An agricultural pest and disease monitoring system and method based on cloud-edge collaboration

Abstract

This invention discloses an agricultural pest and disease monitoring system and method based on cloud-edge collaboration, belonging to the field of pest and disease monitoring technology. The system includes a data acquisition module, a segmented processing module, a matching and comparison module, a spatiotemporal evolution module, and a feedback adjustment module. Its key technical points are: accessing an edge monitoring terminal to acquire crop image sequences; introducing a segmented rule engine to identify the fluctuation amplitude of local textures; extracting effective feature sequence segments and uploading them to the cloud; constructing an agricultural dynamic graph network; extracting node state attributes; executing a tree-like hierarchical matching strategy to obtain pest and disease identification results; injecting the pest and disease identification results into the agricultural dynamic graph network; calculating the disease spread probability through a preset inference training model; generating a spatiotemporal evolution trajectory; determining the predicted risk level; and generating parameter adjustment instructions to feed back to the edge monitoring terminal to update the segmented rule engine. This invention ensures cloud-edge collaboration and achieves effective early warning.

Description

Technical Field

[0001] This invention relates to the field of pest and disease monitoring technology, specifically to an agricultural pest and disease monitoring system and method based on cloud-edge collaboration. Background Technology

[0002] With the rapid development of smart agriculture, the use of computer vision technology for automatic monitoring of crop diseases and pests has become an important means of ensuring agricultural production. In recent years, deep learning technologies, represented by neural networks, have been relied upon in agricultural disease and pest monitoring tasks. Currently, the monitoring method mainly combines cloud processing and edge monitoring terminal processing architectures, but the following shortcomings still exist: On the one hand, because the types and forms of crops change with time and season, the model needs to be continuously updated and iterated in order to ensure recognition accuracy. Existing technologies mostly use the method of full model replacement, which results in a large update package size. Agricultural monitoring equipment is mostly deployed in farmland, where communication networks are often in a weak state, making download failures very easy. At the same time, existing edge monitoring terminals blindly upload massive amounts of redundant data for three months or more, and transmit the massive amounts of collected image data back to the cloud indiscriminately, or only sample and upload based on simple time intervals. This not only wastes valuable uplink bandwidth, but also prevents the cloud from effectively optimizing the texture fluctuations of the edge monitoring terminals, resulting in low overall monitoring efficiency. On the other hand, current cloud-edge collaborative systems are unidirectional data flows, meaning data flows from the edge to the cloud, but lacks a reverse control flow from the cloud back to the edge. Under high-risk conditions, this may lead to missed detection of minute lesions due to insufficient sampling, making it impossible to achieve a dynamic adaptive balance between resource load and matching accuracy. At the same time, existing cloud processing often only processes the collected image sequences to obtain pest and disease identification results, ignoring the spatial topological relationships of crops in the field. In the face of complex crop environments, relying solely on single feature comparisons and lacking tree-like hierarchical screening leads to a certain degree of waste of computing resources and reduced identification accuracy, making it difficult to effectively control and warn against pests and diseases. Summary of the Invention

[0003] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides an agricultural pest and disease monitoring system and method based on cloud-edge collaboration. It connects to an edge monitoring terminal to collect crop image sequences in real time, introduces a segmented rule engine to identify the fluctuation amplitude of local textures, and extracts effective feature sequence segments. This enables cloud-based methods to effectively target and optimize texture fluctuations at the edge monitoring terminal, improving monitoring efficiency. Simultaneously, an agricultural dynamic graph network is constructed in the cloud, and a tree-like hierarchical matching strategy is executed to obtain pest and disease identification results. A preset inference training model is used to calculate the probability of disease spread, thereby obtaining a predicted risk level. Parameter adjustment instructions are generated and fed back to the edge monitoring terminal to update the segmented rule engine. This effectively ensures cloud-edge collaboration, achieves accurate early warning, and solves the problems mentioned in the background technology.

[0004] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: Firstly, this application provides an agricultural pest and disease monitoring system based on cloud-edge collaboration, the system comprising: Acquisition module: Used to connect to the edge monitoring terminal and acquire crop image sequences in the target area in real time; Segmentation processing module: Based on crop image sequences, it introduces a segmentation rule engine, identifies the fluctuation amplitude of local textures, extracts effective feature sequence segments, compresses them to obtain compressed feature vectors, and uploads them to the cloud; Matching and comparison module: used to build an agricultural dynamic graph network in the cloud, mapping compressed feature vectors to node state attributes of the agricultural dynamic graph network; executing a tree-like hierarchical matching strategy, combining node state attributes to build a target feature tree, and comparing it with the preset standard disease classification tree layer by layer to obtain the disease and pest identification results; Spatiotemporal evolution module: used to inject the results of pest and disease identification into the agricultural dynamic graph network, drive the iterative training of the agricultural dynamic graph network through the preset inference training model, output the probability of disease spread, and generate the spatiotemporal evolution trajectory. Feedback Adjustment Module: Based on the spatiotemporal evolution trajectory, the module determines and predicts the risk level and generates parameter adjustment instructions to be fed back to the edge monitoring terminal to update the segmentation rule engine.

[0005] Furthermore, real-time acquisition of crop image sequences in the target area, including: The edge monitoring terminal synchronously triggers and controls the imaging unit and the PTZ pan-tilt unit; When triggered by the imaging unit, and in conjunction with the preset scanning trajectory of the PTZ pan-tilt unit, the crop in the target area is continuously exposed and photographed to acquire crop image data; Edge computing is used to preprocess crop image data to generate crop image sequences.

