Electronic Sensor Technologies in Monitoring Quality of Tea: A Review

Tea, after water, is the most frequently consumed beverage in the world. The fermentation of tea leaves has a pivotal role in its quality and is usually monitored using the laboratory analytical instruments and olfactory perception of tea tasters. Developing electronic sensing platforms (ESPs), in terms of an electronic nose (e-nose), electronic tongue (e-tongue), and electronic eye (e-eye) equipped with progressive data processing algorithms, not only can accurately accelerate the consumer-based sensory quality assessment of tea, but also can define new standards for this bioactive product, to meet worldwide market demand. Using the complex data sets from electronic signals integrated with multivariate statistics can, thus, contribute to quality prediction and discrimination. The latest achievements and available solutions, to solve future problems and for easy and accurate real-time analysis of the sensory-chemical properties of tea and its products, are reviewed using bio-mimicking ESPs. These advanced sensing technologies, which measure the aroma, taste, and color profiles and input the data into mathematical classification algorithms, can discriminate different teas based on their price, geographical origins, harvest, fermentation, storage times, quality grades, and adulteration ratio. Although voltammetric and fluorescent sensor arrays are emerging for designing e-tongue systems, potentiometric electrodes are more often employed to monitor the taste profiles of tea. The use of a feature-level fusion strategy can significantly improve the efficiency and accuracy of prediction models, accompanied by the pattern recognition associations between the sensory properties and biochemical profiles of tea.


Introduction
Tea, as one of the most popular beverages, is consumed by hundreds of millions of people around the world, due to its pleasant flavor and healthcare effects. The tea industry's importance in the world is very high, and such that the United Nations has designated May 21 as "International Tea Day" [1]. This non-alcoholic beverage is produced from the young shoots of the tea plant (Camellia spp., var. sinensis, and var. assamica) [2]. Tea is cultivated in over 35 countries around the world. Most tea producing countries are in Asia (e.g., China, India, Indonesia, Sri Lanka, Iran, Japan, Bangladesh, Taiwan, Thailand, Malawi, Malaysia, etc.) and Africa (e.g., Kenya, Tanzania, Uganda, etc.), while the largest importing countries of this strategic crop are the UK, the USA, Russia, Pakistan, and Egypt [3,4]. Based on the data released by the Food and Agriculture Organization (FAO), the total global production of tea in 2018 was 8.96 million tons [5]. Although the global market value of tea during 2018-2019 increased from USD 200. 105 to 215.148 billion [6], the continuation of this

Non-Sensing Techniques for the Quality Assessment of Tea
Overall, the organoleptic assessment by trained test panels is a common method to evaluate the quality factors of tea type in commercial markets [29]. As mentioned, the quality sensory attributes (e.g., liquor color, flavor, bitterness/astringency, strength) are highly linked with biochemical ingredients, which can be analyzed using spectrophotometer, HPLC, and GC/GC-MS systems. For instance, theaflavins and thearubigins are responsible for color and brightness (Figure 1), while the strength of liquor tea can be a result of the quantities of caffeine and amino acids. All these chemicals can be quantified using spectrophotometer and HPLC systems [30].
Peres et al. [31] previously applied reduced flow micellar electrokinetic chromatography to determine catechins in green tea. GC/GC-MS instruments can be used to evaluate the content of the terpenoid/non-terpenoid aroma profile of various teas, according to their geographical origins [32]. Nonetheless, the high cost of these analytical systems does Recently, analytical techniques such as spectroscopy, machine vision systems, and hyperspectral imaging have been substituted with conventional techniques to monitor chemical compounds involved in the taste, odor, and color of tea and its products (e.g., chromatographic-based techniques of high-performance liquid chromatography (HPLC) and gas chromatography-mass spectrometry (GC-MS)). These methods are very sensitive and selective; however, they are not only expensive, but also require several pretreatment steps for sample preparation. Therefore, it is essential to employ valid and fast alternatives in modern tea industries [27]. The single and combined use of an electronic nose (e-nose), electronic tongue (e-tongue), or electronic eye (e-eye) compared to conventional methods can promote the quality assessment of tea [28]. To the best of our knowledge, there are few reviews on the application of electronic sensing platforms (ESPs) for monitoring the chemical-based organoleptic attributes of tea and its products. As an update, this overview highlights the latest advancements and critical assessment of the state-of-the-art in the field of ESPs usage, to monitor the quality parameters of tea in the industry.

