Enhanced Crop Leaf Disease Classification Using Layer Freezing and Convolutional-Pooling Extensions

The automation of leaf disease detection has attracted growing attention from the research community, driven by its significance in modern agriculture. Multiple studies have explored the automatic diagnosis of plant diseases, and although challenges remain, researchers continue to employ advanced image processing methods and deep learning architectures to enhance classification performance. Rajpoot et al. [13] have utilized deep learning techniques for the autonomous identification of leaf diseases in multiple crops.

For instance, Ahmad et al. [14] implemented a CNN-based approach for plant disease detection, specifically addressing class frequency issues in imbalanced datasets. Their method involved stepwise transfer learning to minimize the convergence time of the CNN architecture [15], and, to optimize for handheld devices, a MobileNet model without a dense layer was employed. Another notable study by Zhou et al. [16] focused on grape leaf diseases and introduced a data augmentation method surpassing previous Generative Adversarial Networks (GANs). By integrating a faster CNN with a fine-grained GAN, the model accuracy of grape leaf disease classification was improved, particularly in identifying local spot areas through data augmentation.

Liu et al. [17] implemented a lightweight CNN for better fine-grained classification performance. SqueezeNet, used in this context, employed a multiscale convolution kernel and a coordinate technique to identify features. Experiments were performed on the plant disease recognition dataset from the 2018 AI Challenger challenge to assess the proposed approach. Yu et al. [18] proposed a hybrid approach that integrates K-means clustering with deep learning for maize leaf disease prediction. In this method, K-means clustering was employed to segment diseased regions in maize leaves, and the segmented outputs were then fed into a deep learning model for classification. The effectiveness of the approach was evaluated against established models, including ResNet and VGG19.

Bukhari et al. [19] developed datasets for wheat stripe rust and applied various segmentation techniques, such as Watershed, GrabCut, and U2-Net. The influence of segmentation on classification performance was assessed using ResNet-18 as the deep learning backbone. Cap et al. [20] developed LeafGAN, a data augmentation technique employing an attention mechanism to produce distinct diseased images from healthy ones. The LeafGAN model produced images of higher visual quality than the standard CycleGAN, thereby improving the effectiveness of plant disease diagnosis. Wu et al. [21] developed a fine-grained model for identifying diseases in peach and tomato leaves using an attentional deep neural network. During training, a Reconstruction and Generation Model was used to discriminate global features, and the proposed model performed well with minimal memory and computational resource requirements.

Stephen et al. [22] proposed a unique strategy that uses both 3D and 2D Convolutional Neural Networks (CNNs) to extract rich information important for classification tasks. Their methodology included an Improved Backtracking Algorithm (IBS) combined with Generative Adversarial Networks (GANs) to improve classification accuracy. Notably, the settings of the IBS method were carefully chosen to solve typical problems such as instability, vanishing gradients, and model collapse. By combining the power of 3D and 2D CNNs [23], the approach efficiently gathered varied color and texture information required for disease detection. This optimized GAN framework effectively detected three common rice diseases—bacterial blight, leaf blast, and brown spot—as well as healthy crops, with an impressive accuracy rate of 98.7%.

Elfatimi et al. [24] proposed a network architecture designed to accurately classify angular leaf spots and bean rust in bean leaves. The model demonstrated accurate recognition of diseases in their early stages, utilizing the MobileNet architecture for disease identification on a public dataset. By fine-tuning key parameters, including the optimizer, batch size, and learning rate, the model achieved superior performance over conventional approaches.

The AgriDet framework [25] was developed to tackle issues related to unclear imagery and background interference in plant disease detection based on image processing methods. The INC-VGGN neural network architecture, based on the Inception-Visual Geometry Group Network, played a crucial role. Preprocessing techniques were applied to alleviate issues during image capture, and the GrabCut technique was employed to address occlusion problems. The Kohonen-based deep learning algorithm successfully identified multiscale features in plant leaves. Singh et al. [26] presented a CNN-based approach for identifying several types of diseases in cotton leaves, including fungal, viral, and bacterial infections.

