of out-of-sample predictions and identification of peach cultivars (Fig. 4). In the analysis, the peach cultivars were encoded as follows: Class = 1.0 for Dixon, Class = 2.0 for Early Grande, Class = 3.0 for Flordaprince, Class = 4.0 for Flordastar, Class = 5.0 for Flordaglo, Class = 6.0 for Florda 834, Class = 7.0 for TropicSnow, Class = 8.0 for Desertred, and Class = 9.0 for Swelling

Statistical analysis

Peach cultivar measurements were assessed via ANOVA using SAS software. To compare the mean values of skin color, fruit firmness, and peach diameters, Least Significant Difference was used at significance level of p ≤ 0.05.

Characteristic parameters of peaches

Table 1 lists the measured L^∗, a^∗, and b^∗ values of the nine tested peach varieties. Similar patterns were observed, with L^∗ as the highest value, a mid-range a^∗ value, and a high b^∗ value. The Early Grande and Desertred cultivars had the highest and lowest average L^∗ values at 60.04 ± 1.33 and 36.76 ± 0.16, respectively, with the lower value perhaps caused by browner coloration (Juncai et al., 2015). Dixon, Flordastar, Flordaglo, and Swelling cultivars had similar mean L^∗ values (Table 1): 54.33 ± 1.51, 55.78 ± 1.26, 51.72 ± 1.11, and 57.05 ± 0.93, respectively. However, for other peach cultivars, Montero-Prado, Bentayeb & Nerín (2013) reported L^∗ values in the range of 71.8 ± 1.9 to 62.4 ± 2.5.

In the present study, differences were observed in a^∗ and b^∗ values, where a^∗ represents “-green” to “+red” and b^∗ represents “-blue” to “+yellow”. The Dixon and Swelling cultivars had a^∗ values of 17.35 ± 1.64 and 16.35 ± 1.07, Flordastar and Flordaglo had a^∗ values of 28.52 ± 1.57 and 27.11 ± 1.54, respectively. The Desertred cultivar had the highest average a^∗ value of 39.71 ± 0.90, indicating that it was generally more intensely red than the other cultivars, whereas the Early Grande cultivar had the lowest average a^∗ value of 9.81 ± 1.05. The red coloration of fruit skin is mainly due to the accumulation of anthocyanins (Lyu et al., 2017). For b^∗, Flordastar, Flordaglo, and Desertred cultivars had similar average values of 38.41 ± 1.51, 38.77 ± 0.73, and 38.10 ± 0.85, respectively. The average b^∗ of Dixon was higher at 49.48 ± 1.50; as the chlorophyll levels decreased (Juncai et al., 2015), however, positive values of b^∗ indicate yellow color; thus, Dixon cultivar which had high b^∗ values is considered high quality. Finally, none of the a^∗ and b^∗ values for the investigated peach cultivars were negative. Our results were in accordance with those reported by Di Vaio et al. (2015). Additionally, Lyu et al. (2017) studied the characteristics of 18 cultivars of white-flesh peach fruits cultivated in North China and found that peel L^∗, a^∗, and b^∗ values were 34.01–52.97, 8.36–22.9, and 7.76–15.66, respectively. It is difficult identify peach cultivars individually based on L^∗, a^∗, and b^∗ values because they are too similar to allow classification. In a variation examination of the fruit quality indices of the Xiahui 8 peach variety, Zhang et al. (2017) reported mean L^∗, a^∗, and b^∗ values of 69.36 ± 4.78, 15.36 ± 5.71, and 23.31 ± 1.81 when fruit harvesting maturity was degree I and 53.25 ± 4.79, 31.61 ± 3.14, and 18.31 ± 1.6, when it was degree II, respectively. They suggested that the a^∗/b^∗ value represented the true color of the peach, and showed that it was 2.54-fold higher in the degree II fruit relative to the degree I Xiahui 8 peach fruit.

