The images used in the study were collected specifically for this research. The images were obtained from vineyards in Tokat province of Turkey and were obtained from a vineyard where the Narince vine leaf species specific to the region was found. The original images were saved in RGB color space, 2056 × 1024‐pixel resolution, and JPG file format. The images were taken with a mobile phone with a 13‐megapixel resolution. These images are divided into three main categories, and there are 821 in total. Data collection took place during the growing season between May and June 2023. Images were captured in natural light conditions. Average file size is 3.2 MB. The original images in the dataset include a background; however, this background was removed during the preprocessing stage. Examples of the original images with the background and the pre‐processed images are presented in Figure 1. In addition, the number of images belonging to each class is given in detail in Figure 2.

In the dataset, the number of healthy grapevine leaves is 310, the number of grapevine leaves affected by grapevine downy mildew disease is 296, and the number of grapevine leaves affected by grape leaf blister mite is 215.

Several pre‐processing steps were applied to remove the blurring in the original images and to remove unnecessary elements from the background. In this process, firstly, contrast enhancement was performed to solve the blurring problem; then the backgrounds of the images were removed, and finally re‐saved in PNG format with a size of 484 × 515 pixels. The original images and the sample images with the background removed and resized are presented in Figure 1.

In this study, ImageDataGenerator is used for data augmentation operations. Instead of creating new data samples, data augmentation allows the images in the existing dataset to be modified with various transformations to create more diversity. In this context, each transformation (e.g., shear, zoom, flip, etc.) creates different variations of the original images and does not increase the number of images in the dataset. This method allows the model to learn and generalize better by “seeing” different variations of the same image.

The images used in the study were normalized by dividing the pixel values by 255. This normalization process was performed to provide a more stable learning process. Within the scope of data augmentation processes, a 20% shear (shift) was applied to the images, which created a perspective change by extending one side of the image and shortening the other. In addition, a 20% zoom‐in and zoom‐out process was applied to the input images, allowing certain parts of the images to be analyzed in more detail. Finally, a horizontal flip process was performed, which made it easier for the model to recognize images from different angles, increasing generalization performance.

This section describes the proposed method for the classification of two different diseases observed on grapevine leaves and healthy leaf samples.

In this study, we developed an innovative hybrid model integrating Convolutional Block Attention Module (CBAM) and Squeeze‐and‐Excitation Networks for grapevine leaf diseases classification. The proposed deep learning‐based model offers significant advantages over traditional manual diagnostic methods. First, it provides significant benefits in terms of time efficiency and cost effectiveness. Unlike manual disease identification, which requires skilled pathologists to visually inspect each leaf individually, our automated approach can rapidly analyze hundreds of leaf images and reduce diagnostic time from hours to seconds per sample. In addition, it will help in early diagnosis and provide convenience for those who do not know about the diseases. The general structure of the proposed approach is given in Figure 3. As a first step, pre‐processing steps such as contrast correction, background removal, and resizing were applied to the images. These processes aim to increase the quality of feature extraction. Pre‐processing also has the potential to affect the performance of the model. After pre‐processing, the data set was divided into training and test data. This ratio was determined as 80% training and 20% test.

Grape leaf images that were preprocessed and separated into training and test were classified with pre‐trained neural network models VGG16, VGG19, DenseNet121, Xception, and InceptionV3. As a result of the analyses, the DenseNet121 model showed the best classification performance, and subsequent analyses were carried out on this model. The main purpose of choosing the DenseNet121 model is to maximize the classification accuracy. In this process, SENet was applied to the DenseNet121 model, and the CBAM mechanism was integrated. Finally, a strengthened structure was created by combining the DenseNet121 model with SENet and CBAM. The experimental results obtained are presented in Table 4.

In this section, the theoretical background of the methods used in the study is briefly explained.

