Study on the Application of Electronic Nose Technology in the Detection for the Artificial Ripening of Crab Apples

Abstract

Ripening agents can accelerate the ripening of fruits and maintain a similar appearance to naturally ripe fruits, but the fruit flavor and quality will be changed compared to naturally ripe fruits. To find an efficient detection method to distinguish whether crab apples were artificial ripened, the naturally ripe and artificially ripe fruits were detected and analyzed using the electronic nose (e-nose) technique in this study. The fruit quality indexes of samples were determined by the traditional method as a reference. Significant differences were found between naturally ripe and artificially ripe fruits based on the analysis of soluble sugar content, titratable acidity content, sugar–acid ratio, soluble protein content, and soluble solids content. In addition, principal component analysis (PCA), linear discriminant analysis (LDA), support vector machine (SVM), and random forest (RF) analyses were performed on the electrical signals generated by the electronic nose sensor, respectively. The results showed that the RF is the best recognition algorithm for distinguishing which crab apples were naturally ripe or artificially ripe; the average recognition accuracy is 98.3%. On the other hand, the prediction models between the e-nose response data and fruit quality indexes were constructed by partial least squares regression (PLSR), which showed that the feature value of e-nose response curves extracted by wavelet transform was highly correlated with the quality indexes of fruits, the determination coefficients (R²) of regression models were higher than 0.91. The results demonstrated that the detection technology with an electronic nose could be used to test whether the fruit of the crab apple was artificially ripe, which is an economical and efficient method.

1. Introduction

The crab apple is widely planted in the north of China. The ripe fruit (‘Saiwaihong’ apple) is red in color and has fragrance. Whether the fruit is attractive to customers depends not only on its appearance but also on its taste and nutritional value, such as sugar content, sugar–acid ratio, and vitamin content. Sugar content and acidity, in particular, dictate whether an apple is sweet or tart and if it is appealing to customers. Most fruits achieve their best organoleptic, and commercial quality attributes only when they ripen naturally. However, in production, treatments with ripening agents are often used to promote the early ripening of fruits. Ripening agents are synthesized chemicals such as ethephon, thidiazuron, and abscisic acid that are sprayed on crops or fruits and vegetables to speed up their growth and development process and cause them to ripen earlier. Although there are differences in nutritional composition between artificially ripe fruits and naturally ripe fruits, they are similar in appearance, which increases the difficulty of distinguishing. Therefore, providing an efficient and low-cost detection method to distinguish whether the fruit has been artificial ripened is urgently required.

In the past decades, quality testing of apples had relied on physical and chemical indexes assessment as well as sensory evaluation. Nevertheless, some studies on the physical and chemical indexes of apples have focused on the correlation of indexes such as acidity and sugar [1], and the methods of physiological indexes testing are time-consuming and laborious, with high requirements for inspectors, which cannot meet the needs of practical testing. However, sensory evaluation is a labor-intensive process with low objectivity and repeatability. Furthermore, even the best-trained panelists may be influenced by various physiological, economic, and personal factors, and human perception varies over time and among people [2]. Some modern analytical techniques, such as gas chromatography–mass spectrometry, high-performance liquid chromatography, and other techniques, could also solve these problems. Yet, these analytical methods are expensive and time-consuming, with complex sample preparation processes. As such, the non-destructive testing of the internal quality of fruits is becoming increasingly important for industry and consumers [3]. As a result, sensor technologies such as electronic noses that can swiftly assess the quality of fruits are gaining popularity.

