Sensory Perception Systems and Machine Learning Methods for Pesticide Detection in Fruits

Abstract

In this study, an electronic tongue (E-tongue) and electronic nose (E-nose) systems were applied to detect pesticide residues, specifically Preza, Daconil, Curzate, Bricol, Accros, Amistar, and Funlate, in fruits such as cape gooseberries, apples, plums, and strawberries. These advanced systems present several advantages over conventional methods (e.g., GC-MS and others), including faster analysis, lower costs, ease of use, and portability. Additionally, they enable non-destructive testing and real-time monitoring, making them ideal for routine screenings and on-site analyses where effective detection is crucial. The collected data underwent rigorous analysis through multivariate techniques, specifically principal component analysis (PCA) and linear discriminant analysis (LDA). The application of machine learning (ML) algorithms resulted in a good outcome, achieving high accuracies in identifying fruits contaminated with pesticides and accurately determining the concentrations of those pesticides. This level of precision underscores the robustness and reliability of the methodologies employed, highlighting their potential as alternative tools for pesticide residue detection in agricultural products.

1. Introduction

Agriculture is the pillar of human and economic development worldwide [1]. Throughout history, agriculture has been a primary source of food supply for humanity since all countries depend on it for their food. Therefore, as the population increases, the demand for nourishment continues to grow exponentially, which generates a request for land to develop food products and thus meet the food needs worldwide [2,3]. The above allows for us to accomplish the second sustainable development goal (SGD2), which consists of ending hunger, achieving food security and improving nutrition, and promoting sustainable agriculture [4,5,6]. According to Food and Agriculture Organization of the United Nations (FAO) data, global food production must increase by 70% to feed the world population by 2050 [7].

Nowadays, rising crop production often requires external inputs, such as pesticides, which have been widely applied to protect foods against pests to improve their quality production and increase crop efficiency [8]. According to estimates by FAO, approximately 45% of annual food production is lost due to pest infestation, threatening food security and causing economic losses due to the resources used to produce these foods (i.e., water, land, energy, labor, and capital) [9,10]. Therefore, effective pest management through a wide range of pesticides is required to confront them and enhance crop production, as mentioned earlier [11].

Currently, in the world, three billion kilograms of pesticides are consumed yearly, of which 47.5% are herbicides, 29.5% insecticides, 17.5% fungicides, and 5.5% other pesticides [12,13]. However, only 1% of total pesticides are used effectively to control pests on target plants [14]. Additionally, 99% of the pesticides used are lost through spray drift, off-target deposition, runoff, and photodegradation, resulting in pesticides entering the environment, where they negatively affect the environment. Soil, water, and vegetation can remain in crops and finally enter the food chain, posing a threat to humans [15,16].

Human beings are exposed to pesticides directly in the workplace or indirectly through environmental media. When handling and using pesticides to irrigate and fumigate crops, farmers are in direct contact with them through inhalation, ingestion, and dermal contact, which causes damage to the skin, eyes, mouth, and respiratory tract [17,18], having to deal with reactions such as headache, irritation, vomiting, sneezing, skin rashes, dizziness, difficulty breathing, neurotoxicity, and chronic diseases such as cancer, with breast cancer being the most common of all and it is associated with organophosphates (malathion and parathion) that affect cell growth and proliferation and incidences of death [19]. It is essential to highlight that the intensity of the effects of pesticides on humans depends on the exposure time, toxicity, and concentration [15]. According to the World Health Organization (WHO), cases of pesticide poisoning have been identified, generating an actual public health problem in the world, especially in developing countries, where the most significant number of cases occur [20,21]. In Colombia, according to the National Institute of Health and the National Public Health Surveillance System (SIVIGILA), in 2020, 17,270 poisonings were reported, mainly due to psychoactive substances, medications, and, thirdly, pesticides, where 37.7% occurred accidentally, and 26.7% with psychoactive intentionality. Colombia also has no exact data on mortality due to pesticide poisoning [22].

Another way of exposure to pesticides is indirectly through environmental media, such as air, water, soil, and the food chain, which can be contaminated with pesticides [18,23]. Undesirable substances such as pesticides in foods are a critical indicator of food quality and safety [14]. We know that fruits and vegetables are vital elements of a healthy human diet due to their high proportion of fiber, vitamins, and minerals. Therefore, their consumption worldwide has increased considerably in recent years and is crucial to people’s health. However, fresh fruits and vegetables contain higher levels of pesticides, mainly fungicides, and insecticides compared to other plant foods, probably due to the short time interval between harvests and market that results in inadequate testing to detect pesticide residues and creates a risk to the health of consumers, as they are often eaten raw or in the form of processed products (juices, jams, purees, etc.) [24]. Furthermore, the concentration of pesticides in fruits and vegetables depends on the amount sprayed on them, the available content in the soil or irrigation water, and the processing of these foods. The above is a way to guarantee safety, prolong shelf life, and maintain or improve the quality of food, which, according to the specifications of the final product, may be subject to different processes [25]. For example, in the food industry, there are several stages such as pre-treatment (washing, rinsing, peeling, grinding, segregation, etc.), heat treatment (cooking, frying, pasteurization, freeze-drying, baking, sterilization, preservation, etc.), and final treatment (packaging, bottling, labeling, etc.), among others, which allow for the concentration of pesticides to be reduced or eliminated [26].

In 2004, the Environmental Working Group (EWG) began publishing the Dirty Dozen list, which identifies some of the fruits and vegetables that contain the highest levels of pesticide residues, according to data obtained from the United States Department of Agriculture (USDA) and the Food and Drug Administration (FDA). The following foods are on the 2022 Dirty Dozen list: strawberry, spinach, kale, nectarine, apple, grape, bell and hot peppers, cherry, peaches, pears, celery, and tomato [27,28]. Therefore, monitoring and evaluating pesticide levels in fruits and vegetables must be observed. Food industries and stakeholders must ensure that foods comply with legal requirements at all stages of the food chain. To protect human health, legal guidelines have been established to control pesticide levels in food through maximum residue levels, which are set by the Codex Alimentarius Commission (CAC) and the joint FAO/WHO Meeting on Pesticide Residues (JMPR), who have tried to establish, review, and harmonize MRLs for pesticides, where globally, these limits continue to be variable, since many countries have their own regulations [24,29]. In the case of Colombia, there is the Ministry of Agriculture and Rural Development through the ICA (Colombian Agricultural Institute), which, through resolution 2906 of 2007 issued by the Ministry of Social Protection, stipulates the maximum concentrations of commercial pesticides in the food [30].

To evaluate the levels of pesticides in fruits and vegetables, sensitive, highly selective, and precise analytical instrumental techniques are required. Among these most used techniques are high-performance liquid chromatography (HPLC) and gas chromatography (GC) coupled to mass spectrometry (MS). Other methods, such as enzyme-linked immunosorbent assays (ELISA) and capillary electrophoresis (CE) have been applied in the determination of pesticide residues in real samples [31,32]. However, these methods have some limitations since they require sophisticated laboratory equipment, preparation, storage, and transportation of samples, which tends to be tedious, high-cost, time-consuming, and require trained personnel. Nowadays, systems based on arrays of gas sensors (E-nose) and chemical electrodes (E-tongue) have been designed to artificially perceive the active molecules of the odor generated from the headspace of the samples and the compounds involved in the flavor of the aqueous solutions [33,34]. These characteristics of sensory systems have allowed for them to be used in different industries, especially in the food industry, to evaluate the quality of other products through the analysis of odors and flavors [35]. These technologies are typically characterized by their ability to provide simple, rapid, non-destructive, portable, and highly specific detection. They also offer good reproducibility and are cost-effective compared to traditional analytical techniques [36].

