Enhancing Corn Pest and Disease Recognition through Deep Learning: A Comprehensive Analysis

Abstract

Pests and diseases significantly impact the quality and yield of maize. As a result, it is crucial to conduct disease diagnosis and identification for timely intervention and treatment of maize pests and diseases, ultimately enhancing the quality and economic efficiency of maize production. In this study, we present an enhanced maize pest identification model based on ResNet50. The objective was to achieve efficient and accurate identification of maize pests and diseases. By utilizing convolution and pooling operations for extracting shallow-edge features and compressing data, we introduced additional effective channels (environment–cognition–action) into the residual network module. This step addressed the issue of network degradation, establishes connections between channels, and facilitated the extraction of crucial deep features. Finally, experimental validation was performed to achieve 96.02% recognition accuracy using the ResNet50 model. This study successfully achieved the recognition of various maize pests and diseases, including maize leaf blight, Helminthosporium maydis, gray leaf spot, rust disease, stem borer, and corn armyworm. These results offer valuable insights for the intelligent control and management of maize pests and diseases.

1. Introduction

Maize is one of the world’s leading crops and is widely grown because of its extreme adaptability [1]. While a significant portion of corn seeds undergo treatment through soaking and coating to mitigate pests and diseases, economic losses resulting from these factors can still amount to tens of millions of dollars annually. Recent research has identified approximately 40 pests and diseases that impact corn, with frequent occurrences of maize leaf blight, Helminthosporium maydis, gray leaf spot, rust disease, stem borer, corn armyworm, and other factors adversely affecting yield [2]. Because of the complex causes, it is difficult to obtain precise information on pests and diseases. Consequently, pest and disease management has evolved into a significant challenge. The timely detection and accurate identification of these issues are crucial for enhancing agricultural production [3].

As a major crop in agriculture, maize often experiences pest and disease occurrences, typically during the ear stage [4]. Conducting manual diagnosis of pests and diseases during this period in the field proves challenging, given the subjective nature of manual discrimination. This hinders effective pest and disease monitoring and identification, subsequently amplifying the complexities of pest and disease prevention and control [5,6]. With the integration of artificial intelligence technology and agriculture, various technologies such as image classification, target detection based on machine learning, and pattern recognition have found extensive application in the field of plant disease detection. These technologies have proven to be highly effective in enhancing identification efficiency and lowering costs associated with pest prevention and control [7,8]. Computer vision technology holds promising application prospects in the field of crop pest detection and identification and is a hotspot for research at present [9]. Nonetheless, the challenge persists due to the presence of multiple and frequently overlapping pest and disease manifestations, diverse and chaotic pest patterns, the need for scale invariance, and intricate surroundings and backgrounds, as well as the occurrence of various sources of interference. Achieving robust, real-time classification and recognition of pests and diseases within complex scenarios remains a challenging task. Research on pest recognition techniques utilizing computer vision has experienced significant advancements in recent years [10].

Nandhini and Ashokkumar [11] introduced an approach for predicting maize leaf diseases through the utilization of deep convolutional neural networks (CNNs). Linear vector quantization (LVQ)-augmented CNN improved the accuracy of maize leaf disease identification while reducing the network parameters. Esgario, et al. [12] put forth a system based on deep learning for the identification of coffee leaf pests. The outcome of their results revealed a remarkable accuracy rate exceeding 97% for the classification of biological stress. Loti, et al. [13] employed a combination of six traditional feature-based methods and six deep learning feature-based methods to extract crucial features related to pests and diseases from images of pepper leaves. Rahman, et al. [14] proposed a method based on deep learning for detecting pests and diseases in rice. Their approach involves a two-stage small CNN architecture, which achieved a noteworthy accuracy of 93.3%, while also boasting a notably reduced model size. Similarly, in another study, Waheed, et al. [15] presented an approach employing deep artificial neural networks and deep learning techniques for the early detection of diseases, pest patterns, and nutrient deficiencies. Their method achieved an impressive accuracy of 99%, demonstrating its efficacy in accurate detection. Tripathi [16] used a deep learning convolutional neural network model, Alexnet, to identify plant diseases. Butte, et al. [17] introduced a method for analyzing aerial images of potato crops through the utilization of deep neural networks. Their primary goal was to showcase the automated spatial differentiation between healthy and stressed crops at the individual plant level. Yadav, et al. [18] developed a convolutional neural network (CNN) model using deep learning and imaging methods for the detection of bacterial diseases from peach leaf images, and the proposed method achieved 98.75% recognition accuracy.

