Edge-cloud collaborative rice disease monitoring method and system for precision agriculture

Abstract

The invention relates to the technical field of image classification, in particular to an edge-cloud collaborative rice disease monitoring method and system for precision agriculture, and the method comprises the steps: constructing a rice disease recognition network; a squeezing-incentive attention module and a multi-scale convolution-space attention module are introduced into a backbone network and a neck network of the rice disease recognition network; replacing standard convolution in the backbone network and the neck network with deep separable convolution; training the improved and optimized rice disease recognition network to obtain a lightweight student model; soft label distillation loss is constructed based on prediction distribution output by a pre-trained deep teacher model and a student model; a total loss function is constructed in combination with cross entropy loss, training of student models is guided through a back propagation algorithm, and a rice disease lightweight recognition model is obtained and used for recognizing rice diseases. Through edge end deployment of a lightweight model obtained through distillation, rapid reasoning of a high-precision model is realized on low-power-consumption edge equipment.

Description

Technical Field

[0001] This application relates to the field of image classification technology, and more specifically, to an edge-cloud collaborative method and system for monitoring rice diseases for precision agriculture. Background Technology

[0002] Rice, as one of my country's main food crops, is susceptible to various diseases during its growth. Early lesions often appear as small targets, making them difficult to detect and identify in a timely manner, which seriously affects rice yield and quality. Traditional rice disease control methods mainly rely on manual inspections and field experience, which suffer from low efficiency, strong subjectivity, limited monitoring scope, and untimely response, especially a high rate of missed detection for early small lesions.

[0003] With the development of technologies such as the Internet of Things, artificial intelligence, and computer vision, automatic disease identification using cameras and AI models has become possible. However, conventional target detection models cannot effectively identify small targets such as early-stage crop lesions. With the improvement of computer hardware performance, deep learning, with its superior feature self-learning capabilities, is now widely used in the field of crop disease identification. The effectiveness of deep learning-based plant disease identification depends on whether the model can learn rich and useful features during training. Although complex network structures can extract rich image features and achieve high identification accuracy, the high model complexity and large number of parameters typically require significant computing resources during training and inference phases, such as high-performance workstations or servers equipped with powerful GPUs, making deployment difficult in power-constrained agricultural scenarios.

[0004] In response, some scholars have begun to focus on research into lightweight models for target recognition. Patent application number 202311060349.X proposes a crop disease detection method and system based on a lightweight improvement of YOLOv8s. It replaces the C2f module in the backbone network with the FasterNext module, introduces a lightweight convolution Ghostconv to replace the standard convolution in the backbone network, and removes the last target detection head in the neck network to reduce the number of model parameters while maintaining detection accuracy. This method has been applied to the detection of wheat scab. However, while lightweight optimization of deep learning models can reduce model complexity and improve inference speed for deployment on edge platforms, the monitoring and data viewing methods for edge devices are not convenient enough. They lack user-friendly interfaces and convenient local monitoring methods for modern users. Furthermore, rice diseases are diverse, with varying lesion sizes and shapes, and different diseases share similar characteristics. This necessitates achieving effective lightweighting of deep learning models while maintaining more sensitive feature extraction capabilities.

[0005] Therefore, how to improve the model's ability to extract features from strong phenotypic features of rice diseases, and achieve lightweighting of the model while ensuring recognition accuracy, has become an urgent technical problem to be solved. Summary of the Invention

[0006] In view of this, this application provides an edge-cloud collaborative rice disease monitoring method and system for precision agriculture. Based on a target detection architecture, the feature extraction module is optimized to improve the model's ability to detect early-stage small lesions and diseases with small inter-class differences in rice. The model is lightweighted by combining knowledge distillation technology, and the student model obtained after distillation is deployed at the edge. A cloud management platform is designed based on the cloud platform to realize remote monitoring of edge devices, improve the efficiency of equipment management, and provide technical support for monitoring the growth status of rice.

[0007] Firstly, this application provides an edge-cloud collaborative rice disease monitoring method for precision agriculture, including:

[0008] A rice disease identification network was constructed based on the YOLO11n framework.

[0009] Squeeze-excitement attention module and multi-scale convolution-spatial attention module are introduced into the backbone and neck network of the rice disease identification network;

[0010] Replace the standard convolutions in the backbone and neck networks with depthwise separable convolutions;

[0011] The improved and optimized rice disease identification network was trained to obtain a lightweight student model.