There are some reviews and research publications in which the different applications of the E-nose and E-tongue have been reported over the years in the food industry. For instance, these systems have been implemented for the early detection of microorganisms (meat, milk, etc.), authenticity of food and beverages, evaluation of food quality (e.g., fish, wine, meat, beer, milk, water, drinks), monitoring food processing, and the identification of ripening states in fruits such as apples, pears, peaches, apricots, citrus fruits, grapes, strawberries, mangoes, and other tropical fruits [35,37,38,39,40,41,42]. However, using these systems combined with chemometric methods to determine pesticide residues in fruits and vegetables has yet to be explored. Nategh et al., in 2021, used an E-nose based on MOS sensors to detect the diazinon pesticide in cherry samples, obtaining 96% of the variation in the data for toxic and non-toxic sweet cherries [43]. Similarly, Yong and collaborators implemented a commercial system (PEN3) with ten metal oxide semiconductor (MOS) sensors, which was used to detect pesticide residues such as cypermethrin and chlorpyrifos in apple samples, achieving a good classification of the pesticide residues using principal component analysis (PCA), linear discriminant analysis (LDA), and support vector machine (SVM) [44]. Marco et al., used a self-made E-nose to obtain qualitative responses to different pesticides [45]. Likewise, studies have been conducted on detecting pesticide residues in fruits such as strawberries, peaches, and sour cherries [8,46].

Different investigations have been carried out regarding the use of E-tongues, which have been focused on the efficient detection of pesticide residues in fruits. Still, it is worth mentioning that these studies have been carried out using biosensors in most cases. For example, Rajni Bala et al. used an ultrasensitive aptamer biosensor based on cationic polymer and gold nanoparticles for the detection of the organophosphate family pesticide malathion in apples, achieving a detection limit of 5.2 pM (0.001 μg kg⁻¹) [47,48]. In another study, an electrochemical nanosensor based on fluorine-doped tin oxide (FTO) was developed for the detection of chlorpyrifos with gold nanoparticles (AuNP) and anti-chlorpyrifos antibodies (chl-Ab). The FTO-AuNPs sensor was successfully used to detect chlorpyrifos in standard and actual samples up to 10 nM for apple and cabbage and further 50 nM for pomegranate [49]. A monoenzymatic impedimetric acetylcholinesterase biosensor has been reported in the literature to sensitively detect carbamate and organophosphate compounds with a very rapid response [50]. However, with recent advances in nanotechnology, metal-based enzyme-free sensors, and corresponding metal oxides have been widely reported. Recent studies demonstrate that CuO nanostructures have been applied in pesticide detection thanks to their ability to bind to phosphate groups in organophosphates [51].

This study aims to validate the use of E-tongue and E-nose technologies as innovative tools for detecting and classifying pesticide residues in various fruits. The fruits selected are among the most widely cultivated in the Santander region, making them ideal candidates for evaluating the effectiveness and practicality of these devices as alternatives to traditional pesticide detection methods, such as GC-MS. The evaluation focused on the key advantages of the E-tongue and E-nose, including non-destructive analysis, portability, speed, and cost-effectiveness. This study began with training and validation phases, followed by testing with new data to assess the devices’ capabilities. The analysis revealed high accuracy in classifying both the presence of pesticide residues and their concentrations using multivariate analysis and advanced machine learning (ML) techniques. This is particularly significant as it meets the growing demand for more accessible, efficient, and environmentally friendly methods to monitor pesticide residues in agricultural products. Figure 1 shows the methodology developed during the pesticide detection study in fruits. Initially, fruits and pesticides were selected, and later, the E-nose and tongue were conditioned.

2. Materials and Methods

The following section provides a detailed description of the various components and materials used in the experimental trials, offering a comprehensive overview of the resources utilized in this study.

2.1. Selection and Collection of Fruit and Pesticide Samples

The fruit samples for this research were collected in the municipalities of “Cerrito” in the Santander department and the municipalities of “Santa Rosa de Viterbo y Tuta”, located in the Boyacá department (Colombia). The first cape gooseberries were collected in Cerrito, and the second in Santa Rosa de Viterbo. Further, strawberry, apple, and plum samples were collected in Tuta municipality. It should be noted that random batches were taken in each crop to obtain a representative sample for the sample collection of the different fruits.

For this study, two types of samples were used: organic fruits (grown with natural substances without chemicals such as pesticides or artificial fertilizers), which were the control group, and fruits contaminated with the different pesticides used in each region.

The selection of pesticides in this study was conducted through a survey of farmers about the types of pesticides used for cape gooseberry, strawberry, apple, and plum crops in each region, highlighting Preza, Daconil, Curzate, Bricol, Accros, Amistar, and Funlate, as shown in

Table 1. Information about pesticides according to the selected fruits.

Fruit	Location	Pesticide	Active Substance	Classification	Reference
Cape Gooseberry	Cerrito (Santander)	Preza	Cyantraniliprole	Insecticide	[53]
		Daconil	Clorotalonil	Fungicide	[54]
	Santa Rosa de Viterbo (Boyacá)	Curzate	Mancozeb	Fungicide	[55]
			Cymoxamil
		Bricol	Carbendazim	Fungicide	[56]
			Flutriafol
Strawberry	Tuta (Boyacá)	Accros	Difenoconazole	Fungicide	[57]
		Bricol	Carbendazim	Fungicide	[56]
			Flutriafol
Plum Apple		Amistar	Azoxistrobina	Fungicide	[58]
		Funlate	Benomyl	Fungicide	[59]

. The necessary information is given here about the classification of pesticides and the active ingredients, which are the chemical substances in pesticides that control pests [52].

2.2. Sample Preparation

This study was divided into two parts: (1) The first phase consisted of the training of multisensory systems to know the ability of the E-nose and E-tongue to discriminate and classify samples of organic fruit and contaminated fruit with different pesticides, which were evaluated according to the information shown in Table 1. At this stage, a 10 mg/kg solution was prepared for each pesticide, and regarding the selected technique, the fruits were contaminated by spraying in the case of the E-nose, and for the E-tongue, it was added directly to the liquid extract of the previously macerated fruit. (2) In this stage, different solutions were prepared at concentrations of 0.2, 0.8, 2, and 3 ppm for each pesticide, to evaluate the detection limit of two multisensory systems. Likewise, the solutions applied to the fruit were made by spraying or liquid extract, as mentioned above.

It is essential to highlight that the organic and contaminated fruits were weighed to maintain the same proportions for each fruit. In the case of the plum, approximately 8 g (±0.09 g) was used, for the strawberry = 10 g (±0.01 g), the apple = 12 g (±0.01 g), and the cape gooseberry = 9 g (±0.09 g), respectively.

