MSCR-FuResNet: A Three-Residual Network Fusion Model Based on Multi-Scale Feature Extraction and Enhanced Channel Spatial Features for Close-Range Apple Leaf Diseases Classification under Optimal Conditions

Abstract

The precise and automated diagnosis of apple leaf diseases is essential for maximizing apple yield and advancing agricultural development. Despite the widespread utilization of deep learning techniques, several challenges persist: (1) the presence of small disease spots on apple leaves poses difficulties for models to capture intricate features; (2) the high similarity among different types of apple leaf diseases complicates their differentiation; and (3) images with complex backgrounds often exhibit low contrast, thereby reducing classification accuracy. To tackle these challenges, we propose a three-residual fusion network known as MSCR-FuResNet (Fusion of Multi-scale Feature Extraction and Enhancements of Channels and Residual Blocks Net), which consists of three sub-networks: (1) enhancing detailed feature extraction through multi-scale feature extraction; (2) improving the discrimination of similar features by suppressing insignificant channels and pixels; and (3) increasing low-contrast feature extraction by modifying the activation function and residual blocks. The model was validated with a comprehensive dataset from public repositories, including Plant Village and Baidu Flying Paddle. Various data augmentation techniques were employed to address class imbalance. Experimental results demonstrate that the proposed model outperforms ResNet-50 with an accuracy of 97.27% on the constructed dataset, indicating significant advancements in apple leaf disease recognition.

1. Introduction

China is the largest apple producer in the world. According to the Food and Agriculture Organization of the United Nations (FAOSTAT) statistics database [1], China’s apple production reached 47.57 million tons in 2022. As a result, apples hold significant economic value in the agricultural sector. However, the prevalence of apple diseases poses a major challenge to apple production, significantly reducing yield [2]. Leaves are among the primary indicators for identifying apple diseases. Therefore, accurate and rapid identification of apple leaf diseases is crucial not only for distinguishing among various disease types but also for minimizing the manual effort and related costs associated with disease scouting. However, the application of deep learning techniques to the identification of apple leaf diseases faces three significant challenges: (1) the small lesion areas on apple leaves make it difficult for models to capture detailed lesion features; (2) the high morphological similarity between different types of apple leaf diseases, particularly among certain fungal infections, complicates the differentiation process; and (3) images of apple leaf lesions taken in complex backgrounds often exhibit low contrast, which negatively impacts classification accuracy. These challenges significantly impede both research on the classification of apple leaf diseases and their practical application.

The accurate classification of apple leaf diseases is crucial for developing effective disease prevention and control measures. Traditional identification methods primarily depend on experts in the relevant fields who classify apple leaf diseases based on their experience [3]. While these methods can achieve high accuracy, they are susceptible to errors due to subjective judgment and individual differences. Consequently, spectral measurement techniques have been introduced to provide an objective diagnosis of apple leaf diseases [4]. Despite their objectivity, these techniques have drawbacks, including cumbersome equipment, slow recognition speed, high costs, and low efficiency. With the widespread use of smartphones and digital cameras, computer-based automatic classification of apple leaf diseases has been widely applied. This technological advancement not only improves the efficiency of disease diagnosis but also reduces the workload of researchers and agricultural workers, thereby offering somewhat notable research value and practical relevance.

To address these issues, many scholars have employed machine learning techniques to identify plant diseases and pests. For instance, Prasad et al. [5] proposed using the Gabor wavelet transform to extract texture information from the segmented diseased parts of plant leaves and classify them. Jian et al. [6] developed an improved Support Vector Machine (SVM) algorithm to classify cucumber leaf diseases using small samples. Liu et al. [7] introduced an enhanced Random Forest algorithm to extract color and texture features from diseased leaf areas, facilitating the classification of sunflower leaf diseases. Vaishnnave et al. [8] utilized the k-nearest neighbor (KNN) classifier to classify peanut leaf disease images. Shi et al. [9] proposed an apple disease identification method based on two-dimensional subspace learning dimensionality reduction (2DSLDR). They mapped observed sample points from high-dimensional space to low-dimensional subspace to achieve optimal classification features.