In summary, the identification of pests and diseases primarily revolves around tasks such as image segmentation and feature extraction. Technical approaches such as computer vision, deep learning, and migration learning are extensively employed, often in combination with experimental analysis, to enhance and refine models [19]. The above review shows that deep learning methods have certain advantages in the field of pest and disease image recognition: first, they can obtain better accuracy than traditional methods; second, they are more generalizable and robust for different scenes [20]. Nonetheless, there remains a requirement to enhance both the recognition speed and accuracy specifically in the context of identifying maize pests within intricate agricultural environments. In the existing research on pest recognition methods, most researchers have pointed out that the extraction of key features and the construction of deep models are the keys to improving recognition accuracy and robustness [21,22]. Therefore, to achieve precise identification of maize pests and diseases within intricate farming environments, it is imperative to enhance and optimize deep learning models. This advancement holds the potential to elevate the accuracy of maize pest and disease identification. Moreover, these improvements could serve as valuable references for conducting research on the identification and detection of pests and diseases affecting other crops as well.

2. Materials and Methods

2.1. Dataset Construction

The study was conducted on six common pests and diseases: leaf blight, Helminthosporium maydis, gray leaf spot, rust disease, stem borer, and corn armyworm [23]. Obtaining data and pictures in agriculture is not easy due to the various factors involved in the development of different diseases, such as ecology, climate, and maize varieties. Therefore, in this experiment, some data were obtained from publicly available datasets, which were recently collected from the experimental fields of the Northeast Agricultural University of China. Additionally, the categorization of the pictures was confirmed by relevant experts. Before collecting samples, the system was inputted with 500 Northeast Agricultural University of China health images to train the software. After that, 2150 original images were collected, including 391 images of maize leaf blight, 327 images of Helminthosporium maydis, 364 images of maize rust, 370 images of gray leaf spot, 356 images of stem borer, and 342 images of corn armyworm. These six categories of images were placed in six different folders, with each folder representing a category label. Label 0 was for maize leaf blight, Label 1 was for stem borer, Label 2 was for corn armyworm, Label 3 was for gray leaf spot, Label 4 was for Helminthosporium maydis, and Label 5 was for rust disease. The images of corn leaf pests and disease samples are shown in Figure 1.

The image preprocessing method in Python (Python Software Foundation, Wilmington, NC, USA) was used to compress all image data to 224 × 224. The corn leaf pest dataset was divided into a training set, a validation set, and a test set in the ratio of 3:1:1. The results of the validation set can be used to adjust the hyperparameters of the model, determine the structure of the network, and make a preliminary evaluation of the model’s effect; the test set is used to test the generalization ability of the final model; however, the results cannot be used for algorithm-related processes such as tuning.

2.2. Data Preprocessing

Data augmentation can be divided into offline data augmentation and online data augmentation according to the data storage method [24]. Offline augmentation processes the original dataset directly, and the number of new data will become n times the original data (n is the augmentation factor), which is suitable for small datasets [25]. Many machine learning frameworks support online data augmentation and can use GPU to optimize the computation [26].

Typically, the greater the amount of data used for training, the higher the model’s accuracy, as well as its robustness and generalizability. Data augmentation is a common method for data expansion in deep learning. In this study, we used offline enhancement (Figure 2) to perform brightness adjustment, random horizontal flip, value adjustment, color adjustment, and random erasure on the images to obtain a new training set of images that was six times larger than the original training set.

