Detection of Localized Damage in Tomato Based on Bioelectrical Impedance Spectroscopy

Abstract

This paper proposes a method for localized damage detection in tomato, with the objective of enabling the detection of bruises prior to sorting. Bioimpedance spectroscopy technology is employed to assess the extent of localized damage in tomato. An equivalent circuit model is constructed, and the impedance spectroscopy data are obtained by developing a local damage measurement platform for tomatoes using a self-designed circular four-electrode BIS sensor. The electrical parameters are then extracted by fitting the constructed equivalent circuit model to the tomato data. Subsequently, we analyze the variation rules of the electrical parameters in different damage levels. To reduce the dimensionality of the features, including biological variables, fitted electrical parameters, and tomato ripeness, we employ Spearman feature selection. We then classify the reduced features by combining the advantages of the support vector machine and the artificial neural network. The results demonstrate that the designed circular four-electrode BIS sensor can non-destructively measure localized damage conditions in tomato. A localized damage measurement platform for tomatoes has been constructed using this sensor. A comparison of the impedance measurements obtained using the designed circular four-electrode BIS sensor with those obtained using a needle sensor proposed by previous scholars revealed that both sensors exhibited a decrease in impedance with increasing damage degree. This finding indicates that the designed circular four-electrode BIS sensor is an effective tool for characterizing damage conditions in tomatoes. The design of the tomato circular four-electrode BIS sensor is an effective means of characterizing tomato damage. The Spearman-SVM-ANN damage classification algorithm, based on the Spearman feature selection, effectively classified tomato damage with a 98.765% accuracy rate. The findings of this study provide a reference for the grading and transportation of tomatoes after harvest.

1. Introduction

Tomato es are a common vegetable and fruit, which are distributed globally. They are rich in lycopene and a variety of vitamins and can be made into tomato sauce, tomato soup, and other foods. China is the world’s foremost producer of tomato, accounting for approximately one-third of global production. As of 2023, the global tomato market capacity was 38.643 billion yuan, with China’s market capacity reaching 11.968 billion yuan, thereby exerting a dominant influence over global tomato production [1,2]. Tomatoes are characterized by their delicate texture, thin skin, and sensitivity to external pressure, which render it susceptible to damage during transportation [3]. Some of this damage is localized and cannot be detected by the naked eye within a short period of time. Such damage may lead to incorrect grading and packing, which can negatively impact the consumer perception of product quality. Additionally, damaged tomatoes are prone to rapid decay, resulting in a waste of transportation costs. Therefore, it is vital to find a way to test for tomato damage. Since tomato damage is not quickly visible from the outside, it is necessary to use special equipment to identify the extent of tomato damage. The first methods for inspecting tomato damage include the following: the nuclear magnetic resonance (MRI) detection method. This method detects and characterizes the environment of water protons in tomato tissues and analyzes the images to identify different levels of damage in tomato tissues based on the contrast provided by the images [4]. This method involves briefly heating the tomato. Internal browning and core decomposition can be detected in the tissues using MRI images [5]. Spin-echo MRI imaging can identify mesocarp tissues damaged due to heat treatment or deworming. MRI images can provide sufficient contrast to identify different levels of damage in tomato tissues. The second method is spectral detection: the utilization of near-infrared (NIR) spectroscopy technology for the collection of spectral information pertaining to tomato [6]. The objective of this study is to determine the content of the internal components of the tomato using physicochemical analysis methods. Additionally, the study will investigate the correlation between spectral information and the internal quality of the tomato, as well as the effect of different spectral pre-processing methods on the original spectral pre-processing of the tomato. A quantitative model for internal quality in tomatoes should be constructed, the model with the best predictive performance selected, and the internal quality of tomatoes then evaluated in a comprehensive manner [7]. The third method is a microstructure analysis. This involves constructing a micro-compression tester and filming the high-speed micro-compression test of the tomato cells using the P-CAM camera. The relaxation parameters of the tomato cells can then be obtained [8]. Additionally, the tissue structure of the tomato pulp and the cellular arrangement of the tomato can be observed using a scanning electron microscope [9]. This allows for the microstructure of the tomato to be identified with greater accuracy and the presence of any tiny damage to be detected. Methods 1 and 2 are well-suited for large-volume detection. However, the cost of nuclear magnetic resonance (NMR) technology is high, and heating detection is prone to cause secondary damage to tomatoes. Additionally, spectroscopic detection is ineffective in detecting damage hidden under the skin at an early stage. Furthermore, method 3 has high equipment detection costs, and taking samples prior to detection will damage the tomatoes themselves. Therefore, there is an urgent need for a moderately expensive and highly accurate means of detection to make up for the shortcomings of the existing methods.

