The dataset used for the study was the Auburn Soybean Disease Image Dataset (ASDID), which is publicly available and has been extensively studied and validated (Bevers et al., 2022). The dataset contains 9648 soybean leaf images, of which 4,181 were taken in 2020 and 5,467 in 2021. The dataset contains eight main categories: bacterial blight ( Figure 1A ), cercospora leaf blight ( Figure 1B ), downey mildew ( Figure 1C ), frogeye leaf spot ( Figure 1D ), healthy ( Figure 1E ), potassium deficiency ( Figure 1F ), soybean rust ( Figure 1G ), target spot ( Figure 1H ). The ASDID dataset were used to construct soybean disease identification model.

The original 9648 images were randomly divided in the ratio of 9:1, of which 8686 were used for model training and validation, and 962 were used for model testing. Table 1 showed the division of the data according to the 9:1 ratio. Figure 1 showed examples of soybean leaves photographed for all eight disease categories.

In addition, in order to verify the adaptability and local application effect of the constructed model, 164 images of four categories of soybean leaves were collected at Dayuzhuang, Lanqing Township, Zhengyang County, Zhumadian City, Henan Province (32027’ 37’’N, 11420’ 11’’E). There were 48 images of bacterial blight, 52 images of frogeye leaf spot disease, 35 images of healthy leaves and 29 images of soybean rust. Soybean leaf images were collected on August 22, 2020. After completing the construction and training of the MFEF-DCNet model, the collected images of soybean leaves were input into the trained MFEF-DCNet model to test the application effect of the model in practice.

As shown in Table 1 , the number of healthy samples used for training and validation was 1469, while the number of bacterial blight was 436. The ratio of the number of healthy samples to the number of bacterial blight samples was greater than 3:1, indicating that the distribution of samples shown a serious imbalance. In CNN, imbalance in sample distribution may lead to overfitting problems in network models. In addition, sample data with differences between different categories could severely affect the accuracy of network training and limit network performance (Buda et al., 2018). Therefore, a data balancing algorithm was introduced before using the data for training (Gao et al., 2021).

The above sampling method was used to balance the sample data, which enhanced the generalization ability of the network. The method was as follows:

(1) Index building: Firstly, the expression of the datasets was defined as D = { [ M 0 , M l , M 2 , ⋯ M i ] [ N 0 , N l , N 2 , ⋯ N i ] T } , in which M i represented the type of soybean leaf disease, and N i was the number of samples which corresponded to M i . Then the maximum number N i of samples in the datasets and its corresponding M i were counted. Finally, the maximum value N i was expressed by N m a x , while M i { i ∈ [ 0 , 7 ] } marked the type which corresponded to N m a x , in which i expressed the label serial number (Label 0 indicated bacterial blight, Label 1 indicated cercospora leaf blight, Label 2 indicated downey mildew, Label 3 indicated frogeye leaf spot, Label 4 indicated healthy, Label 5 indicated potassium deficient, Label 6 indicated soybean rust, Label 7 indicated target spot).

(2) Calculation of the proportionality constant C: N m a x was chosen as the numerator and N i was selected as the denominator. The constant C was calculated according to Equation 1.

(3) Up-sampling: the proportionality factor C = { C 0 , C 1 ⋯ C i } was obtained from Equation 1. Smaller numbers of samples were up-sampled using different methods based on the proportionality constant. Data balancing was achieved by complex up-sampling, which is implemented by taking different up-sampling measures according to the size of the proportionality factor C i , as shown in Equation 2:

(4) Exportation: D ′ = { [ M 0 , M 1 , M 2 , · · · · M i ] [ N 0 ' , N 1 ' , N 2 ' , · · · · N i ' , ] T } was the balanced output, where   N i ' was calculated from Equation 3.

In the calculated datasets, D={[M₀ indicated bacterial spot, M₁ indicated caecilian leaf blight, M₂ indicated downy mildew, M₃ indicated frogeye leaf spot, M₄ indicated healthy, M₅ indicated potassium deficiency, M₆ indicated rust, M₇ indicated target spot][N₀ = 436, N_l = 1439, N₂ = 587, N₃ = 1386, N₄ = 1469, N₅ = 931, N₆ = 1465, N₇ = 973]^T}, the sample number of bacterial blight was the lowest while healthy sample number leaf sample was the highest.

Since N max was 1469, C 0 was equal to 3 according to Equation 1, 2. That means 3 was calculated by dividing 1469 by 436 using the floor principle. The bacterial blight dataset was balanced by an increase of 200% in the number of bacterial blight samples through rotation and contrast adjustments.

