Reliable Augmentation and Precise Identification of EPG Waveforms Based on Multi-Criteria DCGAN

Abstract

The electrical penetration graph (EPG) technique is of great significance in elucidating the mechanisms of virus transmission by piercing-sucking insects and crop resistance to these insects. The traditional method of manually processing EPG signals encounters the drawbacks of inefficiency and subjectivity. This study investigated the data augmentation and automatic identification of various EPG signals, including A, B, C, PD, E1, E2, and G, which correspond to distinct behaviors exhibited by the Asian citrus psyllid. Specifically, a data augmentation method based on an improved deep convolutional generative adversarial network (DCGAN) was proposed to address the challenge of insufficient E1 waveforms. A multi-criteria evaluation framework was constructed, leveraging maximum mean discrepancy (MMD) to evaluate the similarity between the generated and real data, and singular value decomposition (SVD) was incorporated to optimize the training iterations of DCGAN and ensure data diversity. Four models, convolutional neural network (CNN), K-nearest neighbors (KNN), decision tree (DT), and support vector machine (SVM), were established based on DCGAN to classify the EPG waveforms. The results showed that the parameter-optimized DCGAN strategy significantly improved the model accuracies by 5.8%, 6.9%, 7.1%, and 7.9% for DT, SVM, KNN, and CNN, respectively. Notably, DCGAN-CNN effectively addressed the skewed distribution of EPG waveforms, achieving an optimal classification accuracy of 94.13%. The multi-criteria optimized DCGAN-CNN model proposed in this study enables reliable augmentation and precise automatic identification of EPG waveforms, holding significant practical implications for understanding psyllid behavior and controlling citrus huanglongbing.

1. Introduction

Huanglongbing (HLB), also known as citrus greening disease, is a highly infectious bacterial disease caused by the phloem-limited bacterium Candidatus Liberibacter spp., often referred to as the cancer of citrus [1]. The piercing-sucking feeding behavior of the Asian citrus psyllid exacerbates the spread of HLB, making it the primary vector of this disease [2]. Consequently, analyzing the feeding behavior of the Asian citrus psyllid and exploring control measures have become focal points of research [3]. However, the piercing-sucking behavior of these insects is concealed within plant tissues, making direct external observation challenging. The electrical penetration graph (EPG) technique [4,5], as a precise approach for recording the waveforms generated by insect feeding behavior, has become a primary method for studying the behavior of piercing-sucking insects by accurately interpreting the EPG waveforms [6]. The principle of the EPG technique involves connecting an insect with piercing mouthparts and a plant to electrodes, forming part of an electrical circuit. One electrode is attached to the insect and the other to the plant. These electrodes are connected to an electronic measurement circuit, referred to as the primary circuit. A user-controlled power source (Vs) provides alternating current (AC), direct current (DC), or a combination of AC and DC voltage to the circuit [6]. One lead from the power source is grounded, while the other is connected to the plant, typically the growth medium for the plant, such as soil. The insect on the plant is connected to a fine, flexible gold wire electrode that leads to the signal measurement point (M) of the circuit, which serves as the input to a primary amplifier (with a gain of 50 times for DC systems) [7]. According to Ohm’s law, the current passing through each serial component of the circuit is the same, and the voltage differences corresponding to various insect behaviors are displayed in real time as different voltage curves, each representing distinct insect activities [8].

Currently, EPG technology faces challenges in practical applications due to its labor-intensive and time-consuming data processing tasks [9], which require experienced personnel for waveform interpretation and classification. Although experts have attempted to incorporate machine learning methods for waveform recognition, the accuracy of these methods often falls short of practical requirements [10]. Additionally, the complex data processing methods limit the high-throughput analysis and broader application of this technology. Therefore, there is an urgent need to further explore and develop more efficient data processing methods to enhance the application efficiency of EPG technology in studying the behavior of piercing-sucking insects, thereby providing effective support for controlling diseases such as HLB.

In this respect, extensive research has been conducted on the classification of insect EPG signals. Due to the large volume of data, Willett employed Hidden Markov Models (HMM) and Fourier transform methods to process raw EPG data. They performed Fourier transform on unannotated EPG signals of psyllid feeding states and used machine learning methods to classify various feeding states with an accuracy of 97.4% [11]. However, their study only identified six waveform types and did not address the most complex A-wave. Based on existing research, the waveforms were generally categorized into six feeding states: salivation and stylet pathway (C), first contact with phloem (D), phloem salivation (E1), phloem ingestion (E2), xylem ingestion (G), or no probing (NP). Wu used fractal-dimension-Hilbert–Huang transform and decision trees to classify and recognize aphid EPG signals, achieving an average recognition accuracy of 91.43% [5], which needs further optimization.

In addition, research has shown that the sample size of various waveforms in EPG signals may be a significant factor affecting recognition accuracy [12]. In the EPG signals of psyllids, the E1 waveform appears infrequently, making it costly to obtain a sufficient dataset. Data augmentation techniques could be an effective solution to this problem. Generative adversarial networks (GANs) are unsupervised generative models initially developed to augment datasets in machine vision. These models aim to learn the distribution of real data and generate virtual data that resemble the real data [13,14]. GANs have shown great potential in the field of one-dimensional signal research. Teng introduced GANs for data augmentation, improving the classification accuracy of their support vector machine (SVM) model from 88.89% to 95.33% [15]. Zhang used a deep convolutional generative adversarial network (DCGAN) to generate an extended hyperspectral dataset of maize haploid seeds, achieving 100% classification accuracy when combined with a SVM model [16]. In a study on electrocardiography, the use of a deep learning-based atrial fibrillation (AF) signal classification method benefited more from data augmentation using GANs and GARCH than from oversampling. Ultimately, the accuracy improved by 3% after GAN augmentation [17]. These studies provide effective solutions to the time-consuming and labor-intensive problems of collecting sufficient sample data for EPG signals. However, the methods for assessing the reliability of generated data remain immature and require further research testing.

