SwinPyramidNet: a deep learning network for pear leaf disease classification using hybrid convolutional feature pyramids and swin transformer

1 Introduction

Pears are highly nutritious fruit crops that have been cultivated for thousands of years. Over time, they have continuously evolved alongside human civilization. The genus Pyrus comprises a total of 22 primary species, which have been grown in the regions of northern Africa, Asia, and Europe. Notable varieties such as the European pear (Pyrus communis), Chinese pear (Pyrus bretschneideri), and Japanese pear (Pyrus pyrifolia) are grown in temperate zones worldwide (Itai, 2007). The extensive cultivation of pears in China constitutes approximately 80% of the global pear production area (Yang et al., 2020). Importantly, pears contribute to a production volume of over 25 million metric tonnes according to the U.S. Department of Agriculture. This record underscores the importance of pears in global horticulture, especially in countries like the United States, China, and Italy (U.S. Department of Agriculture, 2025).

Pears have a market value expected to reach $41.47 billion by 2028. This is an increase from an estimated $32.35 billion in 2024, as reported by the Business Research Company. They are rich in dietary fiber, minerals, vitamins, and antioxidants. Despite the significant economic and nutritional value of pears, there is a lack of research on diseases that harm pear leaves. Addressing these diseases is essential to maintain consistent productivity and quality in pear farming By concentrating on the leaves of pear trees, diseases can be detected early before they affect the fruit. This is essential for keeping the tree healthy and maximizing crop production. Dealing with problems early in the leaf stage prevents extensive damage and improves disease control efficiency.

Failure to implement timely corrective measures may lead to fruit cracking and abscission, resulting in substantial economic losses (Zhang et al., 2021). Unfortunately, pear trees are susceptible to various leaf diseases, such as leaf curl, leaf spot, and slug damage, which constitute a growing threat to the productivity of pear orchards. Leaf curl is a fungal disease that hinders photosynthesis through leaf distortion, discoloration, and premature drop (Kaviani et al., 2022). Leaf spot is characterized by small black patches on the leaves that often result in significant leaf loss. Leaves are also susceptible to damage from pear slugs or sawflies. Failure to diagnose and treat these conditions can lead to rapid spread of the disease.

Conventional techniques for plant disease detection depend on a domain expert to meticulously inspect the leaf surface and deliver a diagnosis. In contrast, advice is not accessible to all producers in underdeveloped countries due to its high cost and time-consuming nature. Pathogens and pests are major factors contributing to global crop yield losses. They are estimated to cause annual losses of approximately US$220 billion (Singh et al., 2023). Recent technological advancements, such as hyperspectral imaging, remote sensing, thermal imaging, and machine learning algorithms, have significantly improved the detection of plant diseases. These techniques enable the analysis of plant traits that are unnoticeable to the human eye. They provide more precise and detailed information about the presence of diseases.

The traditional method of manually diagnosing diseases in pear leaves is time-consuming and labor-intensive (Karthik et al., 2023a). Moreover, conventional disease detection methods based on physical examination are prone to human error and are less effective in large-scale farming. This has led to an increased focus on developing more advanced and precise methods for early disease detection. Automated methods utilizing machine learning and deep learning enable rapid, accurate, and cost-effective disease diagnosis. The main focus of this study is to design an automated system based on deep learning specifically for pear leaves. Such approach can be integrated with drones in future to enable remote monitoring and diagnostics (Chin et al., 2023).

2 Related works

The following section reviews studies focused on crop disease classification. The methods employed in these studies can be broadly categorized into Machine Learning and Deep Learning techniques.

2.1 Machine learning-based techniques

Various studies have explored different ML algorithms, highlighting their effectiveness in identifying diseases from leaf images. Alagumariappan et al. presented a plant disease identification system that employs a Support Vector Machine (SVM), Random Forest, and AdaBoost classifiers. It was found that SVM demonstrated superior performance compared to the other classifiers (Alagumariappan et al., 2020). Similarly, Jaisakthi et al. explored the use of machine learning for grape leaf disease identification. Their research highlights the effectiveness of image processing combined with machine learning algorithms. The outcomes demonstrate that both Random forest and SVM yield high accuracy rates (Jaisakthi et al., 2019). Dawod & Dobre utilize machine learning methods to improve classification accuracy by integrating features from both leaf surfaces (Dawod and Dobre, 2022).