[0012] Based on the prediction distributions output by the pre-trained deep teacher model and the student model, a distillation loss function is constructed.

[0013] By combining cross-entropy loss and the distillation loss function, and using the backpropagation algorithm to guide the training of the student model, a lightweight rice disease identification model is obtained, which is used to identify rice diseases based on rice plant images.

[0014] In one possible implementation, the squeeze-excitation attention module is used to process the input feature map X = [x1, x2, ..., x...]. c ]∈R C×H×W A squeezing operation is performed, and a descriptor S for each feature channel is generated through global average pooling.

[0015] An activation operation consisting of a sequence of layers with 1×1 convolution, ReLU, and Sigmoid activation is performed on the feature channel descriptor S to evaluate the relationship between different channels and the output weights representing the importance of different channels, thus generating channel attention weights Z.

[0016] The channel attention weight Z is multiplied by the original input feature X to obtain the recalibrated feature map Y, denoted as:

[0017]

[0018] Z=Sigmoid(W2×ReLU(W1S)) (2)

[0019] y c =z c ×x c (3)

[0020] In the formula, H represents the height of the feature map, W represents the width of the feature map, C represents the number of channels in the feature map, and s C x is the c-th element of S. c ∈R H×W It is the c-th channel of the input feature. Y represents the weights generated for each feature channel, and Y is the output feature map of the SE module.

[0021] In one possible implementation, the multi-scale convolutional-spatial attention module uses 1×1 convolution and ReLU operations to compress the feature map feature Y output by the squeeze-excitement attention module. The compressed feature is then passed through four dilated convolutional layers with different dilation rates and a ReLU activation function to output feature descriptors y with different receptive fields. d ;

[0022] Based on the spatial attention module SA, a feature descriptor y of the hybrid dilated convolution output is learned. d Based on the varying importance of different spaces, a spatial attention map M is generated. d ;

[0023] Spatial attention map M d Compared with the original feature descriptor y d Element-wise multiplication is performed, and the channel number of the features from the four branches is fused using the concat operation to obtain the fused feature map M. Then, M is multiplied element-wise with the feature map Y to obtain the attention feature map F.

[0024] One possible implementation replaces the standard convolutions in the backbone and neck networks with depthwise separable convolutions, including:

[0025] The standard convolution is decomposed into depthwise convolution and pointwise convolution; wherein, the depthwise convolution is used to perform convolution operation on each channel of the input rice plant image independently to extract features within the channel; the pointwise convolution is used to fuse inter-channel information on the feature map output by the depthwise convolution.

[0026] In one possible implementation, a distillation loss function is constructed based on the prediction distributions output by the deep teacher model and the student model, including:

[0027] The collected rice disease dataset is input into the pre-trained deep teacher model and the student model to obtain the prediction distributions output by the deep teacher model and the student model.

[0028] Using the predicted distribution output by the deep teacher model as soft labels, the distillation loss between the soft labels and the predicted distribution output by the student model is calculated based on KL divergence; wherein, the distillation loss includes knowledge distillation classification loss, knowledge distillation bounding box loss, and knowledge distillation confidence loss, expressed as:

[0029]

[0030] In the formula, P teacher P is the predicted distribution output by the deep teacher model. student λ represents the predicted distribution output by the student model. cls , λ bbox , λ obj L is the weighting coefficient. KD_cls L represents the knowledge distillation classification loss. KD_bbox Represents the knowledge distillation bounding box loss, L KD_obj This represents the confidence loss from knowledge distillation.

[0031] In one possible implementation, the total loss function is constructed by combining the cross-entropy loss and the distillation loss function, expressed as:

[0032] Loss=αL CE (P true ,P student )+(1-α)L soft (P teacher ,P student (11)

[0033] In the formula, α is a hyperparameter used to balance the weights between real labels and teacher soft labels; L soft It is the distillation loss; P student P is the predicted distribution output by the student model. true It's a real label, L CE It is the cross-entropy loss between the student model's predictions and the true labels. The cross-entropy loss is expressed as:

[0034]

[0035] Secondly, this application provides an edge-cloud collaborative rice disease monitoring system for precision agriculture, which is applied to the edge-cloud collaborative rice disease monitoring method for precision agriculture as described in any of the first aspects, including edge terminal equipment and cloud management platform;

[0036] The edge device has a built-in lightweight rice disease identification model, which is used to identify diseases in the collected rice plant images based on the lightweight rice disease identification model, and upload the disease identification results to the cloud management platform.

