Precise pesticide application method and system of pesticide spraying unmanned aerial vehicle

Abstract

The invention relates to the technical field of precise pesticide application of agricultural unmanned aerial vehicles, and discloses a precise pesticide application method and system of a pesticide spraying unmanned aerial vehicle. According to the method, field information is dynamically collected according to a field digital base map and a mobile sensor network, and target plaques are identified through iterative comparison with a pathological knowledge map. And according to the plaque features, carrying out adaptive fine-grained region division by adopting a bionic fractal rule, and generating hierarchical pesticide application units. And for each unit, reasoning the liquid medicine demand and planning the nozzle action in parallel to form an instruction set for driving the variant spray boom. And the unmanned aerial vehicle executes the instruction to adjust the attitude and flow of the spray boom, and performs rolling optimization on subsequent instructions by fusing real-time environment data. According to the method, natural distribution forms of diseases and insect pests are matched through fractal division, precise targeted pesticide application in a three-dimensional space is achieved through the variant spraying rod, and the space accuracy and environmental adaptability of pesticide liquid putting are improved.

Description

Technical Field

[0001] This invention relates to the field of precision pesticide application technology using agricultural drones, specifically a method and system for precision pesticide application using agricultural drones. Background Technology

[0002] Current precision pesticide application technologies using drones largely rely on preset flight paths and fixed grid zones for variable spraying. These zones are typically divided using regular geometric shapes or uniformly sized grids, which inherently differ from the actual spatial distribution patterns of pests and diseases in the field. The occurrence and development of pests and diseases often exhibit irregular, clustered, and multi-scale natural characteristics, with complex patterns spreading outward from the central disease site. Regular grid divisions disrupt the integrity of these natural patches, leading to a mismatch between the pesticide application decision unit and the actual stress area, resulting in over-spraying or under-spraying.

[0003] Conventional precision spraying systems have limited coupling between decision-making and execution. Variable-rate spraying primarily adjusts the overall flow rate of the nozzles based on the prescription map, while the spray boom itself is typically of a fixed configuration or can only undergo simple overall attitude adjustments. This approach cannot adapt to irregular and fragmented application areas with varying geometry and spatial positions, thus limiting the accuracy of spray coverage. The planning of the execution actions is relatively static, making it difficult to make local adaptive adjustments based on real-time environmental feedback during dynamic flight, thus limiting the real-time responsiveness of spraying operations to complex field conditions.

[0004] Existing technologies struggle to dynamically divide pest and disease patches into zones that align with their biological distribution characteristics, based on the natural morphology and internal heterogeneity of the patches. Furthermore, spraying actuators lack sufficient spatial freedom and adaptability, hindering the precise application of pesticides to subdivided management units of varying shapes and irregular spatial distributions. This creates bottlenecks in application accuracy regarding spatial adaptability and operational flexibility. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for precise pesticide application using a pesticide spraying drone, in order to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a method for precise pesticide application using a pesticide spraying drone, the method comprising: Based on a pre-generated digital base map of the field, a mobile ground sensor node network is deployed, which dynamically collects multi-dimensional field information related to crop stress. The multidimensional field information is iteratively compared and correlated with a pre-constructed crop pathology knowledge graph to dynamically identify potential target patches for pesticide intervention. Based on the boundary features and internal heterogeneity of the potential target patches, a biomimetic fractal segmentation rule is used to adaptively divide the potential target patches into fine-grained regions, generating a drug application management unit with a hierarchical structure. For each of the aforementioned drug application management units, drug demand reasoning and nozzle action planning are executed in parallel to generate a set of machine-executable drug application instructions containing spatial coordinates, drug formulation, and spraying parameters; The drone's variant spray boom system adjusts its attitude and flow rate according to the machine's executable pesticide application command set, while simultaneously integrating real-time environmental perception data during spraying to continuously optimize subsequent commands.

[0007] Preferably, the deployment of a mobile ground-based sensor node network based on a pre-generated digital base map of the field, wherein the mobile ground-based sensor node network dynamically collects multi-dimensional field information related to crop stress, specifically includes: The digital base map of the field is generated by fusing historical remote sensing images, soil electrical conductivity layers, and topographic elevation models, and defines the distribution of static attributes within the field. The mobile ground sensing node network is carried by several autonomous mobile platforms, each platform integrating a spectral imager, a microclimate sensor and a soil probe; The mobile ground sensor node network is initially deployed based on the attribute gradient bands marked on the field digital base map, and dynamically repositioned along the preset inspection line during the operation cycle. The multidimensional field information includes, but is not limited to, crop canopy spectral reflectance, leaf temperature, air temperature and humidity, and root zone soil moisture and nutrient concentration; The dynamic acquisition refers to the mobile ground sensor node network triggering high-density information acquisition events at non-fixed intervals, based on the disease-prone periods and areas inferred from the crop pathology knowledge graph.

[0008] Preferably, the iterative comparison and association mapping between the multidimensional field information and the pre-constructed crop pathology knowledge graph to dynamically identify potential target patches for pesticide intervention specifically includes: The crop pathology knowledge graph is stored in a graph structure, where nodes represent specific disease or pest states and edges represent the causal relationships and probability weights between different environmental factors and the occurrence and development of diseases. The iterative comparison refers to inputting the real-time acquired multidimensional field information into the crop pathology knowledge graph, activating the matching nodes, and performing reasoning propagation along the edges of the graph to calculate the confidence value of each field location in different pest and disease states. The association mapping refers to binding the inferred pest and disease status confidence value with the corresponding geographic coordinates to generate a dynamically updated pest and disease risk heat distribution layer on the field digital base map. The potential target patches are automatically extracted by setting a confidence threshold on the pest and disease risk heat distribution layer, and their boundaries are composed of continuous pixel regions exceeding the threshold. The dynamic identification process is repeated after each information collection cycle, so that the shape, area and location of the potential target patches can evolve and change over time.

[0009] Preferably, based on the boundary features and internal heterogeneity of the potential target patches, a biomimetic fractal segmentation rule is used to adaptively divide the potential target patches into fine-grained regions, generating a hierarchical drug application management unit, specifically including: The boundary features are quantified by calculating the tortuosity and concavity of the contour of the potential target patch; The internal heterogeneity is assessed by analyzing the statistical variance and spatial autocorrelation of the confidence values ​​on the heat map of the pest and disease risk distribution within the potential target patches. The biomimetic fractal segmentation rule simulates the branching growth pattern of veins or roots in nature. Its core is to recursively divide an irregular region into unequal areas according to the degree of heterogeneity until the internal homogeneity of the sub-region meets the preset standard. The adaptive fine-grained region segmentation process is as follows: taking the entire potential target patch as the initial region, and determining whether to trigger segmentation based on the internal heterogeneity evaluation results; if triggered, determining the direction of the initial segmentation line based on the boundary tortuosity, and dividing the region into two sub-regions; recursively repeating the evaluation and segmentation process for each sub-region. The hierarchical drug application management unit ultimately manifests as a tree structure, with the root node corresponding to the original potential target patch, the leaf nodes corresponding to the final drug application management unit that meets the homogeneity standard, and the intermediate nodes representing the transition management areas at different scales.