Recent advances in plant leaf disease detection have highlighted the limitations of traditional machine learning methods. This led to the development of deep neural networks that demonstrate improved classification performance. Classical machine learning algorithms require hand-engineered features, which demand a certain level of domain expertise. Conversely, deep learning algorithms inherently have many configurable parameters. They incrementally capture features through hidden layers, removing the reliance on intricate feature engineering (Alshammari et al., 2024; Karthik et al., 2023b).

2.2 Deep learning-based techniques

Deep learning methods have been widely investigated for the classification and detection of plant leaf diseases. Pre-trained models, along with the application of transfer learning, are used for the classification of plant leaf diseases. In recent years, advanced architectures in deep neural networks, such as EfficientNet have gained significant attention for plant disease classification.

Chen et al. demonstrated the use of a modified AlexNet architecture for classifying leaf diseases highlighting the model’s adaptability to various plant disease datasets (Chen et al., 2022). Alaeddine & Jihene proposed a deeper version of AlexNet for classifying plant diseases (Alaeddine and Jihene, 2021). Alatawi et al. utiized the VGG16 model for plant disease detection. They emphasize the importance of leveraging deep learning techniques to enhance the accuracy of disease detection (Alatawi et al., 2022). The study by Paymode & Malode explores the efficacy of the VGG model for classifying multi-crop leaf diseases (Paymode and Malode, 2022). Additionally, Nguyen et al. integrated VGG-19 with transfer learning and image segmentation methods to classify plant leaf diseases. This showcased the model’s ability to leverage pre-trained weights for improved feature extraction (Nguyen et al., 2022).

Zhou et al. utilized a deep residual dense network for the identification of crop leaf diseases. This highlighted the model’s capability to effectively extract features from diseased and healthy plant images (Zhou et al., 2021). Kanda et al. explored the use of fine-tuned residual neural networks for recognizing diseases in plant leaves (Kanda et al., 2022). Pratap et al. demonstrated the effectiveness of a customized EfficientNetB4 model for the high-precision multiclass classification of diseases (Pratap and Kumar, 2023). In their study, Atila et al. adopted EfficientNet and conducted a comparative analysis of its performance with other models to categorize plant leaf diseases (Atila et al., 2021). Abayomi et al. proposed the application of MobileNetV2 architecture for the detection of plant diseases. Furthermore, they employed data augmentation techniques on low-quality images to enhance the performance of the network (Abayomi‐Alli et al., 2021). Wu et al. explored the use of MobileNetV2 in their research on pear leaf disease recognition. Additionally, MobileNetV2 was integrated with a modified architecture to increase accuracy and could effectively classify pear leaf diseases even in complex backgrounds (Wu et al., 2023). Andrew et al. explored CNN-based pre-trained models such as DenseNet-121, ResNet-50, VGG-16, and InceptionV4 for identifying plant diseases. Comparative analysis identified DenseNet-121 as the best-performing model (Andrew et al., 2022). Similarly, Fenu & Malloci used five pre-trained architectures–VGG-16, VGG-19, InceptionV3 ResNet50, InceptionV3, MobileNetV2, and EfficientNetB0. These architectures served as feature extractors to classify pear leaf diseases and severity levels (Fenu and Malloci, 2021a).

Apart from using stand-alone pre-trained models, ensemble learning techniques have been applied in the context of plant leaf disease detection. Fenu & Malloci demonstrated that ensemble CNNs outperformed individual models in classifying pear leaf diseases. This approach leverages the strengths of various architectures, such as EfficientNetB0, InceptionV3, MobileNetV2, and VGG18 (Fenu and Malloci, 2023). Lakehal et al. further studied the use of CNN ensembles for estimating the severity of pear leaf diseases. By combining predictions from multiple models, the ensemble method improved the reliability of disease severity assessment (Lakehal et al., 2023). Faisal et al. introduced a hybrid feature fusion approach that integrates MobileNetV3, Swin transformer, and a variational autoencoder (Faisal et al., 2023). Banerjee et al. and Yashu et al. explored the integration of CNNs and Random Forests for pear disease classification. Each highlighted the potential of combining deep learning with traditional machine learning. While the former emphasized the development of a unified hybrid methodology to enhance accuracy in disease detection, the latter focused on improving pear disease recognition. This was done by leveraging CNN’s feature extraction capabilities and Random Forests’ decision-making strengths (Banerjee et al., 2024; Yashu et al., 2023).