Bioimpedance spectroscopy (BIS) is widely used in bioassays as a low-cost, simple-to-operate, and emerging technology. The content of charged particles changes in the different states of tomato, which gives the possibility to measure the quality using the dielectric properties of tomato [10]. In recent years, some scholars have employed impedance spectroscopy to investigate various aspects of tomato biology. In a study conducted by Benavente et al., impedance spectroscopy was employed to investigate the impact of varying concentrations of NaCl solutions on the cuticles of tomato fruits, including both ripe and green varieties. The findings revealed that the cuticles of ripe tomatoes exhibited a single relaxation process, whereas those of green tomatoes displayed two distinct relaxation processes [11]. Zhang, for instance, examined the relationship between the electrical impedance of tomato plants and their water content. As the plant loses water, the free water in the cells declines, the cells become less turgid, the tissue hardness diminishes, the electrical conductivity declines, and the impedance of the plant increases [12]. Li et al. measured the impedance spectra of five groups of tomato plants with different phosphorus and potassium concentrations in the frequency range of 1 Hz–1 MHz. They then analyzed the measured impedance data using an equivalent circuit model and established a regression prediction model between phosphorus, potassium, and electrical impedance. This model demonstrated that electrical impedance can be used in the detection and diagnosis of the nutritional status of phosphorus and potassium in tomato plants [13,14]. These studies have demonstrated that bioimpedance spectroscopy is suitable for tomato measurements.

A multitude of experiments have been conducted with the objective of establishing a correlation between the degree of fruit damage and electrical impedance. Varlan et al. utilized the existing Cole three-element model to derive the values of four electrical parameters (low-frequency resistor RL, high-frequency resistor RH, constant phase angle θ, and eigenfrequency f) using least-squares fitting, with the objective of identifying the numerical variations of the tomato electrical parameters with damage. Their findings indicated that the low-frequency resistor RL and the constant phase angle θ are the most sensitive parameters to damage, yet these parameters had not been correlated with the damage level in tomato [15]. Phillipa J. Jackson placed apples with varying degrees of damage for a period of 24 h. Two Ag/AgCl electrodes, 35 mm apart, were then used to pierce the apples, and the electrical impedance was measured at 50 Hz–1 MHz [16]. It was found that the ∆R50 Hz value was correlated with the degree of apple damage. However, due to the uncertainty that still exists when converting the ∆R50 Hz value to the degree of bruising in the apple factor, the practical significance of this finding was low. Albelda employed a two-needle sensor to ascertain the electrical impedance of citrus fruits. The electrical signals generated by the sensor were analyzed using an artificial neural network (ANN) model, which enabled the differentiation between frostbitten and non-frostbitten oranges [17]. M. Mohsen et al. employed two Ag/AgCl electrodes to ascertain impedance spectral data across three states: fresh, heated, and frozen. These data were then subjected to fitting procedures, enabling the derivation of fitted values for the various models under consideration in each of the three states. This approach facilitated the determination of the state of the tomato in question [18]. Eunyong Jeon proposed a real-time system for measuring tomatoes using a microneedle sensor for measurement, integrating a microneedle MEMS chip and AD5933 on a printed circuit board, and inserting the object to be measured during the measurement. The test results demonstrated that the microcontroller-based dual-probe impedance measurement system, which is low-cost, lightweight, and portable, can also be widely used in multiple scenarios for fruit freshness analysis [19]. These studies have demonstrated that bioimpedance spectroscopy is suitable for tomato damage measurements.

