Multi-agent collaborative decision-making method and system for agricultural insect pest assessment and prevention

Abstract

The invention provides a multi-agent collaborative decision-making method and system for agricultural insect pest evaluation and prevention, and relates to the technical field of AI.The method comprises the steps that multi-modal data are collected through edge sensing agents, insect pest target recognition and physiological stress quantification are conducted, and a standardized triple data structure is formed; the cloud analysis agent performs space-time correlation analysis based on the pest diffusion model, predicts a diffusion trend and calculates a risk level; and the decision-making agent generates a prevention and treatment strategy including region division, time sequence arrangement and resource allocation under the resource constraint. Accurate prediction and scientific decision making of pest control are realized, and the crop protection efficiency is improved.

Description

Technical Field

[0001] This invention relates to the field of AI technology, and in particular to a multi-agent collaborative decision-making method and system for agricultural pest assessment and control. Background Technology

[0002] In modern agricultural production, accurate identification and effective control of pests play a decisive role in ensuring crop yield and quality. Although existing technologies have made some progress in pest monitoring, many shortcomings remain. Most existing agricultural intelligent systems adopt a "single-point intelligence + manual intervention" model, where perception, analysis, decision-making, and execution are performed by different devices or services, lacking a native collaborative mechanism. This model leads to asynchronous task states, inconsistent strategy versions, missing feedback loops, and inconsistent collaborative interfaces. Although some systems have introduced middleware such as MQTT and Kafka, additional development of message formats, state machines, access control, and fault tolerance mechanisms is still required, failing to address the fundamental problem of "lack of native collaborative logic." Especially in complex farmland environments, the ability to monitor and accurately identify pests in real time urgently needs improvement.

[0003] Traditional manual inspection methods rely on the experience and subjective judgment of operators. This approach is inefficient, costly, and easily affected by human factors, making it difficult to meet the real-time and accurate pest monitoring requirements of modern agriculture. Furthermore, automated pest detection technology based on RGB images lacks robustness in complex farmland environments, such as changes in light intensity, crop leaf shading, or the emergence of new pests, leading to decreased detection accuracy.

[0004] Existing studies mostly use multispectral remote sensing to monitor crop growth, but they cannot capture the spectral differences caused by pests in the early stages at the single-leaf scale; while single-image recognition can locate pests, it lacks a quantitative assessment of plant health status. Summary of the Invention

[0005] This invention provides a multi-agent collaborative decision-making method and system for agricultural pest assessment and control, which can solve the problems in the prior art.

[0006] A first aspect of this invention provides a multi-agent collaborative decision-making method for agricultural pest assessment and control, comprising: Multimodal sensing data of the target crop area is simultaneously collected by sensing agents deployed on edge computing nodes. The multimodal sensing data includes image data, spectral data, and environmental data. The perceptual agent identifies pest targets in the image data based on a pre-trained target detection model, outputs pest type and spatial distribution density, and quantifies physiological stress indicators in the spectral data to obtain pest identification results and physiological stress quantification results. The pest identification results, the physiological stress quantification results, and the environmental data are encapsulated into a standardized triplet data structure, which includes pest feature dimensions, physiological feature dimensions, and environmental feature dimensions. The analytical agent, deployed on a cloud computing node, receives the triplet data structure, performs spatiotemporal correlation analysis on the pest feature dimension and the environmental feature dimension based on the pest spread dynamics model, generates pest spread trend prediction results, and calculates the comprehensive risk level based on the physiological feature dimension and the pest spread trend prediction results. The decision-making agent, deployed on a cloud computing node, receives the triplet data structure, the pest spread trend prediction results, and the comprehensive risk level. Under preset control resource constraints, it generates a control strategy plan, which includes the division of control areas, the arrangement of control time sequence, and the allocation of control resources.

[0007] The perceptual agent identifies pest targets in the image data based on a pre-trained target detection model, outputting pest type and spatial distribution density. It also quantifies physiological stress indicators in the spectral data to obtain pest identification results and physiological stress quantification results, including: The perceptual agent performs multi-scale feature extraction on the image data to generate a feature map containing spatial location information and semantic feature information; Based on the pre-trained target detection model, target detection and bounding box regression are performed on the feature map, outputting the spatial coordinates and category confidence of the pest targets, and calculating the number of pest targets per unit area based on the spatial coordinates to obtain the spatial distribution density; The sensing agent performs band separation and feature band extraction on the spectral data to obtain a spectral feature vector reflecting the physiological state of the plant; The spectral feature vector is nonlinearly mapped to output a quantitative value of a physiological stress index characterizing the degree of plant stress response. A spatial correspondence is established between the quantitative value of the physiological stress index and the spatial coordinates of the pest target. The pest type, spatial distribution density, category confidence level, and physiological stress index quantification value are correlated and integrated. Based on the spatial correspondence, the physiological stress index quantification value is assigned to the spatial region where the corresponding pest target is located, thus forming the pest identification result and the physiological stress quantification result.

[0008] Based on the pest spread dynamics model, a spatiotemporal correlation analysis is performed between the pest characteristic dimension and the environmental characteristic dimension to generate pest spread trend prediction results, including: We perform temporal alignment and spatial gridding mapping on pest characteristic data and environmental characteristic data to construct a spatiotemporal feature tensor containing temporal evolution information and spatial distribution information; Based on the aforementioned pest diffusion dynamics model, the spatiotemporal feature tensor is fitted with parameters to estimate the pest diffusion rate parameter and the environmental impact coefficient parameter. The pest diffusion rate parameter characterizes the spread intensity of the pest in the spatial dimension, and the environmental impact coefficient parameter characterizes the regulatory effect of the environmental feature dimension on pest diffusion. Using the pest characteristic dimension and the environmental characteristic dimension at the current moment as initial conditions, and combining the pest spread rate parameter and the environmental influence coefficient parameter, the pest spread dynamics model is used to simulate the spatial distribution evolution of pests in future time periods, resulting in predicted pest spatial distribution data containing multiple future time nodes; Statistical analysis is performed on the spatial distribution prediction data of the pests to extract the spatial location change trajectory and pest density growth trend of the pest risk area, thus forming the pest spread trend prediction result.

[0009] The spatial distribution evolution of pests over future time periods is simulated using the aforementioned pest diffusion dynamics model, yielding predicted spatial distribution data for pests covering multiple future time points, including: The future time period is divided into multiple discrete time nodes according to a preset time interval, forming a time node sequence; Map the pest characteristics and environmental characteristics at the current moment to a spatial grid structure to generate an initial state matrix of pest spatial distribution and an environmental state matrix; For each time node in the time node sequence, the spatial distribution state of the pest at the previous time node is calculated based on the pest diffusion dynamics model, and the environmental state parameters of the corresponding time node are introduced to dynamically regulate the diffusion process, thereby obtaining the evolution state of the pest spatial distribution at that time node. The pest diffusion dynamics model is constructed based on a long short-term memory network. The spatial distribution evolution of pests corresponding to each time node in the time node sequence is integrated in chronological order to construct a three-dimensional data structure containing spatial coordinates, pest density, and time identifier, thus forming the pest spatial distribution prediction data containing multiple future time nodes.

