The Impact of Convolutional Layer Selection in ResNet-50v2 Architecture on Corn Leaf Disease Classification

Agriculture is an important economic sector in most nations, including Indonesia. This sector not only serves as the rural economy's backbone, but it also plays an important role in government [1]. Agriculture employs a significant portion of the rural population, lowers unemployment rates, and alleviates poverty. Furthermore, this industry contributes significantly to food security and economic stability. Corn is one of Indonesia's most important food crops, contributing significantly to the country's economy [2, 3]. Corn is used for food, animal feed, and industrial raw materials. Processed corn goods, such as corn flour and corn oil, offer substantial added value. However, corn plant productivity is frequently hampered by a variety of leaf diseases, which can drastically impair crop yield and quality. These diseases, including leaf blight, leaf spot, and leaf rust, can spread swiftly and damage adjacent plants if not recognized and treated promptly.

Early diagnosis of leaf diseases in corn plants is critical for minimizing the disease's detrimental impact. Agronomists have traditionally detected diseases manually by monitoring visual signs on plant leaves. Although this strategy can be beneficial, it is time-consuming, difficult, and costly, with a high reliance on human abilities and expertise. Thus, a more efficient and automated technique is required to identify leaf diseases in corn plants precisely and rapidly.

Recent breakthroughs in artificial intelligence (AI) and image processing have expanded the potential for automated plant disease diagnosis [4-7]. A particularly interesting strategy is the use of convolutional neural networks (CNNs), a form of artificial neural network specifically developed for pattern identification in visual data that has proved to be very effective in a range of picture classification tasks [8, 9]. ResNet-50v2 (Residual Network 50 version 2) is one of the most powerful and popular CNN designs. ResNet-50v2, developed by Microsoft Research, employs the residual learning idea, allowing for the training of extremely deep networks without encountering the vanishing gradient problem [10]. This can be achieved by the use of shortcut connections that connect certain levels in the network, resulting in a smoother and more efficient flow of information. ResNet-50v2 can overcome the network performance decrease. This makes ResNet-50v2 ideal for complicated picture classification jobs. However, the performance of ResNet-50v2 is determined not just by the depth of its architecture, but also by the selection and tuning of its numerous convolutional components. The research [5, 7] shows significant advances in the use of sophisticated deep learning architectures, such as CNNs, ResNet, and others, for the identification of plant diseases, which can increase the precision of diagnostics. According to the research, deep neural networks may have problems like overfitting, especially if the dataset is too small or undiversified. To overcome the drawbacks of feature extraction, research might concentrate on applying the design of the ResNet50-v2 to efficiently extract important characteristics from plant images. As shown in earlier research, overfitting may be lessened by using strategies like dropout or data augmentation with ResNet50-v2.

This research has significant ramifications in the realms of agriculture and AI. By assessing how the selection and setup of various layers in the ResNet-50v2 architecture affect classification performance. Experiments were carried out by altering the layers and evaluating the models. Understanding the effect of each convolutional component should allow us to optimize the ResNet-50v2 architecture for the goal of corn leaf disease classification. This will not only help farmers diagnose and treat leaf diseases more rapidly, but it will also boost the overall productivity and sustainability of the agricultural industry. The findings of this study can be used to design a more complex plant disease detection system that can be applied to a wide range of plants.

The previous sub-chapter addressed the dataset, which had 4000 records divided over four classes. Table 1 shows that there are 1000 data points in each class. The details of the data collection process include 4000 data images from Kaggle sources with an image size of 4000×3000 pixels taken with a 48 MP cellphone camera. The dataset is scaled to 224×224 pixels. The dataset will be separated into three parts: learning, validation, and testing. The percentage of learning data is 80%, but it is divided by the need for validation data at an 80:20 ratio, with a total of 2560 training data (640 gray leaf spot, 640 common rust, 640 northern leaf blight, and 640 healthy) and 640 validation data (160 gray leaf spot, 160 common rust, 160 northern leaf blight, and 160 healthy). The acquisition of test data is 800 data points (200 gray leaf spot, 200 common rust, 200 northern leaf blight, and 200 healthy) at a 20% rate. In this study, an experiment with layer selection was carried out, and the results are displayed in Table 2.