In conclusion, numerous experiments to assess the electrical impedance of fruits necessitate the insertion of electrodes into the fruits, which inevitably results in damage to the fruits themselves. Fruits that have been detected in this manner are difficult to continue to sell in the market. For this reason, this paper proposes a method for the localized damage detection of tomatoes. This method uses self-designed circular four-electrode sensors to obtain the impedance spectra of tomatoes. The constructed equivalent circuit model of the tomato system is then fitted to obtain the electrical parameters. Spearman feature selection is used to downsize the characteristics of the biological variables of tomatoes, the fitted electrical parameters, and the localized compression pressure applied to construct the research laboratory. The results of the localized injury classification model for tomato are used as the criteria for the tomato injury class. The accuracy of the model is evaluated by the accuracy of classification in order to determine the accuracy of the tomato localized damage model. The accuracy is verified by the test set of data in order to explore a method to infer the damage level of tomato by detecting the electrical impedance, which has a lower cost and higher convenience in detecting the damage level of tomato. The circular sensor described in this paper measures the electrical impedance of tomato surface damage without causing further mechanical damage to the tomato. In terms of the detection method, this paper has the advantage of non-injury compared with existing studies. It also has the advantages of low cost and convenience of detection, allowing for the qualitative analysis of the degree of damage to the tomato.

2. Materials and Methods

2.1. Impedance Measurement Experiments

2.1.1. Principle of Tomato Four-Electrode Impedance Spectroscopy

BIS is a detection technique that obtains impedance information by applying alternating current (AC) signals of varying frequencies to the device under test (DUT). By treating the DUT as a circuit system comprising a multitude of resistive and capacitive elements connected in series and in parallel, the electrical characteristics of the electrical elements are employed to describe the electrical properties of the DUT, including conductivity and the dielectric constant. The bioimpedance spectrum is fitted and the electrical information is extracted from the impedance spectrogram, which allows the distinction of the tissue properties of the test object. The impedance of tomatoes can be measured by the voltage and current using a four-electrode method, as illustrated in Figure 1. In order to perform the measurement, the current is injected through the surface electrodes of one pair of driving electrodes, while the voltage is obtained by measuring through the other pair of sensing electrodes. The magnitude $\left|Z_k\right|$ and phase $\angle Z_k$ size of the tomato impedance at each frequency can be calculated from Equation (1). The real and imaginary parts are calculated based on the impedance information at each frequency to obtain the tomato impedance spectrum (Cole–Cole plot).

$$\begin{array}{l}\left|Z_k\right|={\displaystyle \frac{\left|V_k\right|}{\left|I_k\right|}}={\displaystyle \frac{\sqrt{Re{\left(V_k\right)}^2+Im{\left(V_k\right)}^2}}{\sqrt{Re{\left(I_k\right)}^2+Im{\left(I_k\right)}^2}}}\\ \angle Z_k=\angle V_k-\angle I_k=arctan{\displaystyle \frac{Im\left(V_k\right)}{Re\left(V_k\right)}}-arctan{\displaystyle \frac{Im\left(I_k\right)}{Re\left(I_k\right)}}\end{array}$$

(1)

where: $Z_k$ denotes the impedance amplitude of the tomato impedance under excitation at ${\mathrm{f}}_k$ frequency; $\angle Z_k$ denotes the impedance phase of the tomato impedance under excitation at ${\mathrm{f}}_k$ frequency; $\angle U_k$ and $\angle I_k$ denotes the phase of the excitation current and the phase of the response voltage at ${\mathrm{f}}_k$ frequency; $Re$ and $Im$ are the real and imaginary parts of the complex number, respectively.

The use of four electrodes presents a number of advantages over the use of two electrodes. The four-electrode configuration comprises two outer driving electrodes and two inner sensing electrodes, which are capable of applying a current between the driving electrodes and measuring the potential drop between the sensing electrodes. This configuration enables more accurate measurements with an enhanced stability and anti-interference capability. The two-electrode conductivity sensor has only one set of electrodes for simultaneously applying the current and measuring the voltage drop. This configuration renders it susceptible to factors such as electrode polarization, surface contamination, and circuit resistance, which collectively result in a reduction in measurement accuracy, an inability to maintain sufficient stability, and a lack of anti-interference capability. In conclusion, the four-electrode method is demonstrably more accurate, stable, and capable of withstanding interference than the two-electrode method. The aforementioned advantages render the four-electrode method the optimal choice for the measurement in question.