[0010] Under the preset constraints of prevention and control resources, the generated prevention and control strategy schemes include: Based on the pest spread trend prediction results and the comprehensive risk level, a multi-objective optimization objective set is constructed, which includes a pest control effect objective function and a resource consumption objective function. The pest control effect objective function characterizes the pest control efficiency, and the resource consumption objective function characterizes the cost of using pest control resources. The preset constraints on prevention and control resources are analyzed to extract constraints on the total amount of prevention and control resources, the time window for prevention and control operations, and spatial accessibility, forming a set of constraints. Based on the set of multi-objective optimization objectives and the set of constraints, the spatial boundaries of the prevention and control area, the operation time nodes of each prevention and control area, and the resource allocation of each prevention and control area are determined simultaneously by solving the Pareto optimal solution set. The solution obtained by integrating the spatial boundaries of the prevention and control area is used to form the prevention and control area division. The operation time nodes are arranged in chronological order to form the prevention and control time sequence arrangement. The resource allocation is mapped to the corresponding prevention and control area to form the prevention and control resource allocation. The combination constitutes the prevention and control strategy scheme.

[0011] Based on the multi-objective optimization objective set and the constraint set, by solving the Pareto optimal solution set, the spatial boundary of the prevention and control area, the operation time nodes of each prevention and control area, and the resource allocation of each prevention and control area are simultaneously determined, including: Define a decision variable vector that includes spatial, temporal, and resource variables. The spatial variables represent the spatial boundary coordinates of the prevention and control area, the temporal variables represent the operational time nodes of each prevention and control area, and the resource variables represent the resource allocation of each prevention and control area. Substitute the decision variable vector into the prevention and control effect objective function and the resource consumption objective function to establish a mathematical mapping relationship between the objective function value and the decision variable vector; Multi-objective optimization is performed within the feasible solution space. Pareto optimal solution set is obtained through iterative search, which makes the objective function of prevention and control effect and the objective function of resource consumption reach a non-dominated state. Each solution in the Pareto optimal solution set corresponds to a set of decision variable vectors. A target solution is selected from the Pareto optimal solution set according to a preset decision preference criterion. The decision variable vector corresponding to the target solution includes optimal spatial variables, optimal time variables, and optimal resource variables. The optimal spatial variable is resolved as the spatial boundary of the prevention and control area, the optimal time variable is resolved as the operation time node of each prevention and control area, and the optimal resource variable is resolved as the resource allocation amount of each prevention and control area.

[0012] A second aspect of the present invention provides a multi-agent collaborative decision-making system for agricultural pest assessment and control, comprising: The first unit is used to synchronously collect multimodal sensing data of the target crop area through a sensing agent deployed on an edge computing node. The multimodal sensing data includes image data, spectral data, and environmental data. The second unit is used by the sensing agent to identify pest targets in the image data based on a pre-trained target detection model, output the pest type and spatial distribution density, and quantify the physiological stress index of the spectral data to obtain pest identification results and physiological stress quantification results. The third unit is used to encapsulate the pest identification results, the physiological stress quantification results, and the environmental data into a standardized triplet data structure, wherein the triplet data structure includes pest feature dimensions, physiological feature dimensions, and environmental feature dimensions; The fourth unit is used to receive the triplet data structure through an analytical agent deployed on a cloud computing node, perform spatiotemporal correlation analysis on the pest feature dimension and the environmental feature dimension based on the pest spread dynamics model, generate pest spread trend prediction results, and calculate the comprehensive risk level based on the physiological feature dimension and the pest spread trend prediction results; The fifth unit is used to receive the triplet data structure, the pest spread trend prediction results and the comprehensive risk level through a decision-making intelligent agent deployed on a cloud computing node, and generate a prevention and control strategy plan under preset prevention and control resource constraints. The prevention and control strategy plan includes the division of prevention and control areas, the arrangement of prevention and control time sequence and the allocation of prevention and control resources.

[0013] A third aspect of the present invention, An electronic device is provided, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0014] Fourth aspect of the embodiments of the present invention, A computer-readable storage medium is provided, having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0015] The beneficial effects of this application are as follows: By employing edge computing nodes as sensing agents to synchronously collect multimodal data and combining it with a pre-trained target detection model for pest identification and physiological stress quantification, the accuracy and comprehensiveness of pest identification are improved, providing high-quality basic data for subsequent analysis and effectively solving the problem of one-sided pest assessment caused by traditional single data sources.

[0016] The introduction of a ternary data structure enables a unified expression of pest characteristics, physiological characteristics, and environmental characteristics. The cloud-based analytical agent performs spatiotemporal correlation analysis based on the pest diffusion dynamics model, which can scientifically predict pest diffusion trends, calculate comprehensive risk levels, and improve the foresight and accuracy of pest risk assessment.

[0017] The prevention and control strategies generated by the decision-making agent under preset resource constraints achieve optimized allocation of prevention and control areas, timing, and resources, improve prevention and control efficiency, reduce resource waste, and reduce unnecessary pesticide use. This aligns with the concept of green agriculture development and has significant economic and ecological benefits. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the multi-agent collaborative decision-making method for agricultural pest assessment and control, as described in an embodiment of the present invention. Figure 2 This is a flowchart of the detection method of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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.

[0020] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0021] Figure 1 This is a flowchart illustrating a multi-agent collaborative decision-making method for agricultural pest assessment and control, as described in an embodiment of the present invention. Figure 1 As shown, the method includes: Multimodal sensing data of the target crop area is simultaneously collected by sensing agents deployed on edge computing nodes. The multimodal sensing data includes image data, spectral data, and environmental data. The perceptual agent identifies pest targets in the image data based on a pre-trained target detection model, outputs pest type and spatial distribution density, and quantifies physiological stress indicators in the spectral data to obtain pest identification results and physiological stress quantification results. The pest identification results, the physiological stress quantification results, and the environmental data are encapsulated into a standardized triplet data structure, which includes pest feature dimensions, physiological feature dimensions, and environmental feature dimensions. The analytical agent, deployed on a cloud computing node, receives the triplet data structure, performs spatiotemporal correlation analysis on the pest feature dimension and the environmental feature dimension based on the pest spread dynamics model, generates pest spread trend prediction results, and calculates the comprehensive risk level based on the physiological feature dimension and the pest spread trend prediction results. The decision-making agent, deployed on a cloud computing node, receives the triplet data structure, the pest spread trend prediction results, and the comprehensive risk level. Under preset control resource constraints, it generates a control strategy plan, which includes the division of control areas, the arrangement of control time sequence, and the allocation of control resources.

[0022] In one optional implementation, the perceptual agent identifies pest targets in the image data based on a pre-trained target detection model, outputs pest type and spatial distribution density, and quantifies physiological stress indicators in the spectral data to obtain pest identification results and physiological stress quantification results, including: The perceptual agent performs multi-scale feature extraction on the image data to generate a feature map containing spatial location information and semantic feature information; Based on the pre-trained target detection model, target detection and bounding box regression are performed on the feature map, outputting the spatial coordinates and category confidence of the pest targets, and calculating the number of pest targets per unit area based on the spatial coordinates to obtain the spatial distribution density; The sensing agent performs band separation and feature band extraction on the spectral data to obtain a spectral feature vector reflecting the physiological state of the plant; The spectral feature vector is nonlinearly mapped to output a quantitative value of a physiological stress index characterizing the degree of plant stress response. A spatial correspondence is established between the quantitative value of the physiological stress index and the spatial coordinates of the pest target. The pest type, spatial distribution density, category confidence level, and physiological stress index quantification value are correlated and integrated. Based on the spatial correspondence, the physiological stress index quantification value is assigned to the spatial region where the corresponding pest target is located, thus forming the pest identification result and the physiological stress quantification result.

