This study conducts experiments on two datasets: a public dataset and a self-constructed dataset with natural environmental backgrounds. The specifics of these datasets are detailed in Sections 3.1.1 and 3.1.2, respectively, while the data preprocessing process is explained in Section 3.1.3.

The experimental hardware in this study uses an Intel(R) Core(TM) i7–6700 CPU @ 3.40GHz processor, and the operating system is Windows 10. Model training and testing are accelerated using a GPU, specifically a Tesla V100S-PCIE-32GB graphics card. The software environment includes Python 3.8.19, the PyTorch 2.4.1 framework, and the CUDA toolkit 12.4.

The number of iterations is set to Epoch=60, with a batch size of 32. The model training uses the Stochastic Gradient Descent (SGD) algorithm, which is one of the most commonly used optimization methods in machine learning. The initial learning rate for SGD is set to 0.0001. A cosine annealing schedule is applied for the first 50 epochs, gradually decreasing the learning rate from 1e-4 to 1e-6, and the learning rate is kept constant at 1e-6 from epoch 50 to 60.

In this study, six evaluation metrics are employed to assess the performance of the proposed model: Accuracy, Precision, Recall, F1-Score, Kappa, and Matthews Correlation Coefficient (MCC). Their definitions are as follows: Accuracy refers to the ratio of correctly predicted samples to the total number of samples. Precision is the proportion of correctly predicted positive samples among all samples predicted as positive. Recall represents the proportion of correctly predicted positive samples among all actual positive samples. F1-Score is the harmonic mean of Precision and Recall. Kappa measures the agreement between the model’s predictions and the ground truth, while accounting for agreement occurring by chance, thus providing a more objective evaluation of classification performance. MCC evaluates the overall performance of a classification model, particularly suitable for handling imbalanced class distributions. The mathematical formulations of Accuracy, Precision, Recall, F1-Score, Kappa, and MCC are defined as follows (see Equations 1–6).

TP means the model correctly predicts a positive sample. TN means the model correctly predicts a negative sample. FP means the model wrongly predicts a negative sample as positive. FN means the model wrongly predicts a positive sample as negative. p ₀ represents the observed proportion of agreement, i.e., the percentage of instances where the raters reach consensus across all samples. p_e  represents the expected agreement by chance, assuming that the raters classify independently and randomly. In multi-class classification, macro and micro averages are common for evaluation. This study adopts macro average to compute Recall and Precision, and uses the global average value for Accuracy. For example, the macro-averaged Recall is calculated as shown in Equation 7.

To validate the effectiveness of the proposed method, ablation experiments are conducted on the SCEBD dataset. LCAMNet is an improved architecture based on the ShuffleNetV2 model (Zhang et al., 2018). The overall network structure is shown in Table 3 .

To verify the effectiveness of each proposed module, multiple comparative models are designed for experimentation. The baseline model is ShuffleNetV2. Model1 modifies only the downsampling module (SDSM) of ShuffleNetV2. Model2 modifies only the feature extraction module (SFEM). Model3 modifies both the feature extraction and downsampling modules based on ShuffleNetV2. Model4 further incorporates the improved triplet attention mechanism on top of Model3. The detailed configurations of the networks using different strategies are shown in Table 4 .

The experimental results are shown in Table 5 . GFLOPs measures the floating-point operations (in billions) and is used to evaluate the model’s computational complexity. Parameters refer to the number of trainable weights in the model and are commonly used to assess the model’s size. As seen in Table 5 , Model1 outperforms the baseline, indicating that applying DBDM in the downsampling process of apple leaf disease images better preserves feature information compared to the original SDSM. Model2 also shows better performance than the baseline on the test set, suggesting that MFEM is more effective in extracting features from apple leaf disease images in natural environmental backgrounds than SFEM. Model3 performs better than both Model1 and Model2 on the test set, demonstrating that the combined application of MFEM and DBDM has a synergistic enhancement effect on apple leaf disease image classification, surpassing the performance of each method individually. Model4 shows further improvement over Model3. The addition of the enhanced triplet attention mechanism slightly increases the number of parameters but leads to a significant boost in performance. Additionally, compared to ShuffleNetV2, Model4 reduces both floating-point operations and parameter count while achieving a larger improvement in classification performance. In conclusion, through the stepwise introduction of DBDM, MFEM, and the improved triplet attention mechanism, the classification performance of the model continues to improve, validating the effectiveness of the proposed method.