2.1.2. Experimental Setup for Damage Measurement

Tomato injury is a highly intricate process that encompasses a multitude of physiochemical alterations, including cell demise, alterations in the intracellular contents, respiratory rate, and other metabolic shifts. Consequently, the bioelectrical impedance of tomatoes at a specific frequency band under the influence of an alternating electric signal can be utilized to reflect the alterations in the aforementioned biological state of tomato injury within that frequency band. In this study, 145 tomatoes of the Provence variety are selected, labeled, and weighed. The impedance spectra are then measured for the pressing positions in the pressing experiments. An alternating voltage signal with an amplitude of 1 V and a frequency range of 100 Hz to 200 kHz is applied through a 4092B LCR digital bridge. One hundred frequency points are set, and the signal is added to the circular sensor. Sequential impedance spectrum measurements are conducted on 145 tomatoes with the aforementioned voltage signals, as shown in Figure 2. The impedance amplitude (|Z|) and phase (θ) in the specified frequency range are captured. At the conclusion of each measurement, photographs are taken, the tomatoes are replaced, and the aforementioned procedure is repeated. Finally, all the tomatoes are placed at ambient temperature (20 °C) and pressure.

The end probe of the BIS sensor used in this paper is a two-layer printed circuit board (PCB), which contains four concentric electrode arrays on the front side (the width of each pad is 0.8 mm, and the radii of the concentric circles can be designed according to the requirement) and the pads for connecting the wires on the back side (the size of 1.5 mm). The PCB board is FR-4 (Flame-Retardant Fiberglass Reinforced Epoxy Laminate), and the total thickness is 1.6 mm. Connecting the end probe of the BIS sensor and the 3D printed case together is convenient for measurement, and the total cost of the PCB board and the 3D printed case is less than One hundred Chinese Yuan (CNY), which is a low cost and suitable for some specific scenarios of agricultural applications. The sensor and the 3D printed part housing are connected together for easy measurement, and the total cost of the PCB and the 3D printed part housing is less than one hundred dollars, which is low cost and suitable for some specific scenarios of agricultural applications. In this study, the pressing measurement experiments using a force sensing resister were conducted to assess the efficacy of the BIS sensors in detecting pressure-induced damage. The findings revealed that a pressing pressure below 20 kPa did not result in discernible damage to the tomato. Consequently, this study employs a homemade BIS sensor to perform non-destructive testing of the tomato’s surface. It should be noted that the needle electrode used for surface detection may leave traces on the tomato, and in some instances, the electrode may be directly inserted into the fruit. This particular detection method, which results in damage to the object being measured, is referred to as “lossy detection”. The equivalent circuit model of the tomato is illustrated in Figure 2, and the impedance Z can be expressed as follows:

$$\left\{\begin{array}{l}Z=R_1+(R_2//R_{CPE1})+(R_3//R_{CPE2})\\ R_{CPE1}=1/(j^{n1}{\omega }^{n1}{Q}_1)\\ R_{CPE2}=1/(j^{n2}{\omega }^{n2}{Q}_2)\end{array}\right.$$

(2)

where R₁, R₂, CEP1, and CEP2 represent the internal impedance of the tomato, R₃ denotes the contact impedance of the electrode sheet to the tomato interface, j is an imaginary unit, and w is the angular frequency.

2.2. Pressing Test

Tomatoes undergo three stages of non-destructive elastic deformation, biological yield deformation, and cell rupture deformation when subjected to pressure. Four states occur during placement: decay, skin folds, indentation, and no indentation. The evaluation method for tomato quality proposed by Li et al. [20] employs a four-point grading system, with states ranging from a high to low severity of tomato damage. Level 1 damage (LV1) is defined as rotting, which renders the tomatoes unfit for consumption. Level 2 damage (LV2) is characterized by skin wrinkles, and the tomatoes retain their food value, though they may be used in the production of concentrated tomato paste [21]. Tomatoes with indentations are classified as level 3 damage (LV3), as these defects only affect the appearance of the fruit and do not impact its nutritional value. Conversely, tomatoes with no indentations are classified as level 4 damage (LV4), which is equivalent to the standard, unblemished tomatoes.