Traditional image processing methods primarily rely on expert experience to extract and handle features such as texture, color, structure, and shape. Machine learning approaches, on the other hand, typically utilize predefined datasets where disease characteristics are manually selected. Furthermore, the disease images used in these studies are often captured in controlled laboratory settings with optimal lighting and simple backgrounds. However, in real-world scenarios, images of apple leaf diseases present complex and variable backgrounds with lesions of different sizes, making them challenging to identify using conventional algorithms. Consequently, the disease detection models developed through these traditional methods lack generalizability and exhibit poor transferability.

In recent years, with the advent of deep learning [10] technology, convolutional neural networks (CNNs) have become a popular research focus in the automated identification of plant diseases and pests. For example, Liu et al. [2] utilized an improved AlexNet deep convolutional neural network to detect apple leaf diseases, achieving an accuracy of 97.62%. This method offers higher accuracy, faster processing speed, and reduced time consumption compared with traditional methods. Yan et al. [11] proposed an enhanced VGG16 model incorporating global average pooling and batch normalization layers to identify apple leaf diseases, resulting in a 6.3% improvement in recognition accuracy and a reduction in training time to 0.56% of the original duration. The application of deep learning significantly enhances efficiency over traditional methods. Gao et al. [12] introduced BAM-Net, which employs an aggregate coordinate attention mechanism (ACAM) to boost the network’s focus on disease-specific regions, particularly in complex backgrounds, reaching an accuracy of 95.62%. Additionally, Bi et al. [13] presented a solution using the lightweight MobileNet model to implement the apple leaf disease identification system on mobile devices, achieving low cost, stable performance, and high accuracy.

Although the aforementioned research has largely addressed the limitations of traditional methods in handling small features, poor model generalizability, and the inability to process images with complex backgrounds, several challenges remain. Apple leaf lesions are typically small, and there is a high similarity between different disease types, which leads to low model recognition accuracy. Furthermore, the accuracy of identifying apple leaf images in complex backgrounds is still insufficient for practical application. Currently, the attention mechanism has become a key method for enhancing network performance and is a significant research focus [14,15,16,17,18,19,20]. The attention mechanism aids deep convolutional neural networks by suppressing less important pixels or channels, thus boosting performance. A notable approach for improving network performance through the suppression of irrelevant weights is the Squeeze-and-Excitation Networks (SENet) [15]. SENet enhances network performance by integrating spatial information into channel feature responses and computing attention via two Multilayer Perceptron (MLP) [21] layers. To further improve the network’s ability to extract channel and spatial features, the Convolutional Block Attention Module (CBAM) [17] has been introduced. CBAM strengthens the network’s feature extraction capabilities by aggregating features using both average pooling and max-pooling layers.

In addition to enhancing the network’s ability to classify apple leaf disease images through the attention mechanism, integrating a multi-scale feature fusion pyramid sub-network into the backbone network has also been shown to significantly improve performance [22,23,24]. The Single Shot Detector (SSD) [25] demonstrated the effectiveness of feature pyramids in capturing fine details within the network. Consequently, such feature pyramid networks have gained prominence as a focal point in research [26,27]. Moreover, the winning models in the ImageNet [28] and COCO [29] challenges successfully employed multi-scale feature fusion pyramid networks. These findings collectively indicate that applying feature pyramid networks can substantially enhance the capability of deep convolutional neural networks in classifying apple leaf disease images.

Fusion models represent a significant approach to enhancing the performance of networks in classifying apple leaf disease images. In competitions such as Kaggle [30], numerous entries have successfully employed model fusion, yielding impressive results. Fusion models integrate the strengths of various models that excel in particular domains, leading to optimal classification outcomes [31]. The efficacy of model fusion has been well-documented in numerous studies [18], which highlight its superior performance in experimental results.

The remainder of this paper is organized as follows. In Section 2, we introduce the data used for the experiments and describe the improvements made to the network methods in this study. Section 3 presents the experimental results and their analysis, which validate the superiority of MSCR-FuResNet. Section 4 provides the conclusions and outlines potential future work.