Statistical analysis showed that each color parameter (L^∗, a^∗, and b^∗) of each peach variety was significantly different from that of the other varieties (p ≤ 0.05) (Table 1). The range of the coefficient of variation (CV) among the color values was 0.44%–10.70%, indicating that the degree of color indicators varied. The largest CV was observed for a^∗ (10.70%) while the smallest CV was for L^∗ (0.44%). Similarly, Lyu et al. (2017) revealed that a^∗ and L^∗ had the highest and lowest CV values in 18 cultivars of white-flesh peaches.

Table 2 presents the measures of firmness, polar diameter, and cross diameter of the nine tested peach varieties. Peach size is a vital quality for consumer preference (Khezri, Rashidi & Gholami, 2012), with consumers typically preferring fruits of equal size and shape. The mean cross diameter of samples was 5.54–6.50 cm, whereas the polar diameter was 5.51–6.39 cm (Table 2). Sherman et al. (1982) showed that Flordaprince peach fruits are relatively large, averaging over five cm in diameter. However, the same cultivar was smaller when grown under Egyptian climatic conditions, with Mansour & Stino (1989) reporting an average diameter of 4.8–4.9 cm and Shaltout (1987) reporting an average diameter of 5.2–5.5 cm. Moreover, Mansour & Stino (1986) noted that Early Grande peach fruits were medium–large, averaging 5.5–5.9 cm in diameter and 5.5–5.8 cm in length. In addition, Shama (2000) showed that the Early Grande peach fruit had a significantly larger diameter and length than other cultivars. Furthermore, Neelam & Muhammed (2002) evaluated 15 peach cultivars and showed that Loring peaches had a maximum fruit length of 5.15 cm while Early Grande peaches had the highest fruit diameter of 5.51 cm. Additionally, Zayan et al. (2015) showed that for Flordaprince peach, the fruit length and fruit diameter in season 2007 were 5.27 cm and 5.56 cm, respectively, for Early Grande, they were 5.56 cm and 6.01 cm, respectively and for Desertred, they were 5.47 cm and 5.89 cm, respectively.

The firmness of fruit flesh is also an important factor of fruit quality (Maness, Brusewitz & McCollum, 1992). Firmness in peaches can be an indicator of immaturity or over maturity. For example, excessive firmness indicates an immature peach in which the mesocarp is tightly bound to the stone with little free juice (Armstrong, Stone & Brusewitz, 1997). Fruit firmness is considered low when the value is between 0.40 and 0.62 kg/cm² (Santiago-Mejía et al, 2008). In the present study, fruit flesh firmness ranged from 4.73 to 12.16 Lb/inch² (i.e., 1.65–4.24 kg/cm²) at the time of commercial harvest (Table 2). Maximum levels of fruit firmness for marketing fresh peaches and nectarines are set by the EU at a 6.5 kg (= 63.7 N)/0.5 cm² (eight mm diameter probe) (Commission Regulation (EC) No. 1861/2004 of October 28, 2004). Significant differences in firmness were observed among some of the nine tested peach cultivars (p ≤ 0.05). TropicSnow peaches had the highest firmness values at 12.16 Lb/inch², which was significantly higher than that of other peach varieties. Although there were significant differences in firmness among the cultivars, there was no obvious trend among the nine peach varieties (Table 2), even though the picking period for all cultivars was around mid-May to the end of May. There was less difference in firmness among the cultivars Dixon, Early Grande, Flordaprince, Flordastar, Flordaglo, and Florda 834 compared with the cultivars TropicSnow, Desertred, and Swelling; this was likely due to the cultivars in these groups having the same pattern of ripening despite their differences, indicating that firmness is highly cultivar-dependent. Zayan et al. (2015) showed that for Flordaprince, Early Grande, and Desertred peaches, the firmness were 12.79, 14.86, and 12.96 Lb/inch², respectively in season 2007. In addition, Ahmed, El-Habashy & Maklad (2017) reported the firmness for Flordaprince and Swelling peaches budded on Nemaguard rootstock were 8.914 and 5.82 Lb/inch², respectively in season 2014. However, changes in fruit tissue firmness during ripening have been largely attributed to enzymic dissociation of the cell walls between adjacent cells (Huber, 1983; Fischer & Bennett, 1991). Soft peaches can be excessively juicy (Armstrong, Stone & Brusewitz, 1997), while a decline in peach fruit firmness is observed as background color becomes more yellow and less green (Bonora, Stefanelli & Costa, 2013).

