The network architecture of our model(MobileNet-MFS) is shown in Figure 1A . The design of the model inherits the main modules of MobileNet v3, but in order to obtain better diagnostic efficacy, many modifications were also made to the model. The main body of the model is consistent with MobileNet v3, which consists of a two-dimensional convolutional layer, a series of bottleneck layers with different dimensions, a two-dimensional convolutional layer, a pooling layer, and a one-dimensional convolutional layer in sequence. Through this series of modules, feature information on plant disease-affected areas is extracted, and diseases are classified into 9 types through 1×1 convolution. However, at the front end of the model, in order to further explore the feature information that cannot be captured in the original MobileNet v3, we introduced a multi-scale feature extraction module. The most important change is that we have proposed a new FSCA attention mechanism to replace the SE attention mechanism module used in MobileNeT v3. The FSCA mechanism will be explained in detail in the following chapters.

As shown in Figure 1B , in MobileNet-MFS, the most basic module is the bottleneck layer, which is composed of an inverted residual network containing depthwise separated convolutions. It replaces the standard convolution operation with a depthwise convolution followed by a pointwise convolution. This reduces computational complexity and model size while maintaining accuracy. In addition to depthwise separated convolution, the bottleneck layer also includes expansion convolution, which mainly serves to increase the number of channels in the input feature map using a 1x1-sized convolutional kernel. Projection convolution is a 1x1 convolutional kernel with a significantly smaller number of output channels than the input channels, thus achieving the goal of limiting the size of the model. When the input and output channels are the same, a residual network can be used. The bottleneck layer of the inverted residual structure formed by the above convolution operations is finally activated using ReLU or h-swish function.

Although CNN is very powerful in image expression, they are deficient in expressing spatial information. Therefore, the attention mechanism has been introduced into MobileNet v3, which can improve the learning ability of the model by assigning weights to images. In the original version of MobileNet v3, the SE attention module is placed in the middle of the bottleneck layer, Hu et al. (2020) giving an updated set of weight values through two fully connection layers and the activation function. However, the SE attention module only cares about the dependencies between channels and ignores location information, which is crucial for generating spatially selective attention maps. Therefore, we propose our FSCA attention mechanism to replace the SE module.

The FSCA attention mechanism considers both spatial and channel information of the input layer, thus more effectively guiding the model to focus on effective positions in the image. As shown in Figure 2 , the FSCA attention mechanism consists of two concatenated modules. The first module mainly focuses on aggregating features in the spatial directions of X and Y. By averaging pooling in the X and Y directions and performing concat operation, a 1 × (H+W) × C dimensional array is obtained. Furthermore, we normalized the array through convolution, separated it, and activated it with a sigmoid function to obtain a set of weights containing information in the X and Y directions. Afterward, the weights are multiplied with the original data to obtain a set of directional perception feature layers. These transformations allow the attention module to capture long-term dependencies along one spatial direction and preserve precise positional information along another spatial direction, which helps the network locate interested targets more accurately.

The second module focuses on channel attention. In this module, we will take the maximum and average values of the input feature layers on the channels of each feature point. Afterward, we stack these two values and adjust the number of channels using a convolution with a channel count of 1. Then, we take a sigmoid function and obtain the weights of each feature point in the input feature layer (between 0 and 1). After obtaining this weight, we multiply it by the original input feature layer.

By concatenating and multiplying the two steps, we obtain our FSCA attention mechanism, which focuses on both the X and Y dimensions of input and the fusion of information of channels. Therefore, the obtained results are more comprehensive. Since our attention mechanism fused both spatial and channel information, we named it FSCA attention mechanism, which references CBAM Woo et al. (2018) and CA Hou et al. (2021) attention mechanism. In the following experiments, it was demonstrated that the FSCA mechanism helped our model better identify the characteristics of apple leaf diseases.

For apple leaf diseases, there are two main characteristics that are not easily extracted by machines. One is that there is a significant difference in the size of the disease on the leaves, such as Powdery Mill Draw and Grey spot lesions. Another type of disease is that its color or other details may vary depending on the scope of the disease, such as Grey spot and Rust lesions.

The above features involve dimensions of different sizes and are not easily captured by MobileNet V3, which mainly uses 3 × 3 and 5 × 5 convolution operations. In order to enable the machine to capture more features from different dimensions, Li et al. (2020) we have added a multi-scale feature extraction module to the front end of the input layer.

The structure of this module is shown in Figure 3 . Four dimensions of convolution: 1 × 1, 3 × 3, 5 × 5, and 7 × 7 were applied in the module. After the image is convoluted, it is merged into a new feature map and then placed ahead of the network. Through such feature extraction, the accuracy of disease classification was improved.

The images of apple leaves were collected from both laboratory and outdoor environments, with a total of eight diseases. These leaves were divided into nine categories, and each photo was labeled with the disease type. Our data mainly comes from PlantVillage, PPCD2020, PPCD2021, and ATLDSD datasets. PlantVillage is mainly from laboratory environments, while images from the PPCD2020 and PPCD2021 are collected in natural environments. The total number of samples is 15250, including 12204 for the training set and 3046 for the testing set. The sample ratio for the training and testing sets is 4:1.