[0010] Preferably, for each of the aforementioned application management units, parallel execution of pesticide demand reasoning and nozzle motion planning generates a machine-executable application instruction set containing spatial coordinates, pesticide formulation, and spraying parameters, specifically including: The pesticide demand reasoning is performed independently for each pesticide application management unit. Its input is the aggregated value of the multidimensional field information from all collection points within the pesticide application management unit, as well as the dominant pest and disease types and degrees obtained from the crop pathology knowledge graph. The drug demand reasoning is accomplished through a trained neural network model, which directly outputs recommended drug type, concentration, and application rate per unit area. The nozzle motion planning is carried out synchronously with the liquid demand reasoning, based on the geographical boundary of the application management unit, the current flight speed and altitude of the UAV, and the mechanical constraints of the variant spray boom system. The nozzle motion planning calculation yields the nozzle's switching sequence, lateral swing angle, and flow modulation curve within each swing cycle; The machine can execute a set of application instructions that encodes the spatial coordinates, drug formulation, and spraying parameters into a sequence of instructions ordered by timestamps, wherein each instruction precisely corresponds to the action of the UAV when it flies over a specific application management unit.

[0011] Preferably, the variable spray boom system driven by the UAV adjusts its attitude and flow rate according to the machine-executable drug delivery command set, specifically including: The variant spray bar system is composed of multiple independently telescopic segments hinged together, with an atomizing nozzle controlled by a solenoid valve installed at the end of each segment; Attitude adjustment refers to the variant spray boom system dynamically changing the extension length and hinge angle of each segment according to the boundary of the application management unit specified in the machine's executable application instruction set, so that the spray profile formed by the nozzle array matches the shape of the application management unit; The flow adjustment refers to the control of the on / off state and flow rate of the liquid medicine by each atomizing nozzle controlled by the solenoid valve according to the flow modulation curve specified in the set of executable drug delivery instructions of the machine, with millisecond-level precision. The variant spray boom system has a built-in inertial measurement unit that provides real-time feedback on the actual attitude of the spray boom, forming a closed-loop control with the commanded expected attitude.

[0012] Preferably, the process of simultaneously integrating real-time environmental perception data during spraying to continuously optimize subsequent instructions includes: The real-time environmental perception data is provided by the UAV's onboard downward-looking wind field sensor and real-time imaging spectrometer, including real-time wind speed and direction below the operational flight path and instantaneous changes in canopy spectral reflectance. The rolling optimization is performed in units of sliding time windows. When the UAV executes the machine-executable pesticide application instruction set within the current time window, the real-time environmental perception data of the field area to be covered in the next time window is analyzed simultaneously. The rolling optimization first predicts the trajectory of the liquid drift based on real-time wind speed and direction data, and then corrects the nozzle switching timing and swing angle in the nozzle action planning in advance accordingly. The rolling optimization then compares the real-time canopy spectral reflectance data with the baseline spectral data before the operation. If a significant difference is detected, a rapid recalculation of the inference of the drug demand in the affected area is triggered. Based on the results of predictive drift correction and fast recalculation, the set of machine-executable drug delivery instructions that have not yet been executed is dynamically updated.

[0013] Preferably, the dynamic relocation strategy of the mobile ground sensor node network specifically includes: The dynamic relocation strategy is centrally scheduled by the ground monitoring platform, and its decision-making is based on the dynamic changes of the pest and disease risk heat distribution layer and the operation route that the UAV is about to execute. When the pest and disease risk heat distribution layer shows a sharp increase in the confidence value of a certain area, the dynamic relocation strategy will instruct the nearest mobile ground sensor node to move to the area and collect data more intensively. Before the drone performs spraying operations, the dynamic relocation strategy instructs the mobile ground sensor node network to perform supplementary intensive sampling in the area below the planned flight path to provide the latest field information for final command verification.

[0014] Preferably, the updating and expansion mechanism of the crop pathology knowledge graph specifically includes: The initial structure of the crop pathology knowledge graph was constructed from domain expert knowledge; After each operation cycle is completed, the system will archive the complete multi-dimensional field information collected this time, the final set of machine-executable pesticide application instructions, and the pesticide application effect data evaluated by the mobile ground sensor node network as a complete case. By analyzing historical case databases using offline machine learning algorithms, potential correlation patterns between new combinations of environmental factors and the occurrence of pests and diseases can be discovered. The newly discovered association patterns are integrated into the crop pathology knowledge graph by adding new nodes or edges, or by adjusting the probability weights of existing edges, so as to achieve the autonomous evolution of the graph.

[0015] Preferably, the present invention also includes a precision pesticide spraying drone system, the system including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein when the processor executes the computer program, it implements the steps of the precision pesticide spraying drone method described above.

[0016] Compared with the prior art, the beneficial effects of the present invention are: By employing biomimetic fractal segmentation rules to adaptively divide identified potential target patches, the boundaries and scales of the segmented regions dynamically match the natural aggregation characteristics and spread patterns of pests and diseases. This method generates hierarchical management units with multi-scale, self-similar features, rather than fixed grids. This segmentation approach ensures that the morphology of the pesticide application management units closely matches the complex contours and internal heterogeneity differences of the actual stress area, providing a spatial management foundation consistent with the pathological patterns in nature for subsequent differentiated pesticide application and improving the accuracy of target area definition.

[0017] By parallelizing pesticide demand reasoning and nozzle motion planning, and generating machine commands to drive the variable boom spray system, deep collaboration between decision-making and execution in three-dimensional space is achieved. The variable boom can independently adjust its local posture, configuration, or flow rate according to commands, allowing its physical form and motion patterns to adapt to subdivided management units of different shapes, sizes, and spatial positions. This transforms pesticide spraying from a simple flow rate change based on a fixed boom pattern into precise three-dimensional target projection on irregular patches. Combined with real-time environmental perception and rolling command optimization during spraying, a closed-loop dynamic control capability from perception and decision-making to execution is formed, enhancing the accuracy and adaptability of pesticide application operations in complex field environments. Attached Figure Description

[0018] Figure 1 This is a schematic diagram illustrating the working principle of the pesticide spraying drone precision application method described in this invention. Figure 2 A flowchart illustrating the deployment and information collection of a mobile ground-based sensor node network; Figure 3 A flowchart for parallel execution of liquid demand reasoning and nozzle action planning; Figure 4 Line graph comparing attitude parameters of a variant boom spraying system for pesticide spraying drones; Figure 5 A heatmap of the spatial reliability values ​​for pest and disease risk. Detailed Implementation

[0019] 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] Please see Figure 1This invention provides a method for precise pesticide application using a drone. The method includes: deploying a mobile ground-based sensor node network based on a pre-generated digital map of the field, which dynamically collects multi-dimensional field information related to crop stress. The multi-dimensional field information is iteratively compared and mapped with a pre-constructed crop pathology knowledge graph to dynamically identify potential target patches for pesticide intervention. Based on the boundary features and internal heterogeneity of the potential target patches, a biomimetic fractal segmentation rule is used to adaptively divide the potential target patches into fine-grained regions, generating a hierarchical pesticide management unit. For each pesticide management unit, pesticide demand reasoning and nozzle action planning are executed in parallel to generate a machine-executable pesticide application instruction set containing spatial coordinates, pesticide formulation, and spraying parameters. The drone-mounted variant boom system adjusts its attitude and flow rate according to the machine-executable pesticide application instruction set, while simultaneously integrating real-time environmental perception data during spraying to continuously optimize subsequent instructions.