The model involved two processes: identifying diseased leaves and classifying them into their respective disease types. The CNN-based cotton leaf disease prediction model [27] exhibited good performance and computational efficiency. Shafi et al. [28] focused on wheat rust, dividing it into four stages: moderate, resistant, healthy, and susceptible. U-Net was used to eliminate undesirable background features and extract rust disease-related features. The impact of stripe rust disease in wheat was determined using the Xception model and ResNet-50. Intelligent edge detection devices were also employed to monitor rust disease-affected wheat crops. Sunil et al. [29] obtained a cardamom plant leaf disease dataset in Chinnahalli, Sakaleshpur, India, and utilized the U-Net architecture to remove the background for disease detection using EfficientNetV2. The proposed system was tested on grape plant leaves. Liu et al. [30] introduced PiTLid, an apple detection approach based on Inception-v3 transfer learning. This strategy addressed the challenge of insufficient training data. The apple dataset included four classes: black rot, rust, scab, and healthy. PiTLid outperformed other phenotyping models in classification efficiency and demonstrated applicability with limited grape and peach data.

While deep learning has made progress in leaf disease classification, major challenges remain. CNN-based approaches, such as fine-tuned VGG19, ResNet, and MobileNet, frequently overfit small agricultural datasets. Traditional transfer learning algorithms either freeze layers randomly or fine-tune entire networks, potentially resulting in the loss of transferable low-level features. Many architectures also lack focused adjustments to capture complex, disease-specific patterns, resulting in incorrect identification, and some require considerable preprocessing, increasing computational costs and reducing practical efficiency. Moreover, fine-tuning pretrained models necessitates careful selection and adjustment of hyperparameters, such as learning rate, batch size, and optimizer, to optimize performance effectively.

To address these challenges, we propose Selective VGG19-Net, which includes selective layer freezing of Blocks 1 and 2 to preserve crucial low-level features while allowing deeper layers to adapt, as well as a convolutional-pooling depth extension to capture hierarchical, disease-specific patterns.

In deep learning, transfer learning is an important technique that applies the knowledge gained by pretrained models to solve different tasks [32]. This approach eliminates the need to develop models from scratch and instead capitalizes on features learned from extensive datasets like ImageNet. Through fine-tuning or feature extraction, transfer learning can attain high accuracy across diverse domains, even when training data is limited. These models are particularly advantageous for scenarios with limited labelled data and computational resources.

As illustrated in Figure 4 our study applies transfer learning technique for classifying leaf diseases. we started by collecting a dataset of apple and citrus images from two public repositories and resizing them to 224 × 224 × 3. Divide the dataset into three splits: train, test, and valid set, in the proportions of 70%, 20%, and 10%. Rotation, vertical flip, and horizontal flip are examples of augmentation procedures that improve the dataset. The training data was then merged with some augmented images to develop our proposed Selective VGG19-Net model. Our model involves selectively freezing layers in Block 1 and Block 2 of the VGG19 architecture. By selectively freezing layers, we retain essential low-level features from pretrained weights. Additionally, we enhance the frozen VGG19 model by introducing an extra convolutional and pooling layer as shown in Figure 5, increasing the network's depth to capture intricate patterns and hierarchical representations of leaf diseases. Experimental results demonstrate that our method outperforms other fine-tuning approaches, achieving superior classification accuracy. By combining selective layer freezing and extended feature extraction, the model achieves improved identification of disease-relevant features. To evaluate our Selective VGG19-Net, we compare it to CNN-based pretrained models on datasets containing apple and citrus images.