[0037] The cloud management platform includes a data analysis module, which is used to access the rice disease knowledge base based on a large model agent, generate disease prevention and control strategies based on the received disease identification results, and distribute the disease prevention and control strategies to at least one edge device.

[0038] Compared with the prior art, the technical solution provided in this application has the following beneficial effects:

[0039] This application achieves multi-dimensional benefits by organically combining deep learning technology with lightweight design: leveraging the powerful image feature self-learning capability of deep learning and combining channel and spatial collaborative attention mechanisms, it significantly enhances the model's perception accuracy of rice disease characteristics, thereby greatly improving the recognition accuracy; by using deep separable convolution and knowledge distillation techniques, it achieves lightweight model while ensuring high accuracy, enabling fast inference on low-power edge devices, perfectly balancing performance and resource consumption.

[0040] Based on the edge-cloud collaborative architecture, the system can realize real-time monitoring of rice diseases at the edge and centralized management and in-depth analysis through the cloud. The system also supports flexible expansion of edge nodes, which can easily expand the monitoring range. This provides strong support for early and accurate monitoring and scientific prevention and control of rice diseases, and helps to reduce pesticide use and improve crop yield and quality. Attached Figure Description

[0041] Figure 1 This is a flowchart of a rice disease monitoring method based on edge-cloud collaboration for precision agriculture, provided in Embodiment 1 of this application.

[0042] Figure 2 This is a schematic diagram of the structure of the lightweight rice disease identification model provided in Embodiment 1 of this application.

[0043] Figure 3 This is a schematic diagram of the channel and spatial collaborative attention mechanism module provided in Embodiment 1 of this application.

[0044] Figure 4 This is a schematic diagram of the spatial attention module provided in Embodiment 1 of this application.

[0045] Figure 5 This is a schematic diagram of the structure of a depth-separable convolution provided in Embodiment 1 of this application.

[0046] Figure 6 This is a model diagram of the edge device provided in Embodiment 1 of this application.

[0047] Figure 7 This is a schematic diagram of the interface of the edge device local management software system provided in Embodiment 1 of this application.

[0048] Figure 8 This is a schematic diagram of the structure of a rice disease monitoring system based on edge-cloud collaboration for precision agriculture, as provided in Embodiment 2 of this application. Detailed Implementation

[0049] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the embodiments of this application. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0050] Example 1

[0051] join Figure 1 This is a flowchart of a rice disease monitoring method based on edge-cloud collaboration for precision agriculture, provided in Embodiment 1 of this application. Figure 1 As shown, the specific implementation steps of the above method include:

[0052] Step 101: Construct a rice disease identification network based on YOLO11n.

[0053] like Figure 2 As shown in the figure, the lightweight rice disease identification network constructed in this embodiment includes a backbone network, a neck network, and a detection head.

[0054] Step 102: Introduce the channel-space collaborative attention mechanism module SE-MSCA into the backbone and neck networks of the rice disease identification network.

[0055] like Figure 3 As shown, the aforementioned channel-space collaborative attention mechanism module SE-MSCA includes the squeeze and excitation block (hereinafter referred to as the SE module) and the multi-scale convolutional and spatial attention block (hereinafter referred to as the MCSA module).

[0056] In this embodiment, the SE module is used as a lightweight channel attention module. For the input feature map X = [x1, x2, ..., x...], ... c ]∈R C×H×W The squeezing operation generates each feature channel descriptor S through global average pooling. After that, the activation operation, consisting of a sequence of layers with 1×1 convolution, ReLU and sigmoid activation, performs an activation process on the feature channel descriptors to evaluate the relationship between different channels and the output weights representing the importance of different channels, and generates channel attention weights Z. Higher weights are assigned to feature channels that are closely related to rice diseases, while weakening or suppressing those channels that are less relevant to the recognition task.