Herein, this study aimed to generate and evaluate the difficult-to-collect E1 waveforms of psyllid EPG signals using a multi-criteria DCGAN and to identify the seven types of EPG signals with four machine learning algorithms (i.e., DT, SVM, KNN, and CNN). The main research tasks of this study were as follows: (1) The DCGAN network was used to generate psyllid E1 waveforms, with maximum mean discrepancy (MMD) employed to assess the similarity between generated and real data, singular value decomposition (SVD) to determine the optimal training iterations for DCGAN, and visual inspection to ensure the accuracy of the generated data. (2) Verification of the accuracy of the generated waveforms through visual inspection and comparison of the classification accuracy of DT, SVM, KNN, and CNN models before and after DCGAN optimization to validate the effectiveness of the DCGAN network. (3) Selection of the DCGAN-CNN framework to achieve efficient recognition of seven types of waveforms (i.e., A, B, C, PD, E1, E2, and G) in psyllid EPG signals, providing insights into the feeding behavior of psyllids. The proposed DCGAN sample augmentation strategy, based on parameter optimization and evaluation, combined with the CNN framework, can serve as a rapid, accurate, reliable, and stable method for EPG waveform classification and recognition. This study provides a technical approach for the automatic segmentation and waveform discrimination of actual EPG signals and offers a theoretical basis for research in insect virus transmission mechanisms and crop insect resistance mechanisms.

2. Materials and Methods

2.1. Experimental Data Collection

Adult Asian citrus psyllids tested in the experiment were collected from the wild in Gutian County (Fujian Province) and transferred to a net house at Fujian Agriculture and Forestry University for rearing. They were fed on tender shoots of the virus-free citrus cultivar Jiu Li Xiang (Murraya). The newly sprouted summer shoots with a length of approximately 8–10 cm and containing more than six new leaves were selected. The cut ends of the shoots were wrapped with cling film and inserted into water for moisturizing [18].

For data collection (as shown in Figure 1), 2–3 cm copper wires were soldered to copper pins, and 5 cm gold wires were fixed to the other end of the copper wires using water-soluble silver glue. The psyllid was gently held at the wing tip with tweezers, and a drop of water-soluble silver glue was applied to its back. Under a stereomicroscope, one end of the free gold wire was inserted into the silver glue on the back of the psyllid and left for 1 min to ensure a tight bond, avoiding contact between the gold wire and other body parts of the psyllid. The copper pin was inserted into a micromanipulator, which was adjusted to allow the psyllid to feed freely on the leaf. Finally, the plant electrode was inserted into the nutrient pot, the Faraday cage was closed to shield external electromagnetic interference, and the EPG recording system was activated to record EPG data [4].

A Giga-8 DC-EPG system was used to record the EPG signals. (Wageningen, The Netherlands). The experiment was conducted in a climate-controlled chamber with light conditions (temperature at 25 ± 1 °C, relative humidity at 60 ± 5%) within the Faraday cage. Each experiment lasted for 8 to 20 h. A total of 300 h of psyllid EPG waveform data were collected, excluding weak signals and waveforms where the gold wire had detached [19].

2.2. Data Reading and Segmentation

The original EPG signals were manually inspected and annotated using Stylet++ software (version 1.30), which exported the annotations into an (.ANA) file containing the starting points of each segment. This tab-delimited text file, along with the original EPG data in ASCII format, was processed using MATLAB 2023A. To eliminate annotation errors, the first and last 10 data points of the waveforms were removed. The label information from the (.ANA) file was then used to slice the EPG data, and the results were saved as plain text files for further analysis. Then, different types of waveforms were stored as a dataset with a unified length of 100 data points. This length was chosen because the frequencies of waves A, E1, and E2 were generally 10–15 Hz, 5–8 Hz, and 4 Hz [20], respectively. This ensured that all features of each waveform type were included while facilitating processing. The E1 waveforms were stored as a separate dataset. All data were normalized within a specific range, then expanded by 2¹⁶ times and converted to TIFF format for data augmentation. Table 1 lists the number of each waveform sample, with the training and testing sets divided in a ratio of 8:2 for subsequent modeling.

2.3. Algorithmic Methods for Data Analysis

2.3.1. EPG Data Augmentation Based on DCGAN

• (1). GAN algorithmic model

GANs are a type of unsupervised deep learning model consisting of two components, a generator and a discriminator. These two components work in an adversarial manner while simultaneously learning from each other. The generator receives random noise Z (Equation (1)) and generates data G(z) that is as realistic as possible to compete with the discriminator D. The discriminator is used to distinguish whether the generated data is real, striving to accurately differentiate between the generated data G(z) and the real data X. In this setup, G and D constantly compete, and their parameters are continuously optimized until an ideal result is achieved.

(1)$MinmaxV(D,G)=E_{x~P_{data}\left(X\right)}\left[logD\right(x\left)\right]+Ez-p\left(z\right)\left[log\right(1-D\left(G\right(Z\left)\right)\left)\right]$

In this context, D(x) represents the probability that the sample x is real data, while D(G(z)) denotes the probability that the generated data G(z) is real. These probabilities range from [0, 1], where 1 indicates real data and 0 indicates non-real data. Therefore, the optimal state of this adversarial competition is achieved when D(G(z)) = 0.5, meaning the generator can generate data that is indistinguishable from real data. According to Equation (1), the optimization functions for D (Equation (2)) and G (Equation (3)) can be derived as follows:

(2)$MaxV\left(D,G\right)=E_{x~Pdata\left(x\right)}\left[logD\right(x\left)\right]+E_{z-p\left(z\right)}\left[log\right(1-D\left(G\right(Z\left)\right)\left)\right]$

(3)$MinV(D,G)=E_{z-p\left(z\right)}\left[log\right(1-D\left(G\right(Z\left)\right)\left)\right]$

• (2). DCGAN algorithmic model

The GAN has evolved into many variants, among which the architecture of DCGAN is widely used. As show in Figure 2, DCGAN replaced the fully connected layers in the generator and discriminator of the original GAN with convolutional and transposed convolutional layers. This improvement enhanced the stability of GANs and improved the quality of the generated results [21].