Figure 4. Workflow of proposed architecture for apple and citrus disease classification

Figure 5. Block structure of Selective VGG19-Net

4.1 Preprocessing

Model training begins with preprocessing, which is the act of preparing a set of input images for further analysis. In this work, both citrus and apple leaf datasets were utilized, and several pixel-level adjustments, such as scaling and normalization, were applied to every image. To further expand the dataset, several augmentation techniques were applied, including image rotation, horizontal and vertical flipping, shifting, and shearing. These methods increased the dataset size, reduced overfitting, and improved the model’s robustness to variations in input data. Consequently, the preprocessed images were better suited for efficient model training and evaluation. Figure 6 presents examples of preprocessed images from the apple dataset, while Figure 7 shows preprocessed images from the citrus dataset. Table 2 compares dataset samples before and after augmentation, illustrating how preprocessing enhanced dataset diversity.

Table 2. Augmentation results on apple and citrus leaf images

Figure 6. Samples of preprocessed images of apple dataset with augmentation techniques

Figure 7. Samples of preprocessed images of citrus dataset with augmentation techniques

4.2 Selective layer freezing

In our research, we present an approach to fine-tuning a pretrained model, specifically a Selective VGG19-Net tailored for leaf disease classification. Our approach includes strategically freezing certain blocks of layers to optimize the use of pre-existing knowledge while allowing adaptation to the specific task. Additionally, the architecture was extended with extra convolutional-pooling layers to enhance its ability to capture critical features essential for leaf disease classification. A central aspect of the approach involves selectively freezing specific layers, particularly Blocks 1 and 2 during training, as illustrated in Figure 7. By doing so, aim to mitigate overfitting and prevent catastrophic forgetting. Let $l_f$ is the loss function and $\nabla$ be the gradient, the weight update rule during training, $w_i$ is the old weight, $w_j$ is the new weight,

$w_i \rightarrow w_i-\eta \nabla w_i l_f$     (1)

When Layer frozen then,

$\nabla_{w f l_f}=0, w_{f-i}=-w_{f-j}-\eta \cdot 0=w_{f_j}$     (2)

As expressed in Eqs. (1) and (2), Blocks 1 and 2 are chosen for freezing because they primarily learn generic and low-level features, including edges, textures, and basic shapes. These features prove crucial for effective leaf disease classification.

The selective layer freezing not only facilitates the retention of essential knowledge from the pretrained model but also ensures that the model can adapt specifically to the nuances of leaf disease characteristics.

4.3 Convolutional-pooling extension approach

In section 4.3, we introduce a convolutional-pooling extension strategy to enhance the VGG19 architecture, focusing on specific blocks for improved performance in capturing intricate patterns within leaf images. Blocks 1 and 2 consist of two convolutional layers each, activated by ReLU functions, and are followed by a max pooling layer. Notably, these are made freeze only to these blocks, allowing flexibility in the learning process. In our model, we have made enhancements to Blocks 3 and 4 of the VGG19 architecture. Specifically, added two additional convolutional layers to Block 3 and one additional layer to Block 4. Additionally, we have removed Block 5 entirely from the base architecture and these modifications are illustrated in Figure 7. This extension aims to enhance the network's capacity, resulting in superior performance, particularly in handling complex features in leaf images.

Let x be the input. The output of Block 1 is obtained as shown in Eq. (3):

$z_{\text {block} 1}=\sigma\left(w_{\text {conv} 1} * x b_{\text {conv} 1}\right)$     (3)

The output of Block 2 is then computed as expressed in Eq. (4):

$z_{\text {block} 2}=\operatorname{maxpool}\left(\sigma\left(w_{\text {conv} 2} * z_{\text {block } 1}+b_{\text {conv} 2}\right)\right.$     (4)

Similarly, the extended Block 3 produces the output feature maps given in Eq. (5):

$z_{\text {block} 3}=\operatorname{maxpool}\left(\sigma\left(w_{\text {conv} 3} *\left(\sigma\left(w_{\text {conv} 2} * y_{\text {block} 1}+b_{\text {conv} 2}\right)\right)+b_{\text {conv} 3}\right)\right)$     (5)

Finally, the output of Block 4 is derived as shown in Eq. (6):

$z_{\text {block} 4}=\operatorname{maxpool}\left(\sigma\left(w_{\text {conv} 4} * z_{\text {block} 3}+b_{\text {conv} 4}\right)\right)$     (6)

where, $z_{\text {block} 4}$ is the output feature of blocks, convolution layer represented by $w_{\text {conv} 1}$ and $b_{\text {conv} 1}$, here w is the convolutional filter and b is the bias.