[0057] The channel attention weight Z is multiplied by the original input feature X to obtain the recalibrated feature map Y, denoted as:

[0058]

[0059] Z=Sigmoid(W2×ReLU(W1S)) (2)

[0060] y c =z c ×x c (3)

[0061] In the formula, H represents the height of the feature map, W represents the width of the feature map, C represents the number of channels in the feature map, and s C x is the c-th element of S. c ∈R H×W It is the c-th channel of the input feature. Y represents the weights generated for each feature channel, and Y is the output feature map of the SE module.

[0062] Using the aforementioned multi-scale convolution-spatial attention module as a lightweight spatial attention module, a 1×1 convolution C is employed in the multi-scale convolution. 1×1 The input feature Y is compressed using the (·) and ReLU operations. The compressed feature is then passed through four dilated convolutional layers with different dilation rates and the ReLU activation function, outputting feature descriptors y with different receptive fields. d Hybrid dilated convolutions, without increasing the number of model parameters or computational cost, increase the receptive field of the rice disease identification network and establish long-range dependencies between disease target regions. In addition, such as... Figure 4 As shown, to avoid interference from complex backgrounds during feature extraction, this application introduces a spatial attention module (SA), enabling the rice disease identification model to learn the feature map y output by the hybrid dilated convolution. d Based on the varying importance of different spaces, a spatial attention map M is generated.d This guides the increase in the focus of disease identification models on key areas of rice diseases, and will increase the M... d Compared with the original feature map y d Element-wise multiplication is performed, and the channel number of the features from the four branches is fused using the concat operation to obtain the fused feature map M. Element-wise multiplication of M with feature map Y yields the attention feature map F. The relevant operation definitions of the MCSA module are shown below.

[0063]

[0064] SA(y d )=σ(C 7×7 ([Avgpool(y d ),Maxpool(y d (6)

[0065] In the formula, d (d = 1, 2, 3, 4) is the expansion rate. For a 3×3 dilated convolution with dilation rate d, ReLU(·) is the ReLU activation function, and C 1×1 (·) represents a 1×1 convolution, Y is the input feature map of the spatial attention module, and SA(·) represents the spatial attention operation. For the multiplication of characteristic matrices, C 7×7 (·) is a 7×7 convolution, σ(·) is a 1×1 convolution and is a Sigmoid function. Avgpool and Maxpool are channel max pooling and average pooling operations, respectively.

[0066] Step 103: Replace the standard convolutions in the backbone network and neck network with depthwise separable convolutions.

[0067] like Figure 5 As shown in the embodiments of this application, the network structure of the model is designed to be lightweight, using depthwise separable convolution (DPW) instead of traditional convolution, decomposing traditional convolution into depthwise convolution and pointwise convolution. Depthwise convolution is used to independently perform convolution operations on each channel of the input rice disease image to extract features within each channel. Pointwise convolution is used to fuse inter-channel information on the feature map output by depthwise convolution. Through the above depthwise separable convolution operations, the model's parameter count and computational complexity can be reduced while maintaining the ability to extract rice disease features.

[0068] Step 104: Train the improved and optimized rice disease identification network to obtain a lightweight student model.

[0069] Step 105: Input the rice disease dataset into the pre-trained deep teacher model and the aforementioned lightweight student model, and obtain the prediction distribution output by the deep teacher model and the student model.

[0070] In this embodiment, YOLOv12s, a model architecture that performs better and faster in rice disease dataset testing, is used as a deep teaching model for knowledge distillation training. This allows the student model to learn the generalization ability of the deep teacher model while maintaining low complexity, thereby obtaining a final deployment model with higher accuracy and better performance than the original student model.

[0071] Step 106: Calculate the distillation loss of the soft label and the student model predicted distribution based on KL divergence.

[0072] In knowledge distillation scenarios, the prediction distribution output by the deep teacher model is used as a soft label, while the prediction distribution output by the student model is its class probability estimate for the same input sample. The student model needs to adjust its parameters to make its output prediction distribution as close as possible to the soft label of the teacher model, thereby achieving knowledge transfer.