The definition of tomato ripeness previously established by our research laboratory has been referenced [22]. This definition is as follows: the white-ripening stage (S2), during which the green hue gradually fades, and the fruit’s upper surface displays a light reddish hue of approximately 10% saturation; the color-change stage (S3), during which the fruit displays a light reddish hue and the color develops from 60% to 90% saturation; the pink stage (S4), during which the fruit displays a reddish hue and the color develops nearly to 100% saturation; and the red-ripening stage (S5), during which the fruit displays a deep reddish hue and exhibits a slight softening. In accordance with the definition of tomato quality proposed by Li et al., 111 of the 145 tomatoes included in this experiment are in the red-ripening stage, while 30 are in the pink stage [8]. Additionally, four tomatoes were inadvertently damaged during the course of the experiment and were therefore excluded from the final data set. The extent of tomato damage is referenced to the definition previously established by our research laboratory, as illustrated in Figure 3 [23]. The tomatoes utilized in the experiment are Provence tomatoes procured in May 2024 in Shouning, Shandong. The tomatoes are selected based on their cross-diameter, which ranges from 50 to 80 mm, and their height, which ranges from 37 to 70 mm. The dimensions are measured using electronic vernier calipers with an accuracy of 0.01 mm.

The apparatus employed for the pressing test is the GY-4 Fruit Hardness Tester (Hangzhou, China) manufactured by Aipu Metrology Instrument Co., as illustrated in Figure 4. The tomato is subjected to automated picking by a robot [24], during which two force points are applied at the top of the tomato, which are separated by 180° and located two by two. The top of the tomato is selected for the pressing test, as illustrated in Figure 4. In a previous study [23], it is determined that the application of pressure to the tomato fruit results in a diminished effect when the pressure differential is within a range of 10 kPa. In this study, the pressing pressure (P) is applied at intervals of 10 kPa, ranging from 20 to 400 kPa. A single pressing is performed at the aforementioned pressure on the 290 stress points of 145 tomatoes. The average speed of manual tomato picking is 20 tomatoes per minute. The compression time is set at 3 s, after which the tomatoes are labeled in order, the tomato group replaced, and the aforementioned steps repeated. Based on the data deemed valid in this experiment, the number of tomatoes exhibiting first-degree damage is 54, second-degree damage is 494, third-degree damage is 265, and fourth-degree damage is 803.

2.3. Scanning Electron Microscope Experiment

The instruments used for the SEM test are a scanning electron microscope (PrismaE SEM, Thermo Fisher Scientifific, Waltham, MA, USA), a freeze dryer (CoolSafe 110-4, Labogene, Denmark) and a critical point dryer (Leica EM CPD300, Leica Microsystems, Wetzlar, Germany), as shown in Figure 5. Three specimens of tomato tissue exhibiting damage at varying levels are selected for the analysis. The specimens are obtained by slicing the tomato at the point of pressure and then placing them in a solution of 5% glutaraldehyde to prevent any alterations to the tissue structure. The specimens are then observed under a microscope, and the results are presented in Figure 5a. The treated tomato tissue samples that have undergone damage should be rinsed three times in a phosphate-buffered saline (PBS) solution. The samples should then be left to dry for 15 min each, before being transferred to the critical point dryer (Leica EM CPD300, Figure 5c). The drying process, which lasts 15–20 min, is performed in a step-by-step manner and the samples are finally placed in the freeze dryer CoolSafe 110-4, as shown in Figure 5b. The samples should then be frozen for 15–20 min each. Following this, the prepared samples are affixed to the sample stage with a conductive adhesive and introduced to an ion sputtering instrument for coating. The coating is applied at a thickness of 20 nm of the sputtering metal. Finally, the samples are placed on the PrismaE SEM, as presented in Figure 5d, the scanning electron microscope observation stage under a 20 kV accelerating voltage for observation is performed and photographs are taken.