The sphericity of the peaches (polar diameter/cross diameter) was 0.95–1.07 (Fig. 5). Sphericity values <0.9 indicate that the fruit is oblate, whereas values >1.1 indicate that it is oblong; the remaining fruits with intermediate index values are considered round (Buyanov & Voronyuk, 1985). As shown in Fig. 5, the Dixon and Early Grande cultivars were oblong whereas the Flordaprince, Flordastar, Desertred, Swelling, Flordaglo, Florda 834, and TropicSnow cultivars were round. Thus there was some dissimilarity in the shape of the studied cultivars. In particular, Dixon was the least spherically shaped variety. Zayan et al. (2015) showed that for Flordaprince, Early Grande, and Desertred peaches, the shape rate which is a function of fruit length and diameter of the fruit in season 2007 was 0.95, 0.93, and 0.93, respectively.

Statistical analysis of firmness, polar diameter, and cross diameter showed that each parameter of each variety was significantly different from the parameters of other varieties (p ≤ 0.05) (Table 2). According to Commission Regulation (EC) no. 1861/2004, all studied peach cultivars met the minimum diameter to be considered in the “extra” category (i.e., 56 mm).

Performance of the ANN classifier

As shown in Table 3, 112 cases were assigned to the training dataset and 68 patterns were used as the testing dataset (conforming to rules that 62.2% of data should be used for training while 37.8% of data is used for testing). The training dataset comprised data used to train the neural network; the testing dataset was an independent set of data used to track errors during training and prevent overtraining.

Table 4 provides details of the neural network’s architecture. Six input variables were used as covariates in analysis and the standardized rescaling method was used to adjust the covariates. Scale-dependent variables and covariates were rescaled by default to improve network training. All rescaling was performed based on the training data, even if a testing dataset was defined. The network had an input layer with six input nodes, with the number of nodes in the input layer being the number of covariates. A single hidden layer had six hidden nodes and an output layer had one output with nine nodes. A hyperbolic tangent activation function was used in the hidden layers, taking real valued arguments and then transforming them into a range (−1 to 1). The error was the cross entropy error because the softmax activation function was applied to the output layer. It took a vector of real valued arguments and transformed it to a vector for which the elements fell within a range (0 to 1) and summed to 1. Softmax can be used only if all dependent variables are categorical. Training time was very small (00:00:00.054 s), however, Livni, Shalev-Shwartz & Shamir (2014) pointed out that training time is defined as how much computation time is required to learn the class. On the same line, other research papers presented training time of 00:00:00.001s (Pandurangan et al., 2017), 0:00:00.000 s (Maliki et al., 2011), and 0:00:00.265 s (Alabi, Issa & Afolayan, 2013). Additionally, when training a neural net with one hidden layer of size five and one input node for 100 steps, a variety of training times ranging from around 0.808 to 1.04 s could be recorded (Adlers & Pihl, 2018). Generally, the training time depends completely on the structure of used ANN.

Table 5 shows the results of training and applying the final ANN model to the testing dataset. The cross entropy error is presented because the output layer uses the softmax activation function; the ANN attempts to minimize this error function during the training process. The cross entropy error has a predicted value for each category, where each predicted value is the probability that the case belongs to the category. As shown in Table 5, the cross entropy error was 0.165, which is an allowable level that enables analysis to continue into further steps. The percentage of incorrect classifications can be taken from the classification table, in which 0% of classifications were mismatched with the original observed samples. The estimation algorithm ended because the training error ratio criterion (0.001) was achieved (Table 5). Ideally, training would end when the error had converged. The cross entropy error was 0.140 with 0% incorrect classifications of the testing data.