[0021] In one embodiment of the present invention, see [reference] Figure 2 The digital base map of the field is generated by fusing historical remote sensing imagery, soil conductivity layers, and topographic elevation models, defining the static attribute distribution within the field. A mobile ground-based sensor node network is carried by several autonomous mobile platforms, each integrating a spectral imager, microclimate sensor, and soil probe. The mobile ground-based sensor node network is initially deployed based on the attribute gradient bands marked on the digital base map and dynamically repositioned along preset inspection lines during the operational cycle. Multidimensional field information includes crop canopy spectral reflectance, leaf temperature, air temperature and humidity, and root zone soil moisture and nutrient concentration. Dynamic acquisition refers to the mobile ground-based sensor node network triggering high-density information collection events at non-fixed intervals, based on disease-prone periods and areas inferred from the crop pathology knowledge graph. The crop pathology knowledge graph is stored in a graph structure, where nodes represent specific disease or pest states, and edges represent the causal relationships and probability weights between different environmental factors and disease occurrence and development. Iterative comparison involves inputting real-time acquired multidimensional field information into a crop pathology knowledge graph, activating matching nodes, and performing inference propagation along the graph's edges to calculate the confidence value of each field location under different pest and disease states. Association mapping involves binding the inferred pest and disease state confidence values ​​to corresponding geographic coordinates, generating a dynamically updated pest and disease risk heat map layer on the field's digital base map. Potential target patches are automatically extracted by setting a confidence threshold on the pest and disease risk heat map layer; their boundaries consist of continuous pixel regions exceeding the threshold. The dynamic identification process is repeated after each information collection cycle, allowing the shape, area, and location of potential target patches to evolve over time.

[0022] In practice, the digital base map of the field is generated by spatial registration and data fusion algorithms from historical remote sensing imagery, soil conductivity layers, and topographic elevation models. This defines the static attribute distribution within the field, including soil fertility zoning and micro-topographic undulations. A mobile ground-based sensor node network is carried by several autonomous mobile platforms, each integrating a hyperspectral imager, temperature, humidity, and light intensity microclimate sensors, and a multi-parameter soil probe. The mobile ground-based sensor node network is initially deployed based on attribute gradient zones marked on the digital base map, for example, deploying nodes along transition zones where soil conductivity changes. Throughout the operation cycle, it dynamically repositions itself along pre-defined inspection lines covering different areas of the field. Multidimensional field information includes crop canopy spectral reflectance acquired by the hyperspectral imager, leaf temperature acquired by infrared sensors, air temperature and humidity acquired by weather stations, and root zone soil moisture and nutrient concentration acquired by probe-type sensors. The dynamic acquisition process refers to the operation of the mobile ground sensor node network at a non-fixed cycle. Its acquisition triggering logic depends on the reasoning results of the crop pathology knowledge graph. When the crop pathology knowledge graph infers a disease-prone period and area based on time series and environmental conditions, the system will send instructions to the mobile ground sensor node network near that area to trigger a high-density information acquisition event.

[0023] In some embodiments, the crop pathology knowledge graph is stored in a graph database in a graph structure. Its nodes represent specific disease or pest states, such as "early stage of wheat powdery mildew," and edges represent the causal relationships and probability weights between different combinations of environmental factors and the occurrence and development of the disease. For example, the edge connecting "average humidity greater than 80% for three consecutive days" and "occurrence of wheat powdery mildew" has a high weight value. Iterative comparison operation involves inputting the real-time acquired multidimensional field information vector into the inference engine of the crop pathology knowledge graph. Each dimension value in the multidimensional field information vector activates the matching environmental factor node in the graph. The inference algorithm performs probability propagation and confidence updates along the edges of the graph, calculating the confidence value of each sampling point coordinate in the field under different disease and pest states. Association mapping operation involves binding the disease and pest state confidence values ​​calculated by inference with the corresponding geographic coordinates. A dynamically updated disease and pest risk heat distribution layer is generated on the digital base map of the field using a spatial interpolation algorithm. The color depth of each pixel in the layer represents the confidence value at that location. Potential target patches are identified automatically by setting a confidence threshold on the pest and disease risk heat map. All consecutive pixel regions with confidence values ​​exceeding the preset threshold are delineated, forming the vector boundaries of one or more potential target patches. The dynamic identification process is repeated after each data collection cycle. Each new input of multidimensional field information triggers a new round of iterative comparison and association mapping, updating the pest and disease risk heat map and causing the shape, area, and location of potential target patches to evolve over time.

[0024] It is understandable that, during the iterative comparison process, a specific implementation of calculating the confidence value for a specific pest or disease state can be based on Bayesian network inference, expressed by the following formula: ; Where: symbol This indicates the observation of a multidimensional set of field information evidence. Under these conditions, the state of pests and diseases is as follows: The posterior probability, i.e., the reliability value, is represented by the symbol. Indicates the state of pests and diseases The following is a collection of evidence. The likelihood probability, derived from the edge weights and historical statistics in the crop pathology knowledge graph, is given by the symbol. Indicates the status of pests and diseases Prior probability, sign This indicates all possible states of disease and pests. Summation is performed to normalize the probabilities. The crop pathology knowledge graph provides a network of prior and conditional probabilities required for this formula.

[0025] Optionally, the dynamic relocation path planning of the mobile ground sensor node network can employ an algorithm based on real-time thermal gradients. Based on the inspection line, when the autonomous mobile platform detects an abnormally high increase in the heat value of pest and disease risk in a local area, it autonomously adjusts its path to move towards the center of that area to perform intensive data collection. The triggering of multi-dimensional field information collection events not only relies on inferences from crop pathology knowledge graphs but can also be coordinated with the drone's pre-set operational plan. Before the drone arrives at a certain area, the mobile ground sensor node network can be instructed to complete a round of data collection in that area in advance, providing the latest field information input for the upcoming pesticide demand inference.

[0026] In one embodiment of the invention, boundary features are quantified by calculating the tortuosity and concavity of the potential target patch outline. Internal heterogeneity is assessed by analyzing the statistical variance and spatial autocorrelation of the confidence values ​​on the heat map of pest and disease risk distribution within the potential target patch. The biomimetic fractal segmentation rule simulates the branching growth pattern of veins or roots in nature. Its core is to recursively divide an irregular region into unequal areas according to the degree of heterogeneity until the internal homogeneity of the sub-regions meets the preset standard. The adaptive fine-grained region division process takes the entire potential target patch as the initial region, and determines whether to trigger segmentation based on its internal heterogeneity assessment results. If triggered, the direction of the initial segmentation line is determined according to the boundary tortuosity, dividing the region into two sub-regions. The assessment and segmentation process is recursively repeated for each sub-region. The hierarchical pesticide application management unit is ultimately represented as a tree structure. The root node of the tree corresponds to the original potential target patch, the leaf nodes correspond to the final pesticide application management unit that meets the homogeneity standard, and the intermediate nodes represent transitional management areas at different scales.