Deep learning, inspired by the principles of the human brain, is a subfield of machine learning (ML) and artificial intelligence (AI). Its capacity to identify intricate patterns and extract valuable information from massive data sets attracts attention. This technology learns intricate relationships in data and applies them to provide precise classifications or predictions using multi‐layered artificial neural networks. Deep learning networks can represent increasingly complicated data by learning more abstract representations at each layer.

Deep learning typically uses artificial neural networks with three primary components: the input, hidden, and output layers. After the input layer receives data, the output layer performs the final classification or prediction, and the hidden layers learn increasingly complex features. By learning increasingly intricate and abstract properties at each layer, these networks uncover deep structures in the data (Taye 2023).

Transfer learning attempts to apply the classification ability of a model trained on a large number of images to other datasets.

Attention mechanisms are a crucial technique that enhances the power and efficiency of deep learning networks. These strategies enhance the learning process by enabling the model to concentrate on particular regions or features. The Convolutional Block Attention Module (CBAM) exemplifies these systems, comprising two primary components: the Channel Attention Module and the Spatial Attention Module (Suo et al. 2024).

The Channel Attention Module produces summary data for each channel and utilizes this data to create channel‐level attention coefficients. During this procedure, each channel of the input feature map is examined individually. Initially, summary information for each channel is acquired by global average pooling and global maximum pooling processes. These processes yield statistical summaries that assist in assessing the significance of the channels. Subsequently, this summary information is amalgamated via fully connected layers, resulting in the generation of an attention coefficient for each channel. These coefficients assist the model in identifying the significance of various channels, allowing it to concentrate on more pertinent areas of the data (Woo et al. 2018).

Furthermore, the Spatial Attention Module employs the attention mechanism at the spatial dimension. This module assesses the significance of particular regions within the feature map. The Channel Attention Module and the Spatial Attention Module collaborate, enabling CBAM to execute efficient attention processes at both the channel and spatial dimensions. This combination aids the model in optimizing feature selection and enhancing processing efficiency (Ma et al. 2024) (Figure 4).

SENet (Squeeze‐and‐Excitation Network), a crucial architectural component in Convolutional Neural Networks (CNNs), markedly enhances model performance by emphasizing channel interdependencies. The fundamental element of SENet, the “Squeeze‐and‐Excitation” (SE) block, possesses the capability to adaptively adjust feature responses on a channel‐specific basis, particularly through the modeling of channel dependencies. This approach enhances the significance of the information processed in each layer of a CNN and enables the model to concentrate on more critical areas of the data.

The SE block comprises two primary stages: Squeeze and Excitation. During the compression phase, the spatial dimensions of each channel are reduced through the application of global average pooling. This approach generates a singular scalar value for each channel, encapsulating the overarching significance of the channel. This condensed information enhances the understanding of channel‐level dependencies. During the stimulation phase, the summary data acquired in the compression phase is processed via completely connected layers and activation functions. This approach generates an attention weight for each channel. These coefficients enable the model to discern the significance of various channels and to adjust the responses of the pertinent channels accordingly.

The attention mechanism of SENet facilitates channel‐level information processing alongside the convolution operator present in conventional CNNs. Traditional convolution processes primarily emphasize the extraction of spatial data, but the SE block of SENet integrates both spatial and channel‐level information. This enables the model to acquire more informative and distinguishing traits. Specifically, emphasizing channel interactions enhances the model's generalization capability and yields superior performance (Hu et al. 2020) (Figure 5).

The model was trained for 50 epochs. The confusion matrix was used to evaluate the validation performance of the model. This matrix shows the correct and incorrect predictions of the classification model in detail and is an important tool for analyzing the success of the model. In the evaluation made through the confusion matrix, the following four basic terms are used to determine the accuracy of the model:

TP: when the predicted results are correctly classified as positive.

TN: when the model correctly classifies the negative class as negative.

FP: when the model incorrectly classifies the negative class as positive.

FN: when the model incorrectly classifies the positive class as negative.