2.4. Modeling of Localized Injury Classification in Tomato

In the paper, the characteristics of the tomato biological variables, fitted electrical parameters, and tomato ripeness are employed as input variables for the data dimensionality reduction via Spearman feature selection. Spearman’s correlation coefficient is one of three primary correlation coefficients in statistics, along with Pearson’s and Kendall’s. Spearman’s rank correlation coefficient is a nonparametric statistical parameter of rank, which is a correlation coefficient used to quantify the strength of a monotonic relationship between two continuous variables [25]. Spearman correlation coefficients have a broader range of applications than other statistical techniques, as they do not require the data to be distributed normally or that it be collected at equal intervals. Additionally, they can tolerate outliers, which is advantageous in many contexts. The original features are subjected to dimensionality reduction through the Spearman feature selection. Once the cumulative discriminative ability reaches 85% or more, the corresponding combination of several features is employed as the input for the classification model. In this paper, we define features 1–10 to denote ripeness, R₁, Q1, n1, R₂, Q2, n2, R₃, pressure, and weight, respectively, and class denotes the label of the classification. The results are shown in Figure 6, which expresses the correlation between each feature, the closer to 1, the more relevant.

Table 1. Summary of Spearman feature selection.

Features	Eigenvalue	Cumulative Percentage (%)	Relevance
Features 9	0.7632	40.338	40.338
Features 5	0.2921	55.777	15.439
Features 6	0.2908	71.146	15.369
Features 7	0.1007	77.444	6.297
Features 1	0.0917	82.764	5.320
Features 10	0.0891	87.611	4.847

shows that features 9–10, features 5–7, and feature 1 as new features have been able to contain more than 85% (up to 87.611%) of the original input features, and these six combined features are input into the tomato localized damage classification model to improve the accuracy and speed of the model. Furthermore, the input features of tomatoes are subjected to a statistical analysis, which revealed that the first six features (features 9-pressure, features 5-R2, features 6-Q2, features 7-n2, and features 10-weight) are the most significant. One normal-phase element, pressure, weight, and ripeness of the equivalent circuit model of tomato, demonstrated a strong correlation with the damage level of tomato at the significance level of p < 0.001. Additionally, a multiple commonality analysis is conducted to obtain the variance inflation factor (VIF) of the input features of the model, which is found to be less than 1.3. This indicates that the downsized features of the model could be utilized as a characterization of the damage level of tomato to some extent.

The support vector machine (SVM) is highly accurate for small samples and is capable of handling high-dimensional data but is relatively challenging to apply to large-scale datasets and problems with multiple classification outcomes. Furthermore, it is sensitive to external noise and outliers [26]. The fundamental principle underlying the functionality of ANN is its capacity for nonlinear mapping. This enables the network to learn and extract useful features from the input data, thereby reducing the necessity for the tedious process of manual feature engineering. However, the robustness of the ANN to noise and outliers is limited, and the parameters of the network are difficult to adjust. Furthermore, the ANN performs poorly on test data [27]. Accordingly, this paper proposes the construction of a tomato local damage detection model based on two classification algorithms, namely SVM and ANN. Mathematically, damage classification is a multi-objective nonlinear constrained optimization process. The difficulty of building the classification detection model lies in the optimization objective, which is to find the maximum value of the fitness function within the constraints of the upper and lower limits for variables and the number of iterations. The hyperparameters are treated as independent variables, and the optimal hyperparameters of each model are identified through the hyperparameter optimization of a variety of classification models. The data from the test set are then used to generate a confusion matrix for the final models and assess their classification accuracy, as shown in Figure 7.

2.5. Evaluation Criteria for Localized Impairment Models

A confusion matrix is a performance metric utilized in machine learning classification problems, wherein the output can be either binary or multi-categorical. The confusion matrix output can be classified into one of four categories: a true positive (TP) is defined as a sample whose true category is positive, and which is correctly identified as such by the model; a false positive (FP) is a sample whose true category is negative, but which is incorrectly identified as positive by the model; a true negative (TN) is a sample whose true category is negative, and which is correctly identified as such by the model; an FN (False Negative) is when the true category of the sample is negative, yet the model predicts it to be negative. The calculation of the aforementioned four categories allows for the determination of recall, precision, accuracy, and other pertinent information, which in turn enables the assessment of the model’s detection accuracy and other performance metrics. In this study, we utilize the tomato damage level detection experimental platform to ascertain the impedance information on tomatoes, subsequently transforming these data into a damage condition assessment of the test tomatoes. This assessment is then compared with the specified damage level grading standard for tomatoes, and the resulting data are used to calculate the recall, precision, and accuracy of the model through the coincidence of the two.