To address the challenges mentioned above, this paper introduces a three-residual network fusion model based on the ResNet-50 architecture, designed to accurately identify various types of apple leaf diseases. Moreover, we have constructed a new dataset with complex backgrounds by aggregating data related to apple leaf diseases from Plant Village [32], Baidu Flying Paddle [33,34], and various online sources. Training the model on this dataset is anticipated to significantly improve its generalization capability and robustness.

2. Dataset and Methodology

2.1. Dataset and Data Preprocessing

This paper presents an apple leaf disease image dataset containing a total of seven categories of images, including six different apple leaf diseases and images of healthy leaves. The categories are (a) Alternaria blotch (Alternaria mali), (b) brown spot (Marssonina coronaria), (c) gray spot (Didymella pomorum), (d) mosaic (caused by virus infections), (e) rust (Gymnosporangium juniperi-virginianae), and (f) scab (Venturia inaequalis) [35], as shown in Figure 1. These images are sourced from New Plant Village [32], Baidu Flying Paddle [33,34], and publicly available apple leaf disease datasets on the Internet. After a thorough filtering process, a total of 14,000 images were selected. The dataset was divided into training and test sets in a ratio of 80% to 20%, respectively.

The images of apple leaf diseases in the dataset can be categorized into two primary types. The first type consists of images with a simple background, typically captured in laboratory settings under favorable lighting conditions and without any background interference. The second type comprises images taken in actual production environments, such as orchards, where the backgrounds are often complex and contain significant interference. While these complex-background images can enhance the model’s robustness and generalization capabilities, they also introduce challenges to the training process. The presence of complex backgrounds can disrupt the model’s accuracy in recognizing the diseased leaves, even when the background is unrelated to the disease [36]. Moreover, such complexity can considerably extend the training time, leading to increased and unnecessary computational costs.

Considering the characteristics of the image background, this study employed the U-Net [37] neural network to segment the background and foreground of apple leaf disease images with complex backgrounds within the training set. The processed images, along with the generated weight files, were subsequently incorporated into the training set to enhance the dataset’s content. This approach helps reduce the influence of complex backgrounds and thereby improves the model’s accuracy, as illustrated in Figure 2 and Figure 3.

Additionally, the original dataset has issues with an uneven distribution of images for each type of apple leaf disease and low image diversity. These problems increase the difficulty of classification, making data preprocessing crucial [38]. Initially, the images are filtered, and redundant ones are removed. The redundant images refer to the duplicate images that were obtained from the New Plant Village and Baidu Flying Paddle datasets, and therefore, these duplicate images need to be deleted. Then, the dataset proportions are adjusted to a uniform level to allow the network model to extract features across different categories consistently. Finally, data augmentation is applied to the processed images to enhance their diversity. To prevent overfitting due to a limited number of images and to improve the network’s generalization capability, this study employs seven data augmentation methods: brightness enhancement, contrast enhancement, rotation, shearing, affine transformation, flipping, and HSV augmentation, resulting in a robust test dataset. The accuracy comparison of the model before and after image processing is provided in Section 3.5.

Each data augmentation method is applied to each image, thereby increasing the dataset size to seven times its original number of images (1600 training images and 400 test images for each disease and for healthy leaves). Potential anomalous samples in the dataset may negatively impact model performance. This study conducted preprocessing steps prior to inputting the images into the model. Initially, all images in the dataset were uniformly resized to 120 × 120 pixels. Subsequently, the images were converted to Tensor format, altering the data dimensions from height (H), width (W), channels (C) to channels (C), height (H), and width (W). Following this, pixel values were normalized by dividing by 255, standardizing the data within the range of [0, 1]. Finally, Z-Score normalization was applied to each channel of the images, as described by Equation (1).

$$A_i={\displaystyle \frac{{\alpha }_i-\overline{{\alpha }_i}}{{\sigma }_{{\alpha }_i}}},$$

(1)

In this formula, $A_i$ represents the standardized value of the i-th channel, $a_i$ denotes the original pixel value of this channel, $a_i$ indicates the mean pixel value of the channel, and ${\sigma }_{{\alpha }_i}$ signifies the standard deviation. Through this standardization process, we achieve image pixel data characterized by a zero mean and a unit variance. Consequently, the distribution of pixel values within each channel is normalized to have a mean of 0 and a standard deviation of 1. This normalization ensures that the data falls within the range of [−1, 1].