The data used in this study were essentially 32-bit floating-point numbers. The data were initially scaled up by 2¹⁶ times and then converted from one-dimensional data into TIFF format images. First, the generator network was defined, which took a one-dimensional latent noise vector as input. This latent noise vector was typically randomly sampled from a Gaussian distribution. The generator network consisted of a series of transposed convolutional layers that progressively expanded the one-dimensional input data into images of appropriate size. Each transposed convolutional layer was followed by a batch normalization layer and a Leaky ReLU activation function with a slope of 0.2 to enhance training stability and introduce non-linearity (as shown in Figure 2). A fully connected layer adjusted the output to the desired image size, and a hyperbolic tangent (tanh) activation function ensured the output pixel values were within a suitable range.

Then, the discriminator network was defined, which aimed to distinguish between the real images and the fake images generated by the generator. The discriminator network consisted of a series of convolutional layers that extracted features from the input images. Each convolutional layer was followed by a Leaky ReLU activation function with a slope of 0.2. The final convolutional layer was connected to a fully connected layer with an output size of 1, which used a sigmoid activation function for classification. This setup outputted a scalar value representing the probability that the input image was real.

During the data training process, the real one-dimensional data and fake images generated by the generator were used to train the discriminator network. Simultaneously, the generator network was trained to deceive the discriminator as much as possible, making the generated images as realistic as possible. Through iterative training, the generator gradually learned the data distribution, thereby generating high-quality data.

2.3.2. CNNs Classification Model

CNNs are one of the most widely used feedforward neural network structures in the field of deep learning. They include convolution operations and have a deep structure. CNNs can effectively handle one-dimensional or multi-dimensional data, such as image data, time series data, video data, etc. The main components of CNNs include the input layer, convolutional layer, pooling layer, fully connected layer, and output layer [22,23].

Convolutional Layer: The main function of the convolutional layer is to extract features from the input data through local operation. The convolutional layer contains multiple convolutional kernels, each consisting of a set of weight coefficients and a bias vector. Each element of the convolutional kernel performs a convolution operation with the corresponding region of the input feature, known as the receptive field. Typically, the convolutional kernel is square-shaped. During the operation, the convolutional kernel slides systematically over the input features, performing matrix element multiplication and summation within the receptive field, followed by adding the bias term to generate a single pixel value in the output feature map [24].

(4)$$\begin{gathered} Z^{L+1}\left(i+j\right)=[Z^L\otimes w^{l+1}]\left(i,j\right)+b=\sum _{k=1}^{k_l}\sum _{x=1}^f\sum _{y=1}^f\left[Z_k^l\left(s_{0}i+x,s_{0}j+y\right)w_k^{i+1}\left(x,y\right)\right]+b \\ \left(i,j\right)\in \left\{0,1,\dots .L_{l+1}\right\} \end{gathered}$$

(5)$L_{l+1}{\displaystyle \frac{L_{l+2p-f}}{s_0}}+1$

where b represents the bias term, and I and O represent the input and output of the l + 1 convolutional layer, respectively, which are also referred to as feature maps. L{l + 1} is the size of O. Z(i, j) corresponds to the pixel of the feature map. K is the number of channels in the feature map, and f, S₀, and p correspond to the size of the convolutional kernel, the stride, and the padding number, respectively.

Pooling Layer: The primary purpose of the pooling operation is to perform down-sampling. The principle of the non-linear pooling function is to replace individual pixel values in the feature map with statistical values from their neighboring regions. During the pooling operation, the spatial dimensions of the data are reduced while preserving the key features, gradually decreasing the size of the feature map. In this process, the number of parameters and the computational load are also reduced accordingly, which helps to control overfitting to a certain extent [25].

Fully Connected Layer: The fully connected layer is analogous to the hidden layers in a traditional feedforward neural network. It is typically located at the final part of the network and serves to transmit signals to other fully connected layers.

In CNN, multiple convolutional and pooling layers are usually set up in an alternating manner. Each neuron in the convolutional layer connects to a local region of the previous layer and extracts features from that region. The pooling layer is responsible for local sensitivity and secondary feature extraction. This alternating structure of convolutional and pooling layers helps to reduce feature resolution and the number of parameters that need optimization.

In this study, the CNN network employed a structure where convolutional and pooling layers were alternately stacked. Initially, the input layer received the preprocessed data and performed 16 convolution operations, followed by batch normalization and a ReLU activation function to accelerate training and introduce non-linearity. A max-pooling layer was then applied to reduce the feature map size and extract the most important features. This was followed by 32 convolution operations, batch normalization, ReLU activation, and another max-pooling layer. Finally, the fully connected layer integrated these features and used a Softmax layer for multi-class classification.

2.4. Evaluating Indicators of Generated Data

2.4.1. Quality Assessment of Generated Data

The quality of data generated by GANs directly impacts the performance of the model. Therefore, it is crucial to objectively and accurately evaluate the quality of these data. Currently, there are four mainstream evaluation methods:

• (1). Selecting appropriate metrics to quantify the differences between generated data and real data.

• (2). Visually comparing the original data with the generated data.

• (3). Comparing the performance improvement of classification models after incorporating the generated data.

• (4). Analyzing whether the signal characteristics fall within a reasonable range.

If a large amount of E1 waveforms were generated by DCGAN, it would require extracting numerical features from each segment of the signal and comparing them with the original signals to determine if they fall within a reasonable range. This would be an extremely tedious and challenging task. Therefore, the fourth method is not applicable. Instead, a combination of the first three methods is used to validate the reliability of the generated data.