[0006] Furthermore, a segmentation rule engine is introduced, including: Based on the crop image sequence, the first pixel position is identified as the starting position of the current segment, and a crop image subsequence of the first length is extracted within a preset length range as a reference segment. The starting position of the current segment is then adjusted to the ending position of the reference segment. Starting from the adjusted current segment start position, the process is repeated until the entire crop image sequence is traversed: taking the current segment start position as the starting point, each second length within the preset length range is traversed, and the corresponding subsequence is extracted from the crop image sequence to generate several candidate sequence segments; where the second length is a dynamic quantity; the fluctuation index of each candidate sequence segment is calculated, the candidate sequence segment with the largest value is selected as the optimal sequence segment, and the optimal sequence segment with a fluctuation index greater than the standard fluctuation threshold is selected and marked as a valid feature sequence segment.

[0007] Furthermore, the volatility index is calculated as follows: The sliding window step size is preset; based on the candidate sequence segment, the entropy value and contrast value in each sliding window are extracted, and the local texture value is obtained by weighted summation. The absolute value of the first difference between the local texture value of the current sliding window and the previous adjacent window is calculated as the fluctuation amplitude of the local texture. The arithmetic mean of the fluctuation amplitudes of all sliding windows in the candidate sequence segment is calculated to obtain the fluctuation index.

[0008] Further, a tree-like hierarchical matching strategy is implemented, including: Based on the node state attributes, several effective feature sequence segments are restored and defined as leaf nodes. Based on the spatial location information recorded in the node state attributes of each effective feature sequence segment, a clustering algorithm is used to fuse the features of adjacent leaf nodes to generate parent nodes, and then the features are fused upwards layer by layer to generate root nodes, thus constructing the target feature tree. The target feature tree is compared layer by layer with the pre-set standard disease classification tree; Structural components are determined by identifying node information at each level; texture components are determined by identifying local texture values; where node information includes node depth, number of child nodes, and connection relationships. Cross-similarity is calculated based on texture components, topological similarity is calculated based on structural components, and the matching difference is generated by weighted summation using the optimal weight combination. Compare the matching difference with the optimal matching threshold: if the matching difference is greater than the optimal matching threshold, then prune the branch; if the matching difference is less than or equal to the optimal matching threshold, then retain the branch and proceed to the next level until the matching reaches the leaf node of the standard disease classification tree, thus determining the disease and pest identification result.

[0009] Furthermore, obtaining the optimal weight combination and the optimal matching threshold includes: When performing matching at the current level, the resources of the current matching process are classified and quantified to obtain a matching monitoring vector, including computing power, memory usage, cloud-edge communication bandwidth, and timestamp. By using the set of entity nodes to be matched as the sample set, and taking minimizing the load and maximizing the matching accuracy as the dual objective functions, the particle swarm optimization algorithm is used to iteratively solve the policy encoding set, and output the adaptive response strategy for the current matching level. The adaptive response strategy includes: the optimal weight combination and the optimal matching threshold.

[0010] Furthermore, the pre-defined inference training model is built upon a spatiotemporal graph neural network architecture, including: Collect the identified crop image sequences and identification results, extract node state attributes as feature information, map the identification results of each entity node at each time step into a graph snapshot through time series, and integrate them into the target dataset; Spatiotemporal graph neural networks are used as the inference training model. The target dataset is divided into training set, test set and validation set. The training set is used for iterative training, the test set is used to test the trained inference training model, and the validation set is used to evaluate the inference performance of the model. Using a trained spatiotemporal inference model, the pest and disease identification results obtained at the current moment are injected into the agricultural dynamic graph network as the initial state, driving the agricultural dynamic graph network to perform forward inference, outputting the disease spread probability of each entity node in the next N1 time steps, and connecting the probability distributions of each time step in time sequence to generate a spatiotemporal evolution trajectory; where N1 is a positive integer greater than 0.

[0011] Furthermore, the segmentation rule engine was updated, including: Based on the predicted risk level, the corresponding parameter adjustment instructions are obtained and encapsulated into a control message, which is sent to the edge monitoring terminal through the downlink channel. The edge monitoring terminal receives the control message, and after protocol parsing and integrity verification, extracts a new parameter group containing a new sliding window step size and a new preset length range. When the current frame crop image sequence is processed and the next frame is loaded, the original sliding window step size and preset length range are replaced.

[0012] Furthermore, the predicted risk levels include high risk, medium risk, and low risk.

[0013] Secondly, this application provides a cloud-edge collaborative method for monitoring agricultural pests and diseases, the method comprising: Connect to the edge monitoring terminal to collect real-time crop image sequences in the target area; Based on crop image sequences, a segmentation rule engine is introduced to extract effective feature sequence segments by identifying the fluctuation amplitude of local textures, and then compress them to obtain compressed feature vectors which are uploaded to the cloud. An agricultural dynamic graph network is constructed in the cloud, and compressed feature vectors are mapped to node state attributes of the agricultural dynamic graph network. A tree-like hierarchical matching strategy is executed to construct a target feature tree by combining node state attributes, and then the feature tree is compared layer by layer with a pre-set standard disease classification tree to obtain the disease and pest identification results. The results of pest and disease identification are injected into the agricultural dynamic graph network. The network is then driven to undergo iterative training through a pre-set inference and training model. The network outputs the probability of disease spread and generates spatiotemporal evolution trajectories. Risk levels are predicted based on spatiotemporal evolution trajectories, and parameter adjustment instructions are generated and fed back to the edge monitoring terminal to update the segmentation rule engine.