For the use of the E-nose, 7 measurements were taken for each category, and 10 measurements were taken for each category using the E-tongue. This included organic and contaminated fruit with different concentrations of each pesticide and fruits used in the study (plum, strawberry, apple, and cape gooseberry). It should be clarified that this study was conducted to distinguish organic fruits compared to fruits treated with pesticides, where each category of measurements was made from different samples (e.g., 10 different plum samples with Daconil).

2.3. E-Nose

A multisensory system was developed to detect pesticide residues in fruits. This system consists of a concentration chamber and a measurement chamber. The concentration chamber is constructed from durable methacrylate and has dimensions of 12 cm in length, 12 cm in width, and 12 cm in height, resulting in a total internal volume of 1728 cm³. The measurement chamber, also made from methacrylate, contains 16 metal oxide gas sensors, specifically Taguchi sensors from the Figaro Company, located in Arlington Heights, IL, USA. This chamber measures 8.7 cm in length, 8.7 cm in width, and 5.3 cm in height, with an internal volume of 401.157 cm³. The sensors are capable of detecting a wide range of compounds, including ammonia, amines, monochlorodifluoromethane, hydrogen, methane, volatile gases, water vapor in food, hydrogen sulfide, hydrocarbons, carbon monoxide, iso-butane, ethanol, alcohol, organic solvents, and tetrafluoromethane (as detailed in

Table 2. Sensor description of E-nose.

Channel	Sensor Type	Gas Target	Detection Limit
S1	TGS 826	Ammonia and amines	30 ppm
S2	TGS 831	R-22 (monochlorodifluoromethane)	10 ppm
S3	TGS 821	Hydrogen	50 ppm
S4	TGS 826	Ammonia and amines	30 ppm
S5	TGS 842	Methane and natural gas	350 ppm
S6	TGS 880	Volatile gases, water vapor in food	500 pm
S7	TGS 825	Hydrogen sulfide	1 ppm
S8	TGS 830	R-22 (monochlorodifluoromethane)	10 ppm
S9	TGS 800	Carbon monoxide, methane, iso-butane, hydrogen, ethanol.	50 ppm
S10	TGS 880	Volatile gases, water vapor in food	10 ppm
S11	TGS 822	Alcohol and organic solvents	50 ppm
S12	TGS 821	Hydrogen	50 ppm
S13	TGS 832	R-134ª (tetrafluoroethane)	10 ppm
S14	TGS 842	Methane and natural gas	500 ppm
S15	TGS 831	R-22 (monochlorodifluoromethane)	10 ppm
S16	TGS 813	Hydrocarbons in general	500 ppm

).

Figure 2 illustrates two simple diagrams created to explain the measurement process of the E-nose.

The sampling protocol for the E-nose consists of 3 stages: (a) concentration, (b) measurement, and (c) cleaning. A quantity of organic fruit was taken at first, and with the help of a spray bottle, the sample was contaminated with the pesticide of interest. Subsequently, the sample was placed in the concentration chamber (see Figure 2A) for 5 min to generate a headspace and concentrate the volatile organic compounds (VOCs) emitted by the contaminated fruit. During 5 min of concentration, valves 1 and 2 were closed; however, an electric pump transfers an airflow that passes through valve 3 in parallel to the chamber where the sensors are located. At the end of the first stage, the sample acquisition stage continued for 7 min (see Figure 2B). In this stage, valve 3 was closed, and valves 1 and 2 were opened, keeping the electric pump on, which allows for airflow to pass to the concentration chamber and thus drag the VOCs from the sample to the sensor chamber. For the cleaning stage of the sensor chamber, valves 1 and 2 were closed, and valve 3 was opened again for 5 min to circulate the air through the sensor chamber. This way, the volatile compounds are extracted from the previously acquired sample, avoiding the memory effect in the measurement chamber due to poor cleaning. The total duration of the measurement was 17 min for each sample.

2.4. E-Tongue

For the study with the E-tongue, a portable reference Bi-potentiostat µSTAT200 manufactured by DropSens Company located in Oviedo (Spain) was used, which operates with a maximum current of ±200 µA and a potential of 12 VDC.

A screen-printed platinum electrode (reference: DRP-550) was applied to measure the samples, and it was composed of three electrodes: the working and auxiliary electrodes, which were developed with platinum material. In contrast, the reference electrode is made of silver, as were the electrical contacts. An electrochemical technique called cyclic voltammetry (CV) was used to acquire and monitor the measurements in real-time through the “Dropview” software version 3.0 provided by the manufacturer.

All voltammograms were obtained by implementing the parameters in

Table 3. E-tongue operation parameters.

z	Assigned Value
Econd [V]	0
Edep [V]	0
Tcond [s]	0
Tdep [s]	0
Tequil [s]	0.3
Ebegin [V]	0
Evtx1 [V]	−1
Evtx2 [V]	1
Estep [V]	0.01
Srate [V/s]	0.05
Nscans	1

, and in all categories, 50 µL of previously macerated liquid samples were taken and subsequently deposited on the working electrode.

2.5. Data Processing

To analyze the measurement acquired, data treatment was previously conducted on the information collected; these were normalized to give a similar scaled value and extract the most relevant characteristics (minimum weight and maximum value of signals) to be later implemented by data processing algorithms.

Therefore, for the data discrimination and classification process, different multivariate analysis methods such as PCA and LDA were used, and machine learning algorithms such as decision tree classifiers, linear discrimination, the naïve Bayes method, quadratic SVM, cubic SVM, and the fine K-NN method; the previous two were applied with the respective kernel. The methods outlined in the previous sections were implemented using the “Classification Learner” interface in MATLAB 2020b. This tool significantly streamlined the development process for classification methods, providing an intuitive and efficient platform for selecting features, training models, and evaluating performance metrics.

It was crucial to undertake a comprehensive process involving several key steps to effectively utilize machine learning algorithms through this tool for detecting fruit pesticides. First, a feature extraction technique based on two parameters was obtained for each signal: the maximum (max) and minimum (min) values of the signal (i.e., voltage for gas sensor signals and current for electrochemical responses), which represent the highest and lowest points that the signal reaches over a given period. This value provided a quantitative measure of the signal’s variability, which was subsequently used for signal analysis by applying PCA and LDA methods and classification techniques, as described above. On the other hand, data normalization was applied to the dataset using the “mean-centering” and “Auto-scaled” algorithms to ensure that each feature contributes equally to the model. Then, these parameters were calculated by subtracting the mean or dividing by the standard deviation for each feature. Afterward, it was necessary to optimize the parameters of ML methods (decision tree classifiers, linear discrimination, naïve Bayes, quadratic SVM, cubic SVM, and fine K-NN) and select appropriate hyperparameters (i.e., kernel) for each method, such as, for cubic SVM and fine K-NN, learning rates, and regularization parameters. Consequently, the models were refined through a meticulous process of fine-tuning. This involved iteratively adjusting the models based on feedback from initial training and validation runs to enhance their predictive accuracy and generalization capabilities.

The training data were randomly divided into multiple subsets to validate the model’s performance across different segments, and regularization helped control model complexity. For example, when k = 5, the dataset is split into five equal parts. Therefore, k-fold cross-validation divides the dataset into k subsets, and the model is trained and tested k times, each time using a different fold as the test set and the remaining folds as the training set. Afterward, test data were used with the trained model to assess the model’s performance and generalization ability. These data were not used in any way during the training or validation phases.