Non-Sensing Techniques for the Quality Assessment of Tea
Overall, the organoleptic assessment by trained test panels is a common method to evaluate the quality factors of tea type in commercial markets [29]. As mentioned, the quality sensory attributes (e.g., liquor color, flavor, bitterness/astringency, strength) are highly linked with biochemical ingredients, which can be analyzed using spectrophotometer, HPLC, and GC/GC-MS systems. For instance, theaflavins and thearubigins are responsible for color and brightness (Figure 1), while the strength of liquor tea can be a result of the quantities of caffeine and amino acids. All these chemicals can be quantified using spectrophotometer and HPLC systems [30].
Peres et al. [31] previously applied reduced flow micellar electrokinetic chromatography to determine catechins in green tea. GC/GC-MS instruments can be used to evaluate the content of the terpenoid/non-terpenoid aroma profile of various teas, according to their geographical origins [32]. Nonetheless, the high cost of these analytical systems does not allow them to be set up in tea factories. Likewise, since there is always an oxidation process during transferring the sample from the factory to the laboratory, these instruments have restrictions in adapting to routine quality inspection in a production line [33].
In addition, odor activity values, from determining aroma-active compounds, are also measured using the GC-olfactometry method. Moreover, gravimetric techniques are used to analyze the proximate composition, in terms of moisture, soluble solids, fiber, and lipid contents [34]. Koch et al. [35] used liquid chromatography-electrospray ionizationquadrupole-time-of-flight-mass spectrometry (LC-ESI-Q-TOF-MS) and atomic absorption spectrometry (flame and electrothermal) systems, respectively, to analyze the catechinsrelated antioxidant activity and metal content of green tea infusions.
In recent years, near-infrared spectroscopy (NIRS) with electromagnetic waves in the wavelength range of 780-2526 nm has been utilized to classify special-grade teas by quantifying indicative bioactive compounds. This mechanism is based on the response of the hydrogen-containing groups (e.g., C-H, N-H, and O-H) to NIR light. Accordingly, NIRS can be a beneficial tool for predicting tea polyphenols, catechins, theaflavins, caffeine, and chlorophyll [36][37][38]. Moreover, researchers realized that the use of NIRS or Fourier transform-NIRS (FT-NIRS) coupled chemometrics algorithms can increase the content prediction accuracy of a wide range of tea chemicals, such as phosphorus, potassium, catechins, amino acids, caffeine, theaflavins, etc. [39,40]. Dong et al. [41] found that the use of NIRS in combination with synergy interval partial least-squares regression (SI-PLS) can assess the brightness and taste intensity of a Congou black tea infusion, by predicting the ratio of theaflavins to thearubigins during fermentation. Moreover, Wang et al. [42] and Guo et al. [43] have recently evaluated the quality control of matcha tea with the accurate ratio prediction of polyphenols to amino acids using NIRS coupled with different chemometric algorithms.

Electronic Sensor Technologies: Advantages and Disadvantages in the Tea Industry
An artificial olfactory system or e-nose, similarly to chromatographic methods, can identify volatile compounds of tea using a gas sensor array. Furthermore, this technology, compared to GC-MS, is not time-consuming, high-cost, strenuous, restricted, and troublesome, and does not require complicated sample preparation [44,45]. This non-invasive, intelligent online instrument is usually coupled with a variety of materials/sensing principles, such as quartz crystal microbalance, conducting polymers, amperometry, electrochemical, surface acoustic wave, and metal oxide semiconductor (MOS; e.g., nickel, tin, cobalt, titanium, zinc, and iron oxides) sensors. Hence, the type of materials and sensors used can greatly affect the sensitivity, selectivity, efficiency, and response speed of each e-nose [28]. It was reported that the use of MOS sensors compared to the conventional method had better sensitivity to the aroma compounds of tea with long-term stability [46]. Furthermore, an e-tongue can rapidly imitate the sensory awareness of human taste by utilizing the fingerprints of response signals of tea samples [47]. These intelligent instruments are made of a sensor array to detect chemicals, accompanied by modern pattern recognition systems (PRSs). As a result, the sensor signal data can be spontaneously and authentically processed [48]. Therefore, the efficiency and speed of online quality control and inspection of tea powders and beverages using these ESPs can be significantly improved.
Using an e-tongue can also prevent the accuracy loss of sensory panelists' scores, due to their overwork and neural fatigue [29,45]. An e-eye is a bionics-based diagnostic tool for analyzing overall color and visual data. This fast technology, not only classifies different tea samples, but also determines the product quality based on their appearance-related characteristics [49].
Overall, ESPs as portable micro-instruments can assess each quality factor, without any difficult pretreatment. The same analysis using traditional detection techniques needs detailed operators and chemical reagents [27]. However, one of the disadvantages of the MOS sensor array is the high operating temperature (150-400 • C). Therefore, these devices remarkably increase consumable energy levels and require a relatively long time for heating before measuring [28]. Moreover, potentiometric sensors are the most extensively used type in e-tongues, and these platforms only respond to the charged species in solutions, limiting the count of potential analytes. Therefore, controlling the temperature and washing the electrodes are necessary to minimize the negative effects of temperature variations and the adsorption of components present in the solution on the membrane potential [27][28][29]. Moreover, another drawback of e-tongues is the relatively short lifespan of sensor materials, mainly biomaterials. Under this condition, the frequent examination of e-tongue performance is not only essential, but samples with large sizes are also often required [28,44,47]. The efficiency of e-eye systems is a function of the selection and calibration of the system components. Poor or inconsistent lighting considerably affects the acquired image quality, whereas a high-quality image provides a lower complexity and less time for image processing. Hence, assessments under this status may be more uncertain and the instrumental resolution of a signal worse. It has also been demonstrated that the performance of e-eyes can be highly influenced by processing or environmental conditions [27,28,49].
However, the achievement of precise quantitative data is needed, to find the significant associations between quality parameters and ESP signals, by employing proper algorithms [50,51]. In addition to the necessity for time-consuming trial and error for the identification of the best multivariate supervised statistical techniques, another shortcoming is the lack of optimization studies with ESPs using efficient algorithms to analyze tea quality factors [52]. This is a serious need, because a large quantity of redundant information as noise will result from the sensor array, which in turn can cause some delays due to the modeling complexity and the inadequacy of output models. Some reasons for noise formation in ESPs, such as an e-nose, include (i) the accuracy reduction of the prediction models of the mode, owing to the weakness in the response signal intensity for volatile compounds; (ii) the reduced model stability for low signal to noise ratios of e-noses under harsh conditions, such as high temperature and relative humidity; and (iii) a dissimilar sensor array with high response signal for the various volatile constituents of tea [53][54][55]. Thus, optimizing sensors toward a more stable and accurate assessment, to reduce inaccurate output information, can play an important role in better modeling quantitative data to analyze tea quality factors.