Using these four basic terms (TP, TN, FP, FN), various performance metrics of the model can be calculated:

The percentage of true positive or true negative results in an accurate testing set is known as accuracy. This equation is used to calculate it: (1) Accuracy=TP+TN/TP+FP+FN+TN where the parameters exist, the metrics employed to assess specificity, sensitivity, and accuracy in performance evaluation encompass TP (True Positives), TN (True Negatives), FP (False Positives), and FN (False Negatives). The quantity of accurately anticipated positive classes is represented by TP. The result of correctly predicted negative classes is TN. The quantity of cases anticipated to be positive yet ultimately classified as negative is referred to as false positives (FP). The quantity of cases anticipated to be negative when the results should have been positive is now denoted as FN.

Precision is the ratio of correctly predicted positive instances to the total number of instances projected as positive, reflecting the accuracy of positive forecasts. The subsequent formula for calculation is as follows: (2) Precision=TP/TP+FP The ratio of correctly predicted positive instances to the total number of actual positive instances is referred to as recall (sensitivity). This equation is employed for its calculation: (3) Recall=TP/TP+FN In classification problems, the f‐score is a statistic that evaluates the precision of positive predictions against the recall of successfully identified positive classes. The subsequent formula is the calculation (Sofuoğlu and Birant 2023): (4) Fscore=2×Precision×Recall/Precision+Recall

The research was conducted in a cloud environment with an NVIDIA Turing T4 GPU with 16 GB of memory, utilizing Python programming. The resolution of the input images was adjusted based on the architecture of the pre‐trained neural networks employed: 224 × 224 for VGG16 and VGG19, 299 × 299 for InceptionV3 and Xception, and 224 × 224 for DenseNet121. The Adam optimizer was used during training, with a batch size of 32 for both training and testing. The maximum number of training epochs was set to 50, with early stopping implemented to prevent overfitting. A learning rate of 0.0001 was used throughout the training process. The dataset was partitioned into training and testing sets in an 80:20 ratio. The distribution of training and testing samples for each class is illustrated in Figure 6. The primary evaluation metrics utilized to assess the model's performance included accuracy (Acc), precision (Prec), recall (Rec), and F1‐score.

Several parameters were used to establish the experiment as described in Table 2.

In this study, the learning rate was set to 0.0001 during model training, and the Adam Optimization Algorithm was utilized for the optimization process. The batch size for each mini‐batch was configured to 32, and the ReLU activation function was employed. The training process was conducted over 50 epochs.

The proposed deep learning model increased the classification accuracy of the existing DenseNet121 model from approximately 92% to 96.36% in the detection of grapevine downy mildew, grape leaf blister mites, and healthy grapevine leaves, and provided a promising performance increase in this direction. The results obtained in this section are discussed in detail.

The classification performances obtained as a result of the analyses are presented comparatively in Table 3.

When the accuracy values obtained in the training and validation phases are examined in Table 3, significant performance differences are observed between the models. Traditional VGG architectures (VGG19 and VGG16) showed the lowest success with 84.85% and 85.45% validation accuracy, respectively. More advanced architectures InceptionV3 and Xception significantly increased their performance by reaching 91.52% and 92.12% validation accuracy, respectively. The performance of DenseNet121‐based models is remarkably high in both the basic DenseNet121 and its enhanced versions. In particular, the integration of SE blocks (SENet) and attention mechanisms (CBAM) further increased the validation accuracy of the model. The DenseNet121 + CBAM + SENet model achieved superior success in grapevine leaf disease diagnosis by exhibiting the best performance with 96.36% validation accuracy and 97.87% training accuracy.

Precision, recall, and F1‐score metrics of VGG19 and VGG16 (around 0.84–0.85) are significantly lower than DenseNet121 variants. The same metrics are higher than VGG16 and VGG19 for InceptionV3 and Xception. The DenseNet121 + CBAM + SENet model showed the best performance (0.96) in all three metrics. This shows that the classification performance of the model is balanced and can distinguish both positive and negative classes well.