Each of the methods is briefly described below:

• Principal Component Analysis (PCA): A supervised and unsupervised learning method in which observations or data are reordered by decreasing the correlation of all data and choosing new orthogonal axes for each component [60].
• Linear Discriminant Analysis (LDA): A supervised multivariate method developed to obtain good class separability. It is based on the calculation of maximization of the factors concerning the homogeneity in the covariance matrix of each group, and it is decided whether it fits the linear model or the quadratic model [61].
• Naïve Bayes: A probabilistic learning algorithm that bases its mathematical operation on the Bayes theorem. The algorithm calculates the probability of an event “A” given the occurrence of an event B, for it is necessary to have a group of established data, thus making the naïve Bayes method a supervised method [62].
• Support Vector Machines (SVMs): A support vector machine establishes its operation by finding a hyperplane with the best possible separation between a dataset. To do so, it takes the closest data of each class of data and establishes margins; the above serves as a reference point to define said hyperplane and determine the support vector that shows the distance between the margins [63].
• K-Nearest Neighbor (K-NN): This is a regression method based on assumptions of similarity or low dispersion between data. It is also used as a classifier learning method; for this, each group of measurements regarding a number K of nearby data must be compared with its category, which can classify them according to their similarity [64].
• Decision Trees: This method operates on hierarchical classification. Its structure consists of two elements: nodes and branches. In the nodes, each possible result is assessed and weighted against the input values assigned to the group [65].

3. Results

3.1. Data Analysis with Different Pesticides in Fruits

The following section describes the results obtained with both sensory devices from the data processing methods described above.

3.1.1. E-Nose

Figure 3 shows the results of the PCA analysis applied to organic fruit samples and the fruit with pesticides. According to the results, plum reached a variance of 97.27% (Figure 3A), strawberry 96.37% (Figure 3B), apple 97.64% (Figure 3C), and cape gooseberry 96.07% (Figure 3D), achieving good classification and discrimination of organic fruit and contaminated samples. Each value in question is derived from the cumulative variances associated with the first and second principal components (PC1 and PC2). This confirms that the E-nose can detect significant differences in the volatile profiles of fruits depending on the pesticide used. However, in the case of the plum, there was a misclassified measurement between one of the Funlate samples with Amistar, and the same arises with the cape gooseberry, which presented an overlap between the categories of Curzate, Daconil, and Preza, as observed in Figure 3D and Figure 4D based on PCA and LDA analysis. Although Daconil, Curzate, and Preza are pesticides that belong to different chemical classes (organochlorines, acetamides, and anthranilamides) and contain different active ingredients (Table 1) and mechanisms of action, there are some similarities in terms of their chemical properties and their use in pest and disease control in agriculture. The application of these pesticides on fruits may lead to similar patterns of volatile organic compounds (VOCs), which could result in overlaps in PCA and LDA plots when analyzing volatile profiles with the E-nose. This can occur due to the chemical degradation of the pesticides or secondary compounds formed during the application of these products. Additionally, the fruits may emit their own VOCs in response to the chemical or oxidative stress caused by the pesticides. These VOCs may include alcohols, aldehydes, ketones, esters, terpenes, and other compounds that are produced as part of normal metabolic pathways or as a direct response to stress [66,67,68].

Figure 4 illustrates similar behavior using the LDA classifier, where a separation of the organic fruit category is observed concerning fruit contaminated with the pesticide. In the same way, there was an overlap in the case of cape gooseberry with respect to Daconil, Curzate, and Preza. On the other hand, less dispersion in the data concerning that of the PCA was obtained (see Figure 3). The significant overlap between the samples of cape gooseberry treated with Daconil, Curzate, and Preza in the LDA analysis may be attributed to several factors related to the ripening stage, the chemical composition of the pesticides, the physiological response of the fruit to these treatments, the extraction temperature of the VOCs, or the detection capabilities of the E-nose, as the sensors may not be sufficiently specific to distinguish between these compounds. The LDA method was used together in this study, since PCA analysis may be used to reduce the dimensionality of the data, followed by LDA to further reduce the dimensions and optimize the separation between classes for classification tasks. The classification method resulted in a level of 90% accuracy.

3.1.2. E-Tongue

In the PCA analysis, the data obtained by E-tongue were used to achieve high discrimination and separation in the samples. The results from the E-tongue show less scatter in the data than those from the E-nose. This indicates that the compounds detected by the E-tongue are generally more stable, as the interactions between soluble chemicals and sensors are more direct and susceptible to environmental or experimental variability than VOCs. This means that once a sample is prepared and measured, the detected compounds will likely remain more constant, resulting in more reproducible data. Nevertheless, in the case of the apple, an error was generated due to an “outlier” for one of the samples, between the organic fruit category and that contaminated with Funlate (see Figure 5), causing an alteration in the separation between these groups in the PCA space, making it appear as an overlap between the categories. Potential causes of the outlier could include experimental error (such as measurement errors, data recording issues for that specific sample, equipment problems with the E-tongue, cross-contamination, or human errors in sample preparation), biological variability (significant natural variations in chemical profiles due to factors such as ripening, growth conditions, or storage), and chemical interference (presence of a previously undetected compound or an undesired chemical reaction).

It is important to highlight that in all the results obtained with the E-tongue, PC1 explains more than 97% of the variance, which means that a single feature or a set of features is dominant in differentiating the samples. Figure 6 shows the result in the discrimination of the fruits using the E-tongue, where it is possible to observe that the same behavior is maintained in the classification of the four fruits using LDA.

The overlap between the samples of gooseberry treated with Bricol and Curzate in the LDA may be due to several factors related to the fruit’s ripeness, the chemical composition of the pesticides, the fruit’s physiological response to these treatments, the extraction temperature of the VOCs, or the detection capability of the E-nose, as the sensors may not be sufficiently specific to differentiate between these compounds. On the other hand, the LDA classifier obtained 92.65% accuracy.

3.2. Classification of Pesticides in Fruits with Statistical Methods

The accuracies of the methods evaluated with the data acquired through the two devices are found in

Table 4. Classification of fruits with pesticides through E-tongue and E-nose.

Method	Accuracy
	E-Nose				E-Tongue
	Plum	Strawberry	Apple	Cape Gooseberry	Plum	Strawberry	Apple	Cape Gooseberry
Decision trees	76.2%	52.4%	90.5%	65.7%	86.7%	86.7%	83.3%	96%
Linear discrimination	95.2%	81%	90.5%	85.7%	90%	93.3%	93.3%	94%
Naïve Bayes	66.7%	47.6%	90.5%	62.9%	93.3%	93.3%	93.3%	98%
SVM (quadratic)	85.7%	76.2%	90.5%	74.3%	93.3%	93.3%	90%	94%
SVM (cubic)	81%	71.4%	90.5%	65.7%	93.3%	93.3%	86.7%	96%
K-NN (fine)	85.7%	52.4%	90.5%	60%	90%	93.3%	86.7%	96%

, where it can be seen that the E-nose was able to predict the data behavior of different fruits from the linear discrimination classifier, regarding the classification methods, as the classification of plum obtained a value of 95.2%, followed by the apple with 90.5% success rate, where this value is repeated for the rest of the methods evaluated. The information above confirms the PCA and LDA results of Figure 3A,C and Figure 4A,C. Furthermore, it was possible to classify the organic plum from the others by putting into practice the supervised methods in the same way for the organic apple with each of the pesticides studied (Amistar and Funlate). For the other fruits, 85.7% accuracy was obtained for cape gooseberry and 81% for strawberry with the E-nose. Moreover, a cross-validation method was applied to the measurements with a k-fold = 5 to validate these results, obtaining good accuracy through the different techniques.