2.3. Comparison of Maize Disease Classification Models

Deep learning models, such as AlexNet, VGG, DenseNet, and ResNet, are widely used in the field of image recognition by building deep network structures for end-to-end automatic learning. To identify the model with optimal performance for enhancement, this paper adopted the same training strategy across all experiments. To fit the model size, the images were scaled to 224 × 224 pixels uniformly, the total number of model iterations was 10,000, the initial learning rate was set to 3 × 10⁻², the momentum was 0.9, and the weight decay was 0.01. The AlexNet, VGG, DenseNet, and ResNet50 models were trained on the self-built corn pest dataset, and the model accuracies were 81.26%, 87.18%, 90.67%, and 93.83%, respectively. The accuracy and loss curves of the models are shown in Figure 3. DenseNet could reach 90.67% accuracy, but it gradually stabilized after 8000 iterations; AlexNet and VGG fluctuated more and had lower accuracy. Therefore, this paper is based on the ResNet50 model.

3. Pest and Disease Identification Model Improvement with ECA

3.1. ResNet50 Model Structure

ResNet50 has two basic blocks: the feature block and the convolution block. The identity feature block has the same dimension for the input and output feature matrices, allowing serial operations to deepen the network depth, while the convolution block has a different dimension for the input and output feature matrices; thus, the latter is used to change the dimension of the feature vector to avoid the situation where what is learned becomes too complex as the network depth increases [27].

The ResNet50 network contains a total of 50 layers and is divided into six main parts, i.e., the input module, four group blocks, and the output module [28]. The ResNet50 network input module is a 7 × 7 convolution with a step size of 2 and a 3 × 3 maximum pooling operation for the 224 × 224 feature maps entering the network. The 7 × 7 convolution kernel was chosen to perform the convolution operation so that the model input layer could preserve as much global information as possible related to the semantics of adjacent features and establish complex spatial location relationships, thus improving the performance of the model [29]. After a processing series, a feature map of size 64 × 64 was obtained as the input data for the main part of the model.

The main part of the model consists of four large blocks containing three, four, six, and three residual blocks, respectively, for a total of sixteen residual structures. In this model, two residual blocks and one down-sampling block form a residual structure [30]. The down-sampling block with an initial convolution step of 2 performs down-sampling operations on the feature map to reduce the length and width of the feature map. To avoid network degradation due to the change of the dimension of the input residual block feature matrix, the relevant parameters are set in the residual block to carry out the residual block operation without changing the dimension of the input and output feature matrices to ensure the addition operation between the feature matrices by using the property that the small size convolution kernel can control the size of the input and output feature moments [31]. In this way, deep feature extraction can be better utilized to make a strong correlation between several feature maps in each layer, and then the fast response differentiation and activation model can be achieved by linear superposition of different feature maps in the same layer [32]. Finally, the 32 × 32 feature maps obtained from the intermediate processing were scaled down to 16 × 16, and the deeper features with more differentiation were extracted through the fully connected layer, which can be classified by the classification logic of the Softmax classification function [33].

Residual Network

Convolutional neural networks originated from cortical vision research and have excellent performance in image recognition, target detection, and other fields [34]. However, as the layers of neural networks deepen, the models face the problems of gradient dispersion or explosion with network degradation. The network structure is relatively shallow, and the loss of the training set gradually decreases to a saturation value as the number of layers continues to increase. The loss does not keep leveling off, but gradually increases [35]. The use of residual blocks allows the depth of the network to be further deepened under the premise of ensuring normal operation [36]. The residual operation is formulated as shown by Xu, et al. [36].

$$\mathit{H}\left(\mathit{x}\right)=\mathit{F}\left(\mathit{x}\right)+\mathit{x}.$$

(1)

where x is the constant mapping which represents the input of the residual block or module. It is also referred to as skip connection or identity mapping. F(x) is the residual mapping which is a part of the block capable of capturing the residual facts. It consists of one or more layers of convolutional operations whose purpose is to seize the difference or residual between the preferred output and the input x. Essentially, F(x) attempts to analyze the transformation needed to deliver the approach towards the goal output; H(x) is the output of the residual block, which mixes the outcomes of the identification mapping x and the residual mapping F(x). The combined equation allows the network to analyze incremental modifications to the input. The residual module is shown in Figure 4.