The e-nose is a device that mimics the human olfactory system. A typical e-nose system consists of a sensor array and pattern recognition, which are both imitators of olfactory cells and the brain. [4]. In the presence of particular molecule chains, sensors generate an electrical signal, which is how the e-nose works [5]. Since the sensor arrays of the e-nose have broad and partially overlapping selectivity and the fact that apples produce a variety of volatile organic compounds that give them distinctive aromas and create unique flavor profiles, sensor measurements can provide an odor “fingerprint” of the sample. These “fingerprint” patterns are specific to a group of aromatic chemicals and can indicate the overall olfactory information of the sample identified [6]. In addition, as sensor technology advances, the e-nose has become a piece of simple equipment with great detection accuracy, and these devices are increasingly being considered as alternatives to traditional approaches [7,8]. Due to the ease of sample preparation and the advantages of fast, sensitive, real-time, and non-destructive detection, e-nose offers a quick and non-destructive alternative to traditional approaches that rely on lengthy laboratory processing [9]. E-noses have been successfully applied in many fields, especially in the food industry. In the scientific literature, we can find positive applications of electronic nose technology in the identification of fruit quality, such as classifying apples based on their cultivars [10], detecting the freshness of strawberries during storage [11], and detecting adulterants [12]. In horticulture, e-noses have been employed successfully to monitor the aroma of peaches [13], the quality of lychees [14], pears [15], and other fruits and vegetables [16,17], and e-noses also have been applied to the detection of early fungal contamination and infestation of fruits [18,19].

Several scholars have made some attempts to detect the ripeness of apples with electronic noses. Pathange et al. [20] used ripeness indexes such as a starch index and puncture strength to classify Gala apples into three ripeness groups: unripe, ripe, and overripe. According to the results of the discriminant analysis, the electronic nose can effectively classify Gala apples into three ripening groups with a correct classification rate of 83%. Brezmes et al. [21] used an e-nose to assess the ripeness state of Pink Lady apples during their shelf life. The results clearly showed that the e-nose signals are related to the ripening process of apples. These attempts were conducted in the natural state of the fruit. In this study, on the other hand, it was hypothesized that used electronic noses could distinguish between naturally ripe and artificially ripe fruits. To test this hypothesis, this study proposed to evaluate the possibility of classifying naturally ripe and artificially ripe fruits by using the e-nose in combination with an appropriate pattern recognition method and by measuring the physiological indexes of the samples as a reference.

2. Materials and Methods

2.1. Experimental Design

The experiments were conducted in the orchard of Jilin Agricultural University (43°48′ N, 125°25′ E), Changchun, China. All the crab apple trees used in the experiment were six years old and spindle-shaped; the cultivar used is known as the ‘Saiwaihong’ apple. Two treatment schemes of artificial ripening were designed. Treatment A: spray the apples on the tree at the fruit growth stage of 70% ripeness (15 August 2021) with ethephon solution (120 mg/L) to accelerate the ripening of the fruits; Treatment B: spray the apples on the tree at the fruit growth stage of 80% ripeness (22 August 2021) with ethephon solution (120 mg/L) to accelerate the ripening of the fruits. Naturally ripe fruits were considered as control and defined as CK. Five trees were used for each treatment. The treatment fruits were harvested when their appearances were similar to naturally ripe fruits. The fruit appearances at the ripening stage under different treatment conditions are shown in Figure 1. When harvesting, apple samples were selected with no pests, diseases or damage and from the same orchard. Eighty apples were selected for each treatment. After the fruits of each treatment were picked, the odor information of samples was detected with an electronic nose, then the fruit quality indexes were determined. In order to reduce experimental errors, all samples were tested at room temperature.

2.2. E-Nose Detection

The odor characteristics of the samples were analyzed using a laboratory-developed electronic nose. This system consists of a data acquisition card, a test circuit, a miniature air pump, a sampling chamber, and an array of gas-sensitive sensors; the physical diagram is shown in Figure 2. Each of the 12 metal-oxide semiconductor (MOS) sensors in the sensor array is sensitive to a specific category of volatile chemicals in the sample gas.

Table 1. Application of sensors used in the sensor array.