[0014] (III) Beneficial Effects This invention provides an agricultural pest and disease monitoring system and method based on cloud-edge collaboration, which has the following beneficial effects: 1. This invention introduces a segmentation rule engine to identify the local texture fluctuation amplitude of crop image sequences, focusing feature extraction on the dynamic changes in crop image sequences; by extracting entropy and contrast, it obtains effective feature sequence segments containing pest and disease information; at the same time, it uses compressed feature vectors for data upload, which significantly reduces the network transmission load and latency on the edge side, provides high signal-to-noise ratio feature input to the cloud, and realizes deep collaboration between precise edge perception and cloud computing. 2. This invention constructs an agricultural dynamic graph network in the cloud and receives compressed feature vectors uploaded by edge monitoring terminals, directly mapping them to node state attributes. This achieves a seamless connection from single-point edge perception to networked cognition in the cloud. It not only preserves the spatial topological relationship of the original data but also reduces cloud storage pressure by compressing features. During the matching process, a tree-like hierarchical matching strategy is executed, transforming single image recognition into a hierarchical comparison between the target feature tree and the standard disease classification tree. By using structural components and texture components as the core values ​​for layer-by-layer comparison, the robustness and interpretability of disease and pest information recognition are improved. 3. This invention reconstructs the agricultural dynamic graph network by injecting the results of pest and disease identification into the agricultural dynamic graph network, and uses a preset inference model to drive the state iteration of entity nodes, outputting the probability of disease spread. At the same time, it obtains the spatiotemporal evolution trajectory by analyzing the probability of disease spread, which improves the perception dimension and prediction accuracy of the system for the dynamic spread trend of diseases to a certain extent. 4. Based on the spatiotemporal evolution trajectory, the risk level is predicted, and parameter adjustment instructions are generated to update the judgment logic of the segmented rule engine of the edge monitoring terminal in real time, realizing closed-loop control from cloud risk prediction to edge-level control; to a certain extent, through cloud-edge collaboration, the edge load and network transmission bandwidth of routine monitoring are reduced, and the monitoring accuracy and early warning are improved. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of a module of an agricultural pest and disease monitoring system according to an exemplary embodiment; Figure 2 This is a schematic diagram of the framework of an agricultural pest and disease monitoring system according to an exemplary embodiment; Figure 3 This is a flowchart illustrating an agricultural pest and disease monitoring method according to an exemplary embodiment. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Example 1: This invention provides an agricultural pest and disease monitoring system based on cloud-edge collaboration; Figure 1 This is a schematic diagram of a module of an agricultural pest and disease monitoring system according to an exemplary embodiment. Figure 2 This is a schematic diagram of the framework of an agricultural pest and disease monitoring system according to an exemplary embodiment; please refer to... Figures 1 to 2 The system includes: an acquisition module, a segmented processing module, a matching and comparison module, a spatiotemporal evolution module, and a feedback adjustment module, and the acquisition module, segmented processing module, matching and comparison module, spatiotemporal evolution module, and feedback adjustment module are interconnected. The edge monitoring terminal responds to the monitoring command issued by the cloud and immediately starts the data acquisition module; Acquisition module: Connects to the edge monitoring terminal to acquire real-time crop image sequences of the target area; As the core control terminal at the edge of the field, the edge monitoring terminal has a built-in high-performance AI computing module and a high-speed vision interface that connects the imaging unit and the PTZ pan-tilt unit. The high-speed vision interface is a GPIO interface that directly connects the edge monitoring terminal to the imaging unit and the PTZ pan-tilt unit. It uses frame synchronization signals to ensure that the spatial position coordinates of the pan-tilt unit are precisely aligned with the timestamp of each frame of image acquired by the imaging unit, thus ensuring the accuracy of subsequent image stitching or positioning. The imaging unit integrates a high-resolution visible light module and an autofocus lens to capture high-resolution images of crops. Combined with a PTZ pan-tilt unit, which provides all-around spatial perspective control, the imaging unit is driven to perform cruising motion according to a preset scanning trajectory, such as a zigzag or fan-shaped pattern, to capture discoloration, curling, mold spots, and insect bite marks on the canopy and leaf surfaces, ensuring that the field of view covers the target area and obtains high-definition raw image data. The original image data is transmitted to the memory buffer of the edge monitoring terminal. The front-end preprocessing is performed using the ISP built into the edge computing module. Automatic white balance is performed to eliminate the influence of sunlight color temperature changes, distortion correction is performed to correct the distortion caused by the wide-angle lens, and noise reduction is performed to improve texture clarity. The preprocessed images are reassembled according to timestamps and spatial order, invalid frames are removed, and a standardized crop image sequence is generated to facilitate subsequent use by the segmentation rule engine. The significance of the above analysis lies in the following: accessing the edge monitoring terminal breaks the limitation of traditional agricultural monitoring relying solely on single-frame static images, forcing the establishment of a sensing data source in the form of sequential acquisition; this temporal acquisition mechanism provides a key time dimension benchmark for subsequent segmented processing, enabling the system to capture the dynamic texture breathing effect of lesions in natural environments such as light winds, effectively distinguishing structured disease features from random background noise; at the same time, it provides original data support with strong spatiotemporal correlation for high-fidelity spatiotemporal evolution extrapolation in the cloud.