4.4 Computational complexity

The SelectiveVGG19-Net incorporates multiple convolutional blocks with filter depths progressively scaled to 64, 128, 256, and 512, along with max-pooling layers that compress the feature maps to reduce dimensionality. In the final stage, the network transitions into a fully connected architecture with three dense layers. While the convolutional operations capture hierarchical feature representations, the pooling strategy minimizes spatial resolution and computational overhead in the deeper layers. Although the deep layers add a significant number of parameters, the overall design strikes a compromise between feature richness and computational efficiency. The network includes 74,583,109 trainable parameters, requires 298 MB of weights, and can do inference on a single image in 0.3 to 0.4 seconds on a GPU, making it suitable for agricultural applications. Additional efficiency benefits can be achieved by selective layer freezing, pruning, replacing dense layers with global average pooling.

The Selective VGG19-Net for the apple leaf has a model accuracy of 0.9983, as shown in Figure 11(a), and a model loss, as shown in Figure 11(b). Figure 12(a) and 12(b) indicate that the SelectiveVGG19-Net trained on the apple dataset divided the disease into four classes, which are evaluated by the confusion matrix and measured by evaluation metrics such as precision, recall, and f1-score.

Figure 13(a) depicts the accuracy of the SelectiveVGG19-Net for citrus leaves, whereas Figure 13(b) showcases the model's loss. Additionally, Figure 14(a) and 14(b) show the results of the SelectiveVGG19-Net based on the citrus dataset, which classified diseases into five distinct class. This evaluation utilizes metrics including precision, recall, and f1-score which are evaluated by confusion matrix. Table 4 shows how our model performed on both the citrus and apple datasets, as measured by a variety of performance measures. These metrics provide insight into how well the model did during both the training and validation phases. During training on the citrus dataset, the model acquired a loss function value of 0.0523, which indicates the level of inaccuracy between predicted and actual citrus leaf classifications. This dataset has a training accuracy of 0.9803, which is the number of correctly classified citrus leaves compared to the total during training.

During validation in citrus, the model achieved an accuracy of 0.9896 with a validation loss of 0.1541, reflecting its performance on an independent validation dataset. For the apple dataset, the model reached a training accuracy of 0.9983, representing the proportion of correctly identified apple leaf diseases during training. The corresponding training loss was 0.0162, indicating the discrepancy between predicted and actual classifications. On the validation set, the model attained an accuracy of 0.9851 with a validation loss of 0.1252. These results provide insights into the model’s performance in terms of classification accuracy and loss for both apple and citrus leaf datasets.

Figure 11. Selective VGG19-Net results for apple leaves: (a) Accuracy and (b) loss

Figure 12. (a) Confusion matrix analysis of selective VGG19-Net for apple (b) performance evaluation of Selective VGG19-Net for apple

Figure 13. (a) Model accuracy of SelectiveVGG19-Net for citrus (b) Model loss of SelectiveVGG19-Net for citrus

Figure 14. (a) Confusion matrix of Selective VGG19-Net for Citrus (b) Evaluation of Selective VGG19-Net for Citrus

Table 4. Performance of proposed SelectiveVGG19-Net

In addition to these metrics, statistical analyses were performed to assess the reliability of the proposed Selective VGG19-Net. For citrus leaves, the mean accuracy was 0.9896, with a 95 percent confidence interval ranging from 0.9811 to 0.9980, a t-statistic of 20.25, and a p-value of 0.0024, indicating strong statistical significance. Similarly, for apple leaves, the mean accuracy was 0.9823, with a 95 percent confidence interval of 0.9599 to 1.0046, a t-statistic of 6.21, and a p-value of 0.0250, demonstrating consistent and reliable performance.