Apart from CNNs, studies also presented techniques such as Attention mechanism to address plant disease classification. Alirezazadeh et al. investigated the application of the Convolutional Block Attention Module (CBAM) in conjunction with traditional CNNs for pear leaf disease classification. Their results demonstrate the capability of attention mechanisms in improving the classification accuracy of pre-trained models (Alirezazadeh et al., 2022). Similarly, Liu et al. explored the classification of plant leaf diseases using a multi-scale fusion model that leverages EfficientNet and an attention mechanism. Their research emphasizes the importance of a lightweight module that enhances the extraction of disease features from plant leaves (Liu et al., 2022). Karthik et al. introduced an attention-embedded residual CNN for disease detection in plant leaves. They demonstrated that attention mechanisms can significantly improve the model’s focus on relevant features, thereby enhancing classification accuracy (Karthick et al., 2020). Sugunadevi et al. proposed a model utilizing Vision Transformers for classifying plant leaves. This approach differs from traditional convolutional neural networks by utilizing the self-attention mechanism of transformers to enhance classification accuracy (Sugunadevi et al., 2023). Jia et al. proposed a model based on capsule networks specifically designed for pear leaf disease image recognition. This addresses the limitations of traditional CNNs, particularly in terms of spatial hierarchies and part-whole relationships within images. The capsule network’s architecture allows for improved generalization and robustness against variations in leaf images, which is essential for accurate disease classification (Jia et al., 2022).

In addition to classification, few studies performed segmentation for effective plant disease detection. Maryum et al. proposed a two-stage approach for plant disease detection. Image segmentation was conducted using a U-Net architecture to eliminate background noise. Feature extraction and classification were then performed using a pre-trained EfficientNet-B4 model (Maryum et al., 2021). Wang et al. also designed a UNet-based model for the purpose of pear leaf disease segmentation, which integrated a multi-scale feature extraction module with tokenized multilayer perceptron to use dynamic sparse attention (Wang et al., 2023). The aforementioned approaches illustrate the efficacy of deep learning models in detecting pear leaf diseases.

2.3 Research gaps and motivation

Despite the development of various techniques for pear leaf disease classification, several significant challenges still remain in this field.

• Existing models primarily focus on local or global feature extraction. This limitation can reduce their effectiveness in accurately detecting pear leaf diseases. The reliance on local features often leads to a limited understanding of the image context. Meanwhile, global features may overlook critical details present in localized areas. This dichotomy suggests a need for integrated feature fusion techniques that combine local and global information, thereby providing a more comprehensive representation of the leaf images.
• As deep learning models increase in complexity, the deeper layers tend to contain a greater number of parameters, which can result in reduced efficiency and potential accuracy degradation in pear leaf disease classification. To address this, many models utilize low-resolution feature maps. This approach requires scaling between layers to enable data propagation.
• Existing feature fusion methods often struggle to effectively combine multi-scale features. This can lead to suboptimal performance in fine-grained classification tasks where capturing both local details and global context is essential for precise classification.
• Limited research has successfully tackled the necessity for both channel-wise and spatial attention in fine-grained classification tasks, including pear leaf disease classification. An effective attention module that can capture long-range dependencies while maintaining positional information is essential for enhancing the accuracy and precision of pear leaf disease classification.

2.4 Research contributions

The proposed work presents the following research contributions.

• The proposed SwinPyramidNet aims to enhance pear leaf disease classification by utilizing a dual-track architecture: one used for global feature extraction and the other for local feature extraction. This dual-track design enables the network to contextualize the learned features more effectively, thereby improving classification accuracy.
• The Swin transformer track in the proposed model constructs features hierarchically. It begins with smaller patches and progressively integrates information across scales. This approach ensures the effective capture of global features for precise classification.
• The HCFPNet in the proposed network extracts local features from the input image. It comprise of Mobile inverted Bottleneck Convolution (MBConv) blocks and a Bidirectional Feature Pyramid Network (BiFPN) block. The MBConv blocks focus on capturing detailed local features through depth-wise separable convolutions. This allows for efficient feature extraction with fewer parameters. The BiFPN block is incorporated to integrate multi-scale features to enhance feature learning and classification performance.

3 Proposed methods

The proposed SwinPyramidNet is a hybrid dual-track architecture. It uses a Swin transformer track and an HCFPNet track. Figure 1 presents the workflow followed. The proposed dual-track deep learning network is illustrated in Figure 2. The initial step in the procedure is to resize the image to a shape of 224 × 224 with three color channels. The images are subsequently fed into the aforementioned tracks parallelly. Each track consists of separate blocks which execute specific processes contributing to the overall functionality. The global features are extracted by the Swin transformer track, while the HCFPNet is responsible for local feature extraction. The Swin transformer track consists of different blocks arranged sequentially. These blocks are combined with linear embedding and patch division. These blocks are aimed to improve the overall accuracy and learning parameters of the network by employing hierarchical and shifted window techniques to conduct more effective feature aggregation. The Hybrid Convolution Feature Pyramid Network consists of MBConv and 2D convolution blocks, batch normalization layers, and a BiFPN module. Depthwise concatenation is applied to fuse the features from both tracks. This is followed by feature weighting and reduction using Coordinate attention and global average pooling. Finally, Softmax activation is applied.