3.2. Improvement Strategies

3.2.1. ECA Attention Mechanism Module

The ECA module is fused into the ResNet50 network to improve the quality of image feature extraction. The ECA module is a lightweight channel attention module that allows the neural network to pay more attention to salient features during feature extraction. At the same time, compared with other channel attention modules, the ECA module not only ensures the efficient acquisition of all channel dependencies without dimensionality reduction but also enhances the channel features of the input feature map. Proper cross-channel interaction not only reduces the complexity of the model structure but also ensures the efficiency and reliability of the model. It works by first compressing the spatial information through global average pooling, then learning the channel attention information using 1 × 1 convolution, combining the acquired channel attention information with the original input feature map, and, finally, outputting a feature map of specific channel attention. The principle is shown in Figure 5.

The features are first aggregated by global average pooling to obtain channel global information, and the global average pooling operation is formulated in accordance with Shah, et al. [37] and Sladojevic, et al. [38].

$$\mathit{y}=\frac{\mathit{1}}{\mathit{H}\mathit{W}} {{\displaystyle \sum }}_{\mathit{a}}^{\mathit{H}}{{\displaystyle \sum }}_{\mathit{b}}^{\mathit{W}}{\mathit{x}}_{\mathit{i}} \left(\mathit{a},\mathit{b}\right),$$

(2)

where ${\mathit{x}}_{\mathit{i}}$ represents the ith feature map with the input size HW, and y represents the global feature.

Next, the number of cross channels k is calculated adaptively using the channel dimension C. The formula of the adaptive function is in accordance with Shah, et al. [37] and Sladojevic, et al. [38].

$$\mathit{k}=\mathit{\psi }\left(\mathit{C}\right)={\left|\frac{{\mathit{log}}_{\mathit{2}}\left(\mathit{C}\right)}{\mathit{\gamma }}+\frac{\mathit{b}}{\mathit{\gamma }}\right|}_{\mathit{o}\mathit{d}\mathit{d}},$$

(3)

where ${\left|t\right|}_{\mathit{o}\mathit{d}\mathit{d}}$ denotes the t-nearest odd number; C is the channel dimension; and b and $\mathit{\gamma }$ are constants, where b = 1, $\mathit{\gamma }=2$.

Then, the channel weights are calculated using a 1D convolution with a convolution kernel of size k to obtain the interdependencies between channels; 1D convolution is formulated in accordance with Shah, et al. [37] and Sladojevic, et al. [38].

$$\mathit{\omega }=\mathit{\sigma }\left({\mathit{C}}_{\mathit{1}}{\mathit{D}}_{\mathit{k}}\left(\mathit{y}\right)\right),$$

(4)

where $\mathit{\omega }$ is the channel weight;$\mathit{\sigma }$ is the sigmoid function;${\mathit{C}}_{\mathit{1}}\mathit{D}$ is the 1D convolution; y is the result after global average pooling; and k is the convolution kernel size. Finally, the features with channel attention are obtained by dot product operation of the original input features and channel weights.

3.2.2. Model Proposed

Combining the advantages of convolutional neural networks and attention mechanisms in image classification and feature extraction, this paper proposes an ECA_ResNet-based method for corn pest identification. The residual block allows the depth of the network to be further deepened while ensuring normal operation. In actual farming operations, there is interference from the surrounding environment, which increases data complexity and reduces the key feature extraction ability of the model. ECA-Net, as an effective channel attention network with good embedding, can realize the control of global information, enhance the sensitivity of the model to important features, and improve the model accuracy (Figure 6).