The changes in training and validation accuracy values over time for different deep learning models (VGG19, VGG16, InceptionV3, Xception, and DenseNet121) are presented in Figure 7.

It was observed that the training and validation accuracy rates in all models showed a steady increase from the initial low levels. This situation reveals that the models were able to learn the dataset during the training process. DenseNet121 and Xception models showed a more stable and higher performance in validation accuracy compared to other models. These models seem to be superior to the others in terms of generalization ability.

The change in training and validation loss values over time for deep learning models (VGG19, VGG16, InceptionV3, Xception, and DenseNet121) is presented in Figure 8. The graphs show the learning dynamics of the models during their training processes. In all models, training loss and validation loss values are high at the beginning and tend to decrease as training progresses. This shows that the models are learning from the data.

DenseNet121 and Xception models show the most stable decreasing trend in validation loss values, revealing that they have better generalization capacity.

The complexity matrices in Figure 9 show the classification performance of different deep learning models in detail.

When we examine the complexity matrix for the VGG19 model, a high correct classification rate was observed, especially in the “downy mildew” and “healthy” classes. However, the errors were higher in the “mites” class; some samples were misclassified as “downy mildew” and “healthy.” This shows that the model has difficulty distinguishing the “mites” class. When we examine the VGG16 model, similarly, although it exhibited a strong performance in the “healthy” class, the error rate was higher in the “downy mildew” and “mites” classes. The false positive rate in the “downy mildew” class is particularly striking. When we look at the complexity matrix for the InceptionV3 model, it showed a relatively balanced performance for all classes. While almost perfect accuracy was achieved in the “healthy” class, small errors were observed in the “mites” and “downy mildew” classes. When we look at the complexity matrix for Xception, it provided better accuracy in the “mites” class compared to the other models. However, there are small errors in the “downy mildew” and “healthy” classes. It can be said that it has a good performance in general. DenseNet121 showed the most balanced performance in all classes. Especially in the “downy mildew” and “healthy” classes, the correct classification rate is quite high. It also made fewer errors than other models in the “mites” class.

The training and validation processes of the DenseNet121 + SENet, DenseNet121 + CBAM, and DenseNet121 + SENet + CBAM models are given in Figure 10. Training and validation accuracy and training and validation loss are given for each model.

The upper graphs show the change in accuracy, and the lower graphs show the change in loss. In DenseNet121 + SENet, training and validation accuracy increased rapidly at first and stabilized around epoch 20–30. Training accuracy was slightly higher, but validation accuracy was very well matched. Training loss and validation loss decreased rapidly, and although validation loss showed some fluctuations, it generally remained at a low level. For DenseNet121 + CBAM, training accuracy and validation accuracy increased rapidly again and stabilized after epoch 30. Validation accuracy is quite consistent with training accuracy. Training and validation loss generally reached low levels. However, it is seen that there are small fluctuations in validation loss as in other models. For DenseNet121 + SENet+CBAM, training and validation accuracy follow a similar course with other models, but validation accuracy is observed to be more stable. Validation accuracy is one of the most consistent in this model. Training loss decreased steadily, and validation loss is generally low. However, there is less fluctuation towards the end of the epochs, indicating that the model exhibits better generalization performance. In the analysis, DenseNet121 + SENet+CBAM was the most successful generalization model. The close training and validation accuracies showed that all models did not overfit. The combination of CBAM with SENet gave the most stable results in validation accuracy and loss. This combination better captures complex features.

The variations observed in the comparative analysis plots can be explained by several factors. The first set of plots (Figure 7) shows the training and validation accuracy curves of different deep learning architectures (VGG19, VGG16, InceptionV3, Xception, DenseNet121). The structural differences between the architectures cause variations in learning rate and convergence behavior. The third set of plots (Figure 10) shows that the variations in the accuracy curves decrease, especially with the inclusion of attention mechanisms (SENet, CBAM and their hybrid versions). This is due to the fact that attention mechanisms improve the ability of the model to focus on important features.