3.1 Swin transformer track

The Swin transformer model computes self-attention exclusively for local windows to enhance multi-scale feature learning (Liu et al., 2021). It produces hierarchical feature representation through four successive stages. Figure 3 illustrates the overall architectural design of the Swin transformer track. The Swin transformer block extracts relevant features by processing input RGB images using two-dimensional convolutions. The image is then divided into non-overlapping patches known as tokens by the patch partition block. Each patch is of size H × W where H and W represent the height and width of the patch respectively. The patch embeddings are reduced in dimensionality and cast into a fixed dimension using a linear embedding layer. The shape of the patch is defined by the hyperparameter P, which indicates the spatial dimensions of each patch within the image supplied as input. For a specified P, each patch corresponds to a square area of P-by-P pixels. The Swin transformer block operates on the patch embeddings via interactions between a sequence of subunits. Each sub-unit includes two normalization layers, an attention module, and a Multi-Layer Perceptron (MLP), as shown in Figure 4. The Window Multi-Head Self Attention (W-MSA) module is employed by the initial sub-unit to identify dependencies between patch embeddings. In order to concentrate on the shifted windows of patch embeddings, the next sub-unit implements Shifted Window Multi-Head Self Attention (SW-MSA). This approach allows the architecture to efficiently capture long-range dependencies and also reduce the number of parameters. The resulting output is passed as input for the next Swin block via a skip connection, and the long-range dependencies are captured using the hierarchical feature representation. Element-wise parameter addition takes place between each layer normalization block that incorporates multi-head self-attention and the layer normalization block that includes MLP. The down-sampling module increases the number of channels while concurrently reducing the spatial dimensionality of the feature representation. As the concluding step of each stage, the patch merging layer reduces the feature map resolution by a factor of ‘n’ by combining the feature representations of every pair of n × n adjacent patches in a depth-wise manner. The resultant activation map has 96 channels and is integrated with the output of the HCFPNet track. This complete set of features obtained is then utilized to train the model for the pear leaf disease classification task.

3.2 Hybrid convolution feature pyramid network track

The HCFPNet functions concurrently with the Swin Transformer track facilitating the extraction of features across multiple levels. This section highlights the importance of the HCFPNet block in extracting local features, and the components as showcased in Figure 5. This track processes input images with dimension 224 × 224× 3. The features extracted from this track alongside the Swin transformer track are further combined, processed, and then trained to achieve accurate results for pear leaf disease classification.

The HCFPNet incorporates a sequence of alternatively ordered blocks and each block performs distinct operations on the input features. The layers utilized include 2D convolution, MBConv, and batch normalization. The final step of this track uses a BiFPN which fuses extracted features for better generalizability and adaptiveness. Features extracted from the central segment of the track are merged with the output of the final MBConv block. This process combines low-level details with high-level representations. Following this, the network parameters are combined using the BiFPN block.

The MBConv block is a key component of CNNs designed to enable efficient feature extraction while maintaining high performance. (Sandler et al., 2018). The MBConv block has a unique structure that includes an expansion layer, a depthwise separable convolution, and a projection layer. This design allows for a substantial decrease in the number of parameter counts while preserving a high level of accuracy, making it suitable for devices where computational resources are limited (Poulose et al., 2022; Pan et al., 2022). The MBConv architecture, shown in Figure 6 begins with an expansion layer designed to increase the dimensionality of the input feature maps. Let the input feature map have dimensions H × W × C_in, where H represents height, W represents Width, and C_in represents the number of input channels. The expansion layer applies a 1 × 1 convolution with an expansion ratio r, increasing the number of channels to r × C_in, giving a feature map with dimensions H × W × (r × C_in). This expanded feature map is normalized and passed through an activation function to introduce non-linearity. Subsequently, depthwise convolution layers apply a separate filter to each input channel, capturing spatial features while maintaining low computational costs. The depthwise convolution maintains the spatial dimensions H × W and the number of channels to r × C_in. The output is normalized and activated and the feature map moves to the Squeeze and Excitation block. A sequence of dense layers performs channel recalibration: the first layer reduces the channels by applying ReLU activation and the second layer expands them back with sigmoid activation. This recalibrated channel-wise attention vector is then multiplied with the feature map, enhancing important features and suppressing less relevant ones.