[0018] Segmentation processing module: Based on crop image sequences, a segmentation rule engine is introduced to extract effective feature sequence segments by identifying the fluctuation amplitude of local textures, and then compress them to obtain compressed feature vectors which are uploaded to the cloud. Introducing a segmentation rule engine, including: Based on crop image sequences, the position of the first pixel is identified and marked as the starting position of the current segment; Starting from the current segment start position, crop image subsequences of a first length are extracted within a preset segment length range as reference segments. For example, the first length is 10 frames. The first extracted subsequence is taken as the first processing object, and the current segment start position is adjusted to the end position of the reference segment. At the same time, a sliding window is preset, and the local texture value of the last sliding window in the reference segment is extracted and marked as the preceding feature value. Starting from the adjusted starting position of the current segment, the following segmentation steps are executed repeatedly until the entire crop image sequence has been traversed: Starting from the current segment start position, crop image sequences are extracted at each second length within the preset segment length range to generate a set of candidate sequence segments with different lengths; where the second length is a dynamic quantity; for example, by traversing the preset segment length range, a set of second lengths with different lengths is generated every 2 frames. With a preset sliding window step size, based on a certain candidate sequence segment, the fluctuation amplitude of all sliding windows of the candidate sequence segment is extracted to obtain the fluctuation index of the candidate sequence segment; This includes extracting the fluctuation amplitude of all sliding windows in each segment, including: Under the same preset sliding window, the direction of the crop image sequence is controlled to move with a preset step size, and the grayscale of the crop image sequence is processed. Local image matrices are extracted in sequence to obtain local grayscale co-occurrence matrices. Entropy and contrast are extracted based on the gray-level co-occurrence matrix. The entropy is assigned a first weight and the contrast is assigned a second weight. The entropy is then multiplied by the first weight and the contrast is multiplied by the second weight. The results of the two multiplications are then added together to generate the local texture value of the current sliding window. The first and second weights are derived from a lightweight fully connected neural network built at the edge detection end. The model's input layer has two feature nodes, corresponding to the normalized entropy and contrast values, respectively. Two hidden layers are set, each with 16 neurons, both using the ReLU activation function. The output layer has two weight nodes, corresponding to the dynamic weights of the two numerical indicators, denoted as the first weight and the second weight, respectively, and their sum is equal to 1. In this process, it is necessary to collect a large amount of sliding window sample data from historical crop images, calculate the original entropy and original contrast values ​​based on the gray-level co-occurrence matrix for each sample, and then normalize them. Based on agricultural experts' experience in identifying disease morphology, each historical sliding window sample is assigned a texture saliency label. The range is set from 0 to 1, where 1 represents that the window contains clear, typical, and effective disease textures, and 0 represents that the window is an invalid background or random noise. All historical samples are sorted from low to high according to the texture saliency label to form a historical texture dataset. Every three values ​​are sampled from the historical texture dataset, and the resulting dataset is used as the test set for weight learning. At the same time, by learning the nonlinear mapping relationship between entropy, contrast, and texture saliency label, the first and second weights are adjusted through the backpropagation algorithm to minimize the error between the calculated result (entropy × first weight + contrast × second weight) and the texture saliency label. In this embodiment, the mean square error is used for calculation to obtain the first and second weights that can most accurately represent the saliency of disease textures. Calculate the absolute value of the first-order difference between the local texture value of the current window and the local texture value of the previous adjacent window, and mark this difference value as the fluctuation amplitude of the local texture; calculate the arithmetic mean of the fluctuation amplitudes of all sliding windows in the candidate sequence segment to obtain the fluctuation index; it should be noted that for the first sliding window at the beginning of the loop, the previous feature value is used as the local texture value of its previous adjacent window to participate in the calculation. Compare the fluctuation index of the candidate sequence segments in this group, and select the candidate sequence segment with the largest value as the optimal sequence segment at the current position. This indicates that the larger the fluctuation index, the greater the possibility that the candidate sequence segment contains information about diseases and pests; the smaller the fluctuation index, the less likely the candidate sequence segment contains information about diseases and pests, and the greater the invalid static background and redundant noise. The fluctuation index of the optimal sequence segment is compared with the standard fluctuation threshold: if the fluctuation index is greater than the standard fluctuation threshold, the optimal sequence segment is determined to be a valid feature sequence segment containing pest and disease information, and is then processed for subsequent compression; if the fluctuation index is less than or equal to the standard fluctuation threshold, the optimal sequence segment is determined to be a redundant background segment and is then filtered out; after the comparison is completed, the preceding feature value is updated to the local texture value of the last sliding window of the optimal candidate segment, and the start position of the current segment is updated to the end position of the optimal candidate segment, and the loop continues. Based on the effective feature sequence segments, the corresponding local texture values ​​and original image data are extracted respectively; the original image data is compressed and encoded using a hardware-level coding algorithm to generate a compressed image data packet; the compressed image data packet is associated, mapped and encapsulated with the local texture values ​​and the spatiotemporal metadata corresponding to the sequence segment to construct a compressed feature vector; and the edge-side compressed feature vector is uploaded to the cloud. The significance of the above analysis lies in the following: By introducing a segmentation rule engine, the local texture fluctuation amplitude of crop image sequences is identified, and feature extraction is focused on the dynamic changes in crop image sequences; by extracting entropy and contrast, adaptive segmentation processing is achieved to a certain extent, obtaining effective feature sequence segments containing pest and disease information, which can completely preserve the morphological structure of the pests, while maximizing the compression of low-fluctuation background areas, effectively reducing the data bandwidth consumption of the edge monitoring end to the cloud, and providing the cloud with feature inputs that eliminate noise interference, thereby improving the robustness of the recognition system from the source; at the same time, by using compressed feature vectors for data upload, the network transmission load and latency on the edge side are significantly reduced, providing the cloud with feature inputs with high signal-to-noise ratio, and realizing deep collaboration between accurate edge perception and cloud computing.