3. Results and Analysis

Figure 8 illustrates the use of the tomato damage detection platform to the obtain impedance spectra of the tomato ripening process. It shows how the degree of damage and frequency changes affect the tomato impedance spectral line. The low-frequency signal penetration is weak, while the high-frequency signal penetration is strong. This puts the tomato impedance spectral line at the same damage grade impedance as the increase in frequency decreases. Regardless of whether the non-destructive or destructive sensors were used to measure the four levels of damage, the impedance gradually decreased as the degree of damage to the tomato deepened. The observed gradual decrease in tomato impedance with increasing damage level may be attributed to the following factors: in the four states of tomato damage, namely, no indentation and indentation, the tissue fluid increases in the pressurized part of the fruit tissues. This results in an increase in the concentration of conductive ions in the damaged location, thereby reducing the impedance of the tomato. When the epidermal fold state is present, the internal tissues undergo changes, such as softening of the texture and discoloration. The cells within the damaged portion will undergo apoptosis, resulting in a reduction in the concentration of potassium, sodium, calcium ions, and calcium ions in the intracellular fluids. This, in turn, will lead to a decrease in impedance. Conductive substances, including sodium, calcium, and magnesium ions, as well as other ions, will become part of the extracellular fluid. This will result in an increase in the extracellular concentration of conductive ions, which will in turn lead to an increase in the current density. Consequently, the tomato impedance will decrease once more. In the decay state, the number of dead cells will be greater, and the tomato impedance of the damaged parts will be smaller.

The tomatoes subjected to the pressing test exhibited four states during the placement process: rot, epidermal folds, indentation, and no indentation. From the microstructures of the tomatoes before and after the damage obtained using scanning electron microscopy, it can be observed that the epidermises of the tomatoes ruptured after the force overload, with the cells invaded by the mycelium (Figure 9a). This corresponds to the occurrence of the phenomenon of rot. It can be seen that the stage of rupture deformation of the cells occurred before the epidermal rupture, with the cells rupturing at this time (Figure 9b). This corresponds to the phenomenon of epidermal folds. Finally, it can be observed that the stage of biological yielding deformation occurred before the rupture deformation of the cells, with the cells rupturing at this time (Figure 9c). This corresponds to the phenomenon of indentation. Finally, the stage of non-destructive elasticity deformation occurred before the rupture deformation of the cells, with the cells compressed (Figure 9d). This corresponds to the phenomenon of no indentation.

The experiment tested 145 tomatoes, four of which were accidentally damaged during the test. In these 141 tomatoes, we measured two different positions, each three times before and after the damage, i.e., 141 × 2 × 3 × 2 = 1692. At the same time, we developed a data-filtering program to ensure that the wrong data do not go into the next step and produce 1616 sets of data. The data sets obtained from the experiment were divided into a training set, a validation set, and a test set in accordance with a ratio of 8:1:1. The classification results for the tomatoes at each damage level are presented in

Table 2. Comparison of the training time and the correct rate of different classification algorithms.

Classification Algorithms	Training Time (s)	Training Set ACC	Validation Set ACC	Test Set ACC
Spearman-SVM	5.262	99.065%	98.137%	97.531%
Spearman-ANN	5.748	99.056%	98.758%	98.137%
Spearman-SVM-ANN	3.294	99.381%	98.758%	98.765%

. The Spearman-SVM-ANN classification algorithm has been enhanced in terms of performance and accuracy relative to the other algorithms, as evidenced by the reduction in training time to approximately 3.294 s, which is the shortest among the three algorithms. The test set accuracy of the algorithm is 97.531% for Spearman-SVM, 98.137% for Spearman-ANN, and 98.765% for Spearman-SVM-ANN, in descending order. This represents a classification correctness of more than 0.5% and is comparable with the classification accuracy of the same lossy needle electrode classification [17], which has decreased by 1–2%. However, the overall classification effect is still quite good. The final Spearman-SVM-ANN algorithm assigns a weight of 0.50316 to the SVM and 0.49684 to the ANN.