In the final stage, a pointwise convolution with a 1 × 1 kernel reduces the channel dimensions from r × C_in to the target output channels C_out, resulting in a feature map of dimensions H × W × C_out. This ensures that the block outputs refined feature representations while maintaining the required dimensionality. A residual connection is employed to add the input directly to the output, provided the input and output shapes match and the stride is one. This improves gradient flow and model stability. This efficient design enables MBConv blocks to balance computational efficiency with model performance, making them ideal for resource-constrained environments. By utilizing depthwise separable convolutions, squeeze-and-excitation for channel attention, and residual connections, MBConv reduces computational overhead while preserving the ability to capture complex features.

BiFPN is an architecture designed to enhance feature representation in various computer vision tasks. BiFPN improves upon traditional Feature Pyramid Networks (FPN) by introducing a bidirectional flow of information, allowing for more effective multi-scale feature fusion. This is achieved through a series of weighted connections that facilitate the integration of features from different resolutions, thus enabling the model to leverage both high-level semantic information and low-level details simultaneously (Tan et al., 2020).

The BiFPN block consists of multiple layers that combine features from different scales through depth-wise separable convolutions. The input to the block consists of a list of feature maps with different spatial resolutions, denoted as H₁ × W₁ × C₁ and H₂ × W₂ × C₂. Once the feature maps are dimensionally aligned, they are summed together to merge the information from both resolutions, resulting in a feature map with dimensions H₁ × W₁ × C, where C is the resulting channel dimension after fusion. To enhance the fused features, the block applies two depthwise separable convolutions. Each convolution applies a 3 × 3 filter, succeeded by batch normalization and ReLU activation. The key advantage of using BiFPN over other feature fusion blocks is its efficiency in computational resource usage while maintaining high accuracy. BiFPN is designed to be resource-optimized, which renders it appropriate for applications where computational power is limited.

3.3 Coordinate attention module

The coordinate attention block is designed to enhance feature representation in CNN by integrating spatial and channel attention more effectively (Hou et al., 2021). The primary focus of the Coordinate Attention Block is to allow the network to concentrate on essential details while suppressing insignificant data, thereby addressing the issue of information overload that can hinder model performance. As illustrated in Figure 7, the coordinate attention block processes the input feature map through a structured sequence of operations to separate spatial and channel information. The input feature map is subjected to two parallel global pooling operations based on its spatial dimensions. Here, X represents the width, and Y represents the Height. X-average pooling aggregates information along the width to produce a height-encoded feature map with dimensions C × H × 1. Y-average pooling aggregates information along the width to produce a height-encoded feature map with dimensions C × 1 × W. These pooled outputs isolate spatial information specific to each axis and enable the block to capture distinct positional cues from the input. The outputs are then concatenated along the channel axis giving a feature map of size C/r × 1× (W + H), where r represents the reduction ratio that controls the dimensionality of the compressed representation. These features are processed through the 2D convolutional layer to blend the extracted spatial context. The output is then passed through a batch normalization layer and a non-linear activation to enhance feature discrimination. The activated feature map is split into two paths, each leading to separate 2D convolutional layers that reconstruct the coordinate-specific attention maps. One path generates an attention map for the height dimension C× H× 1, and the other produces an attention map for the width dimension C × 1 × W. These attention maps are then activated using sigmoid functions to scale their values between 0 and 1, ensuring that the attention weights can modulate the input appropriately. Finally, the coordinate attention maps fused with the original input feature map. The attention map generated for the height dimension is broadcasted and multiplied across the height, while the width-specific attention map is broadcasted across the width. This re-weighting mechanism refines the input feature map by emphasizing spatially significant channels and suppressing less relevant ones.

4 Results and discussions

The following section provides a brief overview of the dataset involved in this study. It also highlights the data augmentation techniques implemented. It further outlines the experimental setup and the ablation studies conducted to measure the influence of each component on its performance.

4.1 Dataset description

This study utilized the DiaMOS Plant dataset comprising pear leaf images (Fenu and Malloci, 2021b). The dataset contains a total of 3,006 images distributed across four classes: curl, healthy, spot, and slug as depicted in Figure 8.

4.2 Data augmentation

This step involves creating additional samples to expand the dataset, leading to improved accuracy in classifying leaf diseases. It has proven to be effective in boosting classification performance, particularly when dealing with limited or imbalanced datasets. The data preparation begins by splitting the dataset into training, testing, and validation sets, following a 60:20:20 ratio. Following this, data augmentation techniques were applied to the training and validation sets, which included (1) random horizontal flipping, (2) random vertical flipping, and (3) random 90-degree image rotation. These transformations were applied on-the-fly during training, generating different augment version in each epoch.