[0019] Matching and comparison module: Construct an agricultural dynamic graph network in the cloud, map compressed feature vectors to node state attributes of the agricultural dynamic graph network; execute a tree-like hierarchical matching strategy, construct a target feature tree based on node state attributes, and compare it with the preset standard disease classification tree layer by layer to obtain the disease and pest identification results; Constructing an agricultural dynamic graph network, including: By analyzing the compressed feature vectors, spatial location information, including GPS coordinates and target area number, is obtained, and an agricultural dynamic graph network is constructed, which is represented as: G(V, E, X). Where V = {v1, v2, ... v} i} represents the set of entity nodes, v1, v2, v3, v4, v5, v6, v7, v iThese represent the 1st, 2nd, and 1st monitored crop target area numbers, respectively, where i∈[1,N0], N0 is the total number of entity nodes in the target area division, and N0 is a positive integer greater than 0; at the same time, a one-to-one correspondence is established between each entity node and the GPS coordinates in the compressed feature vector; E={e jk} represents the set of spatially adjacent edges in an agricultural dynamic graph network, calculated by traversing any two entity nodes v j With v k The Euclidean distance between them, when this Euclidean distance is less than a preset spatial neighborhood threshold, at entity node v j With v k An edge is constructed between the two crop areas to represent the physical connectivity between them. X represents the node state attribute matrix of all entity nodes; by extracting the local texture value, fluctuation index, entropy value and contrast of the compressed feature vector, and through geolocation, it is spatially mapped to the corresponding entity node with the same GPS coordinate range in the agricultural dynamic graph network, as the node state attribute of that entity node. Execute the tree-structured hierarchical matching strategy, including: Based on the node state attributes, several valid feature sequence segments are restored and defined as leaf nodes; Based on the spatial location information recorded in the node state attributes of each effective feature sequence segment, a clustering algorithm is used to fuse the features of adjacent leaf nodes to generate a parent node, and then fuse them layer by layer upwards until the root node is generated, so as to construct a target feature tree containing multi-scale feature information. The system calls a pre-defined standard disease classification tree, aligns the root node of the standard disease classification tree with the root node of the target feature tree, and traverses the corresponding nodes of the two trees from top to bottom, starting from the next level below the root node. At the same time, it identifies the node information of each level in the target feature tree, including node depth, number of child nodes and connection relationships, calls the pre-defined hash encoding rules, and constructs the feature expression of the target feature tree. When performing matching at the current level, the resources of the current matching process are classified and quantified to obtain a matching monitoring vector, including computing power, memory usage, cloud-edge communication bandwidth, and timestamp. By using the set of entity nodes to be matched as the sample set, and taking minimizing the load and maximizing the matching accuracy as the dual objective functions, the particle swarm optimization algorithm is used to iteratively solve the problem and output the adaptive response strategy for the current matching level. The adaptive response strategy includes: optimal weight combination and optimal matching threshold. A particle swarm optimization (PSO) algorithm is used iteratively to solve the problem: the optimal weight combination and the optimal matching threshold are represented as particles, symbolizing an adaptive response strategy to the matching monitoring vector in the current matching process, which guides the subsequent tree-like hierarchical matching process; the set of entity nodes to be matched is randomly sampled to construct a sample set; in each iteration, the population is initialized, and the matching monitoring vector carried by each particle is obtained and set as the resource constraint boundary; a virtual matching simulation is performed on the sample set, and the expected resource consumption required to implement the strategy and the expected matching accuracy can be predicted. Simultaneously, the fitness value of each particle is calculated. First, the expected resource consumption is compared with the aforementioned resource constraint boundary; if the expected resource consumption exceeds the resource constraint boundary, it indicates that the particle... The strategy can lead to cloud overload, directly reducing the fitness of the particle. If the expected resource consumption is within the resource constraint boundary, the fitness value is positively correlated with the expected matching accuracy and negatively correlated with the expected resource consumption. Based on the fitness value of each particle, the particle swarm is driven to update its position. After multiple iterations, the algorithm terminates when the particle swarm converges to the position with the highest fitness or reaches the preset maximum number of iterations. The best particle after convergence is selected, and its parameters are output as the optimal adaptive response strategy in the current matching state. The weight combination determined by this strategy will be used for subsequent weighted similarity calculation, and the determined matching threshold will be used for subsequent disease pruning, thereby achieving a dynamic balance between cloud computing resources and disease identification accuracy. Simultaneously, feature vectors of the target feature tree node and the standard disease classification tree node are extracted and separated into texture components representing pathological color depth and structural components representing feature expressions through orthogonal decoupling. Texture component: The local texture values ​​of all valid feature sequence segments covered by the nodes of the acquisition level are obtained, the numerical statistical interval is set, the frequency of each local texture feature value falling into different statistical intervals is counted, a normalized frequency histogram vector is generated, and this vector is used as the texture component. Structural components: The characteristic expression is converted into a hexadecimal numerical sequence, which is then used as the structural components; Cross-similarity calculation based on texture components: Obtain the texture component of the current node of the target feature tree and the standard texture component of the corresponding node of the standard disease classification tree, and perform normalization verification on both to ensure that the sum of the elements in the two component vectors is 1, thereby transforming the frequency distribution into a probability distribution; then, traverse each statistical dimension of the texture component, compare the numerical values ​​of the target texture component and the standard texture component in that dimension one by one, and select the smaller value as the cross-overlap value of that dimension; finally, sum the cross-overlap values ​​calculated in all dimensions to obtain the cross-similarity that represents the degree of common overlap between the two texture distributions; since the agricultural environment is unstructured, changes in light will cause the overall drift of texture values. By analyzing the degree of overlap rather than the absolute numerical distance, it has a strong anti-interference ability against light fluctuations and local occlusion. Calculate topological similarity based on structural components: Obtain the hexadecimal values ​​corresponding to the numerical components of the target feature tree and the hexadecimal values ​​corresponding to the numerical components of the standard disease classification tree; calculate the absolute value of the difference between the two hexadecimal values; extract the maximum value among the hexadecimal values ​​corresponding to all standard disease classification trees, and calculate the ratio of the absolute value to the maximum value; obtain the topological similarity by subtracting the ratio from the value 1. Based on topological similarity and cross-similarity, and combined with the optimal weight combination, the matching difference between node state attributes and standard disease classification tree node features is obtained through weighted summation; the matching difference is then compared with the optimal matching threshold. If the matching difference is greater than the optimal matching threshold, it is determined that the current level node status attribute does not match the standard disease classification tree, and the branch is pruned. If the matching difference is less than or equal to the optimal matching threshold, the current level node's status attribute is determined to match the standard disease classification tree, and the branch is retained to proceed to the next level until a leaf node of the standard disease classification tree is matched. The pre-stored label attribute field of that leaf node is then read, and parameters such as crop name and disease type recorded in that field are extracted as the final disease and pest identification result. See Table 1 for specific parameters. Table 1 Tag Attribute Fields