[0023] In practical applications, the image data acquisition device can be a high-resolution camera mounted on an agricultural drone, with a resolution of 4000×3000 pixels, and the spectral data acquisition device can be a hyperspectral imager with a wavelength range of 400-1000nm and 256 bands.

[0024] The multi-scale feature extraction process for image data employs a feature pyramid network structure, specifically including feature extraction at five scale levels. The base network uses a ResNet50 structure, outputting feature maps at layers C2, C3, C4, and C5, corresponding to scales of 1 / 4, 1 / 8, 1 / 16, and 1 / 32 of the original image, respectively. A top-down feature fusion path integrates high-level semantic features with low-level spatial detail features, generating feature maps at five levels, P2 to P6. For example, the P4 feature map is upsampled and fused with the C3 feature map using pixel-level addition. The fused features are then adjusted using 3×3 convolutional kernels to obtain a feature map containing rich spatial location and semantic information.

[0025] The pre-trained object detection model is based on an improved Faster R-CNN architecture and was pre-trained using a dataset containing 20,000 images of 12 common crop pests, including rice planthoppers, rice leaf rollers, and armyworms. When performing object detection on the feature maps, candidate boxes are first generated through a region proposal network. Anchor box sizes are set to 32×32, 64×64, and 128×128 pixels, with aspect ratios of 1:1, 1:2, and 2:1, resulting in nine different anchor box shapes. A sliding window approach with a stride of 16 pixels is used to generate the candidate boxes. The generated candidate boxes undergo non-maximum suppression, retaining those with an IoU threshold below 0.7.

[0026] The bounding box regression uses a four-parameter representation: center point coordinates (x, y) and width and height (w, h). The regression target is the offset between the anchor box and the ground truth box. In a specific example, the bounding box coordinates of the detected rice planthopper target are (256, 312, 48, 52), indicating that the center point is at (256, 312), the width is 48 pixels, the height is 52 pixels, and the class confidence is 0.94.

[0027] Spatial distribution density is calculated based on the number of pest targets per unit area. In practical applications, 1 square meter is used as the basic statistical unit, and the spatial coordinates of the image are converted into actual geographic coordinates. Based on the image capture altitude, camera parameters, and ground resolution, the correspondence between pixels and actual distances can be calculated. For example, at a flight altitude of 5 meters and a ground resolution of 0.5 cm / pixel, the pixel range corresponding to a 1 square meter area is approximately 200 × 200 pixels. The spatial distribution density is obtained by calculating the number of pest targets within this area. In a real-world case, 18 rice planthopper targets were detected within a 1 square meter area, resulting in a spatial distribution density of 18 planthoppers / square meter.

[0028] In terms of spectral data processing, the spectral data of 256 bands were first separated, with a focus on extracting characteristic bands such as 550nm (green light reflectance peak), 680nm (chlorophyll absorption band), 750nm (red edge position), and 800nm ​​(near-infrared plateau). By calculating the reflectance differences and ratios of these characteristic bands and performing normalization, a spectral feature vector reflecting the plant's physiological state was generated. In practical applications, 12 key features were extracted from the original 256-dimensional spectral data to form a 12-dimensional spectral feature vector.

[0029] The quantification of physiological stress indicators employs a nonlinear mapping method, mapping spectral feature vectors to multiple physiological stress indicators. Specific indicators include chlorophyll content, water content, and photosynthetic efficiency. In practical applications, by establishing a pre-defined correspondence between spectral features and physiological parameters, physiological stress indicators of varying degrees can be quantified. For example, a chlorophyll content index of 0.72 (normal range 0.8-1.0) in a certain region indicates moderate chlorophyll degradation; a water content index of 0.85 (normal range 0.9-1.0) indicates mild water stress.

[0030] Spatial correspondence is established through geographic information registration. The acquisition location of spectral data is matched with the spatial coordinates of pest targets in the image data, accurate to the sub-meter level. In a real-world example, the spatial coordinates of a rice planthopper aggregation area range from UTM coordinates (500123.45, 4500678.90) to (500124.45, 4500679.90), with corresponding physiological stress indicators of chlorophyll content index of 0.68 and water content index of 0.82.

[0031] The final pest identification results and physiological stress quantification results are integrated through data correlation to form a complete report. In a practical application case, the correlation integration results show that rice planthoppers were identified in the area near coordinates (500123.45, 4500678.90), with a category confidence level of 0.94, a spatial distribution density of 18 planthoppers / square meter, a chlorophyll content index of 0.68 and a moisture content index of 0.82 for the corresponding plants, and an overall assessment of moderate pest infestation.

[0032] In one optional implementation, a spatiotemporal correlation analysis is performed on the pest characteristic dimension and the environmental characteristic dimension based on a pest spread dynamics model to generate pest spread trend prediction results, including: We perform temporal alignment and spatial gridding mapping on pest characteristic data and environmental characteristic data to construct a spatiotemporal feature tensor containing temporal evolution information and spatial distribution information; Based on the aforementioned pest diffusion dynamics model, the spatiotemporal feature tensor is fitted with parameters to estimate the pest diffusion rate parameter and the environmental impact coefficient parameter. The pest diffusion rate parameter characterizes the spread intensity of the pest in the spatial dimension, and the environmental impact coefficient parameter characterizes the regulatory effect of the environmental feature dimension on pest diffusion. Using the pest characteristic dimension and the environmental characteristic dimension at the current moment as initial conditions, and combining the pest spread rate parameter and the environmental influence coefficient parameter, the pest spread dynamics model is used to simulate the spatial distribution evolution of pests in future time periods, resulting in predicted pest spatial distribution data containing multiple future time nodes; Statistical analysis is performed on the spatial distribution prediction data of the pests to extract the spatial location change trajectory and pest density growth trend of the pest risk area, thus forming the pest spread trend prediction result.

[0033] Obtain a pest monitoring dataset containing pest characteristic dimension data and environmental characteristic dimension data. The pest characteristic dimension data includes pest species identifiers, pest density values, and corresponding timestamps and geographical coordinates for different monitoring points. The environmental characteristic dimension data includes temperature, humidity, wind direction, wind speed, precipitation, and vegetation cover type, along with corresponding timestamps and geographical coordinates.

[0034] In the spatiotemporal data preprocessing stage, pest characteristic dimension data and environmental characteristic dimension data from different sources are aligned in time dimension. Specifically, for data with different temporal sampling frequencies, a time window aggregation method is used to unify all data to the same time granularity, such as alignment on a daily basis. For data with different spatial distribution characteristics, the observation area is divided into uniformly sized grid cells, with each grid cell having a side length of 2 kilometers, forming a regular grid with a spatial resolution of 2×2 square kilometers. Each original data point is mapped to the corresponding grid cell according to its geographic coordinates, and data falling within the same grid cell are aggregated to calculate the average and maximum pest density and the average environmental factor within the grid cell. Through the above processing, a spatiotemporal feature tensor containing temporal evolution information and spatial distribution information is constructed. The tensor has dimensions of T×M×N×K, where T represents the length of the time series, M×N represents the number of rows and columns of the spatial grid, and K represents the number of feature dimensions.