[0027] The boundary features and internal heterogeneity of potential target patches are quantitatively analyzed. Boundary features are quantified by calculating the tortuosity and concavity of the patch outline. Internal heterogeneity is assessed by analyzing the statistical variance and spatial autocorrelation of the reliability values ​​on the heat map of pest and disease risk distribution within the potential target patch. The biomimetic fractal segmentation rule simulates the branching growth pattern of veins or roots in nature. The core of the biomimetic fractal segmentation rule is to recursively divide an irregular region into unequal areas according to the degree of heterogeneity until the internal homogeneity of the sub-regions meets the preset criteria. The adaptive fine-grained region division process uses the entire potential target patch as the initial... In the initial region, the decision to trigger a segmentation operation is made based on the internal heterogeneity assessment results of the potential target patches. If the internal heterogeneity assessment results exceed a preset threshold, segmentation is triggered, and the direction of the initial segmentation line is determined according to the boundary tortuosity. The potential target patch region is divided into two sub-regions. The internal heterogeneity assessment and segmentation process is recursively repeated for each sub-region. The hierarchical drug administration management unit ultimately manifests as a tree structure. The root node of the tree structure corresponds to the original potential target patch, the leaf nodes of the tree structure correspond to the final drug administration management unit that meets the homogeneity standard, and the intermediate nodes of the tree structure represent the transition management areas at different scales.

[0028] It is understandable that the recursive segmentation process in the biomimetic fractal segmentation rule is implemented in the computer algorithm as a depth-first search or breadth-first search traversal. Each segmentation operation is based on the real-time internal heterogeneity assessment result and boundary tortuosity direction of the current sub-region. The direction of the initial segmentation line can be determined based on the principal direction of the line segment with the maximum local tortuosity on the contour of the potential target patch, or by calculating the principal component direction of the confidence value gradient within the potential target patch region, so that the segmentation line passes through the region with high heterogeneity as much as possible. The termination condition of the adaptive fine-grained region division is defined by the internal homogeneity criterion of the sub-region. The internal homogeneity criterion can be set by the variance of all confidence values ​​in the sub-region being less than a fixed threshold or the spatial autocorrelation coefficient of the confidence values ​​in the sub-region being greater than a fixed threshold.

[0029] Optionally, the hierarchical tree structure of the drug application management unit can be implemented in the software system using a data structure of pointers or object references. Each node in the tree structure stores the sequence of boundary polygon vertices of the corresponding region, the internal heterogeneity index, and references to child nodes. During the recursive segmentation process, the generation of segmentation lines can employ computational geometry methods such as splitting based on region skeleton lines or partitioning based on weighted Voronoi diagrams to ensure that the segmented sub-regions maintain geometric connectivity and have relatively smooth boundaries. The spatial autocorrelation calculation in the internal heterogeneity assessment can use the global Moran's index formula, which requires the confidence value and its spatial weight matrix of each pixel within the potential target patch.

[0030] In some embodiments, the quantification of boundary concavity / convexity can be achieved by calculating the area difference ratio between the convex hull of the potential target patch contour and the original contour; a larger area difference ratio indicates more significant boundary concavity / convexity. The statistical variance calculation in the internal heterogeneity assessment uses the unbiased estimation variance formula of the reliability value sample, and the spatial autocorrelation calculation uses the absolute value of the global Moran index formula. The unequal area segmentation of the biomimetic fractal segmentation rule is reflected in the fact that the areas of the two sub-regions generated by each segmentation are not necessarily equal; the segmentation objective is to make the comprehensive internal heterogeneity scores of the two sub-regions as close as possible and lower than that of the parent region.

[0031] It is understandable that the resulting hierarchical tree structure of pesticide application management units provides an independent spatial unit foundation for subsequent parallel pesticide demand reasoning. Each leaf node pesticide application management unit will serve as the input unit for pesticide demand reasoning and nozzle action planning. The computational resource consumption of the recursive segmentation process can be controlled by setting a maximum recursion depth or a minimum region area threshold to prevent the generation of pesticide application management units that are too small or lack significance for spraying operations. The intermediate node information of the tree structure can be used to achieve multi-granularity decision-making in spraying operations, such as more frequent sensing monitoring of intermediate node regions with high heterogeneity.

[0032] Optionally, the preset threshold for triggering segmentation based on internal heterogeneity assessment results can be dynamically adjusted according to historical field data or crop type. For example, a lower threshold can be used during peak disease seasons to generate finer-grained pesticide management units. The adaptive nature of the biomimetic fractal segmentation rule is reflected in the fact that the segmentation decision and direction depend entirely on the real-time quantitative characteristics of the current region, without the need for predefined fixed grids or fixed segmentation patterns. Specific methods for determining the initial segmentation line direction based on boundary tortuosity can include connecting the extreme points of contour curvature or fitting the principal axis of the contour, with the segmentation line perpendicular to the principal axis or along the direction of the fastest descent of the confidence value gradient.

[0033] In one embodiment of the present invention, see [reference] Figure 3 The pesticide demand inference is performed independently for each application management unit. Its input consists of the aggregated value of multi-dimensional field information from all collection points within the application management unit, as well as the dominant pest and disease types and severity inferred from the crop pathology knowledge graph. Pesticide demand inference is accomplished through a trained neural network model, which directly outputs recommended pesticide type, concentration, and application rate per unit area. Sprinkler motion planning is performed synchronously with pesticide demand inference, based on the geographical boundaries of the application management unit, the current flight speed and altitude of the drone, and the mechanical constraints of the variant boom system. Sprinkler motion planning calculates the nozzle's on / off timing, lateral swing angle, and flow modulation curve within each swing cycle. The machine-executable application instruction set encodes spatial coordinates, pesticide formulation, and spraying parameters into a timestamp-ordered sequence of instructions, where each instruction precisely corresponds to the drone's action when flying over a specific application management unit.

[0034] In practice, pesticide demand reasoning and nozzle motion planning are performed in parallel for each pesticide application management unit. Pesticide demand reasoning is performed independently for each pesticide application management unit. The input for pesticide demand reasoning is the aggregated value of multi-dimensional field information from all collection points within the pesticide application management unit, as well as the dominant pest and disease types and severity inferred from the crop pathology knowledge graph. Pesticide demand reasoning is completed through a trained neural network model, which directly outputs the recommended pesticide type, concentration, and application rate per unit area. Sprayer motion planning is performed synchronously with pesticide demand reasoning. The basis for sprayer motion planning is the geographical boundary of the pesticide application management unit, the current flight speed and altitude of the UAV, and the mechanical constraints of the variant spray boom system. Sprayer motion planning calculates the on / off sequence of the nozzle, the lateral swing angle, and the flow modulation curve within each swing cycle. The machine-executable pesticide application instruction set encodes the spatial coordinates, pesticide formulation, and spraying parameters into a sequence of instructions ordered by timestamps. Each instruction in the instruction sequence precisely corresponds to the action of the UAV when flying over a specific pesticide application management unit.