As shown in Figure 4 , there are a total of eight apple diseases in our sample, namely Alternaria leaf spot, Brown spot, Frogeye leaf spot, Grey spot, Mosaic, Powdery Mildew, Rust and Scab. The number of samples was collected in Table 1 . Both Brown spot and Mosaic form large spots on the leaves, but the former will first cause the diseased parts of the leaves to turn yellow in a large area. Powdery Mildew can turn the veins of the leaves white and stain the leaves with white spots. Many other plants also suffer from similar diseases, such as strawberries. Other diseases can cause various types of spots on the leaves, such as Rust causing red spots on the leaves, while Gray spots causing gray spots, and Frogeyes causing yellow-brown spots on the center, similar to those on the outer ring of a frog’s eye. In order to distinguish these different types of spots, neural networks need to first be able to capture these spots and further distinguish the different features of color and shape in the spots.

Accuracy is the most commonly used indicator, which represents the proportion of the true value of a model in the overall population. However, measuring the quality of a model cannot be solely based on accuracy. Some other indicators also reflect the quality of the model. For example, precision focuses on the model’s ability to avoid false positives, while recall focuses on the model’s ability to identify all positive instances. At the same time, when the dataset of the model is imbalanced, the f1-score balances the results of recall and precision, which better reflects the advantages and disadvantages of the model. The area under curve (AUC) shows the trade-off between the true positive rate and the false positive rate. Higher AUC values indicate better discriminability of the model. Therefore, accuracy is used with other performance metrics like precision, recall, f1-Score, and AUC. The definition of accuracy is:

where TN = true negative, FN = false negative, TP = true positive, and FP = false positive.

The expression of precision, recall, and f1-score are equations (2–4), respectively.

The accuracy and loss values of the model are shown in Figure 5 . By analyzing the images, we can conclude that the accuracy of training and testing has improved to over 97% after 20 epochs. For the training data, the loss is around 0.5, while for the test data, the loss stabilizes below 0.1 after 20 epochs. When epochs approach 80, the model achieved a maximum accuracy of 98.7%.

The confusion matrix of the experiment is shown in Figure 6 , where the horizontal and vertical coordinates represent the disease predicted by the model and the real disease respectively. Therefore, when the prediction is consistent with the actual situation, the axis data of the matrix will be added by one. When the predicted disease is inconsistent with the actual disease, the increased value of the matrix appears in the nondiagonal region. Take ‘Rust’ as an example, 534 cases of Rust were accurately identified, but 4 cases were misdiagnosed as Frogeye, 2 cases were misdiagnosed as health, and 10 cases were misdiagnosed as Scab. The 10 misdiagnosed cases were also the most common in the model, due to the similarity in size and color between rust and scab. Next, we want to further modify the model to better distinguish between the two diseases.

We have depicted the Receiver Operating Characteristic (ROC) curve of each disease, as shown in Figure 7 . It should be noted that the true positive rate of various diseases is high, resulting in a very steep ROC curve. The curve of Gery spot is different from several other diseases, as it initially reaches around 0.95. When the false positive rate reaches over 0.6, the true positive rate further increases to over 0.98. The steep ROC curve shows that the model can distinguish various diseases very well. In contrast, the ROC of general models is only diagonal.

In order to visually display the impact of different attention mechanisms, we calculated and compared the accuracy of different attention mechanisms (SE, ECA, CBAM, CA, FSCA, MFS) within the MobileNet v3 framework. As shown in Figure 8 , our proposed FSCA attention mechanism and combined multi-scale MFS attention mechanism grow rapidly with epochs but are slightly slower compared to other types. But when the epoch increases to 20, their stability and maximum value are the best. In contrast, the fluctuation amplitude of other attention mechanisms is relatively large, while the accuracy of the MFS and the FSCA mechanism fluctuates at the highest point, demonstrating special stability.

The accuracy of different CNNs and MobileNet-MFS are also compared. As shown in Figure 9 , the light gray curve represents the accuracy curve of MobileNet-MFS. Compared with other models, it also rises very quickly and gradually reaches its high-level platform after 20 epochs. At the 28th epoch, MobileNet-MFS has an accuracy of around 98%, which is better than other models at the same epoch. Finally, when the epoch reaches 75, the MobileNet-MFS reaches its maximum accuracy of 98.7%, surpassing all other models.

In order to comprehensively compare our model with other classic models, we calculated several indicators such as precision, recall, f1 score, and AUC. These indicators can measure the model’s capabilities from different aspects. From Table 2 , we can note that the MobileNet-MFS has the highest metrics in precision, recall and f1-score. However, in terms of AUC, it is not as good as a group of models such as EfficientNet-B0 and MobileNet-VIT.

Regarding the comparison of model performance, in addition to the above indicators, it is also necessary to consider the computational resources used by the models. MobileNet-MFS is based on MobileNet v3 and belongs to a lightweight CNN. The lightweight of the model will help it be applied to a wider range of scenarios. In addition, the computational complexity of the model is also a very important indicator, and the FLOPs provide an effective method to measure the computational complexity of the model. The indicators provided in Table 3 help us measure various aspects of the model more comprehensively. Taking into account parameter counts, memory size, and FLOPs counts, The MobileNet-MFS has more advantages over EfficientNet-B0, ResNet-34, and DenseNet-121, consuming slightly more computing resources than MobileNet v3, but not as streamlined as ShuffleNet v2.

In summary, through the comparison of various indicators, parameter quantities, and computational complexity, we can conclude that although many excellent models have emerged for image classification, MobileNet-MFS is still a state-of-the-art model.