[0035] In the dynamic parameter fitting stage, a pest diffusion dynamics model was used to fit the parameters of the constructed spatiotemporal feature tensor. This dynamics model describes the spatial diffusion process of pests and the influence of environmental factors on the diffusion process. A pest diffusion dynamics model based on partial differential equations was designed, which considers the natural reproduction rate, diffusion and migration characteristics of pests, and the regulatory role of environmental factors. Historical observation data was input into the model, and the pest diffusion rate parameter D and the environmental influence coefficient parameter α were estimated by minimizing the mean square error between the model prediction and the actual observation. In a practical case, the parameters of the rice planthopper diffusion process in a certain area were fitted, and the diffusion rate parameter D was found to be 0.85 square kilometers / day, indicating that the average spatial diffusion area of ​​rice planthoppers is approximately 0.85 square kilometers per day; in the environmental influence coefficient parameter α, the temperature influence coefficient is 0.12, indicating that for every 1°C increase, the reproduction rate of rice planthoppers increases by approximately 12%; and the humidity influence coefficient is 0.08, indicating that for every 1% increase in relative humidity, the reproduction rate of rice planthoppers increases by approximately 8%.

[0036] In the pest spread simulation phase, current pest and environmental characteristic data are used as initial conditions. Combined with estimated pest spread rate parameters and environmental impact coefficient parameters, a pest spread dynamics model is used to simulate the spatial distribution changes of pests over future time periods. Numerical calculation methods are employed, discretizing the spatial region into grid cells and the time axis into equally spaced time steps. The pest density change in each grid cell within each time step is iteratively calculated. Specifically, a time step of one day is set. Within each time step, the following factors are considered in relation to the pest density within the grid cell: natural reproduction and growth based on the current pest density, pest migration and spread between adjacent grid cells based on the spread rate parameter, and the regulatory effect of environmental factors based on the environmental impact coefficient parameter. Through continuous iterative calculations, predicted spatial distribution data of pests for each day over the next 15 days are obtained. In the rice planthopper case mentioned above, it is predicted that within the next 15 days, varying degrees of rice planthopper spread will occur within a 5-kilometer radius of the initial outbreak area, with the southeast direction experiencing a relatively faster spread due to wind direction.

[0037] In the trend analysis phase, statistical analysis is performed on the predicted spatial distribution data of pests to extract the spatial location change trajectory of pest risk areas and the pest density growth trend. First, based on a preset pest density threshold (e.g., 10 pests per square meter), high-risk pest areas at each time point are identified, and the spatial expansion path of these areas over time is tracked to form the dynamic trajectory of the pest spread front. The time derivative of pest density within each grid cell is calculated to determine the areas with the fastest pest density growth rate; these areas typically represent potential new outbreak hotspots. The urgency of control measures in different areas is calculated based on the time lead before pest density exceeds the economic threshold. Based on the analysis results, a pest spread trend prediction result is generated, including the direction and speed of pest spread, key control areas, and time windows. This result is then visualized using a spatial heatmap to intuitively present the spatiotemporal changes in pest spread.

[0038] In one optional implementation, the spatial distribution evolution of pests over a future time period is simulated using the pest diffusion dynamics model to obtain predicted spatial distribution data of pests containing multiple future time points, including: The future time period is divided into multiple discrete time nodes according to a preset time interval, forming a time node sequence; Map the pest characteristics and environmental characteristics at the current moment to a spatial grid structure to generate an initial state matrix of pest spatial distribution and an environmental state matrix; For each time node in the time node sequence, the spatial distribution state of the pest at the previous time node is calculated based on the pest diffusion dynamics model, and the environmental state parameters of the corresponding time node are introduced to dynamically regulate the diffusion process, thereby obtaining the evolution state of the pest spatial distribution at that time node. The pest diffusion dynamics model is constructed based on a long short-term memory network. The spatial distribution evolution of pests corresponding to each time node in the time node sequence is integrated in chronological order to construct a three-dimensional data structure containing spatial coordinates, pest density, and time identifier, thus forming the pest spatial distribution prediction data containing multiple future time nodes.

[0039] The future time period is divided into multiple discrete time nodes according to a preset time interval, forming a time node sequence. Specifically, if the future prediction period is set to 30 days and the time interval is 1 day, a sequence containing 30 time nodes is formed, denoted as T_1, T_2, ..., T_30, where T_1 represents the first day of the prediction, and T_30 represents the 30th day of the prediction. This time discretization transforms the continuous time prediction problem into a state prediction problem with multiple discrete time points, making it easier for computers to process.

[0040] The current pest characteristics and environmental characteristics are mapped to a spatial grid structure to generate an initial state matrix for the spatial distribution of pests and an environmental state matrix. In practical applications, assuming the study area is a 100 square kilometer agricultural region, it is divided into a 10×10 grid, with each grid cell representing 1 square kilometer. The pest characteristics dimension includes information such as pest species and current density. For example, taking corn borers as an example, at the current time T_0, the pest density recorded in each grid cell ranges from 0 to 100 pests / square meter. The environmental characteristics dimension includes parameters such as temperature, humidity, wind direction, wind speed, precipitation, and vegetation cover. For example, the temperature data in the study area is distributed as 22-28℃, the relative humidity is 60%-85%, the prevailing wind direction is northeast, the wind speed is 2-5 m / s, the precipitation in the last 7 days is 15-30 mm, and the vegetation cover is 65%-90%. These data were acquired through field sampling, meteorological station data, and satellite remote sensing data, and interpolated into a 10×10 grid structure to form an initial state matrix.

[0041] For each time point in the time-node sequence, the spatial distribution of pests at the previous time point is calculated based on a pest diffusion dynamics model. Environmental state parameters for the corresponding time point are introduced to dynamically regulate the diffusion process, thus obtaining the evolutionary state of the pest spatial distribution at that time point. This pest diffusion dynamics model is constructed based on a long short-term memory network, which can effectively capture the spatiotemporal characteristics of pest diffusion.

[0042] In this model, the Long Short-Term Memory (LSTM) network comprises an input layer, hidden layers, and an output layer. The input layer receives pest density values ​​and environmental parameters for the current grid cell and its surrounding grid cells; the hidden layer consists of 64 LSM units, responsible for memorizing long-term and short-term pest spread patterns; the output layer generates the predicted pest density value for the next time point. The model was trained using pest monitoring data from the past three years and corresponding environmental data, totaling approximately 1095 daily samples. The model parameters were optimized using a stochastic gradient descent algorithm, and the root mean square error was used as the learning loss function. The model converged after 500 training iterations, achieving a prediction accuracy of 87.5% on the validation set.