Number	Sensors	Main Applications	Manufacturer and Country
S1	TGS2612	Hydrocarbons, Methane, Liquefied Petroleum Gas	Figaro, Osaka, Japan
S2	GSBT11	Volatile Organic Compounds (VOC), e.g., Acetone	Ogam, Jeollanam-do, Korea
S3	WSP2110	Benzene, Toluene, Formaldehyde	Winsen, Zhengzhou, China
S4	MP135	Ethanol, Cigarette Smoke, Air Pollutants	Winsen, Zhengzhou, China
S5	MS1100	VOC, Toluene, Benzene, Formaldehyde,	Ogam, Jeollanam-do, Korea
S6	MP901	Alcohol, Cigarette smoke, Formaldehyde, Toluene, Benzene, Acetone	Winsen, Zhengzhou, China
S7	TGS2611	Natural Gas, Methane	Figaro, Osaka, Japan
S8	TGS2620	VOC, Alcohols, Organic Solvents Steam	Figaro, Osaka, Japan
S9	TGS2602	Ammonia, Hydrogen Sulfide (high sensitivity to VOC and odorous gases)	Figaro, Osaka, Japan
S10	TGS2610	Alcohols, Butane, Liquid Petroleum Gas, Propane	Figaro, Osaka, Japan
S11	TGS2600	Ethanol, Hydrogen, Hydrocarbons, etc.	Figaro, Osaka, Japan
S12	TGS2603	Trimethyl Amine, Methyl Mercaptan, Hydrogen Sulfide, etc.	Figaro, Osaka, Japan

covers the 12 MOS sensors and summarizes their principal uses as well as some of their properties.

The e-nose was activated 120 min prior to detection in this experiment to ensure the surface of the sensor was heated to operating temperature and the gas path was cleaned with clean air. The detection process started after the electronic nose was initialized. The schematic diagram of the e-nose measurement process in the experiment is shown in Figure 3. The fruits of each treatment were randomly divided into 40 groups of two fruits each. Each group was put into a glass beaker with a 1 L volume and sealed with plastic wrap for 15 min to guarantee that the volatile chemicals in the apples filled the beaker and reached equilibrium. During the cleaning process, clean air was pumped into the gas path and sensor chamber to control the normalization of the sensor signal. Subsequently, the sample gas was pumped in and flowed through the sensor array at a rate of 1 L/min. When the sensor array was exposed to this sample gas, the test circuit converted the sensor conductance change into a voltage change, employing the voltage (V) as the response of the electronic nose sensor, and the typical response graphs (one sample for each treatment) of the sensor for different treatment detection are shown in Figure 4. During the detection process, 12 signal curves of the voltage were plotted, and the measurement phase lasted for 50 s. At the end of the measurement, the data was automatically recorded for the following analysis. After each measurement, the time for the electronic nose sensor to be reset and recalibrated back to baseline levels by clean gas was 300 s before the second round of headspace sampling.

2.3. Determination of Fruit Quality Indexes

After e-nose detection, the quality indexes of the samples were determined. The soluble sugar content of the sample was determined by the anthrone colorimetric method. By titrating with sodium hydroxide until phenolphthalein changed color, the titratable acid content of samples was determined [22]. The sugar–acid ratio was determined by dividing the amount of soluble sugar by the amount of titratable acid. The content of vitamin C and soluble protein was determined by the molybdenum blue method and the Komas Brilliant Blue G-250 method, respectively, and the soluble solids content of the samples were determined by a hand-held refractometer. For the determination of the physiological indexes, samples from each treatment were randomly selected, and three duplicates were performed for each index, with the average values of the three duplicates being used as the original values for each sample group. The experimental data were processed and analyzed using Microsoft Excel 2019 and SPSS 26.0 statistical software, and the significance of differences between treatments was tested by Duncan’s new complex polar difference method (p < 0.05).

2.4. Analysis of E-Nose Data

2.4.1. Feature Extraction Methods

Previous literature has shown that different feature extraction methods have different classification performances. Based on the analysis of the response curve of the electronic nose, five methods were selected for feature extraction, including maximum value, integral value, Fourier transform, wavelet transform, and Gaussian curve fitting [23].

2.4.2. Classifiers

The classification models mentioned in this study were used for classification between three categories. In principal component analysis (PCA), observations of correlated variables are converted into a set of linearly uncorrelated variables called principal components using orthogonal transformations. This transformation is defined in such a way that the first principal component has the maximum possible variance, and subject to the constraint of being orthogonal to the previous component, each subsequent component, in turn, has the maximum possible variance. PCA is a sensitive and highly accurate method for finding principal variables and is one of the most commonly used methods for data analysis and dimensionality reduction in multivariate systems [24].