Table 5 shows a sequential overview of the Selective VGG19-Net architecture, beginning with an input image dimension of 224 × 224 × 3 and follows the classical VGG19 block design [33]. The initial layers are made up of two convolutional blocks, each of which applies 64 filters with a kernel size of (3,3) and stride of 1 × 1.

Table 5. Layer summary of SelectiveVGG19-Net

Each convolutional block is followed by a MaxPooling2D layer with a pool size of (2, 2), which reduces the feature map dimensions to (112, 112). Subsequent layers follow a similar pattern but with additional convolutional layers in blocks 3 and 4, utilizing our strategy of convolutional and pooling extension while selectively freezing certain layers. These added convolutional layers progressively increase the number of filters, aiding in the extraction of more complex features. Throughout, the stride remains consistent, and the kernel size is unchanged. In the final layers before block 4, the output from the preceding convolutional layers is flattened, followed by a fully connected dense layer with 4096 units to prevent overfitting. Dropout regularization is then applied. The layer that produces the result is made up of a dense layer with softmax activation that calculates probabilities for each class in the classification task. Our Selective VGG19-Net was used to classify five types of citrus and four types of apple leaf diseases, utilising the mentioned layer structure to get accurate classification results.

6.1 Comparative evaluation

This study used several kinds of deep learning techniques to assess and evaluate the proposed SelectiveVGG19-Net. The accuracy of this method was compared with existing techniques, including VGG16 [34], VGG19 [35], MobileNetV3Large [36], DenseNet-121 [37], ResNet [38], and EfficientNet [39].

6.1.1 Various experimental results

The SelectiveVGG19-Net was trained on apple and citrus datasets at epochs 20, 30, and 40. During training, metrics of performance like as accuracy and loss were calculated for both the training and validation datasets. As the number of epochs increased, the model's performance became easier to analyse and explain. Figure 15 and Figure 16 depicts how the model's performance changed with different training times, revealing insights into its ability to generalise and effectively classify leaf diseases in apples and citrus. In addition, we tested our model with two distinct optimizers, SGD and RMSprop, as shown in Table 6. These optimizers modify the network parameters. For the apple dataset, SGD obtained a training accuracy of 0.96, whereas RMSprop achieved 0.95. For the citrus dataset, SGD produced an accuracy of 0.96, whereas RMSprop achieved 0.94. These changes minimised the loss and increased training accuracy. However, utilising the Adam optimizer enhanced the SelectiveVGG19-Net's accuracy.

Figure 15. Variation in apple classification performance with changing epochs

Figure 16. Variation in citrus classification performance with changing epochs

Table 6. Comparison of accuracy and loss for different optimizers

Table 7. Evaluation of proposed Selective VGG19-Net against traditional and modern deep learning architectures for leaf disease classification

Table 7 summarizes recent deep learning and transformer-based modeling methods for plant leaf disease classification. Chen et al. [40] presented MtConvNeXt with ternary attention mixed with an SVM classifier for tomato leaf disease detection in ten classes, achieving 96.15% accuracy. Wang et al. [41] used Efficient Channel Attention (ECA) and ConvNeXt was employed to detect diseases in rice leaves across six disease classes and a healthy class, achieving an accuracy of 94.82%. Soeb et al. [42] suggested a YOLOv7 light weight detection framework (YOLO-T) for tea leaves, achieving 97.3% classification accuracy with a high F1-score of 0.965 and mean Average Precision of 98.2%. Hossain et al. [44] investigated a number of transformer models for tomato leaf disease identification, including MaxViT, EANet, CCT, and PVT, but found no accuracies. Yang et al. [45] achieved 97.44% accuracy using DHLC-DETR, a detection-transformer model, using the IDADP dataset containing three rice diseases.