[0043] In the specific diffusion calculation process, taking grid cell (i,j) as an example, the pest density D(i,j,T_k) at time node T_k is jointly determined by the density at the previous time node T_(k-1) and environmental factors. The model considers three main factors: natural reproduction, mortality, and diffusion. Among them, the natural reproduction rate is affected by temperature and humidity. When the temperature is between 25-28℃ and the relative humidity is between 70%-80%, the reproduction rate reaches its highest level, reaching 0.15 (increasing by 15% per day). The mortality rate is affected by precipitation and extreme temperature. Heavy precipitation (>30 mm / day) or excessively high temperature (>35℃) can increase the mortality rate to 0.2 (decreasing by 20% per day). The diffusion factor is mainly affected by wind direction and wind speed. The amount of pest migration from the surrounding eight grid cells to the central cell is calculated according to the wind direction weighting method. For every 1 m / s increase in wind speed, the diffusion distance increases by approximately 0.5 km / day.

[0044] For example, if the pest density of a certain grid cell (5,5) at time T_1 is 20 pests / m², the ambient temperature is 26℃, the relative humidity is 75%, the wind direction is northeast, and the wind speed is 3 m / s, then the pest density of this cell at time T_2 will be affected by its own reproduction (an increase of 20 × 0.15 = 3 pests / m²), natural death (a decrease of 20 × 0.05 = 1 pest / m²), and the migration in and out of surrounding grid cells (assuming a net migration of 2 pests / m²). The final predicted density value at time T_2 is 24 pests / m².

[0045] For each time point T_k, a complete 10×10 grid matrix of pest density distribution is generated. For example, the prediction results for T_10 (day 10) show that in the northeastern part of the study area, due to suitable temperature and humidity conditions, the pest density has increased from the initial average of 15 pests / m² to 45 pests / m², while in the southwest, due to lower temperatures and unfavorable wind direction, the pest density has only increased from the initial 10 pests / m² to 18 pests / m².

[0046] Finally, the spatial distribution evolution of pests corresponding to each time point in the time sequence is integrated in chronological order to construct a three-dimensional data structure containing spatial coordinates, pest density, and time markers, forming pest spatial distribution prediction data for multiple future time points. This three-dimensional data structure can be represented as D(i,j,T_k), where i∈[1,10], j∈[1,10], and k∈[1,30], representing the grid horizontal coordinate, vertical coordinate, and time node index, respectively. The data is stored in JSON format, containing coordinate information, timestamps, and corresponding pest density values, facilitating subsequent visualization and early warning analysis. Through this three-dimensional data structure, agricultural managers can intuitively understand the spatial distribution trend of pests over the next 30 days and formulate targeted control measures in a timely manner.

[0047] In one optional implementation, under preset resource constraints, generating a prevention and control strategy includes: Based on the pest spread trend prediction results and the comprehensive risk level, a multi-objective optimization objective set is constructed, which includes a pest control effect objective function and a resource consumption objective function. The pest control effect objective function characterizes the pest control efficiency, and the resource consumption objective function characterizes the cost of using pest control resources. The preset constraints on prevention and control resources are analyzed to extract constraints on the total amount of prevention and control resources, the time window for prevention and control operations, and spatial accessibility, forming a set of constraints. Based on the set of multi-objective optimization objectives and the set of constraints, the spatial boundaries of the prevention and control area, the operation time nodes of each prevention and control area, and the resource allocation of each prevention and control area are determined simultaneously by solving the Pareto optimal solution set. The solution obtained by integrating the spatial boundaries of the prevention and control area is used to form the prevention and control area division. The operation time nodes are arranged in chronological order to form the prevention and control time sequence arrangement. The resource allocation is mapped to the corresponding prevention and control area to form the prevention and control resource allocation. The combination constitutes the prevention and control strategy scheme.

[0048] The starting point for generating pest control strategies is constructing a multi-objective optimization set. Based on the predicted pest spread trend and comprehensive risk level, an objective function for control effectiveness and a objective function for resource consumption are established. The objective function for control effectiveness quantifies pest control efficiency and can be represented by calculating the reduction rate of pest density or the reduction rate of pest-affected area after control. For example, for pine wilt disease, the objective function for control effectiveness can be set as the percentage reduction in the number of affected pine trees. If the initial number of infected pine trees is 100 and the number drops to 10 after control, the effectiveness rate is 90%. The objective function for resource consumption quantifies the cost of control resources, including inputs in manpower, materials, and finances. In practical applications, labor costs, pesticide costs, and equipment usage fees can be included in the calculation. For example, if a forest area requires 20 workers for pest control operations, with each worker earning 200 yuan per day, pesticide costs of 5,000 yuan, and equipment usage fees of 2,000 yuan, the total cost is 11,000 yuan. The two objective functions form a multi-objective optimization set, providing evaluation criteria for subsequent optimization.

[0049] The formation of the constraint set is a crucial limiting factor in formulating prevention and control strategies. First, the pre-set constraints on prevention and control resources are analyzed to extract the total resource constraints. This includes the total amount of available pesticides, the number of prevention and control equipment, and the number of professional personnel. For example, if a region can utilize 10 spraying drones, each capable of covering 100 mu (approximately 6.7 hectares) per day, then the maximum daily coverage is 1000 mu (approximately 67 hectares). Second, the time window constraint for prevention and control operations is extracted. For instance, due to weather conditions, a region may only have 10 days suitable for ground pesticide spraying in May. Third, spatial accessibility constraints are extracted, considering the impact of terrain and transportation conditions on prevention and control operations. For example, in mountainous areas with slopes exceeding 30 degrees, mechanized equipment cannot be used, requiring manual operation. These constraints collectively constitute the constraint set, setting the boundary for the optimization solution.

[0050] Based on a multi-objective optimization of the objective set and constraint set, the spatial boundaries, operational time points, and resource allocation of the control area are simultaneously determined by solving the Pareto optimal solution set. During the solution process, a non-dominated sorting genetic algorithm is used for multi-objective optimization calculations. Through iterative optimization, a series of non-dominated solutions are obtained, forming the Pareto optimal solution set. From this solution set, the spatial range of the control area, the optimal operational time, and the required resource amount for each area can be obtained simultaneously. For example, for the control of pine wilt disease in a certain forest area, the Pareto solution set indicates that the forest area should be divided into three control areas: high-risk area A (1000 mu), medium-risk area B (2000 mu), and low-risk area C (3000 mu); the optimal operational times are May 1-5, May 6-15, and May 16-25, respectively; the resource allocation is 200 kg of pesticide and 10 staff members for area A, 300 kg of pesticide and 15 staff members for area B, and 400 kg of pesticide and 20 staff members for area C. This simultaneous optimization method can find the optimal combination of prevention and control strategies while taking into account both prevention and control effectiveness and resource consumption.

[0051] The final formulation of the prevention and control strategy requires the integration of the aforementioned optimization results. The spatial boundaries of the prevention and control areas obtained from the solution are integrated to form a prevention and control area division, such as dividing a forest farm into four prevention and control areas: east, west, south, and north. The boundary of each area is defined by a sequence of geographic coordinate points. The operation time nodes are arranged chronologically to form a prevention and control schedule, such as May 1-3 for the east area, May 4-7 for the west area, May 8-12 for the south area, and May 13-17 for the north area. A mapping relationship is established between resource allocation and the corresponding prevention and control areas to form a prevention and control resource allocation, such as allocating 100 kg of pesticide, 5 spraying machines, and 10 workers for the east area; and 150 kg of pesticide, 7 spraying machines, and 15 workers for the west area. These three parts combined constitute a complete prevention and control strategy, providing systematic guidance for actual prevention and control work.