Linear Discriminant Analysis (LDA) is a common classification method where the classification results of each group are linearly correlated. Principal component analysis and linear discriminant analysis differ in that linear discriminant analysis computes using category information. The principal component analysis aims to minimize data noise and multicollinearity between different variables, while linear discriminant analysis aims to minimize intra-class ratios and maximize inter-class ratios [25].

Support vector machine (SVM) is part of the supervised training methods used for regression and classification, and the SVM algorithm is classified as a pattern recognition algorithm. SVM was originally developed for the linear classification of separable data but applies to nonlinear data with the use of kernel functions. The main goal of SVM is to define decision boundaries for different classes of data points using hyperplanes, with edges separating classes from a hyperplane to the data set distance to the nearest point in the data set. It has been extensively used and studied due to its capabilities in classification and regression prediction [26].

Random Forest (RF) is a non-parametric and non-linear classification and regression algorithm. The algorithm is a collection of several decision trees in which each tree is classified from a randomly selected subset of attributes. Then, majority voting is used to obtain the final classification results, where the tree with the highest number of classifications will be selected [27]. It has received increasing attention due to the speed with which decision trees can be made; training hundreds of them will be much faster than training an artificial neural network, and it has strong robustness, high learning speed, and accuracy rate compared to a single classifier [28]. In this research, the above-mentioned pattern recognition methods were selected to analyze the samples in combination with previous literature [29,30]. The feature values of all data were executed by MATLAB 2013 software, and pattern recognition was performed using R language 4.0.2 software.

2.4.3. PLSR

The partial least squares regression method (PLSR) is a multivariate statistical data analysis method that combines the features of principal component analysis (PCA) and multiple linear regression (MLR). When there are high linear correlations between variables, the PLSR approach can be used to create high-performance models. The PLSR approach is suited for solving metric prediction issues based on e-nose data with these qualities [7]. In this study, the PLSR method was applied to build the prediction model of fruit quality indexes. Minitab was used to accomplish the PLSR approach.

3. Results

3.1. Effects of Artificial Ripening Treatment on Fruit Quality Indexes

3.1.1. Soluble Sugar Content, Titratable Acid Content and Sugar–Acid Ratio of Fruit

The effects of artificial ripening on soluble sugar content, titratable acid content, and sugar–acid ratio of fruit are shown in Figure 5. As can be seen in Figure 5A, both treatment A and treatment B were significantly lower than naturally ripe fruit (CK), and the difference between treatment A and treatment B was also significant (p < 0.05). Compared with the CK, the contents of titratable acid of treatment A and treatment B were increased significantly (Figure 5B). Therefore, the sugar–acid ratio of fruit was changed significantly with the treatment of artificial ripening (Figure 5C). Compared to treatment A and treatment B, the soluble sugar content and sugar–acid ratio of the naturally ripe fruit were highest, which will show better taste.

3.1.2. Content of Soluble Protein, Soluble Solids, and Vitamin C

The effects of artificial ripening on the content of soluble protein, soluble solids, and vitamin C are shown in Figure 6. As can be seen in Figure 6A,B, both the soluble protein content and soluble solids content of fruit were decreased compared to CK, and there were significant differences between treatments (p < 0.05). It can be seen that the artificial ripening treatment had little effect on the vitamin C content of fruit in Figure 6C, and none of the differences between the treatments reached a significant level (p > 0.05).

In summary, by combining the quality indexes of each treatment, it can be concluded that although the artificially ripe fruits were similar to the naturally ripe fruits in appearance, the internal quality was not as good as the naturally ripe fruits, proving that the naturally ripe fruits were more suitable for consumption. In addition, some quality indexes such as soluble sugars and titratable acids also reflect the ripeness of the fruits; the results indicated that the internal quality of the artificially ripe fruits does not reach the level of natural ripeness, and the measurement methods of these quality indexes were considered destructive and time-consuming.