[0020] Table 1 lists only a portion of the reference indicators, outlining the names of diseases in rice, wheat, and corn, and the specific features that the algorithms can identify. Specifically, the matching difference determination logic is as follows: The system has a pre-set standard disease feature library, which includes standard disease classification trees for rice sheath blight, wheat stripe rust, and corn leaf blight. When the matching comparison module runs, it calculates the matching difference between the target feature tree and the standard disease classification tree. This difference is obtained by weighting the structural and texture components; for example, the optimal matching threshold for the current level is 0.3. If the matching difference is less than or equal to 0.3, it is determined that there is disease / insect information, and the disease type is directly output. If the matching difference is greater than 0.3, it is determined to be interference or an unknown disease, and it is marked as an abnormal sample and uploaded to the cloud-based manual review pool. The significance of the above analysis lies in the following: This invention constructs an agricultural dynamic graph network in the cloud, receiving compressed feature vectors uploaded by edge monitoring terminals and directly mapping them to node state attributes. This achieves a seamless connection from single-point edge perception to networked cognition in the cloud, preserving not only the spatial topological relationships of the original data but also reducing cloud storage pressure through feature compression. During the matching process, a tree-like hierarchical matching strategy is executed, transforming single image recognition into a hierarchical comparison between the target feature tree and the standard disease classification tree. Through layer-by-layer feature comparison and pruning mechanisms, the computational load of unnecessary features is significantly reduced without sacrificing recognition accuracy, significantly improving the cloud's throughput efficiency in high-concurrency scenarios. In this process, using structural and texture components as the core values ​​for layer-by-layer comparison improves the robustness and interpretability of information-level disease and pest identification.

[0021] Spatiotemporal evolution module: Injects the results of pest and disease identification into the agricultural dynamic graph network, drives the iterative training of the agricultural dynamic graph network through a preset inference training model, outputs the probability of disease spread, and generates the spatiotemporal evolution trajectory; The pre-defined inference training model is built upon a spatiotemporal graph neural network architecture, including: Collect the identified crop image sequences and recognition results, extract the node state attributes in the agricultural dynamic graph network at each historical moment as feature information, and map the recognition results of each entity node in the agricultural dynamic graph network at each moment into graph snapshot data through time series, and integrate them into the target dataset; Spatiotemporal graph neural networks are used as the inference training model. The target dataset is divided into training set, test set and validation set. The inference training model is iteratively trained using the training set. The backpropagation algorithm is used to mine the hidden inter-node propagation dependencies and time series evolution laws in the agricultural dynamic graph network. The trained inference training model is tested using the test set. The inference performance of the model is evaluated using the validation set through prediction accuracy and root mean square error. Using a pre-trained spatiotemporal extrapolation model, the pest and disease identification results obtained at the current moment are injected into the agricultural dynamic graph network as the initial state, and instantiated as the current graph network state vector; the agricultural dynamic graph network is driven to perform multi-step forward extrapolation, outputting the disease spread probability of each entity node in the next N1 time steps, and the probability distributions of each time step are connected in time sequence to generate a spatiotemporal evolution trajectory representing the future spread trend of pests and diseases; where N1 is a positive integer greater than 0; It should be noted that the probability of disease spread is only a reference value. The specific calculation needs to be set according to the actual situation. For example, the actual situation may need to take into account environmental parameters such as wind direction and wind speed, which will not be elaborated here. The significance of the above analysis lies in the following: by injecting the results of pest and disease identification into the agricultural dynamic graph network, the agricultural dynamic graph network is reconstructed, and the state iteration of entity nodes is driven by the preset inference model to output the probability of disease spread. At the same time, the spatiotemporal evolution trajectory is obtained by analyzing the probability of disease spread, which to a certain extent improves the system's perception dimension and prediction accuracy of the dynamic spread trend of diseases.

[0022] Feedback adjustment module: Based on the spatiotemporal evolution trajectory, it determines the predicted risk level and generates parameter adjustment instructions to feed back to the edge monitoring terminal to adjust the segmentation rules; The predicted risk level of each node in the agricultural dynamic graph network is determined based on its spatiotemporal evolution trajectory, including: Based on the spatiotemporal evolution trajectory, the maximum and minimum disease spread probabilities of each entity node in the agricultural dynamic graph network are extracted within the next N2 time steps. For the same entity node, the difference between the disease spread probability and the minimum disease spread probability at the same moment is calculated, and the differences are aggregated to construct a risk data set; where N2 is a positive integer greater than 0. The K-Means clustering algorithm is used to cluster and analyze the risk dataset, calculate the cluster center value of each cluster, and divide all entity nodes into high-risk, medium-risk, and low-risk clusters. For example, the cluster with the highest cluster center value is defined as a high-risk cluster, the cluster with the lowest cluster center value is defined as a low-risk cluster, and the intermediate clusters are defined as medium-risk clusters. Based on the risk cluster category to which a node belongs, the corresponding predicted risk level is defined, and the corresponding parameter adjustment command is triggered. At the same time, the parameter adjustment command is encapsulated into a control message and sent to the edge monitoring terminal through the downlink channel. The edge monitoring terminal receives control messages, and after protocol parsing and integrity verification, extracts a new parameter set containing the new sliding window step size and the new preset length range. When the current frame crop image sequence is processed and the next frame is loaded, the original sliding window step size and preset length range are replaced. High-risk clusters correspond to high-risk levels, triggering the first adjustment command to adjust the sliding window step size to a preset minimum value to improve the spatial resolution of texture feature extraction; and to adjust the lower limit of the feature sequence segment length to a preset minimum length to preserve the independent lesion features at a small scale; The medium-risk level cluster corresponds to the medium-risk level, triggering the second adjustment command. A linear interpolation algorithm is used to calculate intermediate parameters based on the preset minimum and maximum values. The sliding window step size and the lower limit of the length of the feature sequence segment are set as these intermediate parameters to maintain a dynamic balance between feature capture precision and edge computing load. Low-risk clusters correspond to low-risk levels; trigger the third adjustment command to adjust the sliding window step size to the preset maximum value to reduce sampling frequency and computing power consumption; and adjust the lower limit of the feature sequence segment length to the preset maximum length to forcibly filter unstructured background noise and reduce data upload volume; Hierarchical control of the edge monitoring terminal is achieved through the first adjustment command, the second adjustment command, and the third adjustment command, enabling cloud-edge collaboration; The significance of the above analysis lies in the following: based on the spatiotemporal evolution trajectory, the risk level is predicted, and parameter adjustment instructions are generated to update the judgment logic of the segmented rule engine of the edge monitoring terminal in real time, thereby realizing closed-loop control from cloud-based risk prediction to edge-level control; to a certain extent, through cloud-edge collaboration, the edge load and network transmission bandwidth of routine monitoring are reduced, and the monitoring accuracy and early warning are improved.