4.3 Environment setup

Experiments involving the proposed network utilized a cloud-based infrastructure. This setup ensured consistent computational resources throughout the process. The primary hardware for model training included an NVIDIA Tesla P100 GPU with 16 GB of memory and 29 GB of system RAM. This configuration enabled the efficient handling of deep learning workloads and large model architectures.

4.4 Training configuration and hyperparameter tuning

Hyperparameter tuning improves architecture correctness, generalization, and convergence. It boosts model performance without changing its architecture. Table 1 summarizes the hyperparameter search space and optimal values determined through systematic experimentation. The dropout rates in each dropout layer, along with the learning rate and gradient update optimizer are adjusted to improve the model. Dropout rates were varied from 0.4 to 0.7 in increments of 0.1, adjustments to the learning rate were made from 0.00001 to 0.01 in multiples of 10, and the gradient optimizers Adam, SGD, and AdamW were employed to evaluate the model. The best dropout rate was 0.5, which regularizes the model to reduce overfitting by randomly deactivating neurons during training and adding noise to develop more generalizable representations. The best convergence was achieved with Adam, with a learning rate of 0.0001. To improve model correctness by minimizing learning rates during training, a ReduceLROnPlateau scheduler with factor 0.5, patience 3, and a minimum learning rate of 0.000001 was used.

TABLE 1. Hyperparameter tuning results of the proposed network on the Diamos dataset.

4.5 Ablation studies

This section summarizes the ablation studies conducted to determine how each component of the model affects the performance of the proposed architecture. The focus of this study was to find the most effective arrangement of model components for improved performance. Table 2 summarizes the ablation studies that assess the contribution of every component of the proposed network. The effectiveness of the proposed network is assessed through multiple evaluation metrics, such as accuracy, precision, recall, and F1 score. Furthermore, the accuracy and loss curves were generated to elucidate the learning progression of each network. It is important to note that certain training and validation curves exhibit temporary fluctuations during the initial training epochs. These variations are attributed to early-stage optimization instability, where rapid parameter updates and evolving feature representations may lead to short-term oscillations in validation accuracy and loss. As training progresses, the models demonstrate stable convergence and consistent performance improvement.

TABLE 2. Summary of the ablation studies to validate the effectiveness of the proposed network.

4.5.1 Analysis of the five-layered convolution neural network

In this section, we analyze the performance of a basic five-layer CNN architecture. The model consists of five consecutive 2D convolutional layers that capture spatial features from the input image through convolution operations. It is then passed to global average pooling and fully connected layers to classify the image. The five-layered CNN was trained for 50 epochs using ReduceLROnPlateau and Adam optimizer. The observations recorded are illustrated in Figure 9. The five-layered CNN architecture yielded an accuracy of 82.88%.

4.5.2 Analysis of the CNN integrated with the MBConv module

This section examines the effectiveness of using MBConv blocks to construct a CNN track to capture local features. The input image is processed through an initial 2D convolutional and batch normalization layer followed by five consecutive MBConv blocks. Each MBConv block is followed by a batch normalization layer except the final block. The initial convolution layer reduces the spatial dimensions of the images and the batch normalization layers are used to stabilize training. The normalized output is passed through a sequence of MBConv blocks with increasing channel depth. These blocks are essential for computational efficiency while maintaining high accuracy. Each MBConv block is subjected to depthwise separable convolutions to decrease the number of parameters and computations in comparison to standard convolutions. The model was trained for 50 epochs using ReduceLROnPlateau scheduler and Adam optimizer. The observations recorded are illustrated in Figure 10. The baseline CNN integrated with MBConv resulted in an accuracy of 83.45%.

4.5.3 Analysis of the hybrid convolution feature pyramid network (HCFPNet)

In this section, we examine the performance of the HCFPNet. The HCFPNet is a custom CNN track comprising MBConv and BiFPN blocks. Five MBConv blocks are sequentially arranged and the outputs of the third and the fifth block are passed to the BiFPN block. The BiFPN block combines the feature maps of the MBConv blocks. This allows the network to leverage both fine-grained and coarse-grained information, leading to more accurate predictions. A global average pooling layer aggregates the spatial dimensions of the feature map obtained from the BiFPN block. The HCFPNet track was trained for 50 epochs using ReduceLROnPlateau and Adam optimizer. The observations recorded are depicted in Figure 11. The HCFPNet architecture yielded a classification accuracy of 85.68%.