To further validate the effectiveness of the LCAMNet model in disease region identification, this study employs the Grad-CAM (Selvaraju et al., 2017) technique to visualize and analyze the model’s prediction results, comparing the performance before and after the improvements, as shown in Figure 9 . Figure 9A presents the original image, Figure 9B shows the visualization result of the baseline model, and Figure 9C shows the visualization result of the LCAMNet model. Taking a rust disease image as an example, the baseline model focuses only on partial features of the diseased area, leading to missed and false detections, and fails to fully capture the lesion regions. In contrast, the LCAMNet model accurately localizes the key diseased regions of the apple leaf, with more comprehensive and discriminative feature extraction. These results demonstrate that the integration of the MFEM module, DBDM module, and the improved triplet attention mechanism in LCAMNet significantly enhances the model’s ability to focus on diseased areas, thereby improving its feature learning capacity and classification accuracy.

In this experiment, we compare LCAMNet with several classic convolutional neural network models, including VGG (Simonyan and Zisserman, 2014), ResNet (He et al., 2016), ResNext (Xie et al., 2017), and DenseNet (Huang et al., 2017), on SCEBD. The experimental results are shown in Table 6 . These models exhibit issues when applied to disease classification tasks in natural environments, which are manifested in the following aspects: First, while VGG network has a simple structure, it suffers from a large number of parameters due to its deep architecture, making it prone to overfitting. As a result, its performance is relatively poor in the scenarios with small datasets or complex backgrounds. Second, ResNet and ResNext use residual connections to address the vanishing gradient problem. However, due to the deep network layers, they still struggle to effectively capture features in the presence of high noise and irregular lesions. DenseNet, despite having advantages in feature propagation, has a dense connectivity structure that leads to a large number of parameters and computational overhead, making it inefficient when processing high-resolution images. Thus, these four networks suffer from low computational efficiency and high model complexity.

GoogLeNet (Szegedy et al., 2015) designs the Inception module, which applies multiple convolutional kernels of different sizes in parallel to capture features at various scales. This improves both feature representation and computational performance. Although InceptionResNetV2 (Szegedy et al., 2016) incorporates optimization techniques such as multi-branch convolutions, residual connections, and depthwise separable convolutions to improve performance, the network’s complexity and large number of parameters limit its advantages in some application scenarios.

EfficientNetB0 (Tan and Le, 2020), ShuffleNetV2 (Zhang et al., 2018), MobileNetV2 (Howard et al., 2017), and GhostNet (Han et al., 2020) are lightweight models that have been widely used in recent years. These models optimize the network structure to reduce computational overhead while maintaining high accuracy. They incorporate various lightweight techniques, such as depthwise separable convolutions, channel reparameterization, and residual connections, achieving a good balance between computational efficiency and model accuracy. However, these models may face performance bottlenecks when handling more complex tasks due to limited model capacity and expressive power.

LCAMNet, based on ShuffleNetV2 model, optimizes the feature extraction and downsampling modules. By using depthwise separable convolutions and structural reparameterization, it reduces network parameters while still extracting sufficient features. Compared to classic convolutional neural networks, LCAMNet has fewer parameters than ShuffleNetV2 and achieves the highest accuracy. This demonstrates that LCAMNet is a high-performance, lightweight model for apple leaf disease classification. The performance comparison of LCAMNet and several classical CNN models on SCEBD is shown in Table 6 . The three-dimensional visual analysis of accuracy, GFLOPs, and parameter count on SCEBD is presented in Figure 10 .

This study compares the performance of LCAMNet with similar crop disease image classification models on the SCEBD. Re-GoogLeNet (Yang et al., 2023) is a network designed for rice image classification in natural environmental backgrounds, based on a series of improvements to GoogLeNet. First, the 7×7 convolution kernel in the first layer of GoogLeNet is replaced with three consecutive 3×3 convolutions. Then, the Inception module is enhanced by adding the ECA attention mechanism and optimized residual connections to strengthen information flow. Lastly, LeakyReLU replaces the ReLU activation to better capture irregular features in diseased leaves. ALS-Net (Liu et al., 2022a) adds an Inception structure after 3×3 convolution in the ShuffleNetV2 model for multi-scale feature extraction. Additionally, the 3×3 convolutions in the ShuffleNet block are replaced with 5×5 depthwise convolutions to obtain the ALS module. The ELU activation function also replaces ReLU, addressing gradient vanishing and neuron death issues. To further improve classification performance, ALS-Net uses DenseNet161 as a teacher network to guide training and enhance the model’s classification ability. LBMRNet (Li et al., 2023) is a lightweight algorithm for recognizing tomato leaf diseases, aimed at tackling the problem of significant variation within the same class and minimal variation between different classes in tomato leaf disease images. LBMRNet consists of alternating complementary group dilation residual (CGDR) modules and visual enhancement modules. The CGDR module uses a multi-branch design to capture various features of tomato leaf diseases from different receptive fields. It incorporates multiple residual connections to facilitate better information flow across the network layers. The visual enhancement module combines average pooling, max pooling, and 1×1 convolutions as downsampling strategies, effectively fusing the visual enhancement effects and preventing information loss during the downsampling process, thus improving classification accuracy.