2.4.2. Accuracy Analysis of Generated Data

The loss curve can validate the effectiveness of the model but cannot evaluate the quality of the generated data. Therefore, an evaluation metric is needed to assess the quality of the generated data. In this study, MMD was used as the evaluation metric to quantify the differences between the generated data and the real data. MMD determines whether two distributions are identical and is commonly utilized to assess the similarity between distributions, such as vectors, strings, sequences, and images [26]. It is widely applied in various fields, including neural network optimization, domain adaptation, intelligent fault detection, estimating class proportions in unlabeled sample sets, data adaptability assessment, image feature extraction, regression, and time series prediction tasks.

In DCGAN, the discriminator is trained to identify the differences between real samples (training set) and generated samples, validating the effectiveness and accuracy of MMD for measuring the discrepancy between generated and real data [27]. Therefore, MMD was used to analyze the similarity between the generated data and the original data in this study. A larger MMD value indicates greater dissimilarity between the generated data and the original data, whereas a smaller MMD value indicates stronger similarity between the generated data and the original data. The calculation for MMD is as Equation (6):

(6)$MM{D}^2={\displaystyle \frac{1}{m^2}}\sum _{i,j=1}^{m}k\left(x_{i,},x_j\right)-{\displaystyle \frac{2}{mn}}\sum _{i,j=1}^{m,n}k\left(x_i{,y}_i\right)+{\displaystyle \frac{1}{n^2}}\sum _{i,j=1}^{n}k(y_i,y_j)$

The Gaussian kernel function, which served as a similarity function between data points in the MMD calculation, was chosen as the kernel function. It is a monotonic function of the Euclidean distance between two vectors, as shown in Equation (7):

(7)$k\left(x,x^{\prime }\right)=e^{\frac{{‖x-x^{\prime }‖}^2}{{2\sigma }^2}}$

where σ is the bandwidth, which controls the local influence range of the Gaussian kernel function. In this study, the Gaussian kernel within MMD performed well in capturing the non-linear relationships between data points within the distributions, which was essential for evaluating the similarity between the generated and experimentally obtained data of EPG signals.

2.4.3. Visual Discrimination of Generated Data

To further verify the quality of the generated data, a visual evaluation was conducted by three experts specializing in psyllid EPG signals. The experiment involved three repeated random selections of 175 E1 waveforms from the dataset, including 75 original and 100 E1 waveforms generated by the DCGAN model. The real and generated waveforms were randomly assigned identifiers, with the source of each waveform corresponding to its identifier. The experts were tasked with determining whether each waveform was an E1 waveform and whether it was real or generated. Here, $AC{C}^d$ represents the accuracy rate of the expert visual discrimination, TP$R^d$ is the proportion of true E1 waveforms correctly identified as real, and $FP{R}^d$ is the proportion of generated E1 waveforms mistakenly identified as true. The formulas for calculating $AC{C}^d$, TP$R^d$, and FP$R^d$ are shown below:

(8)$AC{C}^d={\displaystyle \frac{T{T}^1+F{F}^1}{T{T}^1+T{F}^1+F{T}^1+F{F}^1}}$

(9)$TP{R}^d={\displaystyle \frac{T{T}^1}{T{T}^1+T{F}^1}}$

(10)$FP{R}^d={\displaystyle \frac{F{T}^1}{F{T}^1+F{F}^1}}$

where, if the waveform is identified as E1 and considered real, it is marked as ‘True’ with the symbol T. If the waveform is identified as generated or not genuinely E1, it is marked as ‘False’ with the symbol F. T1 and T2 represent the numbers of real and generated E1 waveforms, respectively.

2.4.4. Optimizing the Training Iterations of DCGAN by SVD

Singular value decomposition (SVD), an effective dimensionality reduction technique, can map high-dimensional data into a lower-dimensional subspace, effectively reducing the E1 waveform to a single point [21]. SVD rapidly and reliably compresses EPG data from hundreds of features into a few composite metrics, facilitating the visualization and differentiation of data variations.

Generally, the quality of the generated data in a generative model is evaluated based on two criteria: the similarity to real data and the diversity of the generated samples. The Frechet Inception Distance (FID) was commonly used to assess both aspects simultaneously [28]. However, to deal with one-dimensional data, there is a higher risk of producing overly similar data with insufficient diversity during training, in contrast to two-dimensional data. To address this, in this study we used MMD to ensure the generated one-dimensional data closely resembled the real data, while SVD was applied to guarantee the diversity of the generated samples [29,30]

2.4.5. Receiver Operating Characteristic (ROC) Curve

The ROC curve is a graphical representation of the performance of a binary classifier. It is created by plotting the true positive rate (TPR) against the false positive rate (FPR) at various classification thresholds. TPR, also known as sensitivity or recall, represents the proportion of positive instances that are correctly classified as positive, while FPR represents the proportion of negative instances that are incorrectly classified as positive. Each point on the ROC curve corresponds to a specific threshold setting.

Area under the ROC curve (AUCROC) is a widely used evaluation metric in machine learning for assessing the performance of binary classification models. It provides a single scalar value representing the overall quality of the model predictions. AUCROC quantifies the overall performance of a classifier across all possible threshold settings. The AUCROC value ranges from 0 to 1, with higher values indicating better classification performance. An AUCROC of 1 denotes a perfect classifier, while an AUCROC of 0.5 indicates a random classifier, signifying poor performance [31,32].

2.4.6. Evaluation of Model Performance

To validate the generalization capability of the model, we employed four metrics to evaluate performance, including accuracy, precision, recall, and F1-score. Accuracy measures the proportion of correctly classified instances, as calculated by Equation (11):

(11)$\mathrm{A}\mathrm{c}\mathrm{c}\mathrm{u}\mathrm{r}\mathrm{a}\mathrm{c}\mathrm{y}={\displaystyle \frac{1}{n_{\mathrm{s}\mathrm{a}\mathrm{m}\mathrm{p}\mathrm{l}\mathrm{e}}}}\sum _{i=1}^{n_{\mathrm{s}\mathrm{a}\mathrm{m}\mathrm{p}\mathrm{l}\mathrm{e}}}1\left(y_{ai}=y_i\right)$

where $y_{ai}$ represents the predicted class for the i-th sample, $y_i$ is the true class, and n is the number of tested samples [33].