3.2. E-Nose Detection Results of Fruit

3.2.1. The Classification Results Based on Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA)

The maximum value of the electronic nose response curve is the stable value in the whole response process, which reflects the stable response of the sensor to the sample gas and the maximum change of the electrical signal. The maximum value of the electronic nose response data was selected for PCA and LDA of the samples. The PCA score plot based on the maximum values from the samples obtained from the e-nose was demonstrated in Figure 7. As shown in the figure, the samples could be broadly divided into three groups (treatment A, treatment B, and CK), and two principal components were identified that explained 70.57% and 14.61% of the variation in the dataset, respectively, together reaching 85.18%. Although the naturally ripe samples data can be better differentiated from other treatments, the overlap between the artificial ripening samples and the naturally ripe samples was not well-clustered. It indicated that the data may drift in the classification, which results in the overlap of the sample data. Figure 8 illustrates the linear discriminant analysis of the sample data. It can be seen from Figure 8 that the first two discriminant functions explained 88.01% and 11.99% of the effect, respectively, with an overall contribution of 100%. LDA emphasizes the distribution of apple aroma components in space and the distance between them. The more dispersion between the data collection points, the higher the group differentiation. In general, the samples from different treatments of crab apples were separated, but the data collection points of samples between the two artificial ripening treatments still partially overlapped, indicating that the volatile gases of the artificially ripe fruits were similar after maturation, and these samples might have been incorrectly classified as adjacent groups in the classification analysis.

3.2.2. The Classification Results Based on SVM and RF

In this study, the SVM and RF nonlinear recognition algorithms were used by applying different feature extraction methods (maximum value, integral value, wavelet transform, Fourier transform, and Gaussian curve fitting), and the accuracy and selection of feature values were used as an evaluation of the recognition performance of the SVM and RF algorithm. When analyzing the data and constructing the model, the tenfold cross-validation method was applied to the sample data. The recognition results of the SVM and RF based on different feature extraction methods are shown in Figure 9. As can be seen from Figure 9, no significant level was reached between SVM and RF for recognition based on maximum value (p > 0.05), while there was a significant difference between SVM and RF for recognition based on integral value, wavelet transform, Fourier transform and Gaussian curve fitting (p < 0.05). According to the results in Figure 9, it can be found that RF was superior to SVM in identifying artificially ripe samples and naturally ripe samples when based on different feature extraction methods, the average recognition accuracy is 98.3%.

3.2.3. Predicting the Fruit Quality Indexes Based on E-Nose Response Data

According to the pattern recognition results of SVM and RF based on different feature extraction methods mentioned above, the integral value and wavelet transform were selected as the feature data of the PLSR method to predict the quality indexes. The 120 datasets (40 samples per group) were classified randomly into calibration and validation subsets, with 96 datasets (32 samples per group) as the calibration set and 24 datasets (8 samples per group) as the validation set. The quality indexes were defined as dependent variables in the PLSR method, and the feature values extracted by the integral value and wavelet transform feature extraction methods were defined as independent variables. In this study, all the potential variables that replaced the original variables were selected to build the PLSR model. The determination coefficients (R²) were calculated to evaluate the accuracy of the PLSR model.

The analysis results of PLSR based on the integral value of e-nose response curves were shown in Figure 10, and the results showed that the determination coefficient (R²) of the regression model for soluble sugar, titratable acid, soluble protein, and soluble solids was 0.89, 0.93, 0.92 and 0.87, respectively. Figure 11 presented the results of PLSR based on the feature value extracted by wavelet transform, which indicated that the R² of the regression model for soluble sugars, titratable acids, soluble proteins, and soluble solids was 0.94, 0.91, 0.95, and 0.92, respectively.

The above analysis results of PLSR indicated that the quality indexes model based on the electronic nose data of crab apple achieved the ability of nondestructive prediction and performed well to some extent.