In this paper, ECA-Net is embedded into the residual network module, and ⨁ denotes the element product. The attention module is inserted into two blocks with a residual bottleneck structure which constitute block A and block B of the ECA-ResNet basic unit, respectively. When the number of input and output channels of a block is different, a convolutional layer is added to the proximity connection to change the feature map dimension, and the structure is like that of block A. When the number of input and output channels of a block is the same, the proximity connections can be added directly, and the structure is like that of block B. The input features are downscaled to reduce the number of feature channels, which reduces the computational effort and improves the model training speed. By improving the residual network module, we strengthen the connection between the feature channels within the residuals and enhance the model’s ability to extract key features at a deeper level.

3.2.3. Adam Optimizer

The role of the optimizer is to use the gradient to update the parameters so that the loss is continuously reduced. The Adam (adaptive moment estimation) optimizer adjusts the learning rate based on the historical gradient information and normalizes the parameter updates so that each parameter update has a similar magnitude, thus improving the training effect.

The formula for the Adam algorithm is as follows [39,40]:

1. Calculate the gradient information:

$$g_t={\nabla }_{\theta }J\left(\theta ,X_t,y_t\right),$$

(5)

where $J\left(\theta ,X_t,y_t\right)$ denotes the objective function and $X_t$ and $y_t$ denote the features and labels of a small batch of samples, respectively.

2. Compute the first-moment estimate vector $S_t$ and the second-moment estimate vector $R_t$:

$$S_t={\beta }_1{S}_{t-1}+\left(1-{\beta }_1\right)g_t,$$

(6)

$$R_t={\beta }_2{R}_{t-1}+\left(1-{\beta }_2\right)g_t^2,$$

(7)

where $S_t$ denotes the first-moment estimation vector and $R_t$ denotes the second-moment estimation vector; ${\beta }_1$ and ${\beta }_2$ are the decay rates.

3. Calculate the bias-corrected gradient information $g_t^{\prime }$:

$${\widehat{S}}_t=\frac{S_t}{1-{\beta }_1^t},$$

(8)

$${\widehat{R}}_t=\frac{R_t}{1-{\beta }_2^t},$$

(9)

$$g_t^{\prime }=\frac{\alpha {\widehat{S}}_t}{\sqrt{{\widehat{R}}_t}+\epsilon },$$

(10)

where ${\widehat{S}}_t$ and ${\widehat{R}}_t$ denote the bias-corrected first- and second-moment estimation vectors, respectively, α denotes the learning rate, and $\epsilon$ is the smoothing term.

4. Update the model parameters:

$${\theta }_{t+1}={\theta }_t-g_t^{\prime }$$

(11)

The Adam optimizer performs well in many practical problems, especially when training deep neural networks on large datasets.

3.3. Model Building

3.3.1. ECA Attention Mechanism Module

The ECA_ResNet-based maize pest identification model mainly consists of a convolutional layer, a pooling layer, the improved residual block A and block B, and a fully connected layer. The network structure is shown in Figure 7.

The images are input to the model in this paper, Algorithms 1: firstly, after two convolution and maximum pooling operations, where the convolution operation obtains the shallow-edge features of the time-frequency map, and the maximum pooling retains the main features of the data and removes the redundant information to improve the model prediction speed; and secondly, after the improved residual modules’ block A and block B deepen the number of network layers. Embedding ECA-Net inside the residual module not only prevents the phenomena of gradient dispersion, gradient explosion, and network degradation caused by deepening the number of network layers, but can also effectively realize cross-channel interaction, enhance the feature channels useful for the current task, suppress the feature channels of little use for the current task, and adaptively acquire deep key features. Finally, dropout is added between two full connections to improve the model generalization ability and suppress overfitting. A Softmax layer is used as a classifier to achieve corn pest and disease identification.