[0023] For example: In region XX, the edge monitoring terminal's acquisition module starts as planned, capturing image sequences of wheat leaves. The edge monitoring terminal is initially configured in low-risk mode, with a sparse preset sliding window step size and small data processing volume (e.g., 16), and a large preset lower limit for the length range (e.g., 30). In a certain frame of the image sequence, sporadic bright yellow powdery vesicles appear. The segmentation rule engine extracts the effective sequence segment containing information about this pest and disease, compresses it into a feature vector, and uploads it to the cloud. By identifying that the texture is arranged in a dotted line pattern parallel to the leaf veins in space, and matching it with the typical characteristics of wheat stripe rust (e.g., a matching difference of less than 0.2), it is determined to be wheat stripe rust. The disease is identified by using the diseased node as the source of infection and inputting it into the agricultural dynamic graph network. The pre-set inference training model outputs that stripe rust spores will spread to neighboring nodes with a high probability, indicating a high risk level. The feedback adjustment module changes the predicted risk level from low risk to high risk and issues parameter adjustment instructions. In reverse mapping, the preset sliding window step size is reduced to 4 pixels, and the lower limit of the preset length range is increased to 5 frames. A control message is generated and sent to the edge monitoring terminal. The edge terminal receives the message and uses a mutex lock to suspend the segmented rule engine, instantly replacing the parameters 16 pixels and 30 frames in the local register with 4 pixels and 5 frames, ensuring cloud-edge collaboration and achieving effective early warning.

[0024] Example 2: This invention provides a cloud-edge collaborative method for monitoring agricultural pests and diseases; Figure 3 This is a flowchart illustrating an agricultural pest and disease monitoring method according to an exemplary embodiment; please refer to [link / reference]. Figure 3 The method includes the following steps: Connect to the edge monitoring terminal to collect real-time crop image sequences in the target area; Based on crop image sequences, a segmentation rule engine is introduced to extract effective feature sequence segments by identifying the fluctuation amplitude of local textures, and then compress them to obtain compressed feature vectors which are uploaded to the cloud. An agricultural dynamic graph network is constructed in the cloud, and compressed feature vectors are mapped to node state attributes of the agricultural dynamic graph network. A tree-like hierarchical matching strategy is executed to construct a target feature tree by combining node state attributes, and then the feature tree is compared layer by layer with a pre-set standard disease classification tree to obtain the disease and pest identification results. The results of pest and disease identification are injected into the agricultural dynamic graph network. The network is then driven to undergo iterative training through a pre-set inference and training model. The network outputs the probability of disease spread and generates spatiotemporal evolution trajectories. Risk levels are predicted based on spatiotemporal evolution trajectories, and parameter adjustment instructions are generated and fed back to the edge monitoring terminal to update the segmentation rule engine.

[0025] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0026] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0027] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. An agricultural pest and disease monitoring system based on cloud-edge collaboration, characterized in that, The system includes: Acquisition module: Used to connect to the edge monitoring terminal and acquire crop image sequences in the target area in real time; Segmentation processing module: Based on crop image sequences, it introduces a segmentation rule engine, identifies the fluctuation amplitude of local textures, extracts effective feature sequence segments, compresses them to obtain compressed feature vectors, and uploads them to the cloud; Matching and comparison module: used to build an agricultural dynamic graph network in the cloud, mapping compressed feature vectors to node state attributes of the agricultural dynamic graph network; executing a tree-like hierarchical matching strategy, combining node state attributes to build a target feature tree, and comparing it with the preset standard disease classification tree layer by layer to obtain the disease and pest identification results; Spatiotemporal evolution module: used to inject the results of pest and disease identification into the agricultural dynamic graph network, drive the iterative training of the agricultural dynamic graph network through the preset inference training model, output the probability of disease spread, and generate the spatiotemporal evolution trajectory. Feedback Adjustment Module: Based on the spatiotemporal evolution trajectory, the module determines and predicts the risk level and generates parameter adjustment instructions to be fed back to the edge monitoring terminal to update the segmentation rule engine.

2. The agricultural pest and disease monitoring system based on cloud-edge collaboration according to claim 1, characterized in that, The real-time acquisition of crop image sequences in the target area includes: The edge monitoring terminal synchronously triggers and controls the imaging unit and the PTZ pan-tilt unit; When triggered by the imaging unit, and in conjunction with the preset scanning trajectory of the PTZ pan-tilt unit, the crop in the target area is continuously exposed and photographed to acquire crop image data; Edge computing is used to preprocess crop image data to generate crop image sequences.