2.2. Apple Leaf Disease Classification Model

This study introduces MSAR-FuResNet, a fusion model combining three distinct residual networks based on ResNet-50. The model leverages multi-scale feature fusion and enhances features through both channel and spatial domains to improve the classification accuracy of apple leaf disease images. The fusion model comprises three sub-networks: Network A, which is a multi-scale feature fusion pyramid network; Network B, which integrates a mechanism for enhancing channel features and spatial features and then reinforces channel features again; and Network C, a residual network optimized with modified residual blocks and activation functions.

2.2.1. Model Structure

The texture and structural information differences among apple leaf disease images are minimal; therefore, achieving fine-grained classification of these diseases necessitates more sophisticated deep recognition techniques. This study involves modifying the ResNet-50 architecture to enhance its depth. The aim is to improve the performance of the three sub-networks within the fusion model, thereby enabling more accurate classification of apple leaf disease images.

The overall structure of the fusion model is illustrated in Figure 4. Initially, the dataset comprising images of apple leaf diseases undergoes data augmentation. Subsequently, these augmented images are processed by three sub-networks, labeled A, B, and C (as depicted in Figure 4). The final classification outcome for the apple leaf disease images is derived from the joint output of these three sub-networks.

To address the issue of varying classification difficulty among samples, this study employs the loss function specified in Equation (2). This loss function assigns weights to the samples based on the classification difficulty: it allocates a smaller weight ${\alpha }_1$ to easily classified samples and a larger weight ${\alpha }_2$ to samples that are difficult to classify.

$$L_{sum}={\alpha }_1\times L_1+{\alpha }_2\times L_2,$$

(2)

In this context, $L_1$ represents the loss function for easily classified samples, while $L_2$ represents the loss function for difficult-to-classify samples. Given that ${\alpha }_1$ is relatively smaller and ${\alpha }_2$ is relatively larger, $L_2$ becomes the dominant loss function. Consequently, the primary focus of the loss function is directed toward the difficult-to-classify samples. The corresponding expression for the loss function is given by Equation (3).

$$L_{\left(P_t\right)}=-{\left(1-P_t\right)}^{\beta }log\left(P_t\right),$$

(3)

Here, $P_t$ denotes the probability of positive samples, ${\left(1-P_t\right)}^{\beta }$ serves as the modulation factor, and $\beta$ acts as the tunable focusing parameter. As $\beta$ increases, the loss associated with easily classified samples decreases. This adjustment encourages the model to concentrate more on the harder-to-classify samples, thereby mitigating the issue of class imbalance in the dataset.

2.2.2. Multi-Scale Feature Fusion Mechanism of SubNetwork A

To tackle the challenge posed by the small lesion areas on apple leaves, we incorporated a multi-scale fusion pyramid into subnetwork A. This modification aims to enhance the model’s ability to capture the detailed features of the disease. The detailed architecture of the multi-scale fusion pyramid is illustrated in Figure 5.

This mechanism operates by applying convolutions with kernels of varying sizes to obtain feature maps at different scales. Smaller kernels are effective at capturing fine details in the image, which helps in accurately identifying small lesions. In contrast, larger kernels are better suited for capturing the overall structure of the image, making them more effective for recognizing larger objects. For a comprehensive description of the architecture of subnetwork A and the specific technical details of this mechanism, please refer to Appendix A.