4.5.4 Analysis of the swin transformer track

In this section, the efficacy of the Swin transformer track is evaluated. Input images, divided into tokens, are successively processed by Swin transformer blocks. These blocks subsequently execute self-attention and layer normalization to extract local and global feature representations from the input patches. The four subsequent phases improve the hierarchical depiction of the input image. The output is first processed through global average pooling, then flattened, and finally undergoes a linear transformation. Ultimately, it is sent to the classifier for identifying the diseases. The Swin transformer underwent training on the dataset for 50 epochs utilizing the ReduceLROnPlateau scheduler and the Adam optimizer. The documented observations are presented in Figure 12. The Swin transformer track yielded an accuracy of 84.86%.

4.5.5 Analysis of the integrated dual track network without coordinate attention

In this section, the performance of the proposed network without incorporating coordinate attention is examined. The feature maps extracted by the Swin transformer and HCFPNet tracks are concatenated to combine local and global representations. These concatenated features are processed through global average pooling and fully connected layers. The model was trained over the course of 50 epochs using ReduceLROnPlateau scheduler and Adam optimizer. The documented observations are presented in Figure 13. This dual-track architecture resulted in a classification accuracy of 84.21%.

4.5.6 Analysis of the proposed SwinPyramidNet

The following section analyzes the performance of the proposed architecture. The proposed architecture is a dual-track network comprising a Swin transformer track and an HCFPNet track. The Swin transformer track comprises four Swin blocks. These blocks extract global features of the input image. The HCFPNet track focused on local feature extraction and multiscale feature fusion. The extracted features from each track are fused and directed to the coordinate attention block. The coordinate attention block enhances important spatial regions, enabling the model to concentrate on important regions of the image for classification. The model was trained over a course of 50 epochs using ReduceLROnPlateau and Adam optimizer. The documented observations are presented in Figure 14. The evaluation metrics obtained on training the dual branch model were as follows: accuracy of 92.69%, precision of 93.02%, recall of 92.69%, and F1 score of 92.78%. Figure 15 illustrate the confusion matrix and ROC curve showcasing the performance of the proposed model.

4.6 Performance analysis

This section presents a comparative analysis of the proposed model’s performance with the existing methods in classifying pear leaf deficiencies. To ensure a fair and accurate assessment, studies conducted on the same dataset were exclusively considered. This approach enabled a direct and equitable comparison of model capabilities under identical conditions. By limiting the comparison to studies with consistent data, we aim to highlight the proposed model’s strengths. This approach ensures that any performance improvements can be attributed to model design rather than variations in data.

4.6.1 Comparison with the state-of-the-art networks

A comparison was conducted using widely adopted CNNs for image classification: Xception, ResNet50, VGG16, MobileNetV2, and DenseNet121. Each of these architectures is known for unique design features: Xception leverages depthwise separable convolutions; ResNet50 uses residual connections to prevent gradient vanishing; VGG16 is known for its depth and simplicity; MobileNetV2 is optimized for efficiency on mobile devices; and DenseNet121 utilizes dense connections for feature reuse. Table 3 presents the accuracy, precision, recall, and F1 score of each model. The results indicate that the proposed network outperformed these existing CNN architectures across all metrics.

TABLE 3. Quantitative performance comparison of the proposed model with other state-of-the-art CNN architectures on the DiaMOS plant dataset.

Transformers exhibit greater flexibility compared to CNNs. They utilize self-attention mechanisms to effectively capture global relationships and long-range dependencies within image Table 4 provides a comparative analysis of the model’s performance against prominent transformer architectures. Vision Transformer represents one of the earliest architectures to process images as a sequence of patches. Data-Efficient Image Transformer is optimized for performance with limited training data. Transformer-in-Transformer hierarchically models fine-grained features within image patches. Pyramid Vision Transformer employs a hierarchical structure to capture multi-scale features and the Swin transformer incorporates a shifted window approach to capture global dependencies. The findings demonstrate that the proposed network outperforms current sophisticated transformer designs on all performance parameters.

TABLE 4. Quantitative performance comparison of the proposed model with state-of-the-art transformers on the DiaMOS plant dataset.