Electronic Nose Sensors
E-nose comprises a sensor array and a PRS. Similar to the sensory perception of human smell, the sensor array with MOS gas sensors produces a characteristic fingerprint after receiving and reacting aromatic compounds. In addition, the PRS, similarly to the human brain, can differentiate samples with diverse fingerprints [56]. Therefore, this technology, after integrating mathematical methods (Table 1), can concurrently identify volatile compounds as complex odors in the headspace of food samples such as tea. In this context, the volatile compounds present in the food aroma can be converted into electronic signals from the e-nose array, in the form of digital output. Since each volatile organic compound shows an exclusive pattern, a set of these patterns can be applied to diagnose quality changes and for discrimination purposes. Table 1 shows a set of e-noses used to evaluate the sensory quality of various tea samples. E-noses were used to identify and discriminate different tea types, according to the origin, the optimum fermentation time, quality grades, online monitoring of bitterness and astringency levels, and adulteration degree based on the mix ratios ( Table 1). Most of the gas sensors used to diagnose the tea quality were PEN 2 (10 MOS sensors), PEN 3 (10 MOS sensors), Figaro gas sensors (Osaka, Japan; 5 MOS sensors) of different models, two Alpha-MOS e-nose instruments (Fox 4000 (18 MOS sensors), and a Fox 3000 (12 MOS sensors)). Recently, Hidayat et al. (2019) designed a lab-made e-nose with eight MOS gas sensors (Taguchi, TGS, series), to assess tea quality during processing in the factory production line. These newly developed sensors, during the three stages of delay, sampling, and purging, could perform real-time analysis of tea samples at the temperature of 21 • C and relative humidity of 10% ( Figure 2) [57]. Tseng et al. [58] also fabricated a new e-nose for the quality control of oolong tea. This sensor consisted of a sensor array, an adsorbent, a micro-pump with solenoid valves, and temperature and humidity sensors ( Figure 3). This e-nose could be effectively used to differentiate volatile compounds and control the fermentation process in the production line of the tea factory.
Biosensors 2022, 12, x FOR PEER REVIEW 6 of 26 relative humidity of 10% ( Figure 2) [57]. Tseng et al. [58] also fabricated a new e-nose for the quality control of oolong tea. This sensor consisted of a sensor array, an adsorbent, a micro-pump with solenoid valves, and temperature and humidity sensors ( Figure 3). This e-nose could be effectively used to differentiate volatile compounds and control the fermentation process in the production line of the tea factory. , the main part of e-nose device equipped to a sampling system having two electronic valves (three-way system) for the airflow control (B), and the sample chamber (C), and (b) a graphic diagram of the e-nose tool (DAQ is the data acquisition unit). Reprinted from Hidayat et al. [57].

Figure 3.
A developed gas-sensing system for smell discrimination and fermentation monitoring in the on-line production of oolong tea. The image was retrieved from Tseng et al. [58] with some modifications.