According to the above method, the proportionality constants for each disease type sample were calculated based on the data in Table 2 . On the basis of Equation 3, N 0 " , N 1 " , N 2 " , N 3 " , N 4 " , N 5 " , N 6 " , N 7 " were obtained as follows:

Table 2 displayed the sample distribution of each dataset type after data balancing. As shown in Table 2 , the maximum number of healthy leaves after data balancing was 1469 and the minimum number of potassium deficient leaves after data balancing was 931. The ratio of the two sample datasets was approximately 1:1.57, which indicated a more balanced distribution of sample data.

When the number of images in each category was roughly the equal, the model could better learn the features of each category and avoid overfitting or underfitting of some categories.

Image enhancement was a conventional data preprocessing method used to improve inter-class parity and increase the size of training samples. This method was accomplished by employing various transformations such as rotating, scaling, reflecting, adjusting brightness, and adding noise to the original image. ImageDataGenerator in Keras was able to enhance the data in a fast way (Abadi et al., 2016). Therefore, in the study, ImageDataGenerator was used to perform shift, brightness change, flip, rotate, and normalization operations on the data in the training set and normalization operations on the data in the validation set. Figure 2 showed the effect comparison before and after image enhancement of soybean leaves. Figure 2A showed the original image of soybean leaf with bacterial blight; Figure 2B showed the original image of soybean leaf with cercospora leaf blight; Figure 2C showed the original image of soybean leaf with downey mildew; Figure 2D showed the original image of soybean healthy leaf; Figure 2E showed the image obtained after flip for original image of soybean leaf with bacterial blight; Figure 2F showed the image obtained after rotation for original image of soybean leaf with cercospora leaf blight; Figure 2G showed the image obtained after moving vertically for original image of soybean leaf with downey mildew; Figure 2H showed the image obtained after moving horizontally for original image of soybean healthy leaf.

Table 3 displayed the training results and model parameter information based on the MFEF-DCNet, MobileNetV3_small, EfficientNetB0, VGG16, DenseNet201, Xception, and ResNet50 models. The training of all these models was trained from the scratch. In the table, the total parameters were the sum of trainable and non-trainable parameters, which was a reference indicator of the model capacity. The non-trainable parameters were usually set when the model was built and remained constant throughout the training process. During the training process, the trainable parameters were adjusted according to the training data to enable the model to fit the data better. It could be seen from the table that MobileNetV3_small had the lowest single epoch training time among all models. However, the final performance of MobileNetV3_small was poor with an accuracy of 0.9013. Therefore, MobileNetV3_small was not able to meet the practical requirements. Although MFEF-DCNet was longer than MobileNetV3_small in terms of training time for a single epoch, the model achieves a test precision of 0.947 due to the well-designed and optimized structure of MFEF-DCNet. In addition, although the MFEF-DCNet model was the most lightweight with a minimum total parameter of 1,200,104, the values of precision, recall and F1 score in MFEF-DCNet were exceed 0.95. Therefore, the comprehensive performance of MFEF-DCNet was the best among the all mentioned models.

The accuracy curves comparison for the training process of all models were depicted in Figure 5 . By analyzing the accuracy curve of MobileNetV3_small, it can be found that the training accuracy was increasing at the beginning. After 30 epochs, the training accuracy tended to stabilize. However, the final accuracy of MobileNetV3_small was low, which cannot satisfy the requirements of real production environment. In addition, it can be observed that MFEF-DCNet shown the fastest convergence rate in the first 10 epochs from Figure 5 . Meanwhile, the MFEF-DCNet model also achieved excellent results in the later stages of training with the accuracy around 0.9551. The loss value curves for different models were shown in Figure 6 . The loss values shown in the vertical coordinates in Figure 6 indicated the difference between the model recognized outputs and the actual labels. A smaller loss value means a better performance of the model, while a larger loss value means a larger difference between the model recognized results and the real results. As can be seen from the training loss curve, MFEF-DCNet decreased the fastest in the first 5 epochs, and the loss value reached its minimum in the final stage of model training. After 100 Epochs of training, MFEF-DCNet had an accuracy of 0.9551 on the training set.