3. Results and Discussion

3.1. Validity Analysis of Generated Waveforms by DCGAN

The GAN framework functioned with its generator G, and the discriminator D continuously competed until the generated data closely resembled real data. Therefore, an effective GAN model should have exhibited fluctuating loss curves for D during training, making it increasingly difficult for D to distinguish between real and generated data. To obtain comprehensive observations, a high number of iterations was set in this study. Under all parameter settings, the MMD values were extremely small (<0.1), indicating there was a minimal difference between the generated data and the original data. Furthermore, the SVD plots under different epochs (Epoch) were compared as shown in Figure 3. For Figure 3a with Epoch = 500, the spatial distribution of the generated waveform was not limited to the first singular value component. Compared to Figure 3b with Epoch = 1000, the distribution was more concentrated in the first singular value component but still less distributed than the real waveform. As shown in Figure 3c, with Epoch = 2000, the distribution was more uniform than that with Epoch = 1000. When Epoch reached 2000, the spatial distribution of the generated waveform was similar to the real waveform, covering a broader range in the entire three-dimensional space. Finally, based on Figure 3d, it was observed that with Epoch = 4000, the spatial distribution of the generated waveform became even more extensive than the real waveform. This indicated that with the increase in training Epoch, the distribution of the generated waveform became closer to that of the real waveform. The DCGAN network completed its training in an average of 0.77 s per round, totaling 500 rounds. This training was facilitated by an i7-12700H processor and a 3060 graphics card. These results suggested that the DCGAN generation strategy was effective. However, whether the generator was overfitting or underfitting remained to be further verified.

3.2. Accuracy Evalution of Generated E1 Waveform

The characteristics of E1 waveforms included upward peaks, relatively low voltage levels, and frequencies ranging from 4 to 7.5 Hz. As shown in Figure 4, four randomly selected E1 waveforms generated with Epoch = 2000 are presented. The generated waveforms in Figure 4 exhibited typical E1 waveform features. In Figure 4a, the waveform showed upward peaks with four prominent peaks, and the waveform length was 1 s, consistent with the frequency characteristics of E1 waves. The most prominent feature was observed in Figure 4c, where the upward curve of the waveform peak was nearly vertical, followed by a slightly descent, with a frequency ranging from 4 to 7.5 Hz, and the tail of the peak exhibited a certain shape. Consistent with Figure 3, the SVD plot of the generated data showed a more balanced and widespread distribution with Epoch = 2000. Therefore, considering the MMD indicator, SVD plot, and the corresponding classification accuracy of Epochs, the waveforms generated from training at Epoch = 2000 were selected as the required samples for subsequent experiments.

The visual inspection accuracy of the generated data was further analyzed (as shown in Table 2). The visual inspection accuracy rates of the three experts were 40%, 42.9%, and 40%, respectively, indicating that it was difficult for experts to distinguish between real and generated data. Additionally, it was found that the average rating of the generated data was higher than that of the real data, suggesting that the probability of generated data being judged as real was higher than that of real data. This indicates that the generated data can be used as a supplement to real data. Furthermore, the inspection results of the three experts were consistent for 113 samples (out of 175 samples), achieving a consistency rate of 64.6%, further validating the effectiveness and usability of the generated data.

3.3. Comparison of Model Accuracies Based on Data Generation

Following the evaluation of the MMD metric, validation with SVM, and visual inspection, we selected data generated at 2000 epochs as the final training dataset. We then compared the accuracy of the training results with prior results, as shown in Table 3. The experimental outcomes demonstrated that the accuracy of the CNN on the test set increased significantly from 86.3% before data augmentation to 94.2%, an enhancement of 7.9%, making it the most notable among the four classification models. Deep learning models generally require substantial amounts of data for training, and the additional data provided by augmentation enabled these models to better capture the underlying data features, achieving near-optimal accuracy. The test set accuracy of KNN improved by 7.1%, rising from 77.5% to 84.6%; DT improved by 5.8%, from 87.2% to 81.4%; and SVM improved by 6.9%, from 78.7% to 85.6%. These significant improvements can be attributed to data augmentation, which not only increased the diversity of training samples but also enhanced the generalization capabilities of the models, leading to better performance on new data.

Further analysis indicated that data augmentation enabled deep learning models to learn data features more comprehensively, thereby reducing the risk of overfitting. Particularly for CNN, the augmented dataset provided a larger pool of training samples, allowing the model to generalize more effectively rather than being constrained by the characteristics of the training set. This increased diversity and richness in the data significantly enhanced the model’s stability and predictive accuracy.

Moreover, the improvements in accuracy for KNN, DT, and SVM on the augmented dataset further validated the effectiveness of data augmentation in addressing data imbalance issues. For high-dimensional data, the increased sample diversity from data augmentation enhanced the accuracy of KNN’s neighborhood search; DT benefited from reduced bias due to more diverse samples, improving overall model performance; and SVM achieved better-defined decision boundaries with more support vectors, thereby increasing classification accuracy.

In summary, data augmentation not only enhanced the training outcomes of the models but also significantly improved the performance of all classification models on the test set. This validated its effectiveness in resolving data imbalance and improving model generalization. These findings offer robust support and reference for the further optimization and application of deep learning models.

Figure 5 illustrates the ROC curves of four classification models for comparing their performance. The KNN model performed best for the A-wave, with an AUC value of 0.9309. This result may have been due to the irregular nature of the A-wave, allowing the DT to adapt well to the complexity and non-linear relationships of the data. All four classifiers achieved good AUC values for the B-wave. For C-wave recognition, KNN and SVM showed relatively good performance. Both PD-wave and E1-wave, which have distinct features, were well classified by all four classifiers. SVM performed the best for G-wave classification, with an AUC of 0.9819. Except for the DT model, the other three classifiers had AUC scores above 0.9. This result may be attributed to the advantage in handling high-dimensional and linearly separable data of SVM. The large dataset size suited the CNN model, which required extensive training data and computational resources.