Figure 2. (a)
A lab-made e-nose with eight MOS gas sensors used in the production line of a tea factory, consisting of a computer with chemometric tools (A), the main part of e-nose device equipped to a sampling system having two electronic valves (three-way system) for the airflow control (B), and the sample chamber (C), and (b) a graphic diagram of the e-nose tool (DAQ is the data acquisition unit). Reprinted from Hidayat et al. [57].
Biosensors 2022, 12, x FOR PEER REVIEW 6 of 26 relative humidity of 10% ( Figure 2) [57]. Tseng et al. [58] also fabricated a new e-nose for the quality control of oolong tea. This sensor consisted of a sensor array, an adsorbent, a micro-pump with solenoid valves, and temperature and humidity sensors ( Figure 3). This e-nose could be effectively used to differentiate volatile compounds and control the fermentation process in the production line of the tea factory.

Figure 2. (a)
A lab-made e-nose with eight MOS gas sensors used in the production line of a tea factory, consisting of a computer with chemometric tools (A), the main part of e-nose device equipped to a sampling system having two electronic valves (three-way system) for the airflow control (B), and the sample chamber (C), and (b) a graphic diagram of the e-nose tool (DAQ is the data acquisition unit). Reprinted from Hidayat et al. [57].

Figure 3.
A developed gas-sensing system for smell discrimination and fermentation monitoring in the on-line production of oolong tea. The image was retrieved from Tseng et al. [58] with some modifications.

Figure 3.
A developed gas-sensing system for smell discrimination and fermentation monitoring in the on-line production of oolong tea. The image was retrieved from Tseng et al. [58] with some modifications.  Currently, gas-phase e-nose (GP-e-nose) technology, along with multivariate statistical analyses (MVSA), has been utilized to determine the objective sensory evaluation of teas, with analysis of their aroma quality. Yan et al. [82] analyzed the volatile compounds of green tea using GP-e-nose and GC-ion mobility spectrometry (GC-IMS), coupled with MVSA (such as PLS-discriminant analysis (PLS-DA) and principal component analysis (PCA)). Although 33 aromatic compounds were identified, the most important aromatic ingredients were 3-methylbutanal, methyl benzoate, isopropyl alcohol, and heptanal. They also reported that the profile of volatiles was dynamically changed with an enhancement in the drying temperature. However, the GP-e-nose rapidly and successfully characterized the dynamic changes induced by high drying temperatures and complemented the findings obtained from the GC-IMS analysis. In addition, the integration of GP-e-nose and MVSA (PLS-DA and Fisher discriminant analysis (FDA)) has been used to evaluate the aroma quality of Dianhong black tea infusions [80]. They showed that aldehydes were the most frequent compounds identified among 61 volatiles. Accordingly, aromatic compounds of furan, methyl acetate, 2,3-pentanedione, limonene, and linalool had a positive correlation with the aroma quality of tea infusions, where the aroma quality was negatively influenced by the presence of 3-ethylpentane, 1-pentene, (E)-2-hexene, and methyl eugenol compounds [80]. Yan et al. [81], using an e-nose coupled with GC-MS, could analyze the major volatile aromatic components and aroma variances of black tea obtained from two tea cultivars (i.e., Fuyun 6 (mainly, sweet aroma) and Jinguanyin (mostly, floral aroma)). There were nine different aroma compounds between the two cultivars, including heptaldehyde, carveol, hexanal, furfural, 2,2,6-trimethylcyclohexanone, 2-pentadec-2-enylfuran, cis-3hexenyl acetate, methyl salicylate, and 5-methyl-2-hexanone. The e-nose results proved that five out of the ten sensors could differentiate the two cultivars, particularly, S2 (sulfide and hydrogen sulfide), S6 (biogas, methane, and hydrocarbons), S7 (combustible gases), and S10 (combustible gases and alkanes). Recently, Song et al. [83] successfully applied an intelligent fixation as the first step in the appearance and fragrance formation of green tea. The online measurement of aromatic compounds using an e-nose and a set of other sensors coupled with fuzzy logic algorithms was performed to assess the fresh fragrance, enzymatic reaction, and charring occurrence. Moreover, Sanaeifar et al. [84], with the integration of an e-nose and confocal Raman microspectroscopy, could guarantee the tea quality, by monitoring its safety in terms of determining the pesticide residue concentration of chlorpyrifos in tea processing. Thus, it was demonstrated that the use of an e-nose coupled with chemometric strategies represented a useful tool to concurrently monitor the safety and quality of tea.