2.2.3. Channel and Spatial Feature Enhancement and Subsequent Channel Feature En-Hancement Mechanism of SubNetwork B

To address the challenge of distinguishing apple leaf diseases due to their high visual similarity, we implemented a channel feature enhancement mechanism within subnetwork B, aimed at strengthening channel features after image compression. Following this, we developed mechanisms for spatial feature enhancement and channel feature re-enhancement at the end of the residual block to reduce the loss of channel-specific features. These two strategies help the model more effectively differentiate between various types of diseases.

Subnetwork B, shown in Figure 6, enhances the structure of image channel features, while Figure 7 illustrates the enhancement of spatial features and the re-enhancement of channel features.

As illustrated in Figure 6, when enhancing the features of image channels, the feature map first undergoes global average pooling. This operation transforms the feature map from a matrix of [H, W, C] (height (H), width (W), channels (C)) to a vector of [1, 1, C] (channels (C)). This approach not only avoids dimension reduction but also captures cross-channel interactions.

As illustrated in Figure 7, the upper section labeled A depicts the structure designed to enhance the spatial features of an image. Following the application of batch normalization (BN) [39] to the feature image, a scaling factor is computed. Subsequently, the weights for each spatial dimension are determined. These weights are then multiplied with the image that has been adjusted using the scaling factor, and the result is processed through a Sigmoid function to yield a spatially weighted feature map, denoted as Ms. The lower section labeled B in the figure represents the structure for enhancing image channel features, which is analogous to the structure of section A. The key difference lies in the fact that, in section B, the computation is performed on the channels rather than the spatial dimensions, resulting in a channel-weighted feature map, Mc.

For a comprehensive description of the architecture of subnetwork B and the specific technical details of this mechanism, please refer to Appendix B.

2.2.4. Optimization of Convolution Kernels and Activation Functions of SubNetwork C

Due to the low contrast of apple leaf spot images against complex backgrounds, which adversely affects classification accuracy, we have optimized the convolution kernels and activation functions in sub-network C to improve its sensitivity in detecting fine features. This enhancement effectively increases the classification accuracy. The underlying rationale of this optimization mechanism is detailed as follows.

As shown in Figure 8, configuration a represents the unmodified structure, while configuration b represents the modified structure. The residual block in configuration a follows the standard ResNet-50 structure. In contrast, configuration b introduces a 7 × 7 convolutional kernel in the first layer of the residual structure. This modification allows the residual block to retain more image information during the initial stages of processing, thereby reducing information loss.

In the residual block a, ResNet-50’s convolution sequence is 1 × 1 convolution → depthwise convolution → 1 × 1 convolution. For the residual block b, the sequence is altered to depthwise convolution → 1 × 1 convolution → 1 × 1 convolution, inspired by the design of the Swin Transformer [40], where the MSA [40] module precedes the MLP [21] module. This adjustment enhances model accuracy.

Furthermore, unlike the ReLU activation function used in Figure 8a, the residual blocks in subnetwork C employ the GELU activation function. The GELU activation function mitigates the neuron death issue that ReLU may cause during training when the activation input is negative, thereby improving gradient flow.

For a comprehensive description of the architecture of subnetwork C and the specific technical details of this mechanism, please refer to Appendix C.

2.2.5. Fusion Model

Although the three methods discussed above have been integrated into ResNet, each still presents certain limitations. Thus, there is a pressing need for an approach that synthesizes the strengths of these methods, effectively offsetting their weaknesses, to further improve network performance. To this end, we implemented a network fusion strategy. Research indicates that network fusion is a key approach for enhancing model performance. In fact, in Kaggle and other data science competitions [30], ensemble models consistently outperform individual models, demonstrating superior overall performance. By integrating the three improved sub-networks, A, B, and C, we aim to achieve a fusion model that outperforms individual sub-networks in the classification of apple leaf diseases.

Fusion models are used to enhance model accuracy. Within a fusion model, different sub-models might produce varying predictions for the same image due to their distinct focus areas determined by their design mechanisms and structures. The essence of a fusion model lies in selecting the best or most accurate prediction from these sub-models to serve as the final output. This study employs a fusion model strategy to optimize subnetworks, aiming to enhance the classification performance of apple leaf disease images within a deep learning framework. The fusion method used in this research is illustrated in Figure 9.