3.4. Model Performances for Classification of EPG Waveforms

In this study, we performed a comprehensive analysis of the performance of four models across multiple metrics (as shown in Table 4). The CNN model exhibited exceptional performance in terms of accuracy, recall, and F1 scores, nearly reaching optimal levels in all categories. In contrast, the KNN model, configured with Euclidean distance, equal weights, and a k-value of 1, showed relatively unstable performance in certain categories, suggesting a need for further feature engineering. The SVM model, using an RBF kernel function, weight equilibrium, and an auto-adjusted kernel scale, maintained a certain level of stability. However, it demonstrated insufficient recall rates in some categories, indicating the potential for further adjustments. Additionally, the DT model, with parameters including the maximum number of splits, split criterion, and minimum parent node size of 10, exhibited excellent precision but fell short in recall rates and F1 scores for certain categories. This may necessitate adjustments to the pruning operations to enhance performance.

These results may be attributed to the following factors: firstly, external interference and insect disruptions of EPG signals, which are biological signals. Despite efforts to mitigate and minimize individual psyllid variations, data disparities persist. This leads to attribute value dissipation and consequent misclassification, particularly evident in C-waves. Secondly, inconsistent data sample quality due to the large number of data samples, resulting in varying levels of data quality. Deep learning methods are better suited to learning features under such circumstances. Additionally, slicing the waveforms results in the loss of time-series information. Ultimately, all four classifiers require more training samples to enhance their recognition capabilities.

4. Conclusions

In this study, acquiring sufficient E1 waveforms from the citrus psyllid EPG signal collection proved challenging, leading to a limited quantity of E1 waveforms and an uneven distribution of datasets, which ultimately affected the modelling recognition performance. Consequently, a data augmentation method based on DCGAN was employed to generate EPG data of E1, aiming to optimize sample data and enhance the recognition performance. Additionally, to assess the reliability of the EPG data generated by DCGAN and quantitatively evaluate the error between the generated and real data, reliable waveforms were selected based on visual discrimination results, SVD plots, and MMD evaluation metrics. Four classification models (CNN, KNN, DT, and SVM) were established to evaluate the performance of DCGAN-based data augmentation and identify the seven types of EPG signals. The experimental results demonstrated that the proposed data augmentation method could generate reliable samples for imbalanced datasets, thereby improving the modelling performance. Furthermore, continuously adding samples to the training set further improved the classifiers discrimination accuracy. The CNN learned more EPG data features from the incremental data, leading to a significant improvement in classification accuracy on the test set. However, it is still challenging to integrate time-series analysis to further enhance classification accuracy and improve the visualization of classification results to clarify relationships between different waveforms. The DCGAN-CNN model based on reliable evaluation metrics could effectively mitigate dataset imbalance or limitations in nucleus recognition, thus significantly improving the classification accuracy of EPG signals. Even when classifiers lacked sufficient data fitting and classification capabilities, such as the SVM and DT models, the improvement in classification accuracy with data augmentation still showed efficacy. This study provides methodological guidance and data support for further exploration of the probing behavior and transmission mechanisms of citrus psyllids.

Author Contributions

Conceptualization, X.K., C.W. and L.Z.; Methodology, X.K., C.W., L.Z. and F.Q.; Software, C.W., W.Z., N.H. and S.C.; Validation, X.K., C.W. and T.Z.; Formal Analysis, X.K., C.W. and F.Q.; Investigation, X.K., C.W., W.Z. and L.Z. Data Curation, X.K., C.W. and F.Q.; Writing—Original Draft Preparation, X.K., C.W. and F.Q.; Writing—Review and Editing, X.K., C.W. and F.Q.; Visualization, X.K., L.Z. and S.C.; Supervision, H.W., X.K., T.Z. and F.Q.; Project Administration, L.Z. and F.Q.; Funding Acquisition, T.Z. and F.Q. All authors have read and agreed to the published version of the manuscript.

Institutional Review Board Statement

Not applicable.

Informed Consent Statement

Not applicable.

Data Availability Statement

Data will be made available on request.

Conflicts of Interest

The authors declare no conflicts of interest.

Footnotes

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Figures and Tables

Figure 1. Schematic diagram of EPG data collection.

Figure 2. DCGAN Network Framework.

Figure 3. The SVD graphs of DCGAN with different Epoch values: (a) Epoch = 500, (b) Epoch = 1000, (c) Epoch = 2000, and (d) Epoch = 4000.

Figure 4. Generated E1 waveform: (a) E1-1, (b) E1-2, (c) E1-3, and (d) E1-4.

Figure 5. ROC curves of four models for classification of EPG waveforms: (a) KNN, (b) SVM, (c) CNN, and (d) DT.

Table 1

Training set and testing machine sample quantity table.

Waveform Style	Number of Training Sets	Number of Testing Sets
A	1021	221
B	1036	208
C	1087	218
PD	561	113
E1	232	116
E2	1000	200
G	906	182

Table 2

Expert visual inspection results of generated E1 waveforms.

Experts	Labels	Expert Inspection			Results
		T1	F1	Total Number	$\mathbf{A}\mathbf{C}{\mathbf{C}}^{\mathbf{d}}$	$\mathbf{T}\mathbf{P}{\mathbf{R}}^{\mathbf{d}}$	$\mathbf{F}\mathbf{P}{\mathbf{R}}^{\mathbf{d}}$
Expert A	T	45	30	75	40%	61%	76%
	F	74	26	100
Expert B	T	48	27	75	42.9%	67%	75%
	F	77	23	100
Expert C	T	46	29	75	40%	65.7%	76%
	F	76	24	100

Note: T rows and T1 column: The number of E1 that were originally true and judged to be true, represented by TT1. T rows and F1 column: The number of E1 that were originally genuine but were determined to be forged, represented by TF1. F rows and T1 column: The number of E1 that were originally forged but judged to be genuine, represented by FT1. F rows and F1 column: The number of E1 that were originally forged and judged as forged, represented by FF1. $\mathrm{A}\mathrm{C}{\mathrm{C}}^{\mathrm{d}}$, $\mathrm{T}\mathrm{P}{\mathrm{R}}^{\mathrm{d}}$, and $\mathrm{F}\mathrm{P}{\mathrm{R}}^{\mathrm{d}}$ were calculated by Equation (8), Equation (9), and Equation (10), respectively.