Electronic Tongue Sensors
An e-tongue not only can imitate the human tongue, but also can present a more sensitive, automated, and unbiased representation of tastes than human tasters. This intelligent technology comprises an array of liquid sensors to analyze liquid-based food products (such as vegetable oils) and beverages, an appropriate method of pattern recognition, and multivariate calibration for data processing. This emerging tool can be utilized after transforming gas and solid materials into their liquid forms or extracts. Therefore, the sensors can easily respond to solubilized tastes, by transducing a signal or a pattern of signals recognized by the software system [28,85]. Recent two-decade studies on the use of e-tongues in measuring the quality of tea types are summarized in Table 2. In general, the e-tongue systems were voltammetric [68,[86][87][88][89][90][91][92][93][94], potentiometric [29,95], and fluorescent [45] sensor arrays. In recent years, the commercial SA-402B [96,97] and TS-5000Z [98][99][100] taste sensing systems (Insent Inc., Japan) have been applied to detect the taste quality of different tea samples (Figure 4a,b). These sensors are constructed based on Prof. Kiyoshi Toko's idea, with different potentiometric electrodes with lipid-polymeric membranes. One of the other e-tongue systems used to analyze tea taste is the Alpha MOS ASTREE (Figure 4c) [27,47,101]. The sensing electrodes work based on the potentiometric measurement principle, to analyze the taste of liquid-based products or solids dissolved in a liquid [29]. Thus, the sensors used in e-tongue systems are mostly potentiometric to assess the taste of a broad range of teas.
Overall, the maximum number of cross-sensitive electrodes is seven, to detect tastes of umami, richness, sourness, saltiness, sweetness, bitterness, and astringency [27,95,101]. However, He et al. [29] used a potentiometric sensor array along with PCA to classify tea samples by evaluating ten organoleptic characteristics, including heavy, thick, sweet, fresh, mellow, fired, and stale flavors, as well as astringency, tenderness, and purity. The application target of e-tongue systems is the quality prediction of tea from various geographical origins and grade levels, with much better discrimination compared to the conventional methods. Yan et al. [101] showed that the integration of an e-tongue with seven independent sensors and chemometrics could efficiently diagnose the geographical origins of Anji-white tea. In addition, He et al. [29] discriminated black and green teas from various geographical origins and quality grades using an e-tongue technology based on the prediction of sensory attributes. Zhang et al. [102] recently reported the analysis of etongue data to classify tea according to semi-supervised learning of generative adversarial networks. The results proved that this approach could more effectively promote the accuracy of classification than a multi-class support vector machine (SVM), PLS-DA, and decision tree. Huo et al. [103], using an artificial nose and tongue with to colorimetric sensor arrays, could successfully classify green tea harvested from the same geographical origin using a hierarchical cluster analysis (HCA) algorithm.
Cheng et al. [104] assessed the use of an e-tongue and GC-MS to evaluate the effect of storage age on the metabolic profile and taste quality of a 20-year sequence of aged Qingzhuan tea. They identified 47 bioactive compounds responsible for the age variation of tea quality. They were formed through different biochemical mechanisms, such as methylation of catechins, glycosylation of flavonoids, formation of theabrownins, biosynthesis of triterpenoids, and degradation of flavoalkaloids. Although the content of polyphenols and flavonoids was enhanced at the 10th year of aging, the amount of theabrownins reached the maximum level at the 15th year of aging. Moreover, not only new triterpenoids were formed at the end of aging time, but also the quantity of initial triterpenoids over the aging time was significantly increased. Accordingly, the authors reported the improved taste of Qingzhuan tea after 10 years of storage with a decrease in bitterness and astringency values and an increase in the main quality-related bioactive compounds. Li et al. [105], using an e-tongue system and an ultrahigh performance liquid chromatograph-Orbitrap-mass spectroscopy, discriminated the taste quality and chemical profiles of 28 crush-tear-curl black tea samples, collected from six geographical regions. The results showed that teas collected from Sri Lanka, North India, China, South India, and Kenya had tastes being more umami and astringent; more umami, more sweetness, and astringent; moderate umami and sweetness; and moderate umami and astringent, respectively. The multivariate statistical analysis also revealed that teas contained 78 metabolites such as amino acids, pigments, and flavone/flavonol glycosides.
The use of e-tongues and e-noses has led to the improved performance and accuracy of tea grading and classification, using the simultaneous analysis of odor and taste. It was concluded that these combined ESPs, rather than the single systems, can significantly promote the clustering and classification rates of tea [47,68,71,90,92,103,106]. However, the classifier type for determining the best prediction model was varied ( Table 2). The concurrent application of an e-tongue and e-eye based on deep learning has recently led to great achievements in rapidly discriminating Pu-erh tea stored for different times, from 0 to 8 years [92]. Zhi et al. [47] earlier realized that a synergy of modeling feature fusion (fuse the time-domain and frequency-domain based features) and decision fusion (D-S evidence to associate the classification findings from multiple classifiers) of an e-nose and e-tongue could meaningfully improve the accuracy rate for quality prediction of teas with four different grades.      Black tea "qi men" Self-designed e-tongue device, 6 various cylindrical metal electrodes (outside), and a Ag/AgCl reference electrode (inside) SRD, PLS-DA, SRD-PLS-DA High efficiency and capability to identify the tea sample grade using e-tongue data The potential and effectiveness of the PLS-DA-SRD model for tea grade classification [110] Black tea Cyclic voltammetry e-tongue (CVET) with an glassy carbon/platinum electrode Negatively association between the bitterness and aftertaste-bitterness and the content of polyphenols, flavonoids, and polysaccharides of dark teas [98] Congou black tea SA402B (Insent, Japan), 6 taste sensors array, an Ag/AgCl electrode ACO, ELM, LS-SVM, PLS-DA, SVM The taste assessment potential of tea products in the actual production process Introducing ACO optimization algorithms for the best combination of taste features of the sensor array [96] Yellow tea TS-5000Z (Insent, Japan), 5 taste sensors array PCA, PLS-DA, HCA The correlation determination of taste types and biochemical compositions of tea The exact evaluation of taste properties (i.e., sweetness, umami, bitterness, astringency, and richness) [99] Autumn green tea TS-5000Z (Insent, Japan), 6 lipid membrane sensors OPLS-DA, HCA Detecting the improved taste of tea during fermentation The dominant taste (strong umami taste) assessment due to the presence of theabrownins [100] Pu-erh tea   In addition, the use of various pattern discriminant algorithms has recently contributed to the recognition of associations between the tastes and biochemical profiles of tea [27,98,99]. Using a combination of non-targeted metabolomics and an e-tongue, Cheng et al. [98] found some associations between different tastes (i.e., bitterness, aftertastebitterness, astringency, and aftertaste-astringency) and the 49 chemical compounds of five dark teas. The results showed that the bitterness and aftertaste-bitterness decreased in the presence of polyphenols, flavonoids, and polysaccharides, whereas the content of theabrownins increased these tastes in dark teas. On the other hand, the aftertasteastringency had a negative relationship with theabrownins, while this aftertaste was directly associated with the content of polyphenols and flavonoids. Wei et al. [99] assessed changes in chemical ingredients in yellow tea during its processing. Based on the results of PLS analysis, they found that the umami and sweetness tastes were more influenced by amino acids and sucrose, but the bitterness and astringency of yellow tea were enhanced in the presence of catechins and phenolic acids. Moreover, the red value of tea leaves was positively associated with the content of phaeophorbides. Moreover, Xu et al. [27], using a triple ESP of an e-nose, e-tongue, and e-eye, not only could qualitatively assess the quality of tea, but also efficaciously predicted the content of the main chemical compounds, including amino acids, catechins, polyphenols, and caffeine, in Longjing green teas.