Table 2. Test scenario

We often set the hyperparameters to an Adam optimizer, a learning rate of 0.0001, a batch size of 32, and epochs of 20. To optimize the model's performance, these hyperparameters must be appropriately set because they directly affect the model's stability, generalization ability, and rate of convergence. Every convolutional layer in the ResNet50-v2 architecture is known to play a specific purpose. The model's ability to learn from the data is influenced by the activation function selection. ReLU, for instance, can aid in resolving the vanishing gradient issue, resulting in improved performance and quicker convergence in deeper networks. Therefore, employing an unsuitable activation function may impede training and result in a slower rate of convergence. However, batch normalization enables greater learning rates and can enhance the model's overall performance. It also contributes to the prevention of overfitting. Additionally, the model's capacity to generalize is impacted by the pooling technique. Although max pooling keeps the most important features, it may miss important information that average pooling may have picked up. These specific convolutional layers and their configurations are crucial for ResNet-50v2 to perform effectively in image classification tasks while maintaining efficient training dynamics. These decisions also help ResNet-50v2 tackle difficult challenges.

The performance of the ResNet-50v2 model on corn leaf classification was measured using accuracy, precision, recall, and F1-score calculations. Accuracy refers to how accurately the measurement findings match the actual value [25]. Figure 5 shows that accuracy measurement indicates how near the measurement findings are to the actual value, or in other words, the fraction of right predictions from all data (Eq. (1)) [26]. Precision refers to the accuracy of repeated measurements. Precision is a proportion of correct positive predictions among all positive predictions (TP and FP) (Eq. (2)) [27]. Recall refers to the model's ability to detect all positive cases that exist. Recall lists a number of positive cases identified by the model (Eq. (3)) [28]. The F1-score is a combination of precision and recall used to evaluate the model's performance in classification. A high F1-score value suggests that the model has strong accuracy and recall (see Eq. (4)) [29].

5.png

Figure 5. Confusion matrix

Accuracy $=\frac{T P+T N}{T P+T N+F P+F N}$     (1)

Precision $=\frac{T P}{T P+F P}$     (2)

Recall $=\frac{T P}{T P+F N}$     (3)

$F 1_{\text {Score }}=\frac{2 T P}{2 T P+F P+F N}$     (4)

Table 2 serves as a guideline for evaluating the developed model. Table 3 shows a summary of the outcomes of the model performance metrics, including the accuracy, precision, recall, and F1-score. Table 4 summarizes the findings of the model test performance measures.

The accuracy of validation data is used to track training and compare models. This is done to evaluate the model's performance during learning. When we analyze the model assessment results in Table 3, we see that the batch normalization layer achieves the highest accuracy value of 94.6% compared to the other layers. This is directly related to the outcomes of precision, recall, and F1-score assessments, in which batch normalization outperforms other layers. Although the results from one layer to the next are not very different, this demonstrates that batch normalization dominates overall performance outcomes by predicting the results of the model class with the actual class.

6.png

Figure 6. Comparison of accuracy results

On the other side, testing is used to determine real performance once learning is completed. The purpose is to assess the model's ultimate performance. Table 4 demonstrates that the batch normalization layer achieves the highest accuracy, recall, and F1-score values of 98.0% and 98.1%, precision. This number is 2.3% lower than the "None" layer, which has a value of 95.7%, 3.5% in the pooling layer, and 5.9% in the activation layer. Figure 6 shows a comparison of validation and test accuracy results.

In artificial intelligence, loss is a quantifiable statistic that indicates how much the model's prediction deviates from the actual value. The model's ability to predict data improves as the loss value decreases. Figure 7 shows that the batch normalization layer has the minimum loss, which is 16.4%, as found during validation. Similarly, in the test, the least loss is 8.7%, indicating that the model correctly learnt the pattern in the data and generated accurate predictions in that layer. Comparing earlier research using the ResNet-50v2 model demonstrates that this approach can complete its classification goal. The information on several examples of previous studies is detailed in Table 5.

7.png

Figure 7. Comparison of loss results

Table 3. Results of model performance

Table 4. Performance results on test data

Table 5. Comparison studies