[0035] In some embodiments, the neural network model used for pesticide demand inference is a multi-layer feedforward network. The input layer nodes of the neural network model correspond to the aggregated feature vector of multi-dimensional field information and the encoding vector of the dominant pest and disease type. The output layer nodes of the neural network model correspond to the application concentration and dosage of various candidate pesticides. The training data of the neural network model comes from historical pesticide application records, historical multi-dimensional field information of the corresponding field, and post-application effect evaluation data. The aggregated value of multi-dimensional field information can be calculated using the arithmetic mean, weighted average, or maximum value of data from all collection points within the pesticide application management unit. The weights are determined by the distance between the collection point and the geometric center of the pesticide application management unit or the confidence value of the location of the collection point. The dominant pest and disease type and severity obtained from the crop pathology knowledge graph are represented in vector form, where each element represents the occurrence probability and severity level of a specific pest and disease.

[0036] It can be understood that the output of the neural network model for drug demand reasoning can be represented by the following mapping relationship: ; Where: symbol This represents the output vector of the neural network model. The dimension corresponds to the number of preset drug solution formula parameters, and the symbol is... This represents the forward propagation function of a neural network with a specific activation function and layer structure, denoted by [symbol]. This represents the aggregated feature vector of multidimensional field information within the pesticide application management unit, with the symbol... This represents the encoding vector of dominant pest and disease types and severity obtained from the crop pathology knowledge graph, with the symbol... This represents the vector concatenation operation, with the symbol... This represents the set of weight parameters obtained by the neural network model through training. The synchronous execution of nozzle motion planning and pesticide demand inference is reflected in the fact that the two computational processes share the geographical boundary data of the pesticide application management unit as initial input, but run independently on different computational threads or processor cores. During the computation process, the nozzle motion planning module needs to obtain the current flight speed and altitude estimates provided by the UAV flight control system in real time. The calculation results of the nozzle motion planning module, including the switching timing, lateral swing angle, and flow modulation curve, must meet the physical motion limits and flow control accuracy of the variant spray boom system.

[0037] In some embodiments, the nozzle motion planning calculation of the switching timing involves determining the opening and closing times of each atomizing nozzle controlled by an independent solenoid valve in the nozzle array based on the geometric relationship between the boundary polygon of the drug administration unit and the preset flight path of the UAV. The accuracy of the switching timing reaches the millisecond level to match the flight speed of the UAV. The lateral swing angle planning is based on the width of the drug administration unit and the swing range that the variant spray boom system can achieve, calculating the periodic swing angle sequence of the nozzle in the vertical plane of the flight direction. The flow modulation curve defines the functional relationship of the liquid flow rate with time within a nozzle swing cycle. The shape of the flow modulation curve is designed to compensate for the difference in the amount of deposited in the edge area caused by the nozzle swing. The encoding format of the machine-executable drug administration instruction set adopts a structured data protocol. Each instruction includes a timestamp field, a spatial coordinate field, a drug formulation identifier field, and a spray parameter structure. The spray parameter structure encapsulates the nozzle switching timing array, the swing angle array, and the flow modulation curve parameters.

[0038] It is understandable that the logic of sorting the machine-executable pesticide application instruction set by timestamp is based on the geographical order and estimated time of the drone's flight over each pesticide application management unit. The timestamp is calculated by adding a fixed offset to the estimated time of the drone's arrival at the boundary of the pesticide application management unit. The spatial coordinate field stores the key vertex coordinate sequence of the boundary polygon of the pesticide application management unit or the geographic grid code that can uniquely identify the pesticide application management unit. The pesticide formulation identifier field points to a lookup table of predefined pesticide mixing ratios and types. The flow modulation curve parameters in the injection parameter structure can be a set of control point coordinates or polynomial coefficients used to reconstruct the curve.

[0039] Optionally, the neural network model for pesticide demand inference can be trained with different instances for different crop types or growth stages, and the appropriate neural network model instance can be selected and loaded during inference based on the crop information of the current field. When the distribution of collection points within the application management unit is uneven, the aggregated value of multidimensional field information can be obtained by integrating and averaging continuous surface data generated based on spatial kriging interpolation. The flow modulation curve in the nozzle motion planning can be combined with the application rate per unit area output by pesticide demand inference, and integral calculations can be used to ensure that the total amount of pesticide sprayed during the flight over the application management unit is consistent with the recommended dosage.

[0040] Optionally, the generation of the machine-executable pesticide application instruction set can be completed at the UAV ground control station and uploaded to the UAV flight control system, or it can be generated in real time on the UAV's onboard computing unit. Each instruction in the instruction sequence precisely corresponds to the action of the UAV when flying over a specific pesticide application management unit. This means that the validity period of an instruction starts from when the UAV enters the boundary of the corresponding pesticide application management unit and ends when the UAV leaves the boundary of that pesticide application management unit. The parallel execution of pesticide demand inference and nozzle action planning relies on the support of a multi-tasking operating system or parallel computing framework. The two processes exchange the identifier and geographical boundary information of the pesticide application management unit through shared memory or message passing mechanisms to maintain data synchronization.

[0041] In one embodiment of the invention, the variant spray boom system is composed of multiple independently extendable segments hinged together, each segment having an atomizing nozzle controlled by a solenoid valve at its end. Attitude adjustment refers to the variant spray boom system dynamically changing the extension length and hinge angle of each segment according to the boundary of the application management unit specified in the machine-executable application instruction set, so that the spray profile formed by the nozzle array matches the shape of the application management unit. Flow rate adjustment refers to each solenoid valve-controlled atomizing nozzle controlling the on / off state and flow rate of the pesticide solution with millisecond-level precision according to the flow rate modulation curve specified in the machine-executable application instruction set. The variant spray boom system has a built-in inertial measurement unit that provides real-time feedback on the actual attitude of the spray boom, forming a closed-loop control with the expected attitude of the instruction. Real-time environmental perception data is provided by the UAV's onboard downward-looking wind field sensor and real-time imaging spectrometer, including real-time wind speed and direction below the operating flight path, and instantaneous changes in canopy spectral reflectance. Rolling optimization is performed in sliding time windows; while the UAV executes the machine-executable application instruction set within the current time window, it simultaneously analyzes the real-time environmental perception data of the field area to be covered in the next time window. The rolling optimization process first predicts the pesticide drift trajectory based on real-time wind speed and direction data, and accordingly adjusts the nozzle switching timing and swing angle in the nozzle action plan. Second, it rapidly compares real-time canopy spectral reflectance data with baseline spectral data before operation. If a significant difference is detected, it triggers a rapid recalculation of pesticide demand inference for the affected area. Based on the results of the predicted drift correction and rapid recalculation, the set of machine-executable pesticide application instructions that have not yet been executed is dynamically updated.