In Figure 11, the confusion matrices created for DenseNet121 + SENet, DenseNet121 + CBAM, and DenseNet121 + SENet + CBAM are given. These matrices show the classification performances of the models. Each cell represents the number of class examples that the model predicted correctly or incorrectly.

Looking at Figure 11, the classification performance of the models is analyzed for three classes (downy mildew, healthy, and mites). Looking at the confusion matrix for DenseNet121 + SENet, for the downy mildew class, there are 58 correct predictions and 2 incorrect predictions (classified as mites). For the healthy class, there are 59 correct predictions, and 2 incorrect predictions (classified as 1 mites, 1 downy mildew). For the mites class, there are 39 correct predictions and 4 incorrect predictions (classified as 3 downy mildew, 1 healthy). When we look at the overall performance of the DenseNet121 + SENet model, the performance of the model is generally good, but there are more incorrect predictions in the mites class compared to other classes. Looking at Figure 11 for DenseNet121 + CBAM, for the downy mildew class, there are 59 correct predictions and 1 incorrect prediction (classified as mites). For the healthy class, there are 60 correct predictions and 1 incorrect prediction (classified as mites). For the mites class, there are 38 correct predictions and 5 incorrect predictions (2 classified as downy mildew, 3 as healthy). When we look at the overall performance of the DenseNet121 + CBAM model, it is seen that it is more successful in the downy mildew and healthy classes. However, there are still some incorrect classifications in the mites class. When Figure 11 is examined for DenseNet121 + SENet + CBAM, there are 59 correct predictions and 1 incorrect prediction (classified as mites) for the downy mildew class. There are 59 correct predictions, and 2 incorrect predictions (2 classified as downy mildew) for the healthy class. There are 41 correct predictions and 2 incorrect predictions (1 classified as downy mildew, 1 as healthy) for the mites class. In terms of overall performance, it is the model that made the least incorrect predictions in the mites class. When we look at the overall performance of DenseNet121 + SENet + CBAM, it exhibited a more balanced performance compared to other models in all classes.

The graphs regarding the training and validation processes of the pre‐trained models used in the study and the proposed models are presented in Figure 12.

As a result, the findings show that methods that take into account attention mechanisms and channel relationships provide a significant advantage in grape leaf disease classification. In particular, CBAM and SENet components built on the DenseNet121 architecture provided superior accuracy in both training and validation stages by providing better feature extraction from visual data. These findings can serve as a reference for future studies and guide model optimization. In addition, the performance of the supported models (CBAM, SENet) has yielded results at a level that is competitive with many studies in the literature. Our DenseNet121 + CBAM + SENet model has shown a performance close to or superior to other studies in the literature with 96.36% validation accuracy. The results of the performance improvement are given in Table 4.

The proposed method integrates DenseNet121, CBAM (Convolutional Block Attention Module), and SENet (Squeeze‐and‐Excitation Networks) for the classification of grapevine leaf diseases. The model was trained for 50 epochs with a learning rate of 0.0001, using the Adam optimizer, a batch size of 32, and the ReLU activation function. This configuration resulted in an improvement in accuracy, achieving 92.73% accuracy, which represents an enhancement of 3.66% compared to previous results. The use of attention mechanisms such as CBAM and SENet likely contributed to better feature selection and improved model performance, highlighting the effectiveness of combining these advanced techniques.

Although the integration of the CBAM and Squeeze‐and‐Excitation (SE) modules into the baseline model leads to a slight increase in computational time during both training and inference, this overhead remains minimal and within acceptable limits for most practical use cases. The computational complexity increase introduced by the proposed hybrid attention mechanism is entirely reasonable and justifiable considering the performance improvement and classification accuracy gain it provides. Moreover, with modern GPU hardware, this overhead does not pose a significant obstacle in practical applications.