When analyzing the samples, the data may change due to the nature of the sample, even if it is from the same fruits and pesticides, taking into account the new variables added, such as sugars, minerals, or acids that are not volatile at ambient conditions. The classification percentages reach 95.2% concerning the results achieved by the classification methods, as seen in Table 4.

Moreover, adding new variables improves the classification percentage by changing the order of the data; they can be adjusted to a model that improves the classification. With the E-tongue, the method that best predicts the behavior of the data is naïve Bayes, since success percentages of 98% were obtained with cape gooseberry and 93.3% for plum, strawberry, and apple. Regarding cape gooseberry, if compared with Figure 4D and Figure 6D, it was possible to determine that with the E-tongue, better classification and data discrimination were obtained. This explains that there is a satisfactory linear correlation applying the linear discrimination method, where it managed to achieve 80% classification in all fruits. The E-tongue appears more effective than the E-nose for classifying pesticide-treated fruits. This is evidenced by its greater accuracy and consistency across all classification methods and fruit types. The E-nose, although a useful system for agro-food applications, shows a more variable performance and, in this study, was less effective on fruits such as strawberry and cape gooseberry, which means that volatile compounds detected by the E-nose are more difficult to use for classification compared to compounds detected by the E-tongue.

Table 5. Metrics obtained from the confusion matrix (linear discrimination function) for fruit pesticide identification using the E-nose.

Metrics	Plum L. Discrimination			Strawberry L. Discrimination			Apple L. Discrimination			Gooseberry L. Discrimination
	Org	Am	Fu	Org	Ac	Br	Org	Am	Fu	Org	Br	Cu	Da	Pr
Precision	100%	100%	87.5%	83.3%	75%	85.7%	85.7%	100%	87.5%	87.5%	100%	83.3%	83.3%	75%
Sensitivity	85.7%	100%	100%	71.4%	85.7%	85.7%	85.7%	85.7%	100%	100%	100%	71.4%	71.4%	85.7%
Specificity	100%	100%	92.9%	92.9%	85.7%	92.9%	92.9%	100%	92.9%	96.4%	100%	96.4%	96.4%	92.9%
Accuracy	95.3%	95.3%	95.3%	81%	81%	81%	90.5%	90.5%	90.5%	85.7%	85.7%	85.7%	85.7%	85.7%
F1 score	92.3%	100%	93.3%	76.9%	80.%	85.7%	85.7%	92.3%	93.3%	93.3%	100%	76.9%	76.9%	80%
NPV	93.3%	100.%	100%	86.7%	92.3%	92.9%	92.9%	93.3%	100%	100%	100%	93.1%	93.1%	96.3%

depicts the metrics used to evaluate the performance of an E-nose in discriminating between organic fruits and fruits contaminated with various pesticides, based on the selection of the best classification model obtained in Table 4 and the confusion matrix. The metrics presented include precision, sensitivity, specificity, accuracy, F1 score, and negative predictive value (NPV). The linear discrimination model, in most cases, achieved classification values above 90%, with precision rates reaching up to 100% in identifying pesticides in some fruits. The E-tongue seems more effective than the E-nose for classifying pesticide-treated fruits. This is evidenced by its greater accuracy and consistency across all classification methods and fruit types. Although the E-nose is a useful system for agro-food applications, it shows a more variable performance and, in this study, was less effective on fruits such as strawberries and cape gooseberries, indicating that the volatile compounds detected by the electronic.

Likewise,

Table 6. Metrics obtained from the confusion matrices of different ML models for fruit pesticide identification using the E-tongue.

Metrics	Plum SVM (Quadratic)			Strawberry L. Discrimination		Apple L. Discrimination			Gooseberry Naive Bayes
	Org	Am	Fu	Org	Ac	Br	Org	Am	Fu	Org	Br	Cu	Da	Pr
Precision	100%	90.9%	90.9%	100%	90%	90.9%	90%	100%	90.0%	100%	100%	90.9%	100%	100%
Sensitivity	80%	100%	100%	90%	90%	100%	90%	100%	90%	100%	90.0%	100%	100%	100%
Specificity	100%	95%	95%	100%	95%	95%	95%	100%	95%	100%	100%	97.5%	100%	100%
Accuracy	93.3%	93.3%	93.3%	93.3%	93.3%	93.3%	93.3%	93.3%	93.3%	98%	98%	98%	98%	98%
F1 score	88.9%	95.2%	95.2%	94.7%	90%	95.2%	90%	100%	90%	100%	94.7%	95.2%	100%	100%
NPV	90.9%	100%	100%	95.2%	95%	100%	95%	100%	95%	100%	97.6%	100%	100%	100%

illustrates the metrics derived from the confusion matrix of the SVM, linear discriminant, and naïve Bayes models, which show precision values reaching up to 100% in identifying nearly all pesticides in fruits. This result confirms that the E-tongue outperformed the E-nose, demonstrating superior accuracy and effectiveness in pesticide detection across various fruit samples.

3.3. Classification of Pesticide Concentrations in Fruits with Multivariate Analysis Methods

This study also evaluated the ability of the E-nose and E-tongue to differentiate between different concentrations of pesticides existing in each fruit, based on the measurements made and the application of different data processing methods. Therefore, different concentrations were prepared between 0.2 and 3 ppm values in each pesticide to determine the discrimination and classification between organic fruit samples and those contaminated with pesticides.

3.3.1. Plum

Figure 7 shows the results of a PCA and LDA analysis for different concentrations of the pesticides Amistar and Funlate in plums, using an E-tongue, where the graphs illustrate the discrimination between different concentrations of pesticides and organic fruits, showing a high selectivity and repeatability in the measurements. On the other hand, the variance explained by PC1 alone is 92.09% for Amistar and 91.45% for Funlate, indicating a strong capacity of the E-tongue system to capture most of the variability in the data.

Likewise, in Figure 8, the data acquired by the E-tongue are projected from the organic sample to concentrations of 3 mg/kg, with a variance of 99.69% (Figure 8A) for Amistar and 98.88% (Figure 8A) in the Funlate for PC1. Figure 8A illustrates that the organic fruit samples are clearly separated from the samples treated with different concentrations of Amistar.

In the case of organic plums and plums treated with Funlate and Amistar pesticides, it can be seen from the responses in Figure 7 and Figure 8 that the E-tongue is more effective than E-Nose in discriminating between organic fruits and fruits treated with different concentrations of pesticides. Therefore, the E-tongue provides a better separation between the groups, indicating that there is greater sensitivity and precision in detecting changes in the chemical profiles, where the E-nose shows more variability in the data, especially at higher concentrations.