[0052] In practical applications, the plan can be dynamically adjusted based on feedback information during the control process. For example, if the pest situation in a certain area exceeds expectations during the control process, resource input can be appropriately increased or operation time adjusted to ensure the overall control effect. Through this closed-loop feedback mechanism, the adaptability and effectiveness of the control plan can be further improved.

[0053] In one optional implementation, based on the multi-objective optimization objective set and the constraint set, the spatial boundary of the prevention and control area, the operation time nodes of each prevention and control area, and the resource allocation of each prevention and control area are simultaneously determined by solving the Pareto optimal solution set, including: Define a decision variable vector that includes spatial, temporal, and resource variables. The spatial variables represent the spatial boundary coordinates of the prevention and control area, the temporal variables represent the operational time nodes of each prevention and control area, and the resource variables represent the resource allocation of each prevention and control area. Substitute the decision variable vector into the prevention and control effect objective function and the resource consumption objective function to establish a mathematical mapping relationship between the objective function value and the decision variable vector; Multi-objective optimization is performed within the feasible solution space. Pareto optimal solution set is obtained through iterative search, which makes the objective function of prevention and control effect and the objective function of resource consumption reach a non-dominated state. Each solution in the Pareto optimal solution set corresponds to a set of decision variable vectors. A target solution is selected from the Pareto optimal solution set according to a preset decision preference criterion. The decision variable vector corresponding to the target solution includes optimal spatial variables, optimal time variables, and optimal resource variables. The optimal spatial variable is resolved as the spatial boundary of the prevention and control area, the optimal time variable is resolved as the operation time node of each prevention and control area, and the optimal resource variable is resolved as the resource allocation amount of each prevention and control area.

[0054] In this implementation, the spatial boundaries of the prevention and control area, the operational time nodes of each prevention and control area, and the resource allocation of each prevention and control area are determined simultaneously by solving the Pareto optimal solution set. During implementation, a decision variable vector containing spatial, temporal, and resource variables needs to be defined. Spatial variables represent the spatial boundary coordinates of the prevention and control area, and can be represented using a latitude-longitude coordinate system or a UTM coordinate system. Taking a quadrilateral prevention and control area as an example, the boundary can be represented by the coordinates of four points (x1, y1), (x2, y2), (x3, y3), and (x4, y4). Temporal variables represent the operational time nodes of each prevention and control area, and can be represented as the start and end times of the prevention and control operation, such as t_start and t_end. Resource variables represent the resource allocation of each prevention and control area, including human resources, equipment resources, and pesticide resources, and can be represented by h_resource, e_resource, and m_resource, respectively.

[0055] In the specific implementation process, the decision variable vector is substituted into the control effect objective function and the resource consumption objective function to establish a mathematical mapping relationship between the objective function value and the decision variable vector. The control effect objective function can consider indicators such as pest suppression rate and crop protection rate. For example, the control effect E can be defined as the weighted sum of the percentage decrease in pest population density and the percentage increase in crop yield. The resource consumption objective function can consider labor costs, equipment costs, pesticide costs, etc., and their weighted summation yields the total resource consumption R. For a specific decision variable vector, the pest population density change trend under a specific control strategy can be predicted through a pest and disease spread model, and then the control effect E value can be calculated; at the same time, the resource consumption R value can be calculated based on the resource unit price and usage.

[0056] When performing multi-objective optimization within the feasible solution space, an improved non-dominated sorting genetic algorithm is used for iterative search. This algorithm first randomly generates an initial population of 100 individuals, with each individual corresponding to a decision variable vector. New candidate solutions are generated through crossover and mutation operations, with a crossover probability of 0.8 and a mutation probability of 0.1. In each iteration, all solutions are evaluated according to the non-dominated sorting principle, and the Pareto optimal solution is selected. After 200 iterations, a stable Pareto optimal solution set is obtained, where each solution achieves a balance between the two objectives of prevention and control effectiveness and resource consumption.

[0057] To illustrate with a practical example, suppose there is a 1000-hectare farmland that needs pest and disease control, with an initial insect population density of 5 insects per square meter. Without control, the expected yield loss is 30%. Using the optimization method described above, a set of Pareto optimal solutions were obtained. One equilibrium solution is as follows: Divide the farmland into 3 control zones. The first zone has an area of ​​400 hectares, with spatial boundary coordinates (0,0), (0,2000), (2000,2000), (2000,0), and the operation period is from day 5 to day 7. 200 kg of pesticide, 10 personnel, and 2 pieces of equipment are allocated. The second zone has an area of ​​300 hectares, with spatial boundary coordinates (2000,0), (2000,1500), (4... The first area has the following coordinates: (000,1500), (4000,0). The operation will take place from day 3 to day 4, with 150 kg of pesticide allocated, 8 personnel, and 1 piece of equipment. The second area has an area of ​​300 hectares, with spatial boundary coordinates of (2000,1500), (2000,3000), (4000,3000), (4000,1500). The operation will take place from day 7 to day 9, with 150 kg of pesticide allocated, 8 personnel, and 1 piece of equipment. This plan is expected to reduce pest density by 85%, reduce yield loss to within 5%, and consume a total of 25,000 yuan in resources.

[0058] The target solution is selected from the Pareto optimal solution set according to a preset decision preference criterion. The decision preference criterion can be determined based on the target weight method, the ideal point method, or the fuzzy decision method. In this embodiment, the target weight method is used, setting the target weight for prevention and control effect to 0.6 and the target weight for resource consumption to 0.4. The solution with the highest comprehensive score is selected as the final target solution by calculating the comprehensive score. The decision variable vector corresponding to the target solution includes the optimal spatial variable, the optimal time variable, and the optimal resource variable.

[0059] Finally, the optimal spatial variable is analyzed as the spatial boundary of the control area, specifically the set of boundary point coordinates for each control area; the optimal temporal variable is analyzed as the operational time nodes for each control area, represented by specific dates and times; and the optimal resource variable is analyzed as the resource allocation for each control area, including the specific quantities of manpower, equipment, and pesticides allocated. These analytical results can be directly used to guide actual pest and disease control operations, providing precise control decision support.

[0060] The above methods can achieve a balance between pest and disease control effectiveness and resource consumption under limited resource constraints, improve control efficiency, reduce resource waste, and realize precise pest and disease control management.

[0061] Figure 2 The flowchart of the detection method of the present invention specifically includes: Experimental Data: This embodiment uses a self-built field dataset covering typical pests such as rice planthoppers and rice stem borers. The pest images in the dataset include information on different image sizes, lighting conditions, occlusion levels, shooting angles, and target pest sizes. Spectral data uses a self-built rice spectrum-MDA dataset. The data is divided into three parts: training set 70%, validation set 20%, and test set 10%. Preprocessing: The spectra are denoised using Savitzky-Golay smoothing (window 11, second-order polynomial); the image resolution is uniformly scaled to 640×640 pixels and illumination is normalized; environmental factors are linearly interpolated to the same time point as the spectrum-image.