After 100 epochs of training from scratch, the best trained weights were selected for all models and these weight coefficients were tested on a total of 962 test images that the models had never trained before. The test results of different models for different soybean leaf disease categories were shown in Table 4 . It was evident from the table that the proposed MFEF-DCNet performs better than other models in the vast majority of recognition results. And the accuracy, macro accuracy, and weighted accuracy of the MFEF-DCNet model all greater than 0.94. In terms of the recognition results for the different soybean leaf disease categories, the MFEF-DCNet model had relatively poor recognition precision for the health category. In terms of recall, the MFEF-DCNet model had relatively poor recognition accuracy for frogeye leaf spot, and the model was able to correctly recognize 139 out of 154 disease images with a recognition recall of 0.9026. While the MFEF-DCNet model performed the best for downey mildew and potassium deficiency. All images of the two categories for soybean leaf diseases were correctly recognized with a recognition precision of 1.0000. In addition, the MFEF-DCNet model showed a small difference in recognition results between the different disease categories of soybean leaves. This indicated that the model achieved good performance for each category without overfitting or underfitting. Among the 962 test images, the MFEF-DCNet model correctly recognized a total of 911 images eight categories with an accuracy of 0.947. This result was the best among the all models used in this study. Figure 7 showed the confusion matrix drawn by the MFEF-DCNet model after testing on 962 images. The numbers on the diagonal of the confusion matrix represented the counts that the recognitions of the model matched the true labels. From the confusion matrix, it was evident that the proposed MFEF-DCNet model achieved satisfactory results in all categories of test images.

The ablation experiments were conducted as shown in Table 5 . As can be seen from the table, the standard convolutional model had 6.66 million parameters, with an accuracy rate of 91.79% and an F1-score of 91.97%. Using the standard convolutional model as a baseline, when the CBAM module was introduced separately, the number of model parameters increased by 0.09M, yet the accuracy and F1-score improved by 0.52% and 0.55%, respectively. When replacing the standard convolution module with a separable convolution module, the number of model parameters was significantly reduced by 83.4%, with the parameters of 1.11 million. Meanwhile, model accuracy and F1-score improved by 2.08% and 1.92%, respectively. When the separable convolution and CBAM modules were introduced simultaneously, the model parameters reduced by 82.0%, while the model performance was the best, with an accuracy and F1-score of 94.70% and 94.73%, respectively. The accuracy and F1-score were improved by 2.91% and 2.76%, respectively. The above results indicated that separable convolutions had a good effect on model parameter compression, while the CBAM module can further enhance the model’s recognition capabilities through feature enhancement. Therefore, the MFEF module constructed by combining separable convolution and CBAM could significantly improve the recognition performance of the model, which provided technical support for the deployment of lightweight disease recognition models in practical scenarios.

In order to further verify the adaptability and actual application effect of the MFEF-DCNet model, the 164 local images of soybean leaves, which were collected in Dayuzhuang, Lanqing Township, Zhengyang County, Zhumadian City, Henan Province, were input into the MFEF-DCNet model. The recognition results of the model were shown in Table 6 . As shown in the table, the average accuracy of the MFEF-DCNet model for four categories of soybean leaves was 0.9024, and the macro-mean and weighted mean of the model were around 0.9. The results indicated that the MFEF-DCNet model had a satisfactory recognition effect for the local soybean leaves. At the same time, it was shown that MFEF-DCNet had better application effect in practical applications. It also showed that the MFEF-DCNet model had a favorable application effect in practical applications.

Feature map visualization was a technique for understanding deep learning models, which is useful for observing the intermediate results produced by the model while processing the input data. By visualizing the feature map, the model’s ability to interpret the input data and extract features could be quantified and demonstrated, which in turn contributed to optimizing the performance of the model. Figures 8 and 9 illustrate the process of visualizing the feature map of downy mildew using MFEF-DCNet. Figure 8 showed the image of downy mildew as input to the neural network, and Figure 9A , Figure 9B and Figure 9C showed the feature maps of downy mildew at different depths of the model during the forward process [Early stage ( Figure 9A ), Mid stage( Figure 9B ), Last stage ( Figure 9C )]. According to the feature maps of different periods, it can be seen that the shallow information extracted by the model is mostly edges and textures. With the depth of the network, what MFEF-DCNet learns became more and more abstract, and the extracted features became richer and richer. However, the visual interpretability of these features was also getting decreasing.

By mapping the visualized feature map to the original image, the change of the network focus during the network training process could be observed. Figures 10A – C were the heat maps of different stages of the model by mapping the feature maps after visualization to the original image [Early stage ( Figure 10A ), Mid stage ( Figure 10B ), Last stage ( Figure 10C )]. From the figures, we can see that the attention of the network gradually approaches towards the direction of the disease features during the learning process of the neural network.

According to the heat maps of different periods, it can be seen that the focus of the model was gradually approaching towards the direction of downy mildew disease characteristics in the process of neural network learning.