3.3.2. Strawberry

On the other hand, the behavior of the strawberry using the E-nose through PCA and LDA analysis of the Across and Bricol concentrations can be seen in Figure 9. They were correctly discriminated and classified, generating a variance of 94.08% (Figure 9A) and 93.96% (Figure 9B). Despite the PCA and LDA plots showing clear separation between the different concentrations of Bricol and Across with respect to organic strawberry, there is a lack of linear distribution in the concentrations using the Bricol pesticide, as observed in Figure 9, where the concentration clusters are located in an ascending manner from the concentration of 0.2 to the concentration of 3 ppm. The lack of a linear distribution in the multivariate analysis graphs may be due to different factors such as the way the pesticides interact with the different components of the fruit (such as sugars, acids, and water), the natural variability in the chemical composition of the fruits (due to differences in ripening, storage, growth conditions, etc.), and the effects of mixing VOCs, among others.

Figure 10 depicts the responses of the PCA and LDA models in identifying pesticide concentrations in strawberries using the E-tongue. For the pesticide Across, the device demonstrated a strong ability to distinguish between organic strawberries and those treated with varying concentrations of Across, achieving a variance of 99.85%, which differentiates the detected chemical profiles (see Figure 10A). In contrast, for the pesticide Bricol, the measurements showed some similarities, likely due to the presence of outliers that emerged during experimental tests or partial sensor saturation caused by surface interactions, among other factors. Despite these challenges, the system achieved a variance of 99.91%, as depicted in Figure 10B. This high level of variance means that, even with the presence of outliers, the system retains substantial sensitivity in detecting chemical differences between treated and organic simples.

3.3.3. Apple

Figure 11 demonstrates that the E-tongue effectively distinguishes between organic apples and those treated with varying concentrations of Amistar and Funlate pesticides. The PCA plots reveal a generally good separation, though some overlap is observed between closely related concentrations. In contrast, the LDA plots demonstrate clear and distinct separation without overlap. This means that while both PCA and LDA are valuable for classification, LDA could classify between different pesticide concentrations in apple samples.

For this data analysis, the apple samples obtained by E-tongue were included. It is noticed that for Amistar, a variance of 97.34% was achieved (see Figure 12A). Additionally, there is a slight scatter at the 3 ppm concentration, where the data overlap with the 2 ppm concentrations. The dispersion observed in some concentration groups represents that the accuracy of the E-tongue may vary depending on the pesticide concentration and sample composition, or due to limitations of the platinum screen-printed sensors used, as these have a limited capacity to adsorb molecules on their surface. At higher concentrations, such as 2 ppm and 3 ppm, the active sites on the electrode may begin to saturate, meaning that a further increase in pesticide concentration does not result in a proportional increase in the electrochemical signal. Consequently, the sensor’s electrode may not accurately distinguish between closely related concentrations, leading to overlap in the data.

Consequently, the same behavior is represented in the LDA analysis (Figure 12C). Additionally, in the PCA analysis of the Funlate pesticide was reached a discrimination of 98.52% in PC1, which denoted a discrimination between the organic sample and the pesticide concentrations, although there was a slight dispersion between the data for the organic apple.

3.3.4. Cape Gooseberry

Finally, the behavior of the concentrations applied in the cape gooseberry for the pesticides Bricol (A), Curzate (B), Daconil (C), and Preza (D) was evaluated by applying PCA, depicted in Figure 13, where a sequential dispersion of the data is generated according to the increase in the pesticide concentration.

Except for Curzate, all pesticides have more than 90% of the variance explained by the first principal component (PC1), indicating that most of the variability in the data is captured within a single principal dimension. Curzate, with only 68.54% of the variance explained by PC1, shows greater data dispersion, which means less differentiation between concentrations compared to the other pesticides. One possible explanation for the observed variability in the detection of Curzate is chemical degradation. This type of pesticide may be more unstable or reactive on the fruit’s surface or during measurement. Rapid degradation or secondary reactions with the components of the gooseberry can produce additional volatile products or unpredictably modify the original VOCs [69].

Moreover, Figure 14 shows the behavior of pesticide concentrations applied to cape gooseberry, which was acquired through the measurements made with the E-tongue, where there is less dispersion compared to the results shown in Figure 13. The above was because the E-tongue obtained variance values greater than 98%, compared with the E-nose.

Consequently, the variation values achieved with the E-tongue for pesticides are 99.74% in Bricol, 89.87% in Curzate, 99.02% for Daconil, and 99.32% in Preza. Figure 15 illustrates the data classification with the LDA analysis method, where it was also compared with the PCA analysis, achieving good repeatability in the responses with the E-nose. However, the close clustering of samples with 0.2 ppm of pesticides (Curzate and Preza) and the organic fruit samples in Figure 15 can be owing to the low sensitivity of the sensor at low concentrations, as the amount of pesticide present may be near the detection limit of the platinum sensor used in the E-tongue.

Figure 16 represents the results of gooseberry concentrations where the measurements were classified from the LDA analysis using the E-tongue, achieving different variance percentages of 99.74% in Bricol, 89.87% in Curzate, 99.02% in Daconil, and 99.32% in Preza.

3.4. Classification of Pesticide Concentrations in Fruits with Machine Learning Methods

For data processing, different supervised machine learning methods were used, which were applied to calculate the accuracy and compare the measurements acquired with the E-nose and E-tongue.

Table 7. Classification of pesticide concentrations in fruits through data acquired using E-tongue and E-nose.

Method	Accuracy Percentage (%)
	E-Nose										E-Tongue
	Plum		Strawberry		Apple		Cape Gooseberry				Plum		Strawberry		Apple		Cape Gooseberry
	Am	Fu	Ac	Br	Am	Fu	Br	Cu	Da	Pr	Am	Fu	Ac	Br	Am	Fu	Br	Cu	Da	Pr
Decision Trees	77.1	85.7	88.6	85.7	71.4	82.9	80.0	80.0	71.4	82.9	96.0	64.0	96.0	92.0	82.0	82.0	90.0	96.0	80.0	96.0
LDA	82.9	71.4	71.4	85.7	77.1	82.9	65.7	80.0	82.9	60.0	92.0	92.0	98.0	92.0	96.0	96.0	98.0	90.0	96.0	92.0
Naïve Bayes	91.4	88.6	77.1	97.1	85.7	97.1	91.4	88.6	94.3	82.9	98.0	80.0	94.0	92.0	80.0	84.0	94.0	98.0	92.0	100
SVM (quadratic)	97.1	80.0	71.1	91.4	91.4	88.6	88.6	88.6	97.1	85.7	96.0	84.0	96.0	92.0	84.0	90.0	94.0	94.0	96.0	98.0
SVM (cubic)	97.1	85.7	74.3	94.3	94.3	88.6	88.6	88.6	94.3	88.6	96.0	86.0	94.0	94.0	94.0	88.0	90.0	96.0	98.0	100
K-NN (fine)	97.1	94.3	82.9	100	88.6	97.1	82.9	94.3	80.0	80.0	92.0	90.0	96.0	90.0	94.0	92.0	96.0	96.0	96.0	100