As illustrated in Figure 9, the image is fed into three sub-networks: A, B, and C. These sub-networks classify the image and yield their respective accuracies: ACC(A), ACC(B), and ACC(C). The fusion model compares these accuracies and, if they differ, it employs the highest accuracy among the subnetworks as the output of MSCR-FuResNet. In cases where two subnetworks have identical accuracies that differ from the third, the fusion model adopts the accuracy of the two matching subnetworks. This fusion strategy ensures that the optimal or most accurate prediction is retained when predicting the same image, thereby enhancing image recognition performance compared to using independent sub-models.

3. Experimental Results and Analysis

In all the experiments in this chapter, accuracy refers to the ratio of correct predictions to total predictions without considering whether the predictions are for positive or negative cases. As shown in Equation (4) below.

$$\mathrm{A}\mathrm{c}\mathrm{c}\mathrm{u}\mathrm{r}\mathrm{a}\mathrm{c}\mathrm{y}={\displaystyle \frac{\mathrm{T}\mathrm{P}+\mathrm{T}\mathrm{N}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P}+\mathrm{F}\mathrm{N}+\mathrm{T}\mathrm{N}}},$$

(4)

If an instance is positive and is predicted as positive, it is classified as a True Positive (TP). If an instance is negative and is predicted as negative, it is classified as a True Negative (TN). If an instance is negative and is predicted as positive, it is classified as a False Positive (FP). If an instance is positive and is predicted as negative, it is classified as a False Negative (FN).

3.1. Environmental Configuration and Experimental Setup

The methodology presented in this study utilizes the Pytorch 1.12.1 deep learning framework. Experiments were executed on a system equipped with an Intel Core i9-7960X CPU (Intel Corporation, Santa Clara, CA, USA) and an NVIDIA GeForce RTX3080Ti GPU (NVIDIA Corporate, Santa Clara, CA, USA). The operating system employed was Windows 10, version 22H2. For our experimental setup, the learning rate (RL) was configured to 0.0001, with a batch size of 16 and seven distinct sample categories. To optimize model performance, the number of epochs was set to 280. Detailed configuration parameters are provided in

Table 1. System configuration and experimental setup.

Number	Name	Description
1	Operating System	Windows 10 22H2 For WorkStation
2	Programming Language	Python 3.10
3	PyTorch Version	1.12.1
4	CUDA Version	12.0.76
5	CPU	Intel Core i9-7960X CPU @ 4.2GHz
6	GPU	NVIDIA GeForce RTX 3080 Ti
7	Memory	Crucial Ballistix BL16G32C16U4W.16FE 64 GB (Micron Technology, Inc., Boise, ID, USA)
8	Solid State Drive	SAMSUNG MZPLL3T2HAJQ-00005 3.2 TB (Samsung Semiconductor, Inc., Yongin, South Korea)
9	Motherboard	Asus TUF X299 Mark 2 (Asus Technology Co. Ltd., Shanghai, China)

.

3.2. Sub-Network Fusion Comparison Experiment

To verify the effectiveness of subnet fusion in this study, we employed ablation experiments. According to the characteristics of the fusion model, it must contain three or more subnets; fusion is not possible with only two subnets. Thus, our experiments are divided into four parts: Part A, a model with only subnet A; Part B, a model with only subnet B; Part C, a model with only subnet C; and Part D, a fusion model incorporating subnets A, B, and C. The experimental results are presented in Figure 10 and Figure 11.

The results of the subnetwork fusion comparison experiments are illustrated in Figure 10 and Figure 11. The fused model demonstrates a higher validation accuracy in classifying apple leaf disease images compared with individual subnetworks. Additionally, as shown in Figure 10B, the fused model also converges faster during training.

The highest classification accuracy achieved by the fused model was 97.27%. This was superior to the accuracy of subnetwork A at 96.36%, subnetwork B at 96.31%, and subnetwork C at 93.62%. These results underscore the enhanced effectiveness of sub-network fusion.