Electronic Eye Sensors
Nowadays, the use of accurate and non-contact color measuring tools is necessary to assess the processing and quality of food products. Monitoring image information via digital detection and analysis using an e-eye instrument is possible. This technology is able to extract quantitative color data from selected areas of digital images, so that, via image processing, it can successfully distinguish inhomogeneous shapes and colors [111]. An e-eye undertakes techniques including acquisition, processing, and analysis of images. Overall, a camera takes the evaluated object's reflected light and transfers it into electrical analog signals. After that, a computer processing system obtains the target characteristic information, selects the areas of interest, and classifies them into background and target images. The image segmentation process, which makes it possible to obtain the region of interest containing chemical information, can be performed by thresholding, edgebased segmentation, or region-based segmentation. The color information is the basis for analytical factors, whereas qualitative or quantitative analytical data can be extracted by using single calibration, pattern recognition, and multivariate analysis [27,111,112].
In various studies, e-eyes along with other ESPs (i.e., e-nose and e-tongue) have been applied to assess tea quality. Yang et al. [92] recently evaluated the simultaneous application of an e-tongue and e-eye based on deep learning, for differentiation of Pu-erh tea storage time. The applied e-eye system comprised four main parts of a five-million electronic eyepiece, the support, a light-emitting diode (LED) lamp, and an adaptor for the LED lamp. They proved that feature-level fusion, based on a deep learning algorithm, had a more accurate and robust classification performance, leading to a rapid detection, to discriminate the storage time of Pu-erh tea. Wang et al. [49] monitored withering conditions of the leaf stacks during black tea processing using a fusion of an e-eye, micro-NIRS, and colorimetric sensing array. Although the multi-technology fusion system designed in this study could not be utilized for in situ and intelligent control of the withering process of black tea leaves, the performance of this process in maintaining the quality of black tea was significantly improved.
Xu et al. [27] earlier revealed that the concurrent use of an e-nose, e-tongue, and e-eye integrated with suitable algorithms would contribute to detecting tea quality qualitatively and quantitatively. They explained that the prediction algorithms of SVM and random forest (RF) had superior performance compared to partial least squares regression (PLSR). Moreover, the regression models of RF, according to fusion signals could lead to the best prediction findings for the content of chemical ingredients. Xu and Wang [112] assessed the aroma, taste, and color signals of tea using the combined technology of an e-nose, e-tongue, and e-eye. Better performance in analyzing tea quality was obtained using fusion signals compared to the individual signals in SVM, RF, and PLSR models. In a comprehensive study, Wei et al. [99] recently monitored dynamic color changes and the significant presence of components related to umami and sweetness tastes (such as amino acids, phenolic acids, catechins, and sucrose) in tea leaves during the first 18 h of the yellowing process using an e-eye, e-tongue, and metabolomics. The yellowing process analysis after 24 h showed decreased amounts of the abovementioned compounds and increased levels of betaine, piperidine, theasinensin B, and lysophosphatidylcholines. Furthermore, the e-nose system exhibited a significant reduction in the levels of glycosidically-bound volatile compounds (i.e., benzylethyl primeveroside, linalool primeverosides, geranyl primeveroside, and linalool oxide primeverosides) in the tea yellowing step [99]. Overall, it can be concluded that the combined ESP-based technological strategy can provide an alternative system for the online management of sensory quality attributes of various food and beverage products [27,99,112].