illustrates that the E-nose responded better with the measurements of plum contaminated with Amistar by applying quadratic the SVM, cubic SVM, and fine k-NN methods, with an accuracy of 97.1%. In the case of plum with Funlate, the fine k-NN method was applied to the dataset, obtaining a value of 94.3% success. Moreover, in strawberry, the decision trees method reached a value of 88.6% and 85.7% in the use of Across and Bricol pesticides, where the latter obtained 100% accuracy through the classification method. For this reason, the Bricol accomplished higher percentages than Across pesticide applying other methods because the gas sensor array could identify the chlorine compound present in the active compound of this pesticide. Regarding the apple, 94.3% success rate was reached by using the cubic SVM method with the pesticide Amistar, and with Funlate, an accuracy of 97.1% through the naïve Bayes and fine K-NN methods, respectively. Hence, the E-nose has greater accuracy in identifying the compounds present in Funlate. On the other hand, the cape gooseberry with the Bricol pesticide obtained a better classification using the naïve Bayes method, which achieved a success rate of 97.1%; Similarly, the pesticide Curzate was classified with 94.3% success rate applying the fine k-NN method. Consequently, for the Daconil pesticide in cape gooseberry fruit, 97.1% was accomplished by using the SVM method and the same classification with the Preza pesticide. Additionally, the cubic SVM method achieved an 88.6% success rate. According to the findings found through the classification methods of four pesticides studied with the E-tongue, greater accuracy was accomplished in the cape gooseberry fruit in identifying the Preza pesticide.

In the same way, these classification methods were also applied to the data acquired by E-tongue, where better success percentages were obtained. For example, in plums, the Amistar pesticide was identified with a success rate of 98% through naïve Bayes, and 92% in Funlate with the linear discrimination method. Concerning strawberries and using Across pesticide, the best classified method was linear discrimination, achieving an accuracy of 98%. In the same way, the Bricol pesticides with strawberries gained a 94% success rate through the cubic SVM. The above is compared with what was obtained with the E-nose, since the pesticide could be identified in the strawberry because the Bricol pesticide has an active ingredient that is mainly made up of nitrogen, oxygen, carbon, and hydrogen molecules, which have an electrochemical behavior, where the detection is possible with the platinum electrode of the E-tongue. With respect to apples, the method with the greatest accuracy in classifying the data was linear discrimination, with a value of 96% in the classification of the Amistar pesticide. For cape gooseberry, the best methods that classified pesticides were linear discrimination with a success rate of 98% with the pesticide Bricol, the naïve Bayes method with Curzate, and the cubic SVM method for the Daconil pesticide. Finally, in detecting the Preza pesticide, the best results were achieved with 100% accuracy in the data classification obtained with the naïve Bayes, cubic SVM, and fine K-NN methods. It should be clarified that a cross-validation method was applied to the measurements with a k-fold = 5 used for the different methods. Table 5 describes the results of the data classification methods.

3.5. Results with Sensory Perception Systems Using Test Data

Figure 17, Figure 18, Figure 19 and Figure 20 demonstrate the results of the validation analysis using the E-tongue and E-nose to determine the presence of pesticides in fruits directly harvested from pesticide-treated crops. Initially, models were created (see Figure 3 and Figure 5) using machine learning methods using organic fruit samples and fruits contaminated in the laboratory. As previously explained, these models were trained and validated to identify distinctive features of each group using pattern recognition techniques such as LDA and PCA applied to the data obtained from the two electronic devices. Subsequently, test fruit samples, whose condition (organic or contaminated) was previously unknown, were analyzed to validate the created models. If the test fruit samples are closer to the group representing fruits contaminated with a specific pesticide, they are inferred to contain that pesticide. Conversely, if the test fruit samples are closer to the organic fruit group, they are considered pesticide-free. These results were further validated through interviews with the crop owners, who provided additional confirmation of the findings obtained using the E-nose and E-tongue.

3.5.1. Plum

Figure 17A,B illustrate the LDA and PCA analysis results with the plum data obtained with the E-nose, where a separation can be seen between the organic fruit and the fruit obtained from the crop.

Through the PCA, it is observed that the plum samples contaminated with the Funlate pesticide at a concentration of 3 ppm and the newly cultivated plum sample generate an overlap of the measurements with each other, which indicates that the plum sample cultivated crop has a greater incidence of the Funlate pesticide (see Table 6). The above confirms what was mentioned by the farmers, where the incidence of the active principle of this pesticide is detected during the pest control stage, which was previously applied.

However, the Amistar pesticide samples (concentration of 3 ppm) do not group closely with the cultivated fruit. Further, the Amistar may be in a lower concentration because it is applied in an initial stage of the ripening process, which causes its presence in the fruit to decrease over time, either due to natural washing conditions or crop irrigation. In analyzing the E-tongue response with the plum data, Figure 17C,D show excellent behavior in separating the categories. Likewise, through the results of the LDA analysis (Figure 18D), an approach between the cultivated fruit and Funlate is illustrated.

3.5.2. Strawberry

Figure 18 illustrates the multivariate analyses for the strawberry with the sensory devices. It was determined that similar behavior was maintained concerning the plum. In this way, it was possible to compare the data from the training stage and those acquired from the cultivated fruits.

The results with the E-nose, as seen in Figure 18A,B, represent a separation between the organic strawberry and the cultivated strawberry regarding the strawberry contaminated with the Bricol and Across pesticides to the concentration of 3 mg/kg. In the same way, proximity is noted between the sample contaminated with Bricol and the cultured sample, which is observed more clearly using the E-tongue, as depicted in Figure 18C,D. Therefore, the overlap presented between the samples is assumed to be that the cultivated strawberry is found at a higher concentration of the pesticide; furthermore, it is essential to highlight that this same behavior was not obtained with the Across sample since it is located in a more distant region concerning the cultivated strawberry.

The above confirms that with the E-tongue, it is possible to detect the predominant incidence of Bricol on the fruit. The responses are represented through LDA and PCA analysis, which indicates that the contribution of non-volatile substances is more significant than that of volatile substances.

3.5.3. Apple

In the study with the apple (see Figure 19), a comparative analysis could be made to the plum because the same pesticides were applied to both fruits. Therefore, in the results with the E-nose (Figure 19A,B), it can be seen there is a more significant influence by the pesticide Funlate concerning Amistar, because one of the measurements of the sample of the cultivated apple has a similarity with the model contaminated with Funlate, which shows that the results achieved by the E-tongue confirm this information since it is in accordance with Figure 19C,D, where a grouping is presented between these samples.

3.5.4. Cape Gooseberry

In the same way, to the three previous fruits, it is observed through multivariate analysis that it is possible to maintain a good separation of two types of organic and cultivated fruits. For this reason, Bricol was detected as the pesticide that obtained the highest concentration in the cape gooseberry to analyze pesticides in the fruit. In the E-nose, it can be observed in the PCA and LDA (Figure 20A,B) that the categories are separated, which does not confirm what pesticide type of the fruit samples could have cultivated. However, using the E-tongue, it is possible to observe the separation of the categories (Figure 20C,D), in which there is a similarity between the cultivated cape gooseberry sample and the cape gooseberry contaminated with Bricol pesticide at a concentration of 3 ppm.