3. The agricultural pest and disease monitoring system based on cloud-edge collaboration according to claim 2, characterized in that, The introduction of the segmentation rule engine includes: Based on the crop image sequence, the first pixel position is identified as the starting position of the current segment, and a crop image subsequence of the first length is extracted within a preset length range as a reference segment. The starting position of the current segment is then adjusted to the ending position of the reference segment. Starting from the adjusted current segment start position, the process is repeated until the entire crop image sequence is traversed: taking the current segment start position as the starting point, each second length within the preset length range is traversed, and the corresponding subsequence is extracted from the crop image sequence to generate several candidate sequence segments; where the second length is a dynamic quantity; the fluctuation index of each candidate sequence segment is calculated, the candidate sequence segment with the largest value is selected as the optimal sequence segment, and the optimal sequence segment with a fluctuation index greater than the standard fluctuation threshold is selected and marked as a valid feature sequence segment.

4. The agricultural pest and disease monitoring system based on cloud-edge collaboration according to claim 3, characterized in that, The volatility index is calculated as follows: The sliding window step size is preset; based on the candidate sequence segment, the entropy value and contrast value in each sliding window are extracted, and the local texture value is obtained by weighted summation. The absolute value of the first difference between the local texture value of the current sliding window and the previous adjacent window is calculated as the fluctuation amplitude of the local texture. The arithmetic mean of the fluctuation amplitudes of all sliding windows in the candidate sequence segment is calculated to obtain the fluctuation index.

5. The agricultural pest and disease monitoring system based on cloud-edge collaboration according to claim 4, characterized in that, The execution of the tree-structured hierarchical matching strategy includes: Based on the node state attributes, several effective feature sequence segments are restored and defined as leaf nodes. Based on the spatial location information recorded in the node state attributes of each effective feature sequence segment, a clustering algorithm is used to fuse the features of adjacent leaf nodes to generate parent nodes, and then the features are fused upwards layer by layer to generate root nodes, thus constructing the target feature tree. The target feature tree is compared layer by layer with the pre-set standard disease classification tree; Structural components are determined by identifying node information at each level; texture components are determined by identifying local texture values; where node information includes node depth, number of child nodes, and connection relationships. Cross-similarity is calculated based on texture components, topological similarity is calculated based on structural components, and the matching difference is generated by weighted summation using the optimal weight combination. Compare the matching difference with the optimal matching threshold: if the matching difference is greater than the optimal matching threshold, then prune the branch; if the matching difference is less than or equal to the optimal matching threshold, then retain the branch and proceed to the next level until the matching reaches the leaf node of the standard disease classification tree, thus determining the disease and pest identification result.

6. The agricultural pest and disease monitoring system based on cloud-edge collaboration according to claim 5, characterized in that, The process of obtaining the optimal weight combination and the optimal matching threshold includes: When performing matching at the current level, the resources of the current matching process are classified and quantified to obtain a matching monitoring vector, including computing power, memory usage, cloud-edge communication bandwidth, and timestamp. By using the set of entity nodes to be matched as a sample set, and taking minimizing the load and maximizing the matching accuracy as dual objective functions, the particle swarm optimization algorithm is used to iteratively solve the problem and output the adaptive response strategy for the current matching level. The adaptive response strategy includes: optimal weight combination and optimal matching threshold.

7. The agricultural pest and disease monitoring system based on cloud-edge collaboration according to claim 5, characterized in that, The preset inference training model is built based on a spatiotemporal graph neural network architecture, including: Collect the identified crop image sequences and identification results, extract node state attributes as feature information, map the identification results of each entity node at each time step into a graph snapshot through time series, and integrate them into the target dataset; Spatiotemporal graph neural networks are used as the inference training model. The target dataset is divided into training set, test set and validation set. The training set is used for iterative training, the test set is used to test the trained inference training model, and the validation set is used to evaluate the inference performance of the model. Using a trained spatiotemporal inference model, the pest and disease identification results obtained at the current moment are injected into the agricultural dynamic graph network as the initial state, driving the agricultural dynamic graph network to perform forward inference, outputting the disease spread probability of each entity node in the next N1 time steps, and connecting the probability distributions of each time step in time sequence to generate a spatiotemporal evolution trajectory; where N1 is a positive integer greater than 0.

8. The agricultural pest and disease monitoring system based on cloud-edge collaboration according to claim 7, characterized in that, The updated segmentation rule engine includes: Based on the predicted risk level, the corresponding parameter adjustment instructions are obtained and encapsulated into a control message, which is sent to the edge monitoring terminal through the downlink channel. The edge monitoring terminal receives the control message, and after protocol parsing and integrity verification, extracts a new parameter group containing a new sliding window step size and a new preset length range. When the current frame crop image sequence is processed and the next frame is loaded, the original sliding window step size and preset length range are replaced.

9. The agricultural pest and disease monitoring system based on cloud-edge collaboration according to claim 8, characterized in that, The predicted risk levels include high risk, medium risk, and low risk.

10. A method for monitoring agricultural pests and diseases based on cloud-edge collaboration, characterized in that, The method includes: Connect to the edge monitoring terminal to collect real-time crop image sequences in the target area; Based on crop image sequences, a segmentation rule engine is introduced to extract effective feature sequence segments by identifying the fluctuation amplitude of local textures, and then compress them to obtain compressed feature vectors which are uploaded to the cloud. An agricultural dynamic graph network is constructed in the cloud, and compressed feature vectors are mapped to node state attributes of the agricultural dynamic graph network. A tree-like hierarchical matching strategy is executed to construct a target feature tree by combining node state attributes, and then the feature tree is compared layer by layer with a pre-set standard disease classification tree to obtain the disease and pest identification results. The results of pest and disease identification are injected into the agricultural dynamic graph network. The network is then driven to undergo iterative training through a pre-set inference and training model. The network outputs the probability of disease spread and generates spatiotemporal evolution trajectories. Risk levels are predicted based on spatiotemporal evolution trajectories, and parameter adjustment instructions are generated and fed back to the edge monitoring terminal to update the segmentation rule engine.

Images