3.3.2. Algorithm: ECA_ResNet-Based Maize Pest Identification Model

Algorithms 1 ECA_ResNet-Based Maize Pest Identification Model

Preprocessing:

Resize input pics to 224 × 224 pixels.

Normalize pixel values to [0, 1] range.

Model Architecture:

Convolutional layer for preliminary capabilities.

Max-pooling for substantial features and dimension discount.

Improved residual blocks A and B with embedded ECA interest.

Fully connected layer for deep capabilities.

Dropout for regularization.

Softmax layer for multi-class category.

Training:

Partition dataset into training, validation, and test sets.

Train in the use of cross-entropy loss.

Use Adam optimizer with appropriate hyperparameters.

Apply back-propagation to replace model weights.

Testing and Evaluation:

Evaluate model on check set.

Compute accuracy, precision, bear in mind, F1-rating.

Construct confusion matrix for performance analysis.

CInterpretation:

Visualize misclassified pictures.

Analyze elegance activation maps.

Deployment:

Deploy model for real-time pest identification.

4. Experiments and Analysis

4.1. Ablation Experiment

Ablation experiments were conducted on ResNet50 models with different improvement methods under the same dataset. Among them, ECA-ResNet50 denotes the improved model with the addition of the ECA attention mechanism and Adam-ResNet50 denotes the improved model with the addition of the Adam optimizer; two improved methods of ECA and Adam are superimposed in the improved ResNet50 model. The results of the ablation experiments are shown in

Table 1. Ablation experimental results of improved ResNet50.

Model	Number of Iterations	Validation Set Accuracy/%	Test Set Accuracy/%
ResNet50	10,000	93.05	89.14
ECA-ResNet50	10,000	94.21	91.93
Adam-ResNet50	10,000	93.73	89.86
ECA-Adam-ResNet50	10,000	96.02	93.95

.

From the results of the ablation experiments in Table 1, it can be seen that compared with the original ResNet50 model, the ECA-ResNet50 model improved the accuracy of the validation set and the test set by 1.16% and 2.79%, respectively, which shows that the ECA attention mechanism improved the accuracy of the test machine more significantly. The accuracy of the validation set and the accuracy of the test set of Adam-ResNet50 improved by 0.68% and 0.72%, respectively. Overlaying the two improvement methods improved the ResNet50’s validation set accuracy by 2.97%, and the test set accuracy by 4.81%. It can be seen that both improvement methods are positive improvements and have a significant effect on the model performance.

4.2. Comparison of Detection Results of Different Models

To further verify the effectiveness of the improved ResNet50 model, the improved ResNet50 model was tested against AlexNet, VGG, and DenseNet, respectively, which are most commonly used in the current image classification field, as shown in Figure 8.

As can be seen from the content of Figure 8, the improved ResNet50 model improved by 12.96%, 7.71%, and 4.33% in the validation set accuracy compared to the AlexNet, VGG, and DenseNet models, respectively. The experimental results show that the improved ResNet50 model outperforms other commonly used classification models in maize pest identification.

4.3. Confusion Matrix for Maize Pest and Disease Identification Model

To objectively evaluate the accuracy of the model in various types of pests and diseases and improve the readability of the confusion matrix of the model, 150 images of each of the six types of pests and diseases were selected from the test set as test data in this experiment, and the resulting confusion matrix is shown in Figure 9.

In the Figure 9 confusion matrix, the vertical axis is the real label of corn pest images, the horizontal axis is the model prediction label, the right side is the heat map of prediction accuracy distribution, the main diagonal line is the number of samples correctly predicted by the model, and the darker color represents the higher accuracy rate. By comparing the number of correctly predicted samples on the main diagonal in the confusion matrix, it can be seen that the improved model improves the recognition ability of various diseases. As can be seen from Figure 9, gray leaf spot and rust have the lowest accuracy rates, with test accuracies of 95.3% and 94.0%, respectively. The probability of mutual misjudgment of corn armyworm and gray leaf spot is higher; on analyzing the reason we concluded that corn armyworm and gray leaf spot are more similar in macroscopic features, and the image acquisition equipment of this model has a long observation distance, which cannot capture the subtle features. The external features of maize leaf blight, stem borer and Helminthosporium maydis are more obvious, and the model has the highest recognition rate for these pests and diseases.