[0073] In this embodiment, the difference between the soft label and the student model's predicted distribution is calculated using KL divergence and used as the distillation loss to drive the student model to continuously adjust its parameters during training, making the output distribution approximate the teacher model. Specifically, the distillation loss consists of knowledge distillation classification loss, knowledge distillation bounding box loss, and knowledge distillation confidence loss, corresponding to the three tasks of classification, bounding box regression, and confidence in object detection, respectively. The distillation loss function is expressed as:

[0074]

[0075] In the formula, P teacher P is the predicted distribution output by the deep teacher model. student λ represents the predicted distribution output by the student model. cls , λ bbox , λ obj L is the weighting coefficient. KD_cls L represents the knowledge distillation classification loss. KD_bbox Represents the knowledge distillation bounding box loss, L KD_obj This represents the confidence loss from knowledge distillation.

[0076] The knowledge distillation classification loss uses Kullback-Leibler divergence (KL Divergence) to quantify the difference between the class prediction outputs of the student model and the deep teacher model, expressed as:

[0077]

[0078] In the formula, i represents the category of rice disease, N represents the total number of samples, and T cls (i) represents the class prediction output of the deep teacher model, S cls (i) represents the class prediction output of the student model, D KL This represents the KL divergence.

[0079] The knowledge distillation bounding box loss uses Mean Squared Error (MSE) to compare the difference between the bounding box regression output of the student model and the bounding box regression output of the deep teacher model, expressed as:

[0080]

[0081] In the formula, i represents the category of rice disease, N represents the total number of samples, and T reg (i) represents the bounding box regression output of the deep teacher model, S reg (i) is the bounding box regression output of the student model.

[0082] The knowledge distillation confidence loss uses binary cross-entropy (BCE) to evaluate the difference between the confidence scores predicted by the student model and those predicted by the deep teacher model, expressed as:

[0083]

[0084] In the formula, i represents the category of rice disease, N represents the total number of samples, and T obj (i) represents the confidence score predicted by the deep teacher model, S obj (i) represents the confidence score predicted by the student model.

[0085] Step 107: Construct the total loss function by combining the cross-entropy loss, which includes the cross-entropy loss and the distillation loss mentioned above.

[0086] In this embodiment, the above-mentioned total loss function is used to guide the training of a lightweight student model through backpropagation algorithm to obtain a lightweight rice disease identification model. The total loss function is expressed as:

[0087] Loss=αL CE (P true ,P student )+(1-α)L soft (P teacher ,P student (11)

[0088] In the formula, α is a hyperparameter used to balance the weights between real labels and teacher soft labels; L soft It is the distillation loss, defined as shown in equation (7); P studentP is the predicted distribution output by the student model. true It's a real label, L CE It is the cross-entropy loss between the student model's predictions and the true labels. The cross-entropy loss is expressed as:

[0089]

[0090] Step 108: Based on the above rice lightweight identification model, identify diseases in the collected rice plant images.

[0091] In this embodiment of the application, the final lightweight rice recognition model is deployed on edge devices, such as... Figure 6 As shown in the illustration, the specific hardware platform of the edge device in this embodiment is the NVIDIA Jetson platform. An image acquisition camera, power module, and communication module are integrated with the NVIDIA Jetson platform to form the hardware component of the rice leaf disease recognition system. The image acquisition camera acquires disease image data of rice leaves. The power module supplies power to the hardware platform. A Python development environment is set up on the NVIDIA Jetson platform, and the optimized lightweight rice recognition model obtained from the distillation process is ported to the NVIDIA Jetson platform. This is combined with a visual software system interface developed based on Pyside6, such as... Figure 7 As shown in the diagram, the camera is turned on, the image of the rice plant to be identified is read in, and the disease identification and analysis of the collected rice disease image data is completed, and the results are output and displayed.

[0092] Step 109: Based on the disease identification results, generate disease control strategies using the rice disease knowledge base.

[0093] In this embodiment, a cloud management platform remotely monitors edge devices, stores and analyzes the disease identification and analysis results processed by the edge devices. A large model agent is run to access a rice disease knowledge base, analyze the received disease identification results, and generate disease control strategies or suggestions. Furthermore, the cloud management platform distributes the generated disease control strategies to users of the edge devices via a server. Users can access the cloud management platform via a PC terminal or directly access the edge devices via a mobile terminal within an intranet environment to obtain the aforementioned disease control strategies.