As evidenced by the classification results presented in

Table 3. Classification results of the Spearman-SVM-ANN algorithm.

Forecasting Team Member Information
		LV1	LV2	LV3	LV4	Total
count	LV1	48	6	0	0	54
	LV2	3	490	1	0	494
	LV3	0	1	264	0	265
	LV4	0	0	1	802	803
%	LV1	88.9%	11.1%	0%	0%	100%
	LV2	0.6%	99.2%	0.2%	0%	100%
	LV3	0%	0.4%	99.6%	0%	100%
	LV4	0%	0%	0.1%	99.9%	100%

, there were six classification errors in 54 LV1s, four classification errors in 494 LV2s, and only one classification error in 265 LV3s and 803 LV4s. The overall classification correctness rate for LV1 was relatively low at 88.9%, while the remaining three impairment classes exhibited considerably higher correctness rates. In the test set data illustrated in Figure 10, one LV1 was misclassified, while the remaining eight were correctly predicted, resulting in an overall accuracy of 88.9%. Similarly, one LV2 was misclassified, while the remaining 45 were correctly predicted, resulting in an overall accuracy of 99%. The remaining accuracies were 100%, which demonstrates that the Spearman-SVM-ANN algorithm is effective and that the classification algorithm is applicable when the tomato localized damage level is light and the detection of the application is necessary. Conversely, when the degree of damage is heavy, the damage can be detected by the naked eye. At this juncture, a dual-pronged approach with the visual approach can enhance the detection accuracy.

4. Summary and Outlook

4.1. Summary

This paper presents the establishment of a localized damage detection model for tomatoes and the acquisition of impedance spectrum data through the construction of a tomato localized damage measurement platform, comprising a self-designed circular four-electrode BIS sensor. The electrical parameters are obtained by fitting the constructed equivalent circuit model of tomatoes. In addition, based on Spearman’s feature selection, the features, including the biological variables of the tomato, the fitted electrical parameters, and the ripening degree of the tomato, are downgraded and inputted into the tomato localized damage classification model to classify the damage.

(1)

A circular four-electrode BIS sensor is designed for the nondestructive measurement of localized damage in tomato. A localized damage measurement platform for tomatoes is constructed by combining this sensor. A comparison of the impedance measurements obtained from the sensor with those obtained from the needle sensor proposed by previous scholars reveals a similar trend, with the impedance decreasing with the increasing damage degree. This validates the effectiveness of the circular four-electrode BIS sensor for tomato in characterizing damage.

(2)

Multiple features, including biological variables, fitted electrical parameters, and tomato ripeness, are subjected to Spearman feature selection, resulting in a downscaled feature set comprising 85% or more of the total features. This downscaled feature set is then inputted into the classification model. A total of 1616 sets of data obtained from the experiments are divided into three subsets: the training set, the validation set, and the test set. The ratio of the training set to the validation set and the validation set to the test set is 8:1:1, respectively. The classification accuracies of the tomato in each damage class are as follows, in descending order: the results demonstrate that 97.531%, 98.137%, and 98.765%, respectively, are the optimal classification algorithms, with Spearman-SVM-ANN being the most effective in detecting tomato damage.

4.2. Outlook

The current article examines the damage location caused by humans. In the actual production process, it is possible that the specific location of the damage may not be known. Therefore, the next step involves exploring the damage at different measurement points and determining the distance of the damaged electrical components. This approach allows for the accurate identification of the damage location.

In addition, the current experimental platform for tomato damage detection is using an acquired impedance analyzer that is powered by 220 V and is not portable. Future work will focus on miniaturizing the electrical impedance-based damage detection platform so that it can truly work offline. At the same time, apply to other fruits and vegetables, and establish the corresponding electrical parameter database, so as to use bioelectrical impedance to solve other fruit quality detection problems. In order to further realize the industrial application, we consider adding our sensors to the robotic arm hand claw to screen out the possible damaged tomatoes under the camera, and at the same time, use the robotic arm to clip the tomatoes and read the impedance spectrum data while clipping for quality classification and detection.