4.4. Performance of the Improved Model on Maize Pests and Diseases

Figure 10 shows some examples of disease prediction results and class activation feature map visualization results before and after the model improvement. The shallow convolution of the improved model has an improved ability to extract detailed texture information of leaves, which can better achieve the recognition of fine disease spots. The improved model can more accurately determine the occurrence area of disease spots on leaves; for example, in the results of Figure 6, more attention is paid to the leaf edge features for corn gray spot disease, and more attention is paid to the inner leaf spot area for small spot disease.

5. Discussion

5.1. Challenges Faced

Convolutional neural network technology is an innovative deep learning technology nowadays. The present study achieved outstanding results in the research of corn pest identification technology based on convolutional neural networks. However, the application of this technology in the field of crop pest and disease identification is still in the initial stage. Moreover, because different agricultural pest and disease problems have different degrees of complexity, pest and disease recognition still have considerable drawbacks and limitations in practical scenarios.

At present, for agricultural pest data, some crops do not have a large and rich enough database, and researchers need to spend a great deal of labor and time collecting data by themselves. Researchers often do not study the same diseases and crops and cannot exchange data with each other, which greatly reduces the efficiency of research. Of course, researchers can use data augmentation to expand their own datasets, but inevitably information loss will occur in the process. For example, for certain diseases, leaves are extensively ulcerated and branch parts are covered with a large number of insect eggs. Image cropping can lead to inaccuracies in leaf ulcer area and an imbalance in insect egg density information, resulting in loss of critical information.

The ablation experiments were designed to investigate the effect of different strategies on the overall performance of the ResNet50 model in corn pest and disease identification. Three variations of the model have been considered: ECA-ResNet50 with the ECA interest mechanism, Adam-ResNet50 with the Adam optimizer, and ECA-Adam-ResNet50 combining both upgrades. The results reveal that the performance of pest recognition was significantly improved by each model. However, combining the ECA mechanism and the Adam optimizer model considerably amplified their performance, leading to a significant validation set precision of 2.97% and a highest test set accuracy of 4.81%. These findings highlight the efficacy of each ECA mechanism and the Adam optimizer in improving the model’s predictive efficiency. The combined impact of these models indicate a synergistic courting between the two strategies in optimizing the ResNet50 model for corn pest and disease detection. A recent study compared the Inception V3, VGG 16, VGG 19, CNN, and ResNet 50 models for early detection of a rice disease. The study found that the ResNet 50 model outperformed other models in term of precision. However, the study did not explore the efficacy of different optimization strategies in enhancing the model’s overall performance [37].

To validate the superiority of the progressed ResNet50 model, a comparative assessment was carried out against a well-established deep learning model in image classification, i.e., AlexNet, VGG, and DenseNet. In terms of validation set accuracy, the progressed ResNet50 model overtook the AlexNet, VGG, and DenseNet models. This considerable margin of development underlines the efficacy of the improvements made to the ResNet50 model as the best choice for maize pest and disease recognition [38] compared to the performance of different deep learning models for plant disease classification, where ResNet50 achieved the highest accuracy due to its deeper architecture.