3.3. Ablation Study

To assess the impact of the modified network mechanisms on the performance of MSCR-FuResNet, we conducted a series of ablation experiments. As the fusion model must include three or more distinct subnetworks, the experiments were divided into four parts: Part A, Part B, Part C, and Part D. These are defined as follows: Part A integrates the improved subnetworks A, B, and C, forming MSCR-FuResNet; Part B combines the improved subnetworks A and B with the original, unimproved residual block network C; Part C merges subnetwork A, which lacks a multi-scale feature fusion mechanism, with the improved subnetworks B and C; and Part D fuses the improved subnetworks A and C with subnetwork B, which does not include spatial feature enhancement or the mechanism for secondary enhancement of channel features at the end of the residual block within the residual structure. The results of the ablation experiments are shown in Figure 12 and Figure 13.

As shown in Figure 12, the Part A fusion model achieves the best classification performance for apple leaf disease images and converges faster during training compared with the other fusion models (Part B, Part C, Part D) in the experiment. The ablation experiment results, shown in Figure 13, further illustrate this point. The accuracy of the Part B fusion model is 1.37% lower than that of the Part A model, indicating that subnetwork C, after processing with residual blocks, significantly impacts the fusion model. The Part C fusion model’s accuracy is 1.46% lower than the Part A model, highlighting the importance of the multi-scale feature fusion mechanism in classifying apple leaf disease images. The Part D fusion model has an accuracy that is 2.37% lower than the Part A model, showing that the spatial feature enhancement and secondary enhancement mechanism for channel features at the end of the residual block are beneficial for the fusion model’s classification performance on apple leaf disease images. The ablation experiments confirm the effectiveness of the mechanisms introduced in the modified network on the overall fusion model.

3.4. Comparison Experiment between Sub-Networks and the Original ResNet-50 Network

To ensure that each subnetwork that was improved based on the ResNet-50 network demonstrates higher accuracy in classifying apple leaf disease images compared with the original ResNet-50 network, we conducted comparative experiments. These experiments compared the performance of subnetworks A, B, and C against the original ResNet-50 network. The results, highlighting the comparative accuracy, are illustrated in Figure 14 and detailed in

Table 2. Comparative analysis of subnetworks and the original Resnet-50 network in terms of accuracy for apple leaf disease image classification.

Network Name	Comparison with Resnet-50	Accuracy
Subnetwork A	+4.59%	96.36%
Subnetwork B	+4.54%	96.31%
Subnetwork C	+1.85%	93.62%
Resnet-50	±0.00%	91.77%

.

The analysis indicates that subnetwork A increased the accuracy by 4.59%, subnetwork B by 4.54%, and subnetwork C by 1.85% over the original ResNet-50 network. Consequently, the modified subnetworks A, B, and C not only achieve higher accuracy in apple leaf disease recognition but also demonstrate faster convergence during training compared with the original ResNet-50 network.

3.5. Data Enhancement Experiment on Apple Leaf Disease Images

Due to the uneven distribution and limited diversity of the apple leaf disease image dataset, classifying these images has become more challenging. To address this issue, we applied seven data augmentation techniques: enhancing brightness, enhancing contrast, rotating images, applying shear transformations, using affine transformations, flipping images, and augmenting HSV data. Additionally, we utilized U-Net to process the regions of the images with varying levels of importance. Following these augmentations, our model’s accuracy in classifying apple leaf disease images significantly improved. The experimental results before and after data augmentation are illustrated in Figure 15 and Figure 16.

Figure 15 clearly demonstrates that both the model’s convergence speed and its classification accuracy for apple leaf disease images improved post-data augmentation compared with the pre-augmentation results. Figure 16 further corroborates these findings, indicating that the highest and average classification accuracies after data augmentation exceeded those before augmentation. Specifically, the highest accuracy improved by 1.49%, and the average accuracy improved by 1.77% after data augmentation. These experiments confirm that augmenting apple leaf disease images enhances the model’s classification accuracy.