When compared to these cutting-edge approaches, our Selective VGG19 model had a total accuracy for classification of 98.03% with citrus leaf disease datasets and 99.83% with apple leaf disease datasets. This increase in classification accuracy demonstrates not only the efficacy of selective layer freezing, but also the effect of depth extension in improving model understanding for more consistent feature extraction and generalization performance, particularly when working with small and unbalanced agricultural datasets.

Table 8 summarizes the evaluations of the four apple dataset classes: Apple Black Rot, Apple Healthy, Apple Scab, and Cedar. We tested our model against the models given above and obtained the results. Our proposed model was tested for precision, recall, and f1-score in contrast to existing cutting-edge techniques. The SelectiveVGG19-Net performed well in terms of accuracy, with a score of 0.9983 for apple disease. Table 9 compares our Selective VGG19-Net to other conventional deep neural network models.

Table 8. Performance comparison with existing methods for apple leaf disease detection

Table 9. Comparison with state-of-the-art methods for citrus leaf disease detection

While these models were originally trained on their dataset and performed well on it, they exhibited lower training accuracy when applied to our dataset, specifically the citrus dataset. Among these models, VGG19 outperformed VGG16 with an accuracy of 0.9442 compared to 0.85. Additionally, MobileNetV3Large, DenseNet, ResNet, and EfficientNet achieved accuracies of 0.8523, 0.9032, 0.7611, and 0.7141, respectively. We evaluated the performance of each model on different classes of citrus leaf diseases, including Black Spot, Melanose, Canker, Greening, and Healthy, using metrics such as precision, recall, and F1-score. Our Selective VGG19-Net demonstrates superior performance compared to these state-of-the-art methods.

To compare Selective VGG19-Net model with other existing models, we implemented it on two datasets and achieved impressive results is 98% accuracy for citrus classification and 99% accuracy for apple classification. Figure 17 visually represents the performance of our model alongside other existing CNN-based models.

Figure 17. Accuracy performance evaluation of the proposed SelectiveVGG19-Net against other models for apple and citrus leaf disease classification

The proposed Selective VGG19-Net outperformed both classic convolutional neural networks and recent transformer-based architectures in classifying apple and citrus leaf illnesses, with 99.83% accuracy for apple and 98.03% for citrus. The framework's selective freezing method, which preserves Blocks 1 and 2, allows it to maintain low-level structural feature such as edges, contours, and textures from pretrained ImageNet weights. This approach assures strong feature extraction in the early stages, while the additional convolutional-pooling depth improves hierarchical learning and allows the network to locate tiny, disease-specific patterns that are often difficult to detect in agricultural datasets.

This selective adaptation not only prevents overfitting and catastrophic forgetting, which are common problems with very small and imbalanced datasets, but it also stabilizes convergence during training. By leveraging a deeper hierarchical structure, the model demonstrates superior discrimination of closely resembling disease categories when compared with conventional CNNs like VGG16, ResNet50, and MobileNet, and even with more sophisticated transformer-driven models such as Vision Transformers and ConvNeXt. These findings demonstrate the efficiency of combining selective layer freezing with deeper architecture extensions as a reliable and generalizable technique to plant disease detection.

Beyond predicting performance, there are broader implications for real-world deployment to consider. Ethical considerations may develop as a result of dataset biases, as most images are from specific geographic regions and environmental conditions, thereby limiting generalizability across climates and crop kinds. In practice, the additional layers increase processing complexity, potentially limiting deployment on low-power or edge devices. Optimization solutions including pruning, quantization, and edge deployment can minimize memory usage and inference delay while maintaining accuracy, allowing for scalability in real-world agricultural environments.

To summarize, Selective VGG19-Net not only offers state-of-the-art performance in detecting apple and citrus leaf diseases, but it also provides a technically robust, flexible, and scalable precision agriculture framework. Its adaptation to various crops, disease kinds, and pretrained backbones demonstrates its promise as a dependable and generalizable tool for promoting sustainable farming methods and contributing to global food security.