[0042] In practical implementation, the variable spray boom system mounted on the drone adjusts its attitude and flow rate according to the machine-executable spray command set. The variable spray boom system consists of multiple independently extendable segments hinged together. Each segment is equipped with an atomizing nozzle controlled by a solenoid valve. Attitude adjustment refers to the variable spray boom system dynamically changing the extension length and hinge angle of each segment according to the spray management unit boundary specified in the machine-executable spray command set, so that the spray profile formed by the nozzle array matches the shape of the spray management unit. Flow rate adjustment refers to each atomizing nozzle controlled by a solenoid valve controlling the on / off state and flow rate of the liquid with millisecond-level precision according to the flow modulation curve specified in the machine-executable spray command set. The variable spray boom system has a built-in inertial measurement unit (IMU) that provides real-time feedback on the actual attitude of the spray boom. The feedback data from the IMU and the expected attitude of the machine-executable spray command set form a closed-loop control, while simultaneously incorporating real-time environmental awareness during spraying. The system performs rolling optimization of subsequent instructions based on real-time environmental perception data, which is provided by the UAV's onboard downward-looking wind field sensor and real-time imaging spectrometer. This real-time environmental perception data includes real-time wind speed and direction below the operational flight path, as well as instantaneous changes in canopy spectral reflectance. Rolling optimization is performed in units of sliding time windows. While the UAV executes the machine-executable pesticide application instruction set within the current time window, it simultaneously analyzes the real-time environmental perception data of the field area to be covered in the next time window. Rolling optimization first predicts the pesticide drift trajectory based on real-time wind speed and direction data, and accordingly corrects the nozzle switching timing and swing angle in the nozzle action planning. Secondly, rolling optimization quickly compares the real-time canopy spectral reflectance data with the baseline spectral data before operation. If a significant difference is detected, it triggers a rapid recalculation of the pesticide demand inference for the affected area. Based on the results of the predicted drift correction and rapid recalculation, the system dynamically updates the subsequent machine-executable pesticide application instruction set that has not yet been executed.

[0043] In some embodiments, the attitude adjustment of the variant spray boom system can be achieved through a predefined shape mapping function. The input of the shape mapping function is a description of the boundary polygon of the drug administration unit relative to the UAV's fuselage coordinate system, and the output of the shape mapping function is a sequence of target extension lengths and target articulation angles for each segment of the variant spray boom system. The flow control of the atomizing nozzle controlled by the solenoid valve is driven by a pulse width modulation signal. The duty cycle of the pulse width modulation signal is proportional to the instantaneous flow rate value specified by the flow modulation curve in the machine's executable drug administration command set. The actual attitude data fed back by the inertial measurement unit includes the pitch angle, roll angle, and yaw angle of each segment of the variant spray boom system. The closed-loop control algorithm compares the difference between the actual attitude data and the expected attitude of the command and generates adjustment signals, which are sent to the drive motors and articulated servos of each segment. The downward-looking wind field sensor typically uses an ultrasonic or hot-film anemometer array, and the real-time imaging spectrometer continuously acquires narrow-band spectral images as the UAV flies by. The sliding time window length in the rolling optimization is set according to the UAV's flight speed and the processing delay of the computing system. The sliding time window length ensures sufficient time to complete data analysis, prediction calculation, and command update for the next window area.

[0044] It can be understood that a specific implementation of the shape mapping function in attitude adjustment can be expressed as solving an optimization problem, with the goal of minimizing the geometric difference between the expected spraying area formed by the nozzle array and the boundary polygon of the pesticide management unit. The formula is as follows: ; Where: symbol This represents the geometric loss function that measures the difference between two polygonal regions, with the symbol... This represents the extension / retraction length vector of each segment of the variant boom system. and hinge angle vector The determined polygonal region of the nozzle array projected onto the horizontal plane, symbol This represents the boundary polygon of the drug administration management unit specified by the machine's executable drug administration instruction set. The execution of the flow modulation curve is manifested as the generation of a time-varying control signal function for each solenoid valve-controlled atomizing nozzle. ,function The value range corresponds to the nozzle opening or flow rate. The prediction of the liquid drift trajectory in rolling optimization can use a diffusion equation based on the Lagrange particle model. The equation inputs include real-time wind speed and direction, UAV altitude, and nozzle droplet size distribution.

[0045] In practice, the specific parameter mapping relationship between attitude and flow adjustment can be defined by a lookup table or configuration matrix. For example, for a variant boom system with three independently retractable segments, refer to Table 1 for the typical attitude parameter configurations for application management units of different shape categories.

[0046] Table 1. Mapping table of attitude parameters for the variant boom spray system: In some embodiments, the closed-loop control of the inertial measurement unit can employ a proportional-integral-derivative (PID) control algorithm, with the parameters tuned based on the dynamic model of the variant spray boom system. Real-time wind speed and direction data are used to construct a local two-dimensional wind field model. Pesticide drift trajectory prediction is achieved by simulating droplet movement within this wind field model. The prediction results are used to generate forward compensation or delay commands for nozzle switching timing, and reverse deflection compensation commands for swing angles. Rapid comparison between real-time canopy spectral reflectance data and baseline spectral data can be accomplished by calculating the difference between key vegetation indices at corresponding spatial locations. When the difference exceeds a set threshold, the system determines that the crop condition in that area has changed significantly and triggers an alarm. Rapid recalculation of pesticide demand inference uses the same neural network model as the main inference process, but the input data is replaced with new multi-dimensional field information aggregation values ​​obtained by interpolation or estimation based on the latest real-time environmental perception data. The result of rapid recalculation is used to generate an updated pesticide formulation and dosage command, which replaces the command of the corresponding pesticide management unit in the original machine-executable pesticide application command set.

[0047] It is understandable that the dynamic updating of the set of machine-executable dosing instructions that has not yet been executed is accomplished through the communication interface between the UAV flight control system and the mission management system. Update instructions are issued incrementally and overwrite the corresponding timestamps or spatial coordinates in the original instruction queue. The millisecond-level precision control of the atomizing nozzle controlled by the solenoid valve relies on the support of a high-precision timer and a high-speed digital input / output interface. The drive mechanisms of each segment of the variant spray boom system can use electric push rods or linear motors for extension and retraction, and servo motors for articulated rotation. All computational tasks during the rolling optimization process must be completed before the sliding time window ends to ensure that the updated set of machine-executable dosing instructions takes effect promptly before the UAV enters the next work area.

[0048] Optionally, the shape mapping function for attitude adjustment can also be implemented by training a deep learning model. This model takes the image or coordinate sequence of the application management unit boundary as input and directly outputs the target pose parameters for each segment of the variant boom system. The flow modulation curve can be customized based on the spatial distribution of confidence values ​​within the application management unit, setting higher flow rates for sub-regions with high confidence values ​​within corresponding time periods. Feedback data from the inertial measurement unit can be used not only for closed-loop control but also to record the actual spray trajectory and perform post-event comparative analysis with the commanded expected trajectory. The data sampling frequency of the downward-looking wind field sensor needs to be sufficiently high to capture the impact of instantaneous wind field fluctuations on drift prediction.