Table 3

Comparison of classification accuracy before and after data augmentation.

Model	Without DCGAN		With DCGAN
	Training Set	Testing Set	Training Set	Testing Set
CNN	87.6	86.3	95.3	94.2
KNN	78.1	77.5	85.2	84.6
DT	82.2	81.4	88	87.2
SVM	78.8	78.7	87.5	85.6

Table 4

Comparison of model performances in identifying different types of EPG signals.

Mode	Index	1 (A)	2 (B)	3 (C)	4 (PD)	5 (E1)	6 (E2)	7 (G)
CNN	Precision	0.9573	1.0000	0.9777	1.0000	1.0000	0.9423	1.0000
	Recall	0.9438	1.0000	0.9441	1.0000	1.0000	0.9457	1.0000
	F1 Score	0.9505	1.0000	0.9606	1.0000	1.0000	0.9440	1.0000
KNN	Precision	0.70199	0.96193	0.78922	0.96392	0.99404	0.72532	0.8623
	Recall	0.72674	1.0000	0.60626	1.0000	1.0000	0.84196	0.81450
	F1 Score	0.71415	0.9806	0.68574	0.98163	0.99701	0.7793	0.83772
SVM	Precision	0.73554	0.91304	0.90215	0.85374	0.96899	0.65401	0.84105
	Recall	0.69736	0.89189	0.42410	0.95722	1.0000	0.97135	0.89126
	F1 Score	0.71594	0.90234	0.57697	0.90252	0.98425	0.78170	0.86542
DT	Precision	0.82014	0.98292	0.82222	0.99645	0.99900	0.71622	0.80843
	Recall	0.71009	1.0000	0.68077	1.0000	1.00000	0.88170	0.89979
	F1 Score	0.76115	0.99139	0.74484	0.99822	0.9995	0.79039	0.85166

References

1. Boina, D.R.; Bloomquist, J.R. Chemical control of the Asian citrus psyllid and of huanglongbing disease in citrus. Pest Manag. Sci.; 2015; 71, pp. 808-823. [DOI: https://dx.doi.org/10.1002/ps.3957] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/25491482]

2. Gottwald, T.R. Current epidemiological understanding of citrus huanglongbing. Annu. Rev. Phytopathol.; 2010; 48, pp. 119-139. [DOI: https://dx.doi.org/10.1146/annurev-phyto-073009-114418] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/20415578]

3. Hall, D.G.; Richardson, M.L.; Ammar, E.D.; Halbert, S.E. Asian citrus psyllid, Diaphorina citri, vector of citrus huanglongbing disease. Entomol. Exp. Appl.; 2013; 146, pp. 207-223. [DOI: https://dx.doi.org/10.1111/eea.12025]

4. Walker, G.P.; Fereres, A.; Tjallingii, W.F. Guidelines for conducting, analyzing, and interpreting electrical penetration graph (EPG) experiments on herbivorous piercing–sucking insects. Entomol. Exp. Appl.; 2024; 172, pp. 564-580. [DOI: https://dx.doi.org/10.1111/eea.13434]

5. Xing, Y.; Li, B.; Wu, L.; Yan, F. Waveforms Eavesdropping Prevention Framework: The Case of Classification of EPG Waveforms of Aphid Utilizing Wavelet Kernel Extreme Learning Machine. Appl. Artif. Intell.; 2023; 37, 2214766. [DOI: https://dx.doi.org/10.1080/08839514.2023.2214766]

6. Tjallingii, W.F. Comparison of AC and DC systems for electronic monitoring of stylet penetration activities by homopterans. Principles and Applications of Electronic Monitoring and Other Techniques in the Study of Homopteran Feeding Behavior; Entomological Society of America: Annapolis, MD, USA, 2000; 41.

7. Backus, E.A.; Bennett, W.H. The AC–DC correlation monitor: New EPG design with flexible input resistors to detect both R and emf components for any piercing–sucking hemipteran. J. Insect Physiol.; 2009; 55, pp. 869-884. [DOI: https://dx.doi.org/10.1016/j.jinsphys.2009.05.007]

8. McLEAN, D.L.; Kinsey, M.G. A technique for electronically recording aphid feeding and salivation. Nature; 1964; 202, pp. 1358-1359. [DOI: https://dx.doi.org/10.1038/2021358a0]

9. Tayyab, U.B.; Arif, M.J.; Gogi, M.D.; Akhtar, S.; Abdullah, M.J.; Ali, F. Tracking the Feeding Mechanism of Sap-Sucking Insect-Pests Through Electropenetrography (EPG). J. Insect Behav.; 2024; 37, pp. 58-81. [DOI: https://dx.doi.org/10.1007/s10905-024-09850-1]

10. Adasme-Carreño, F.; Muñoz-Gutiérrez, C.; Salinas-Cornejo, J.; Ramírez, C.C. A2EPG: A new software for the analysis of electrical penetration graphs to study plant probing behaviour of hemipteran insects. Comput. Electron. Agric.; 2015; 113, pp. 128-135. [DOI: https://dx.doi.org/10.1016/j.compag.2015.02.005]

11. Willett, D.S.; George, J.; Willett, N.S.; Stelinski, L.L.; Lapointe, S.L. Machine learning for characterization of insect vector feeding. PLoS Comput. Biol.; 2016; 12, e1005158. [DOI: https://dx.doi.org/10.1371/journal.pcbi.1005158]