4. Discussion

Nowadays, most studies have focused on using the E-nose to detect pesticides in fruits (apples [44], sweet cherries [43], peaches [46]). In the study conducted by Rivera Nategh et al. [46], an E-nose composed of eight chemoresistive sensors was used to detect organophosphate pesticide residues in peaches (Prunus persica) and distinguish between three concentrations (1, 2, and 3 ppm) as well as pesticide-free fruit. The sensor responses were recorded using PCA, achieving a variance of 99.8% with two principal components. Another study by Nategh et al. [43] demonstrated the capability of an E-nose composed of 10 MQ and TGS reference sensors to detect Diazinon pesticide residues in cherries and differentiate between four stages of fruit maturity. The authors employed three data analysis methods: PCA, LDA, and ANN. PCA score plots of PC1 and PC2 captured 90–96% of the variance in the data for toxic and non-toxic sweet cherries, while ANN and LDA achieved 100% classification accuracy. In another study, Tang et al. [44] evaluated the performance of a commercial PEN3 E-nose equipped with 10 metal oxide semiconductor (MOS) sensors for detecting pesticides in apple samples. The fruit samples were treated with two pesticides containing cypermethrin and chlorpyrifos at four different concentrations and one mixture. PCA and LDA methods successfully differentiated between control and contaminated samples at varying concentrations. Although both methods were effective, PCA exhibited the best discrimination ability. The SVM method demonstrated high accuracy, ranging from 94% to 97% for training datasets and from 90% to 93% for testing datasets, concluding that the commercial E-nose is a useful tool for pesticide residue detection. In a more recent study, an E-nose composed of 11 gas sensors was fabricated, allowing for differentiation between pesticide-contaminated cherries (Korban) and non-contaminated cherries. Four classification algorithms (random forest classifier, extra trees classifier, decision tree classifier, and k-nearest neighbor (kNN)) were employed, with the extra trees classifier achieving the most satisfactory results, with a classification accuracy of 94.30%, a sensitivity of 93.00%, and a specificity of 95.60% [70]. Regarding E-tongues, there is limited literature focused on pesticide detection in fruits. However, some biosensors made from various materials have been used to detect pesticides (malathion and cadusafos) in laboratory-prepared matrices (water, food). In terms of classification, sensitivity, and specificity, studies on pesticide detection in fruits using E-noses show promising results. Additionally, the ability of these devices to differentiate between various types of pesticides and concentrations, along with their high sensitivity and specificity, makes them promising tools for enhancing safety and quality control in the food industry [71,72,73]. However, to date, no studies have been reported in the literature that implement both an E-nose and tongue simultaneously to detect pesticides in different fruits and concentrations. Only one study conducted by the same authors [74] managed to discriminate and classify newly cultivated and organic fruits using both technologies. Both devices demonstrated the ability to discriminate and classify the studied samples, although the E-tongue exhibited greater data dispersion and some overlap between the analyzed data classes. Therefore, the distinction between the study proposed in this research and the others mentioned above lies in using both technologies to detect pesticides in various fruits characteristic of the region compared to organic fruit that was free of contamination.

Additionally, different performance metrics of the E-nose and E-tongue were evaluated using the best classification model reported in Table 4, followed by the evaluation of the metrics for the best models. Overall, both devices achieved up to 100% classification accuracy in detecting pesticides in some fruits, but the E-tongue showed better precision in most cases, particularly for fruits like plums, apples, and cape gooseberries treated with various pesticides, indicating a robust capability to correctly identify both organic and pesticide-treated fruits. It is also evident that in other metrics (sensitivity, accuracy, and F1 score), the E-tongue demonstrated better efficiency than the E-nose, as the latter showed more variability in the data, especially for strawberries and cape gooseberries. This could suggest that the E-nose is more sensitive to variations in the fruit matrix or pesticide characteristics. Additionally, different pesticide concentration ranges were evaluated to assess the ability of these devices to detect and classify samples within categories, generally achieving accuracy greater than 80% and reaching 100% in detecting Preza and Bricol for cape gooseberries, depending on the classifier implemented. Similarly, the E-tongue showed better performance than the E-nose in discriminating pesticide concentrations in fruits. This is attributed to its higher precision, sensitivity, accuracy, and specificity, as well as its more effective capability to distinguish between different pesticide concentrations in various fruits. Finally, test samples (whose condition was previously unknown) were used with the model created in the first stage of this study to determine the type of pesticide contained in the test samples, where successful classification results were achieved. Although good results have been obtained, it is necessary to validate the operation of the sensory equipment based on the previously created model and trained with the learning methods, where it is also convenient to use commercial analytical equipment such as GC-MS, since it allows for identifying the active ingredients of pesticides, and in this way verify the results obtained previously. Therefore, the detection of the active ingredients of pesticides employing this equipment would help us to confirm the information provided by farmers and by the responses of sensory perception systems.

Based on the results obtained in this study, we achieved high accuracies, with some classifications reaching 100%. These outcomes demonstrate the effectiveness of our approach and make our study competitive when compared to previous research in the field of pesticide classification.

5. Conclusions

This pilot study was carried out to evaluate the ability to detect pesticides in fruits through an E-nose and E-tongue, where the E-nose was composed of metal oxide gas sensors, while the E-tongue was through platinum electrodes.

Some difficulties were observed in differentiating the categories of Curzate, Daconil, and Preza, since a similarity of the compounds between the characteristic chemicals of those pesticides was noticed. On the other hand, it was possible to identify the pesticide categories in the previously mentioned fruits by applying the platinum electrode since, through the responses of the multivariate analysis, the variance of the dataset was high in the first two PCA components and LDA factors.

It should be clarified that the PCA method was used primarily to visualize the behavior and discrimination of pesticide categories in fruits within a two-dimensional space. It is important to note that PCA was not used to classify the measurements. However, its application has also been demonstrated in optimizing machine learning methods, as it enabled extracting more meaningful and informative features from the data, thereby enhancing the performance of the predictive models.

In the analysis made of data concentrations from the different types of pesticides applied to each of the fruits, it was observed that the E-nose and E-tongue could discriminate concentrations from 0.2 ppm to 3 ppm, and they were able to identify each of the categories of organic fruit and concentrations.

There were slight overlaps in the strawberry samples and plum samples with pesticides, where a better percentage of accuracy was observed in the data classification using each classification method. For example, the cape gooseberry fruit obtained 100% success rate. Nevertheless, the E-nose obtained good success rates in the classification of strawberries and plums, reaching percentages close to those of the E-tongue, indicating that both sensory systems could be useful in evaluating quality in producing different types of fruits.

In this study, the electronic tongue was highlighted as a promising tool for pesticide detection in fruits, providing interesting results across a variety of fruit types and pesticides. Its performance results may make it ideal for applications in quality control and food safety, as it could enable producers and regulators to make informed decisions about the safety and quality of products. On the other hand, the electronic nose also proved to be effective due to its high accuracy and specificity in various fruit–pesticide combinations, which could be a valuable tool, especially in applications requiring rapid detection and a high level of specificity. Therefore, the choice between the electronic tongue and the electronic nose should be based on the specific needs of the application, considering factors such as detection speed, required accuracy, and the type of samples to be analyzed. However, it is important to continue optimizing these types of systems to have a better level of performance.

As future work, it is necessary to make improvements in the analysis of certain specific pesticide substances through GC-MS to optimize the sampling and acquisition protocol with the E-tongue, and also evaluate the behavior of the responses of the gas sensors to achieve better performance in the classification of the samples. On the other hand, a data fusion of the information extracted from both equipment could help improve the results since gas sensors and electrochemical analysis would help to increase the information of the samples acquired by the measurement equipment.