[0049] Optionally, data acquired by the real-time imaging spectrometer can be processed in real-time using onboard edge computing equipment to extract the normalized difference vegetation index or other spectral feature maps, and perform pixel-by-pixel differential operations with the pre-loaded base feature maps to quickly locate areas of change. The priority of predictive drift correction and rapid recalculation in rolling optimization can be set; for example, when a sharp increase in wind speed is detected, drift correction calculation is prioritized; when a large-area spectral change is detected, rapid recalculation of pesticide demand inference is prioritized. The updated machine-executable pesticide application command set can be simultaneously backed up to the ground control station for mission auditing and effect evaluation while being sent to the UAV flight control system. Each solenoid valve-controlled atomizing nozzle in the variant spray boom system can have an independent pesticide pipeline and metering unit to achieve rapid switching and precise delivery of different pesticide formulations.

[0050] See Figure 4 This is a bar chart comparing the attitude parameters of a pesticide spraying drone's variant boom system. It shows the extension length and hinge angle parameters of the three boom segments corresponding to different shaped spraying management units. Different shapes of spraying units correspond to different boom parameters, ensuring that the spray profile of the nozzle array matches the shape of the target area, improving spraying accuracy. It provides a quantitative reference for drone boom attitude adjustment, avoiding missed sprays and double sprays caused by improper parameters. Through differentiated parameter configuration, it achieves efficient coverage of complex shaped areas, reducing drone operation path redundancy. It provides a clear parameter mapping relationship for the dynamic adjustment of the drone's variant boom, simplifying the boom attitude control logic and improving the efficiency and accuracy of attitude adjustment during pesticide application.

[0051] In one embodiment of the invention, the dynamic relocation strategy of the mobile ground sensor node network is centrally scheduled by the ground monitoring platform. Its decision-making is based on the dynamic changes in the pest and disease risk heat map and the upcoming operational flight path of the UAV. When the pest and disease risk heat map shows a sharp increase in the confidence value of a certain area, the dynamic relocation strategy instructs the nearest mobile ground sensor node to move to that area and collect data more densely. Before the UAV performs spraying operations, the dynamic relocation strategy instructs the mobile ground sensor node network to conduct supplementary dense sampling in the area below the planned flight path to provide the latest field information for final instruction verification. The initial structure of the crop pathology knowledge graph is constructed using domain expert knowledge. After each operation cycle, the system archives the complete multi-dimensional field information collected, the final machine-executable pesticide application instruction set, and the subsequent pesticide application effect data evaluated by the mobile ground sensor node network as a complete case. Offline machine learning algorithms are used to analyze the historical case library to discover potential correlation patterns between new combinations of environmental factors and pest and disease occurrence. Newly discovered association patterns are integrated into the crop pathology knowledge graph by adding new nodes or edges, or by adjusting the probability weights of existing edges, thereby enabling the graph to evolve autonomously.

[0052] In practical implementation, the dynamic relocation strategy of the mobile ground sensor node network and the updating and expansion mechanism of the crop pathology knowledge graph are implemented. The dynamic relocation strategy of the mobile ground sensor node network is centrally scheduled by the ground monitoring platform. The decision-making basis of the dynamic relocation strategy comes from the dynamic changes of the pest and disease risk heat distribution layer and the operation flight path to be executed by the UAV. When the pest and disease risk heat distribution layer shows a sharp increase in the confidence value of a certain area, the dynamic relocation strategy will instruct the nearest mobile ground sensor node to move to that area and collect data more intensively. Before the UAV performs spraying operations, the dynamic relocation strategy will instruct the mobile ground sensor node network to conduct spraying in the area below the planned flight path. Supplementing with intensive sampling to provide the latest field information for final instruction verification, the initial structure of the crop pathology knowledge graph is constructed from domain expert knowledge. After each operation cycle, the system archives the complete multidimensional field information collected, the final machine-executable pesticide application instruction set, and the subsequent pesticide application effect data evaluated through a mobile ground sensor node network as a complete case. By analyzing the historical case library through offline machine learning algorithms, potential correlation patterns between new combinations of environmental factors and the occurrence of pests and diseases are discovered. The newly discovered correlation patterns are integrated into the crop pathology knowledge graph in the form of adding nodes or edges, or adjusting the probability weights of existing edges, to achieve the autonomous evolution of the graph.

[0053] In some embodiments, the process of centrally scheduling dynamic relocation strategies by the ground monitoring platform is based on an optimization model. The objective function of the optimization model is to maximize the perception coverage of high-risk areas and planned operation areas within a finite time. Constraints include the energy limits, movement speeds, and communication ranges of each autonomous mobile platform in the mobile ground sensor node network. The dynamic changes in the pest and disease risk heat map layer are quantified by the spatiotemporal gradient of the confidence value in the monitoring layer. When the confidence value of a certain local area increases beyond a preset threshold over two consecutive monitoring periods, the system determines that the confidence value of that area has risen sharply. In the decision logic, the "nearest mobile ground sensor node" is determined based on the Euclidean distance between the current location of the autonomous mobile platform and the center of the risk area. Encrypted data collection is implemented. This means adjusting the sampling interval of the node from the conventional mode to the high-frequency mode, and possibly expanding its spectral imaging scanning range. During the supplementary intensive sampling phase before the UAV performs spraying operations, the dynamic relocation strategy calculates the spatial range and time series of the flight path coverage based on the flight plan uploaded by the UAV, and schedules all available mobile ground sensor nodes under and around the flight path within a certain buffer zone to perform a round of synchronous, high spatial resolution sampling within a specific time window before the operation. The collected multidimensional field information is immediately transmitted to the processing center for final verification of the machine-executable pesticide application instruction set based on the latest data. The verification content includes, but is not limited to, verifying whether the pest and disease risk level within the pesticide application management unit has changed significantly.

[0054] It is understandable that the updating and expansion mechanism of the crop pathology knowledge graph is an offline, periodic process. The initial structure of the domain expert knowledge construction defines the core entity nodes and basic causal relationship edges of the graph. After each operation cycle, the archived complete case includes time-series data of timestamps, geographical range, multi-dimensional field information, details of the executed machine-executable pesticide application instruction set, and crop status assessment data at a specific time point after pesticide application. Pesticide effect data is collected by a mobile ground sensor node network according to plan after pesticide application and is quantified by comparing changes in specific spectral indices or disease symptoms before and after pesticide application. When offline machine learning algorithms analyze the historical case library, association rule mining or graph representation learning techniques are used to extract the statistical associations between frequently occurring environmental factor patterns and disease / pest status or pesticide effects from a large amount of case data. If newly discovered association patterns reach a certain confidence and support threshold, they are submitted to the graph management system. The graph management system decides whether to add them as new nodes, new edges, or perform Bayesian updates on the probability weights of existing edges based on the content of the new patterns.