[0062] Model training: A YOLOv8 model was trained using pest image data for pest target detection; a 1D-CNN model was trained using spectral data for predicting MDA content. Prediction data was output as structured "pest-physiology-environment" data triplets by the perceptual agent via a standardized JSON interface for quality verification by the analytical agent.

[0063] Reasoning and decision-making process: Model inference: The YOLOv8 model is called to output pest information (type, density, confidence level), and the 1D-CNN model is called to output the physiological indicator MDA content; Upload and encapsulation: The data is compressed into 18B fixed-length frames (8B pest information + 8B physiological information + 2B environmental CRC), encrypted with AES-128, and written to / field / rawtopic by the NB-IoT module via WebSocket.

[0064] Data Review: The integrity and validity of the "pest-physiology-environment" triplet uploaded by the sensing agent are verified. Abnormal data is discarded and marked, while valid samples are entered into risk assessment. Risk Assessment: Risk assessment is performed using preset business threshold logic—high risk is triggered when pest density exceeds the economic threshold or MDA content exceeds the severe range; when temperature, humidity, and light conditions meet suitable conditions for pest spread and the density exceeds the baseline, a spread duration warning is output; under high light and high pest density conditions, an additional "increased irrigation" suggestion is added to reduce the stress and harm of pesticides on plants. Result Encapsulation: A standard JSON containing risk level, density, MDA, spread window, light suggestion, and time stamp is generated and sent to the decision-making agent in real time via the Coze message queue, supporting multi-dimensional retrieval and subsequent review based on "pest type-date-plot".

[0065] Data Acquisition: The Coze incremental data stream is used to subscribe to and analyze the agent's "riskData" messages in real time. Simultaneously, external service plugins getNowAndForecastWeather and getAirQuality, as well as field resource interfaces, are invoked to obtain 24-hour weather forecasts, a list of prohibited pesticides, available equipment, and operable time periods, forming a complete constraint package. Decision-Making Solution Generation: The large model autonomously calls the cloud template library based on the risk level (riskLevel) and simultaneously injects meteorological, regulatory, and resource constraints to instantly generate an executable prevention and control plan. Command Issuance: The decision-making agent encapsulates the controlSuggestions structure (including immediateAction, pesticide, dosage, operation time, and follow-up measures) along with deviceID and commandCode into JSON, encrypts it using AES-128, and issues it via CozeBus control_command_topic. MQTT QoS=2 supports breakpoint resumption and up to 3 retransmissions, ensuring a closed-loop process.

[0066] The system architecture uses the Coze scheduled task interface to automatically retrieve the complete chain data of the previous 24 hours every day, including "analysis results - decision-making scheme - execution effect". If the pest mortality rate is less than 80%, the basic optimization process is triggered, and the current environment and decision record is written into the case library.

[0067] This invention provides a multi-agent collaborative decision-making system for agricultural pest assessment and control, comprising: The first unit is used to synchronously collect multimodal sensing data of the target crop area through a sensing agent deployed on an edge computing node. The multimodal sensing data includes image data, spectral data, and environmental data. The second unit is used by the sensing agent to identify pest targets in the image data based on a pre-trained target detection model, output the pest type and spatial distribution density, and quantify the physiological stress index of the spectral data to obtain pest identification results and physiological stress quantification results. The third unit is used to encapsulate the pest identification results, the physiological stress quantification results, and the environmental data into a standardized triplet data structure, wherein the triplet data structure includes pest feature dimensions, physiological feature dimensions, and environmental feature dimensions; The fourth unit is used to receive the triplet data structure through an analytical agent deployed on a cloud computing node, perform spatiotemporal correlation analysis on the pest feature dimension and the environmental feature dimension based on the pest spread dynamics model, generate pest spread trend prediction results, and calculate the comprehensive risk level based on the physiological feature dimension and the pest spread trend prediction results; The fifth unit is used to receive the triplet data structure, the pest spread trend prediction results and the comprehensive risk level through a decision-making intelligent agent deployed on a cloud computing node, and generate a prevention and control strategy plan under preset prevention and control resource constraints. The prevention and control strategy plan includes the division of prevention and control areas, the arrangement of prevention and control time sequence and the allocation of prevention and control resources.

[0068] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0069] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0070] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0071] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-agent collaborative decision-making method for agricultural pest evaluation and control, characterized in that, The method comprises the following steps: Synchronously collecting multi-modal perception data of a target crop area by a perception agent deployed at an edge computing node, wherein the multi-modal perception data comprises image data, spectral data and environmental data; The perception agent performs pest target recognition on the image data based on a pre-trained target detection model, outputs pest type and spatial distribution density, and quantifies physiological stress indicators from the spectral data to obtain pest recognition results and physiological stress quantification results; The pest recognition results, the physiological stress quantification results and the environmental data are packaged into a standardized triple data structure, which contains pest feature dimensions, physiological feature dimensions and environmental feature dimensions; An analysis agent deployed at a cloud computing node receives the triple data structure, performs spatio-temporal correlation analysis on the pest feature dimensions and the environmental feature dimensions based on a pest diffusion dynamics model, generates pest diffusion trend prediction results, and calculates a comprehensive risk level according to the physiological feature dimensions and the pest diffusion trend prediction results; A decision agent deployed at the cloud computing node receives the triple data structure, the pest diffusion trend prediction results and the comprehensive risk level, generates a control strategy scheme under the constraint of a preset control resource, and the control strategy scheme contains control area division, control time sequence arrangement and control resource allocation.

2. The method of claim 1, wherein, The perception agent performs pest target recognition on the image data based on a pre-trained target detection model, outputs pest type and spatial distribution density, and quantifies physiological stress indicators from the spectral data to obtain pest recognition results and physiological stress quantification results, which comprises: The perception agent performs multi-scale feature extraction on the image data to generate a feature map containing spatial position information and semantic feature information; Based on the pre-trained target detection model, the perception agent performs target detection and boundary box regression on the feature map to output the spatial coordinates of the pest target and the class confidence, and calculates the number of pest targets per unit area according to the spatial coordinates to obtain the spatial distribution density; The perception agent performs wave band separation and feature wave band extraction on the spectral data to obtain a spectral feature vector reflecting the physiological state of the plant; The perception agent performs nonlinear mapping on the spectral feature vector to output a physiological stress indicator quantization value representing the degree of plant stress response, and establishes a spatial correspondence between the physiological stress indicator quantization value and the spatial coordinates of the pest target; The pest type, the spatial distribution density, the class confidence and the physiological stress indicator quantization value are associated and integrated, and the physiological stress indicator quantization value is attributed to the spatial area where the corresponding pest target is located according to the spatial correspondence, forming the pest recognition results and the physiological stress quantification results.