4.6.2 Performance analysis with the existing works

Table 5 compares the proposed model with the existing works that utilize the DiaMOS dataset for classifying pear leaf diseases. The following metrics were utilized to analyze and compare performance with previous studies: accuracy, precision, recall, and F1 score. Previous studies predominantly emphasize the use of pre-trained models and their ensembles. In contrast, the proposed network presents a dual-track architecture that combines the strengths of transformers, CNNs, and attention mechanisms. By effectively integrating these components, the model is capable of capturing global and local features with improved performance. The results show that networks proposed by other studies achieved accuracies ranging from 86% to 91%. The proposed network achieved an accuracy of 92.69% improving on the existing results of previous studies.

TABLE 5. Performance analysis of the proposed study with existing works on the DiaMOS plant dataset.

4.7 Evaluation on the external datasets

This section evaluates the performance of the proposed network on two external datasets: PV_ALE: Apple leaf disease classification dataset and PlantVillage datasets. Table 6 summarizes the results obtained on the external datasets.

TABLE 6. Performance of the proposed network on external datasets.

4.7.1 Evaluating with the PlantVillage dataset

PlantVillage is one of the most extensively used public benchmarks in plant disease detection, encompassing a wide range of crop species and disease classes (Hughes and Salathe, 2015). To validate cross-dataset transferability, we evaluated the proposed model on the PlantVillage dataset. The model achieved an accuracy of 80.50%, a precision of 80.73%, a recall of 80.50%, and an F1-score of 80.42%, as summarized in Table 6. The moderate performance relative to within-distribution results is attributable to distributional shift in imaging conditions and class compositions. The results indicate meaningful generalization capability across a large-scale multi-class plant disease dataset.

4.7.2 Evaluation on the PV_ALE dataset

The PV_ALE (PlantVillage Apple Leaf Extended) dataset extends PlantVillage’s apple leaf images with 3,383 images across six classes—Alternaria Leaf Spot, Apple Scab, Black Rot, Cedar Apple Rust, Healthy, and Powdery Mildew—introducing greater diversity and diagnostic complexity (Akinyemi and Adebayo, 2025). As reported in Table 6, our model achieved an accuracy of 94.27%, a precision of 92.00%, a recall of 94.27%, and an F1-score of 93.11%. The stronger performance relative to PlantVillage is likely attributable to the domain-specific nature of PV_ALE, where the narrower focus on apple leaf diseases aligns more closely with the fine-grained feature representations learned by the proposed architecture.

4.8 Limitations and future work

The following section outlines few limitations and potential scope for future research of the proposed network.

• The dataset utilized in this research lacks a representation of environmental factors. This reduces the model’s effectiveness in practical applications where these factors impact disease presentation. The future works need to focus on working with diverse datasets collected from different sources representing different environmental settings.
• Further research in this area can concentrate on lowering the model’s computational complexity. This will accelerate the deployment of the model on resource-constrained devices for real-time plant disease classification.
• A potential direction for future research involves enhancing the model’s architecture. Additionally, improvements in learning strategies may be pursued. The integration of domain adaptation techniques with cross-domain feature learning methods can enhance the model’s usability across diverse datasets and practical applications.

5 Conclusion

Advancements in technology have enhanced agricultural practices, leading to better crop management and increased productivity. The diseases affecting pear can cause significant damage to the crop yield and quality. Traditional methods rely on labor-intensive visual inspection of symptoms, which leads to delayed treatment and potential crop loss. Therefore, it demands the timely detection of leaf diseases to protect fruit quality and prevent substantial economic losses. Deep learning techniques have emerged as a promising alternative for the classification of plant leaf diseases. The proposed network uses a dual-track architecture to enhance pear leaf disease classification. The Swin transformer track extract global features and long-term dependencies from the input images. The HCFPNet track is designed to capture local features and multi-scale features. The coordinate attention mechanism concatenates and refines the feature maps obtained from both the tracks. It enhances the model’s ability to focus on spatially relevant details across multiple channels by encoding precise positional information. The network was trained on the DiaMOS Plant dataset achieving an accuracy of 92.69%. The results obtained indicate that the proposed network outperforms the existing deep learning architectures, demonstrating its effectiveness in classifying pear leaf diseases.

Generative AI statement

The author(s) declared that generative AI was used in the creation of this manuscript. Artificial intelligence (AI) tools, including OpenAI’s ChatGPT, were used to support the preparation of this manuscript. These tools assisted in tasks such as language refinement, grammar correction, and content structuring. All intellectual content, data interpretation, and final decisions were made solely by the authors, who take full responsibility for the integrity and accuracy of the work presented.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Hicham Karmouni, Sidi Mohamed Ben Abdellah University, Morocco

Reviewed by: Bo Hu, Nanchang University, China

Sivaraman Eswaran, Curtin University Sarawak, Malaysia