12. Ramezan, C.A.; Warner, T.A.; Maxwell, A.E.; Price, B.S. Effects of training set size on supervised machine-learning land-cover classification of large-area high-resolution remotely sensed data. Remote Sens.; 2021; 13, 368. [DOI: https://dx.doi.org/10.3390/rs13030368]

13. Goodfellow, I.; Pouget-Abadie, J.; Mirza, M.; Xu, B.; Warde-Farley, D.; Ozair, S.; Courville, A.; Bengio, Y. Generative adversarial networks. Commun. ACM; 2020; 63, pp. 139-144. [DOI: https://dx.doi.org/10.1145/3422622]

14. Wang, Z.; She, Q.; Ward, T.E. Generative adversarial networks in computer vision: A survey and taxonomy. ACM Comput. Surv. (CSUR); 2021; 54, pp. 1-38. [DOI: https://dx.doi.org/10.1145/3439723]

15. Teng, G.E.; Wang, Q.Q.; Kong, J.L.; Dong, L.Q.; Cui, X.T.; Liu, W.W.; Wei, K.; Xiangli, W.T. Extending the spectral database of laser-induced breakdown spectroscopy with generative adversarial nets. Opt. Express; 2019; 27, pp. 6958-6969. [DOI: https://dx.doi.org/10.1364/OE.27.006958]

16. Zhang, L.; Nie, Q.; Ji, H.; Wang, Y.; Wei, Y.; An, D. Hyperspectral imaging combined with generative adversarial network (GAN)-based data augmentation to identify haploid maize kernels. J. Food Compos. Anal.; 2022; 106, 104346. [DOI: https://dx.doi.org/10.1016/j.jfca.2021.104346]

17. Hatamian, F.N.; Ravikumar, N.; Vesal, S.; Kemeth, F.P.; Struck, M.; Maier, A. The effect of data augmentation on classification of atrial fibrillation in short single-lead ECG signals using deep neural networks. Proceedings of the ICASSP 2020—2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP); Barcelona, Spain, 4–8 May 2020.

18. Tjallingii, W. EPG Systems. Available online: https://www.epgsystems.eu/epg-measuring (accessed on 1 June 2024).

19. Moreno, A.; Miranda, M.P.; Fereres, A. Psyllids as major vectors of plant pathogens. Entomol. Gen.; 2021; 41, pp. 419-438. [DOI: https://dx.doi.org/10.1127/entomologia/2021/1289]

20. Tjallingii, W.F. Salivary secretions by aphids interacting with proteins of phloem wound responses. J. Exp. Bot.; 2006; 57, pp. 739-745. [DOI: https://dx.doi.org/10.1093/jxb/erj088]

21. Bonani, J.P.; Fereres, A.; Garzo, E.; Miranda, M.P.; Appezzato-Da-Gloria, B.; Lopes, J.R.S. Characterization of electrical penetration graphs of the Asian citrus psyllid, Diaphorina citri, in sweet orange seedlings. Entomol. Exp. Appl.; 2010; 134, pp. 35-49. [DOI: https://dx.doi.org/10.1111/j.1570-7458.2009.00937.x]

22. Radford, A. Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv; 2015; arXiv: 1511.06434

23. Li, Z.; Liu, F.; Yang, W.; Peng, S.; Zhou, J. A survey of convolutional neural networks: Analysis, applications, and prospects. IEEE Trans. Neural Netw. Learn. Syst.; 2021; 33, pp. 6999-7019. [DOI: https://dx.doi.org/10.1109/TNNLS.2021.3084827]

24. Gonzalez, R.C. Deep convolutional neural networks [lecture notes]. IEEE Signal Process. Mag.; 2018; 35, pp. 79-87. [DOI: https://dx.doi.org/10.1109/MSP.2018.2842646]

25. Khan, A.; Sohail, A.; Zahoora, U.; Qureshi, A.S. A survey of the recent architectures of deep convolutional neural networks. Artif. Intell. Rev.; 2020; 53, pp. 5455-5516. [DOI: https://dx.doi.org/10.1007/s10462-020-09825-6]

26. Stergiou, A.; Poppe, R.; Kalliatakis, G. Refining activation downsampling with SoftPool. Proceedings of the IEEE/CVF International Conference on Computer Vision; Montreal, BC, Canada, 11–17 October 2021.

27. Schrab, A.; Kim, I.; Albert, M.; Laurent, B.; Guedj, B.; Gretton, A. MMD aggregated two-sample test. J. Mach. Learn. Res.; 2023; 24, pp. 1-81.

28. Wang, W.; Li, H.; Ding, Z.; Nie, F.; Chen, J.; Dong, X.; Wang, Z. Rethinking maximum mean discrepancy for visual domain adaptation. IEEE Trans. Neural Netw. Learn. Syst.; 2021; 34, pp. 264-277. [DOI: https://dx.doi.org/10.1109/TNNLS.2021.3093468]

29. Li, C.-L.; Chang, W.C.; Cheng, Y.; Yang, Y.; Póczos, B. Mmd gan: Towards deeper understanding of moment matching network. Advances in Neural Information Processing Systems; MIT Press: Cambridge, MA, USA, 2017; pp. 2203-2213.

30. Bińkowski, M.; Sutherland, D.J.; Arbel, M.; Gretton, A. Demystifying mmd gans. arXiv; 2018; arXiv: 1801.01401

31. Jayasumana, S.; Ramalingam, S.; Veit, A.; Glasner, D.; Chakrabarti, A.; Kumar, S. Rethinking fid: Towards a better evaluation metric for image generation. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition; Seattle, WA, USA, 17–21 June 2024.

32. De Hond, A.A.; Steyerberg, E.W.; Van Calster, B. Interpreting area under the receiver operating characteristic curve. Lancet Digit. Health; 2022; 4, pp. e853-e855. [DOI: https://dx.doi.org/10.1016/S2589-7500(22)00188-1]

33. Tharwat, A. Classification assessment methods. Appl. Comput. Inform.; 2021; 17, pp. 168-192. [DOI: https://dx.doi.org/10.1016/j.aci.2018.08.003]