The construction and evaluation of the confusion matrix provide an understanding of the model’s capability to accurately recognize various types of pests and diseases affecting maize crop. Through visual inspection, it is evident that the improved model version, represented by ECA-Adam-ResNet50, attains greater recognition throughout different classes. The important diagonal of the confusion matrix shows that a wide variety of samples could be efficiently expected through the model. The darker pigmentation on this diagonal indicates better accuracy predictions. A comparison of the diagonal factors between the authentic ResNet50 and the improved ECA-Adam-ResNet50 model exhibits a systematic enhancement in prediction accuracy, declaring that the model has an improved ability to distinguish among numerous pests and diseases. For instance, a study has proposed a CNN architecture called INC-VGGN for the detection of rice and maize diseases, which exploits a pre-educated VGG-19 and replaces its remaining layer with a new one [39]. Another study has proposed a modified model of the MobileNet model for the identification of plant diseases and compares its overall performance with different models together with AlexNet and VGG [40]

Figure 10 show perceptive visualizations of disease prediction and feature map activations before and after improvement of the model. The improved model demonstrates a unique capacity to capture complex texture details present on the corn leaves, thus assisting the identification of disease spots. Apparently, the model refined convolutional layers and enable efficient precise delineation of disease spots, exemplified by way of its emphasis on leaf edges for certain diseases and internal leaf spots for other diseases, respectively.

Agricultural pests and diseases are not unique. Some pests and diseases cannot be identified because the source of the disease or insect cannot be observed at the beginning of the damage. The vast majority of pests and diseases that can be identified are identified too late.

Light is unstable in natural environments, and even when artificial lighting is used, leaves or branches can block the light, resulting in uneven illumination and bias to the image. The present model has the ability to accurately spotlight particular pests and disease spots. The illustrated specimen data underscore the model’s accurate prediction in pest and disease identification.

The results outlined a clear and substantial improvement in corn pest and disease recognition through the combination of deep learning strategies, particularly the ECA and the Adam model. The ablation experiments, version comparisons, confusion matrix evaluation, and performance visualizations together validate the effectiveness of the improved ECA-Adam-ResNet50 version as an innovative solution for precise pest and disease recognition in maize crop.

5.2. Outlook

With the development of computer hardware equipment, convolutional neural networks show great potential for image recognition, and pest identification technology based on the convolutional network has greatly improved the efficiency of agricultural production. Therefore, it is important to further promote pest identification technology and solve the difficulties and limitations in pest identification applications to guarantee agricultural production and promote agricultural economic development.

To overcome these difficulties, it is first necessary to build an abundant and rich database of pests and diseases. To build a rich database of pests and diseases, we can consider developing a ground robot to detect pests on crops at close range and in real-time. The robot could be used to detect pests not only on the canopy of crops but also on the backs of crop leaves. This would also allow worldwide data sharing, permitting research teams around the world to collect images with very different characteristics in different regions. By properly integrating the corresponding datasets to obtain images with more distinctive features and simpler noise, the resulting image sets will be more representative, and the research results will be more meaningful and applicable to the conditions of agricultural production.

6. Conclusions

In this study, we propose an improved ResNet50 model based on the identification of maize pests and diseases, and we take maize leaf blight, Helminthosporium maydis, gray leaf spot, rust disease, stem borer, and corn armyworm as the research objects. Based on the classical ResNet50 model, the ECA attention mechanism and Adam optimizer are introduced to identify six types of pests and diseases of maize. The main findings are as follows:

The improved ResNet50 model achieves 96.02% accuracy of the optimal validation set, with an initial learning rate of 3 × 10⁻² and an iteration number of 6000 when the ECA attention mechanism is introduced and the Adam optimizer is selected. Compared with AlexNet, VGG, DenseNet, and the classical ResNet50 model, the optimal validation set accuracy is improved by 12.96%, 7.71%, 4.33%, and 2.97%, respectively.

Using the confusion matrix to visualize the classification results of maize pests and diseases, the overall test accuracy of the improved ResNet50 model reached 97.77%, and its recognition rates of maize leaf blight, Helminthosporium maydis, gray leaf spot, rust disease, stem borer, and corn armyworm were above 90%, further validating the effectiveness of the improved ResNet50 model for maize pest and disease recognition.