[0094] Compared with the prior art, the technical solution provided in Embodiment 1 of this application has the following beneficial effects:

[0095] This application leverages the strong self-learning capabilities of deep learning technology for image features. Combined with a lightweight channel and spatial collaborative attention mechanism, it enhances the model's perception of rice disease characteristics, improving the accuracy of rice disease identification. Furthermore, based on deep separable convolution and knowledge distillation techniques, a lightweight rice disease identification model is designed, enabling rapid inference of the high-precision model on low-power edge devices. This effectively balances performance and resource consumption. Under an edge-cloud collaborative architecture, it achieves effective collaboration between real-time edge monitoring and centralized cloud management and deep analysis. The system architecture also supports easy addition or removal of edge nodes, facilitating the expansion of the monitoring range. This provides strong technical support for early and accurate monitoring and scientific prevention of rice diseases, helping to reduce pesticide use and improve yield and quality, aligning with the development direction of precision agriculture, smart agriculture, and modern agriculture.

[0096] Example 2

[0097] See Figure 8 This is a schematic diagram of the structure of an edge-cloud collaborative rice disease monitoring system for precision agriculture, provided in Embodiment 1 of this application. Figure 8 As shown, the system includes edge devices and a cloud management platform. The edge devices are used to collect rice image data and process it in real time, while the cloud management platform is used to remotely monitor the edge devices and store and analyze the processed disease identification results.

[0098] Specifically, the aforementioned edge devices are built on a low-power embedded hardware platform, including an image acquisition module, a disease recognition module, a central processing unit, a hardware acceleration unit, and a communication module (network interface).

[0099] The aforementioned image acquisition module is specifically used to acquire images of rice plants.

[0100] The aforementioned disease identification module incorporates a lightweight rice disease identification model, specifically used for disease identification.

[0101] The aforementioned central processing unit is specifically used to execute the aforementioned disease identification module using the hardware acceleration unit, thereby enabling real-time disease identification of the collected rice plant images.

[0102] The aforementioned communication module is specifically used to upload the disease identification results and related data to the cloud management platform, and to receive instructions sent by the cloud management platform.

[0103] The aforementioned cloud management platform includes a data storage module, a data analysis module, and a web management interface. The data storage module receives and stores rice plant image data, disease identification results, and device status data uploaded from at least one edge device via a network interface. The data analysis module runs a large model agent, accesses a rice disease knowledge base, analyzes the received edge data, generates disease control strategies or suggestions, and distributes them to edge users. The web management interface displays the data analysis results and manages the edge device cluster, presenting relevant information to users through the web interface.

[0104] Compared with the prior art, the technical solution provided in Embodiment 2 of this application has the following beneficial effects:

[0105] This application addresses the shortcomings of existing rice disease monitoring technologies, including insufficient detection capabilities for small lesions and inter-disease differences, difficulty in balancing model performance and efficiency due to limitations in computing power and power consumption of edge devices, and inconvenient user interaction. It provides an edge-cloud collaborative rice disease monitoring system for precision agriculture. This system constructs a lightweight rice disease identification model with strong disease feature perception capabilities, improving the model's accuracy in identifying various rice diseases. Simultaneously, it solves the problems of slow processing speed, response latency, and even failure to run the model on embedded devices due to insufficient resources, enabling the deployment and application of the model on mobile edge devices. The designed cloud management platform displays the real-time operating status and data flow of each device, allowing administrators to view detailed device information and operational status, and enabling data push, log backtracking, and remote command issuance. This solves the problems of difficult edge device management and high maintenance costs, ensuring stable device operation and guaranteeing the continuity of disease identification and data collection.

[0106] Although embodiments of this application have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of this application, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for monitoring rice diseases using an edge-cloud collaborative approach for precision agriculture, characterized in that, include: A rice disease identification network was constructed based on the YOLO11n framework. Squeeze-excitement attention module and multi-scale convolution-spatial attention module are introduced into the backbone and neck network of the rice disease identification network; Replace the standard convolutions in the backbone and neck networks with depthwise separable convolutions; The improved and optimized rice disease identification network was trained to obtain a lightweight student model. Based on the prediction distributions output by the pre-trained deep teacher model and the student model, a distillation loss function is constructed. A total loss function is constructed by combining cross-entropy loss and the distillation loss function. The training of the student model is guided by the backpropagation algorithm to obtain a lightweight rice disease identification model, which is used to identify rice diseases based on rice plant images.