Data Analysis and Classification Algorithms
Before the final analysis, raw data are pre-processed using a variety of methods to transform the dataset to present a better input to the PRSs, in terms of averaging, linearization, or normalization [48]. Some structural problems in the design of e-noses, such as the low performance of pumping systems, noisy recordings of gas sensors, and the remaining aroma compounds after the cleaning phase, negatively interfere with the diagnosis of the sensor array's response to the analyte [33]. However, the determination of the best preprocessing method to improve the efficiency of the pattern recognition stage through cleaning or denoising signals is different, from one type of sensor system to another. Some of their most practical are the linear algebraic techniques of singular value decomposition (SVD), standard normal variate (SNV) transformation, background or baseline subtraction, auto scale, and baseline subtraction plus auto-scale [113]. After this necessary stage, the clustering within the datasets is mainly performed using PCA, fuzzy c-means (FCM) algorithm, and self-organizing map (SOM). Meanwhile, FCM and SOM algorithms were applied in unusual and older studies [59,114], while a PCA analysis was conducted in most ESP-based works, to cluster sensory attributes of teas with different origins, processing, and harvest and storage times (Tables 1 and 2). The PCA not only visualizes all the information available in a dataset, but also contributes to differentiating different samples according to price, environmental, chemical, and operational variables [69]. Other unsupervised methods (such as the nonlinear mapping technique of HCA) and supervised algorithms (such as LDA, PLS-DA, SVM, and KNN) were used to classify tea samples during the monitoring of their quality using ESPs. Regression algorithms of PLS and PLSR (linear), artificial neural networks (ANNs; such as MLP, LVQ, PNN, and RBF (nonlinear)), and support vector regression (SVR, nonlinear) in combination with ESPs could well classify and simplify the data of the sensory properties of tea. Furthermore, using a back-propagation neural network (BPNN) and RF models as supervised learning algorithms showed a good ability to lessen the mean square error between the computed and desired outputs of the network, to discriminate teas with dissimilar attributes (Tables 1 and 2).
The quality of food materials such as tea can be quantitatively determined using proper algorithms, by discovering the correlations between ESP signals and quality indices. On the one hand, a high number of chemical compounds are involved in the sensory quality of tea; while, on the other, most of these ingredients are related to each other in producing the aroma, taste, and color quality of tea. Thus, the fusion signals of ESPs might be efficient in assessing the content of chemical compounds in tea types [46,112]. Since the main part of human perception is a consequence of multiple sensory organs, data fusion can deliver more accurate information about organoleptic characteristics compared to a single ESP. In general, the fusion is classified into three groups: low-level (data fusion, DaF), intermediatelevel (feature fusion, FeF), and high-level (decision fusion, DnF). The DaF integrates the raw sensor response signals into a sole signal, whereas the features extracted from the sensor signal amounts in FeF are interconnected with a diversity of feature extraction and selection techniques. Finally, DnF, by combining decisions from several sensory channels, exhibits the output of multiple classifiers, to attain an ending forecast, which is similar to the fusion structure in the human brain. To sum up, fusion contributes to a better description and more accurate classification of data by combining the information obtained from multiple sources [47,115]. The best classification models for discriminating tea quality grades based on fusion signals with an accuracy of over 90% were SVM and RF [27], K-nearest neighbors (KNN) [47], SVM [44,45,96], extreme gradient boosting (XGBoost) [77], kernel-based LDA (KLDA)-KNN [69], fuzzy neural network (FNN) [68], the Bayesian approach [48], and PCA and KNN (71). Moreover, Yang et al. [77] recently determined the feasibility of developing a cross-category model for predicting tea polyphenols based on the feature fusion of an e-nose and hyperspectral imagery (HIS). They confirmed that the accuracy of the fusion features, according to frequency-domain feature (FdF) and time-domain feature (TdF), obtained from the e-nose and HIS systems was much more than the features from each sensor. Among the three regression models of RF, XGBoost, and Grid-SVR, the efficiency of XGBoost as a decision-tree-based ensemble machine learning algorithm, compared to the two other models, was higher due to the increase of estimation accuracy for tea polyphenol content ( Figure 5).
DaF integrates the raw sensor response signals into a sole signal, whereas the extracted from the sensor signal amounts in FeF are interconnected with a div feature extraction and selection techniques. Finally, DnF, by combining decisio several sensory channels, exhibits the output of multiple classifiers, to attain an forecast, which is similar to the fusion structure in the human brain. To sum up contributes to a better description and more accurate classification of data by co the information obtained from multiple sources [47,115]. The best classification for discriminating tea quality grades based on fusion signals with an accuracy of o were SVM and RF [27], K-nearest neighbors (KNN) [47], SVM [44,45,96], extreme boosting (XGBoost) [77], kernel-based LDA (KLDA)-KNN [69], fuzzy neural (FNN) [68], the Bayesian approach [48], and PCA and KNN (71). Moreover, Ya [77] recently determined the feasibility of developing a cross-category model for ing tea polyphenols based on the feature fusion of an e-nose and hyperspectral (HIS). They confirmed that the accuracy of the fusion features, according to fre domain feature (FdF) and time-domain feature (TdF), obtained from the e-nose systems was much more than the features from each sensor. Among the three re models of RF, XGBoost, and Grid-SVR, the efficiency of XGBoost as a decision-tre ensemble machine learning algorithm, compared to the two other models, was hig to the increase of estimation accuracy for tea polyphenol content ( Figure 5).