To the best of our knowledge, almost all research results related to the identification of peach fruit have taken offline images and none of the studies performs in real-world scenario with acceptable accuracy. However, in the present study, the overall classification accuracy was 100% as shown in Table 6 in the training and testing phases and this mean that using the selected features, the peach varieties could be recognized easily. No overfitting is occurred, however, overfitting is a major problem in supervised machine learning which avoids us from perfectly generalizing the models to well fit experimental data on training dataset, in addition to hidden data on testing dataset (Ying, 2019). As a consequence of the existence of noise, the complexity of classifiers, and the limited size of training set, overfitting occurs. To decrease the effects of overfitting, several approaches are suggested to address to these causes as reported by Ying (2019).

Regardless of the ANN models are performing better in its application to human problems, there is an increasing need to address the problem of adopting a regular methodology in ANNs development stage to increase its performance (Abiodun et al., 2018). For instance, a potential way to improve the robustness of the ANN is to increase the volume of training data and use both good and bad quality data as training samples. However, model robustness means predictive capability of ANN types in generalizing range of data like those used for training (Abiodun et al., 2018). On the other hand, the image grading system is simple and efficient and can be considered a suitable first stage to mechanizing the commercial grading of fruits (Momin et al., 2017). Thus, in fruit classification automation systems, which are based on color features extracted from fruit images, the noise in images can be eliminated using image processing algorithm (Momin et al., 2017) to control a process at optimum efficiency (Patel et al., 2012). Moreover, the number of training cases is one of various factors which can affect the optimum performance of an ANN classifier (Sildir, Aydin & Kavzoglu, 2020). The vital idea for dealing with small datasets is to maximize the extension of the training set by a specific facility of the tool, the so-called all-frame or leave-one-out cross validation procedure (Pasini, 2015).

The ANN classifier could be a very useful tool in identification of peach varieties based on skin color, size and firmness. Thus, it will enable to manufacture more easily, simply and economically an equipment in real-world to be developed thanks to this model. It would be possible to construct automation systems, using the ANN classifier developed, for identification peach varieties to use in supermarkets or grading stations.

Sensitivity analysis

Sensitivity analysis (Kakhki, Freeman & Mosher, 2019) can be used to appraise the relationship between the input attributes and outcomes as a feature of the extraction method in ANN models (Harrington & Wan, 2002). Such analysis calculates the significance of each attribute (independent variables) in determining the probability of an output in the neural network (IBM, Released 2010). Figure 6 presents the importance of all attributes used in the present study. However, the importance of an independent variable is a measure of how much the network’s model-predicted value changes for different values of the independent variable (Alabi, Issa & Afolayan, 2013). Additionally, Fig. 7 presents the normalized importance of all attributes used in the present study. By inspection of Fig. 6,the order of the importance of the input variables that had the greatest effect on identification of peach cultivars was as follows: polar diameter (0.215), L^∗ (0.192), b^∗ (0.190), a^∗ (0.169), firmness (0.154), and cross diameter (0.081).

Also, by inspection of Fig. 7, the order of the normalized importance of the input variables that had the greatest effect on identification of peach cultivars was as follows: polar diameter (100%), L^∗ (89.0%), b^∗ (88.00%), a^∗ (78.50%), firmness (71.30%), and cross diameter (37.53%). Thus, polar diameter was significantly important for the potential identification of peaches, which is in agreement with the findings of a previous study by Naskar & Bhattacharya (2015) on the significance of work-related activities in fruit classification, in which geometric and color attributes were used; these authors showed that the color and shape, shape and texture, or texture and color attributes were acceptable variables for fruit recognition. Therefore, as the number of investigated variables increases, there is a greater possibility of identification and the model becomes easier to interpret. The findings of this study have to be seen in light of some limitations. The peach identification accuracy is based on measurements of polar diameter, L^∗, b^∗, and a^∗ of peach skin, firmness, and cross diameter for definite cultivars. These measurements, therefore, subjected to biases and confounding that may have influenced our identification accuracy.