3.6. Comparison Experiment of Network Methods

To evaluate the classification performance of the improved model in this study on the preprocessed apple leaf disease image dataset, we conducted comparative experiments using MSCR-FuResNet, conventional ResNet-50 [41], Vgg16 [42], MobilenetV2 [43], and Swin Transformer [44] networks. The average, highest, and lowest accuracies of these networks are presented in

Table 3. Comparison of average, highest, and lowest network accuracy for apple leaf disease image classification.

Model Name	Average Accuracy	Max Accuracy	Min Accuracy
MSCR-FuResNet (Ours)	96.13%	97.27%	95.45%
Resnet-50 [41]	90.04%	91.77%	89.03%
Vgg16 [42]	90.11%	90.82%	88.61%
Mobilenet V2 [43]	91.82%	93.64%	90.09%
Swim Transformer [44]	93.17%	94.52%	92.86%

. The results confirm that the improved, optimized model developed in this study demonstrates superior performance in classifying apple leaf disease images. Detailed experimental results are illustrated in Figure 17.

Moreover, to conduct a more comprehensive comparison of MSAR-FuResNet’s performance, we applied our model to the apple leaf disease dataset provided by New Plant Village. On this dataset, our model achieved an average accuracy of 98.97%, with a peak accuracy of 99.35% and an average loss of 0.073. Compared with a 2023 study [45] that also classified apple leaf diseases using the New Plant Village dataset [32], the accuracy of our proposed model improved by 1.19%, demonstrating relatively enhanced classification performance and generalization capability.

As illustrated in Figure 18, the results of the network method comparison experiments indicate that our model outperforms other networks in terms of both convergence speed and the accuracy of classifying apple leaf disease images. In Figure 18, the network accuracy results show that our model surpasses others in average, maximum, and minimum accuracy. The closer the blue, red, and yellow lines in the figure are to the outer black line, the higher the model’s accuracy. These findings confirm that our improved model exhibits superior classification performance on the preprocessed apple leaf disease image dataset compared with other networks, showcasing enhanced competitiveness.

4. Conclusions and Future Directions

4.1. Conclusions

In order to facilitate the rapid identification of apple leaf diseases and thereby enhance apple production, we introduce MSAR-FuResNet. This model is a fusion of three-residual networks based on ResNet-50, specifically designed to address the limitations of traditional methods, which often struggle with low accuracy in disease detection, particularly in complex backgrounds. To further improve the model’s performance, we restructured the ResNet-50 backbone, integrating the benefits of multi-scale feature fusion with enhanced channel spatial features, which boosts the extraction of similar features while minimizing the risk of feature loss.

Experimental results demonstrate that the proposed MSAR-FuResNet model achieves impressive recognition performance. Trained and tested on a processed, self-constructed apple leaf disease dataset, the model attained a classification accuracy of 97.27%. When evaluated on the apple leaf disease dataset provided by New Plant Village, our model achieved an average accuracy of 98.97%. In comparison with other approaches, our model not only maintains high accuracy but also significantly improves the ability to recognize apple leaf diseases in complex backgrounds.

The primary contribution of this study is its capability to automatically and efficiently detect apple leaf diseases in natural environments. This reduces the need for manual labor and associated costs in disease identification, thereby enhancing the efficiency of agricultural informatics in the research of apple leaf diseases.

4.2. Future Directions

Our model has shown excellent classification performance, generalization capabilities, and robustness on the apple leaf disease dataset. However, there is still room for improvement in the current network architecture. Future research will focus on the following three aspects:

• Model Efficiency: The suitability of the model is closely linked to its size and the number of parameters. Future work will aim to develop lightweight models that maintain high performance while being easily deployable on resource-constrained mobile platforms.
• Dataset Enhancement: Increasing the diversity and complexity of the dataset by incorporating images that depict multiple types of apple leaf diseases within a single image.
• Algorithm Optimization: Enhancing the underlying algorithms to ensure that the model operates with greater speed and lower complexity. This will involve developing more efficient algorithms to improve the model’s computational efficiency.