Conclusions and Future Trends
This overview showed that the use of ESPs as non-destructive, and autom struments not only accelerates the online monitoring of tea quality but also sign

Conclusions and Future Trends
This overview showed that the use of ESPs as non-destructive, and automated instruments not only accelerates the online monitoring of tea quality but also significantly reduces the assessment time and cost for the specialists and personnel of tea factories globally. The use of these intelligent devices is an efficient alternative for testing experts, to determine tea quality through the organoleptic evaluation of aroma, taste/aftertaste, color, and mouthfeel scores. In general, ESPs were used for the qualitative and quantitative classification of tea samples, with the same/various geographical origins, grades, as well as harvest and storage times. The analysis of output data from each electronic sensor system using machine learning algorithms and statistical prediction models could adequately classify tea based on the target quality properties. The optimal prediction model for tea quality discrimination was varied, because diverse feature selection algorithms were examined on dissimilar datasets. Furthermore, the combination of signals obtained from two or three ESPs (e-nose, e-tongue, and e-eye) with a feature-level fusion strategy could significantly enhance the efficiency of identification and prediction models. This quality classification led to the recognition of biochemical compounds and their relationships with the profile of taste, color, and aroma changes during storage and processing.
According to the promising results of using ESPs in maintaining and improving the quality of tea, similar e-sensors integrated with superior feature representing methods can be used to predict and classify the sensory quality of other food products (such as fruit juices, soft drinks, plant and animal milk, and vegetable oils), for quickly satisfying the demands of both consumers and the market. There has been tremendous growth in presenting machine learning methodologies in classifying tea quality. It is expected that the new, efficient frameworks can extract deep features from various sensor signals, to perform tea authentication with more sensitivity and accuracy in the future. Meanwhile, the capability of diverse kinds of features can be assessed using various signal transformation methods. Since the combination of triple ESPs has been performed less frequently, the simultaneous application of an e-nose, e-tongue, and e-eye may result in all-around data support for a wide-ranging quality assessment of tea and other foods/beverages. Moreover, it is interesting to compare the results obtained from ESPs, (FT-)NIRS, HPLC-diode-array detection (DAD) fingerprints coupled with chemometric analysis, etc., to correctly evaluate the performance of e-sensors in the quality control of tea samples. A serious effort should be made to design and develop advanced micro-and nano-e-sensors, followed by their comparison with smartphone-based micro-NIRS. Another important issue is the production of self-adhesive bioelectrodes in the structure of ESPs, for the stable detection of sensory attributes under harsh environmental conditions, so that they can sensitively simulate human physiological signals for food quality assessment. Last, it is recommended to compare the online variation in healthy-functional compounds and health benefits of tea using single and combined EPSs in the future.