From Table 7 , it could be seen that the performance of ALS-Net and LBMRNet are inferior to LCAMNet (for LBMRNet, only the network structure is restored). This is likely because, although ALS-Net improves ShuffleNetV2, the multi-scale feature extraction is only performed at the initial stages of the network and fails to effectively integrate these features at each stage, limiting its feature extraction capability. Although LBMRNet improves the feature extraction and downsampling layers, its multi-residual structure leads to the continuous retention of irrelevant information, affecting the recognition performance. In comparison, Re-GoogLeNet optimizes GoogLeNet by introducing the attention mechanism to preserve important features and uses residual connections to retain original features. While its performance is better than that of LCAMNet, its higher FLOPs and Parameters may limit its applicability in resource-constrained scenarios. The confusion matrices of the models are shown in Figure 11 .

This study compares the performance of LCAMNet with several classical CNN models on the FGVC8 dataset to verify its effectiveness and superiority. Detailed experimental settings are provided in Section 4.1. Table 8 presents a performance comparison between LCAMNet and several classical CNN models on FGVC8. The results show that LCAMNet achieves a classification accuracy comparable to GoogLeNet and InceptionResNetV2, while significantly reducing FLOPs and Parameters compared to these two models. This indicates that LCAMNet maintains a high recognition accuracy while lowering computational resource consumption. In addition, LCAMNet achieves better classification accuracy than ShuffleNetV2 on the public dataset, while maintaining slightly lower FLOPs and parameters. This demonstrates that LCAMNet effectively integrates techniques like multi-branch feature extraction modules, dual-branch downsampling modules, and attention mechanisms into the ShuffleNetV2 architecture, making it suitable for different datasets. Three-dimensional visual analysis of accuracy, GFLOPs, and parameters on FGVC8 is presented in Figure 12 .

Furthermore, LCAMNet is compared with similar crop disease image classification models. Experimental results show that LCAMNet outperforms ALS-Net and LBMRNet in overall performance. Although its performance is slightly lower than Re-GoogLeNet, LCAMNet has significantly fewer FLOPs and Parameters, demonstrating a notable optimization in computational resource consumption. In conclusion, the experimental results fully validate the effectiveness of LCAMNet on the public dataset. LCAMNet achieves competitive recognition accuracy while significantly reducing computational complexity, showcasing strong practical application value and potential for broader deployment. Detailed results can be found in Table 9 , and the models’ confusion matrices are shown in Figure 13 .

Although LCAMNet achieves outstanding performance on both the SCEBD and FGVC8 datasets, several limitations remain based on the experimental results:

As shown in Table 6 , LCAMNet demonstrates a favorable balance between accuracy and model complexity compared to various CNN models on the SCEBD dataset. However, in the comparative experiments with other crop disease identification models Table 7 , although LCAMNet outperforms ALS-Net and LBMRNet, its accuracy is slightly lower than that of Re-GoogLeNet. This indicates that while LCAMNet holds significant advantages in terms of parameter count and computational cost, there is still room for improvement in feature fusion and deep representation capabilities. Moreover, the staged multi-scale attention mechanism present in Re-GoogLeNet significantly enhances performance, suggesting that LCAMNet could benefit from incorporating richer stage-wise fusion strategies for further optimization.

Furthermore, LCAMNet maintains competitive performance on the FGVC8 dataset (as shown in Tables 8 , 9 ), indicating its generalization capability. However, some classification confusion still occurs among certain categories in FGVC8, as evidenced by the confusion matrix in Figure 13 . This reveals a risk of misclassification, especially when dealing with disease symptoms that have similar morphological patterns and subtle color differences. These observations suggest that LCAMNet could further enhance its fine-grained modeling capacity for complex lesion patterns.

Publicly available datasets were analyzed in this study. This data can be found here: https://www.kaggle.com/competitions/plant-pathology-2021-fgvc8; https://github.com/JasonYangCode/AppleLeaf9; https://www.scidb.cn/en/detail?dataSetId=0e1f57004db842f99668d82183afd578.