3. The method of claim 1, wherein, The perception agent performs multi-scale feature extraction on the image data to generate a feature map containing spatial position information and semantic feature information; Based on the pre-trained target detection model, the perception agent performs target detection and boundary box regression on the feature map to output the spatial coordinates of the pest target and the class confidence, and calculates the number of pest targets per unit area according to the spatial coordinates to obtain the spatial distribution density; The perception agent performs wave band separation and feature wave band extraction on the spectral data to obtain a spectral feature vector reflecting the physiological state of the plant; The perception agent performs nonlinear mapping on the spectral feature vector to output a physiological stress indicator quantization value representing the degree of plant stress response, and establishes a spatial correspondence between the physiological stress indicator quantization value and the spatial coordinates of the pest target; The pest type, the spatial distribution density, the class confidence and the physiological stress indicator quantization value are associated and integrated, and the physiological stress indicator quantization value is attributed to the spatial area where the corresponding pest target is located according to the spatial correspondence, forming the pest recognition results and the physiological stress quantification results. The perception agent performs multi-scale feature extraction on the image data to generate a feature map containing spatial position information and semantic feature information; Based on the pre-trained target detection model, the perception agent performs target detection and boundary box regression on the feature map to output the spatial coordinates of the pest target and the class confidence, and calculates the number of pest targets per unit area according to the spatial coordinates to obtain the spatial distribution density; The perception agent performs wave band separation and feature wave band extraction on the spectral data to obtain a spectral feature vector reflecting the physiological state of the plant; The perception agent performs nonlinear mapping on the spectral feature vector to output a physiological stress indicator quantization value representing the degree of plant stress response, and establishes a spatial correspondence between the physiological stress indicator quantization value and the spatial coordinates of the pest target; The pest type, the spatial distribution density, the class confidence and the physiological stress indicator quantization value are associated and integrated, and the physiological stress indicator quantization value is attributed to the spatial area where the corresponding pest target is located according to the spatial correspondence, forming the pest recognition results and the physiological stress quantification results. Parameter fitting is performed on the spatiotemporal feature tensor based on the pest diffusion dynamics model to estimate a pest diffusion rate parameter and an environmental influence coefficient parameter, the pest diffusion rate parameter representing the propagation intensity of the pest in the spatial dimension, and the environmental influence coefficient parameter representing the regulation of the environmental feature dimension on the pest diffusion; The pest feature dimension and the environmental feature dimension at the current time are taken as initial conditions, the pest diffusion rate parameter and the environmental influence coefficient parameter are combined, and the pest spatial distribution evolution simulation in the future time period is performed through the pest diffusion dynamics model to obtain pest spatial distribution prediction data containing multiple future time nodes; Statistical analysis is performed on the pest spatial distribution prediction data to extract the spatial position change trajectory of the pest risk area and the pest density growth trend to form the pest diffusion trend prediction result.

4. The method of claim 3, wherein, The pest spatial distribution evolution simulation in the future time period through the pest diffusion dynamics model to obtain pest spatial distribution prediction data containing multiple future time nodes includes: The future time period is divided into multiple discrete time nodes according to a preset time interval to form a time node sequence; The pest feature dimension and the environmental feature dimension at the current time are mapped to a spatial grid structure to generate a pest spatial distribution initial state matrix and an environmental state matrix; For each time node in the time node sequence, the pest spatial distribution state of the previous time node is calculated based on the pest diffusion dynamics model, and the environmental state parameter of the corresponding time node is introduced to dynamically regulate the diffusion process to obtain the pest spatial distribution evolution state of the time node, wherein the pest diffusion dynamics model is constructed based on a long short-term memory network; The pest spatial distribution evolution state corresponding to each time node in the time node sequence is integrated in time sequence to construct a three-dimensional data structure containing spatial coordinates, pest density and time identifiers to form the pest spatial distribution prediction data containing multiple future time nodes.

5. The method of claim 1, wherein, Under the constraint condition of the preset control resource, the control strategy scheme includes: According to the pest diffusion trend prediction result and the comprehensive risk level, a multi-objective optimization target set containing a control effect target function and a resource consumption target function is constructed, the control effect target function representing pest control efficiency, and the resource consumption target function representing control resource use cost; The preset control resource constraint condition is analyzed to extract a control resource total amount constraint, a control operation time window constraint and a spatial accessibility constraint to form a constraint condition set; Based on the multi-objective optimization target set and the constraint condition set, a Pareto optimal solution set is solved to simultaneously determine the spatial boundary of the control area, the operation time node of each control area and the resource allocation amount of each control area; The spatial boundary of the control area obtained by solving is integrated to form the control area division, the operation time node is arranged in time sequence to form the control time sequence arrangement, and the resource allocation amount is mapped to the corresponding control area to form the control resource allocation, which are combined to constitute the control strategy scheme.

6. The method of claim 5, wherein, determining the spatial boundary of the prevention and control region, the operation time node of each prevention and control region and the resource allocation amount of each prevention and control region by solving a Pareto optimal solution set based on the multi-objective optimization target set and the constraint condition set, including: defining a decision variable vector containing a spatial variable, a time variable and a resource variable, the spatial variable representing the spatial boundary coordinates of the prevention and control region, the time variable representing the operation time node of each prevention and control region, and the resource variable representing the resource allocation amount of each prevention and control region; substituting the decision variable vector into the prevention effect objective function and the resource consumption objective function to establish a mathematical mapping relationship between the objective function value and the decision variable vector; performing multi-objective optimization solving in the feasible solution space, and obtaining a Pareto optimal solution set in which the prevention effect objective function and the resource consumption objective function are in a non-dominated state by iterative search, each solution in the Pareto optimal solution set corresponding to a group of decision variable vectors; selecting a target solution from the Pareto optimal solution set according to a preset decision preference criterion, the decision variable vector corresponding to the target solution containing optimal spatial variables, optimal time variables and optimal resource variables; analyzing the optimal spatial variables into the spatial boundary of the prevention and control region, the optimal time variables into the operation time node of each prevention and control region, and the optimal resource variables into the resource allocation amount of each prevention and control region.

7. Multi-agent collaborative decision system for agricultural pest assessment and control, for implementing the method according to any one of claims 1 to 6, characterized in that, comprising: A first unit for synchronously collecting multi-modal perception data of a target crop region by a perception agent deployed on an edge computing node, the multi-modal perception data including image data, spectral data and environmental data; A second unit for the perception agent to perform pest target identification on the image data based on a pre-trained target detection model, output pest type and spatial distribution density, and perform physiological stress index quantification on the spectral data to obtain pest identification results and physiological stress quantification results; A third unit for packaging the pest identification results, the physiological stress quantification results and the environmental data into a standardized triple data structure, the triple data structure containing pest feature dimensions, physiological feature dimensions and environmental feature dimensions; A fourth unit for receiving the triple data structure by an analysis agent deployed on a cloud computing node, performing spatio-temporal correlation analysis on the pest feature dimensions and the environmental feature dimensions based on a pest diffusion dynamics model to generate pest diffusion trend prediction results, and calculating a comprehensive risk level according to the physiological feature dimensions and the pest diffusion trend prediction results; A fifth unit for receiving the triple data structure, the pest diffusion trend prediction results and the comprehensive risk level by a decision-making agent deployed on a cloud computing node, generating a prevention and control strategy scheme under a preset prevention and control resource constraint condition, the prevention and control strategy scheme including prevention and control region division, prevention and control timing arrangement and prevention and control resource allocation.

8. An electronic device, comprising: comprising: a processor; a memory for storing processor-executable instructions; wherein the processor is configured to invoke the instructions stored in the memory to perform the method of any one of claims 1 to 6.

9. A computer-readable storage medium having stored thereon computer program instructions, wherein, The computer program instructions, when executed by a processor, implement the method of any one of claims 1 to 6.

Images