2. The edge-cloud collaborative rice disease monitoring method for precision agriculture according to claim 1, characterized in that, The squeeze-excitation attention module is used for performing a squeeze operation on an input feature map X=[x1, x2, ···, x c ]∈R C×H×W to generate a descriptor S for each feature channel through global average pooling. An activation operation consisting of a sequence of layers with 1×1 convolution, ReLU, and Sigmoid activation is performed on the feature channel descriptor S to evaluate the relationship between different channels and the output weights representing the importance of different channels, thus generating channel attention weights Z. The channel attention weight Z is multiplied by the original input feature X to obtain the recalibrated feature map Y, denoted as: In the formula, H represents the height of the feature map, W represents the width of the feature map, C represents the number of channels in the feature map, and s C It is the c-th element of the feature channel descriptor S, x c ∈R H×W It is the c-th channel of the input feature. Y represents the weights generated for each feature channel, and Y is the output feature map of the SE module.

3. The edge-cloud collaborative rice disease monitoring method for precision agriculture according to claim 1, characterized in that, The multi-scale convolution-space attention module adopts 1*1 convolution and ReLU operation to compress the feature map feature Y output by the squeeze-excitation attention module, and the compressed feature is output by four dilated convolution layers with different dilated rates and ReLU activation functions, and the feature descriptor y with different receptive fields is output d ; Based on the spatial attention module SA, learn the feature descriptor y of the mixed dilated convolution output d The importance of different spaces in the middle, generate a spatial attention map M d ; The spatial attention map M d with the original feature descriptor y d Element multiplication is performed on the features of the 4 branches, and the channel number fusion is performed by using the concat operation to obtain the fused feature map M. The feature map M is multiplied by the feature map Y to obtain the attention feature map F.

4. The edge-cloud collaborative rice disease monitoring method for precision agriculture according to claim 1, characterized in that, Replacing the standard convolutions in the backbone and neck networks with depthwise separable convolutions includes: The standard convolution is decomposed into depthwise convolution and pointwise convolution; wherein, the depthwise convolution is used to perform convolution operation on each channel of the input rice plant image independently to extract features within the channel; the pointwise convolution is used to fuse inter-channel information on the feature map output by the depthwise convolution.

5. The edge-cloud collaborative rice disease monitoring method for precision agriculture according to claim 1, characterized in that, Based on the prediction distributions output by the deep teacher model and the student model, a distillation loss is constructed, including: The collected rice disease dataset is input into the pre-trained deep teacher model and the student model to obtain the prediction distributions output by the deep teacher model and the student model. Using the predicted distribution output by the deep teacher model as soft labels, the distillation loss between the soft labels and the predicted distribution output by the student model is calculated based on KL divergence; wherein, the distillation loss includes knowledge distillation classification loss, knowledge distillation bounding box loss, and knowledge distillation confidence loss, expressed as: In the formula, P teacher is a prediction distribution output by a deep teacher model, P student is a prediction distribution output by a student model. λ cls , λ bbox , λ obj are weight coefficients, L KD_cls represents a knowledge distillation classification loss, L KD_bbox represents a knowledge distillation bounding box loss, L KD_obj represents a knowledge distillation confidence loss.

6. The edge-cloud collaborative rice disease monitoring method for precision agriculture according to claim 1, characterized in that, The total loss function is constructed by combining the cross-entropy loss and the distillation loss function, and is expressed as: Loss(αL CE (P true ,P student )+(1-α)L soft (P teacher ,P student ) (11) In the formula, α is a hyperparameter used to balance the weights between real labels and teacher soft labels; L soft It is the distillation loss; P student P is the predicted distribution output by the student model. true It's a real label, L CE It is the cross-entropy loss between the student model's predictions and the true labels. The cross-entropy loss is expressed as:

7. An edge-cloud collaborative rice disease monitoring system for precision agriculture, applied to the edge-cloud collaborative rice disease monitoring method for precision agriculture as described in any one of claims 1 to 6, characterized in that, This includes edge devices and cloud management platforms; The edge device has a built-in lightweight rice disease identification model, which is used to identify diseases in the collected rice plant images based on the lightweight rice disease identification model, and upload the disease identification results to the cloud management platform. The cloud management platform includes a data analysis module, which is used to access the rice disease knowledge base based on a large model agent, generate disease prevention and control strategies based on the received disease identification results, and distribute the disease prevention and control strategies to at least one edge device.

Images