4. Discussion

It is significant to identify whether the horticultural products in the market have been treated with a ripening agent, which is helpful for the management department to evaluate the quality. At this time, few studies have reported that using an electronic nose to detect the horticultural product is naturally ripe or artificially ripe. Most are about using the electronic nose to test the ripeness of horticultural products. Wang et al. [31] used an electronic nose to evaluate the ripeness and monitor the shelf life of tomatoes. Results showed that pink-stage fruit could be distinguished from the light-red stage and red-stage ones based on the E-nose using principal component analysis (PCA) and linear discriminant analysis (LDA). Gómez et al. [32] used an electronic nose to detect the ripening status of tomatoes, and the results demonstrated that the electronic nose could differentiate the ripening states of tomatoes. Using LDA analysis, it was possible to distinguish and classify the different tomato maturity states, and this method was able to classify 100% of the respective total samples for each group. Previous studies had also noted the application of electronic noses in combination with SVM and other methods on fruits. For example, Mahdi et al. [33] used a support vector machine (SVM) approach to characterize the freshness of strawberries in different polymer packages using an e-nose with high classification accuracy. Voss et al. [27] used e-noses combined with random forest algorithms to be able to provide a fast and accurate response to the growth cycle of peaches with a low marketing cost. All these studies indicate that e-nose combined with proper pattern recognition could have good applications in fruits. In other works, Pathange et al. [20] employed an electronic nose to assess the ripeness of the apple cultivar, “Gala”. This instrument allowed to positively divide the fruits into three groups according to their stage of ripeness (unripe, ripe, and overripe) with an accuracy of 83%. The work in this study not only investigated whether artificial ripening reduces the quality of apples but also established a model for detecting quality indexes through the electronic nose technique, obtaining good prediction accuracy. Furthermore, it should be pointed out that the process of measuring the volatile organic compound signals from the electronic nose in this study does not require us to destroy the fruit of crab apple, but rather it is performed in a non-destructive manner, so the process is more suitable for application to the actual fruit supply chain process.

Fruit ripening is the process of accumulation of nutrients, flavor substances, and various enzymes, etc. The comparison results of quality indexes show that the soluble sugar content, sugar–acid ratio, and soluble solids content of the artificially ripe fruits are lower than the naturally ripe fruits. In our future work, the changes of gas volatiles of artificially ripe fruits will be analyzed in more detail by gas chromatography-mass spectrometry (GC-MS) and electronic nose, and then select a smaller number of gas sensors for the artificial ripening monitoring of crab apples and develop a special electronic nose system for the artificial ripening monitoring of fruits. The electronic noses will be practical and cost-effective for fruit quality monitoring.

5. Conclusions

Spraying ethephon solution could accelerate fruit ripening and maintain an appearance similar to naturally ripe fruits, but the fruit quality indexes were changed greatly compared to naturally ripe ones. In this study, an electronic nose was used to detect the odor signal of crab apples that were naturally ripe or artificially ripe. To establish a link between the e-nose signal and fruit quality, the classification algorithms of PCA, LDA, SVM, and RF were applied and compared. The main conclusions are as follows:

1. It can be seen from the results of quality indexes that naturally ripe fruits have better flavor and that it is possible to distinguish between naturally ripe and artificially ripe fruits based on the results, but the operation process is tedious, time-consuming, requires the involvement of professional testers and cannot be achieved in real-time. In contrast, the use of an electronic nose system based on an array of metal oxide semiconductor sensors are faster and more convenient for practical applications.

2. The classification results indicated that the accuracy of LDA was higher than PCA, but both LDA and PCA showed the problem of data overlap. Compare the nonlinear recognition algorithm SVM with RF; RF was more effective when pattern recognition was performed based on different feature values, and the average recognition accuracy was 98.3%.

3. The analysis results of PLSR showed that the feature value extracted by wavelet transform was highly correlated with the quality indexes of fruits, and the determination coefficients (R²) of regression models were higher than 0.91.

Based on the above findings, it was concluded that naturally ripe crab apples and artificially ripe crab apples could be distinguished by the analysis of the e-nose response data. Subsequently, the model of the e-nose to identify naturally and artificially ripe crab apples was successfully established, and a good prediction of the correlation between e-nose and quality indexes of crab apples was made. This study provides evidence that an electronic nose can be used as a nondestructive method to distinguish artificially ripe crab apples.