[0055] In practical implementation, the centralized scheduling algorithm for the dynamic relocation strategy can adopt a task allocation mechanism based on dynamic priorities. Each mobile ground sensor node is assigned a dynamically updated task priority list. The tasks in the task priority list include routine inspection line tasks, emergency monitoring tasks with a sharp increase in response confidence value, and pre-operation sampling tasks for UAV flight paths. The ground monitoring platform assigns priorities to each task according to the global optimization objective and issues movement and data collection commands. The autonomous evolution process of the crop pathology knowledge graph can be modeled using the graph embedding learning framework shown in the following formula to discover new associations between nodes: ; Where: symbol The loss function of the graph embedding model is represented by the symbol. Represents the existing set of edges in the graph. All node pairs To perform summation, the sign and Representing nodes respectively and nodes The low-dimensional vector representation obtained through model learning is called the embedded vector, and its symbol is... This represents the sigmoid activation function, symbol... The dot product of two embedding vectors is used to measure the strength of the association between nodes. (Symbol: ) This represents a pair of nodes obtained through negative sampling that do not have edges. Summation encourages the model to separate the embedding vectors of irrelevant nodes by minimizing the loss function. Learned node embedding vectors It can be used to predict potential new edges in the graph, i.e. new association patterns. Node pairs with high prediction scores will be regarded as candidates for newly discovered association patterns.

[0056] In some embodiments, after a command indicating a sharp increase in the response confidence value is issued to a mobile ground sensor node, the autonomous mobile platform of the mobile ground sensor node plans the shortest path to avoid obstacles and moves towards the target area. During the movement, it continuously uploads its location status. Upon reaching the target area, the mobile ground sensor node operates in encrypted data acquisition mode until it receives a stop command from the ground monitoring platform or the preset data acquisition duration ends. Supplementary intensive sampling tasks for UAV flight paths require each mobile ground sensor node to complete sampling within a specified time window. Therefore, the dynamic relocation strategy will comprehensively consider various factors when dispatching such tasks. Taking into account the current location of the node, the time required to move to the sampling point, and the remaining energy of the node, the feasibility of the task is ensured. The update operation of the crop pathology knowledge graph is carried out in a human-computer interactive review interface. The system presents the new association patterns discovered by the offline machine learning algorithm to agricultural experts in a visual form. The agricultural experts then make the final confirmation on whether to formally integrate them into the crop pathology knowledge graph. When integrating new association patterns, if they are completely new pest or disease types or environmental factors, new nodes are created; if they are relationships between existing nodes that have not been recorded, new edges are added; if they are corrections to the confidence of existing relationships, the probability weights of the corresponding edges are adjusted.

[0057] It is understandable that the transmission of instructions in the dynamic relocation strategy relies on a stable, low-latency wireless communication network. The ground monitoring platform and the mobile ground sensor node network maintain a connection through a wireless LAN or mobile communication network. The mobile ground sensor node network can evaluate the pesticide application effect data by revisiting the application area at fixed time points such as the 3rd and 7th day after application, collecting multi-dimensional field information with the same indicators as before application, and evaluating the effect through difference analysis. The historical case database is stored in a structured database, with each case indexed by a unique operation identifier and associated with the corresponding field digital base map geographic area. The offline machine learning algorithm runs on a computing server isolated from the real-time control system, periodically scanning newly added case data and starting analysis tasks. The output of the analysis task is a structured association pattern report.

[0058] Optionally, the decision-making of the dynamic relocation strategy can also incorporate a predictive model. Based on the evolution trend of the disease and pest risk heat distribution layer, the predictive model can predict the possible risk outbreak areas in the future and schedule the deployment of a mobile ground sensor node network to the predicted area in advance. In addition to a sharp increase in the confidence value, the triggering condition for encrypted data collection can also be determined by combining the rules of the crop pathology knowledge graph about the precursors of specific diseases. For example, when the soil moisture is continuously higher than a certain threshold and the canopy temperature shows a specific pattern, encrypted data collection of the relevant area will be triggered even if the confidence value does not rise sharply. Optionally, the offline machine learning algorithm in the crop pathology knowledge graph update and expansion mechanism can employ various techniques, including but not limited to frequent pattern mining, causal discovery algorithms, or time-series pattern mining algorithms. The association patterns extracted from the historical case library are not limited to the relationship between environmental factors and pests and diseases, but can also include the relationship between pesticide application operations and effect feedback, which can be used to optimize future pesticide demand inference. The probability weight update of graph nodes and edges can adopt the Bayesian network parameter learning method, using historical case data as new observational evidence to update the conditional probability table in the network. The human-computer interaction review interface allows agricultural experts to accept, reject, or modify the association patterns recommended by the system. All expert operations will be recorded and used to improve the recommendation quality of the machine learning algorithm in the future.

[0059] See Figure 5 This is a heatmap of the spatial reliability value of pest and disease risk. It displays the distribution of pest and disease risk within a field area using two-dimensional spatial coordinates. The color intensity corresponds to the risk reliability value (color bar on the right), with higher values ​​(closer to 1.4) indicating higher risk. The location of the risk center at different time points is marked, showing a slight shift of the risk center in the positive X-axis direction over time, reflecting the dynamic spread trend of pests and diseases. This type of map is used for precise pesticide application decisions by agricultural spraying drones. High-risk areas (yellow / bright yellow) require focused application, and the movement of the risk center can help the drone dynamically adjust its operational path. By monitoring the changes in the risk center at different time points, the spread trend of pests and diseases can be monitored in real time, providing a basis for adjusting the drone's operational sequence and path, thus improving the timeliness of pesticide application.

[0060] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0061] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for precise pesticide application using a pesticide spraying drone, characterized in that, Perform the following sequence of operations: Based on a pre-generated digital base map of the field, a mobile ground sensor node network is deployed, which dynamically collects multi-dimensional field information related to crop stress. The multidimensional field information is iteratively compared and correlated with a pre-constructed crop pathology knowledge graph to dynamically identify potential target patches for pesticide intervention. Based on the boundary features and internal heterogeneity of the potential target patches, a biomimetic fractal segmentation rule is used to adaptively divide the potential target patches into fine-grained regions, generating a drug application management unit with a hierarchical structure. For each of the aforementioned drug application management units, drug demand reasoning and nozzle action planning are executed in parallel to generate a set of machine-executable drug application instructions containing spatial coordinates, drug formulation, and spraying parameters; The drone's variant spray boom system adjusts its attitude and flow rate according to the machine's executable pesticide application command set, while simultaneously integrating real-time environmental perception data during spraying to continuously optimize subsequent commands.

2. The method for precise pesticide application using a drone according to claim 1, characterized in that, The deployment of a mobile ground-based sensor node network based on a pre-generated digital base map of the field, specifically including the dynamic collection of multi-dimensional field information related to crop stress by the mobile ground-based sensor node network, includes: The digital base map of the field is generated by fusing historical remote sensing images, soil electrical conductivity layers, and topographic elevation models, and defines the distribution of static attributes within the field. The mobile ground sensing node network is carried by several autonomous mobile platforms, each platform integrating a spectral imager, a microclimate sensor and a soil probe; The mobile ground sensor node network is initially deployed based on the attribute gradient bands marked on the field digital base map, and dynamically repositioned along the preset inspection line during the operation cycle. The multidimensional field information includes, but is not limited to, crop canopy spectral reflectance, leaf temperature, air temperature and humidity, and root zone soil moisture and nutrient concentration; The dynamic acquisition refers to the mobile ground sensor node network triggering high-density information acquisition events at non-fixed intervals, based on the disease-prone periods and areas inferred from the crop pathology knowledge graph.

3. The method for precise pesticide application using a drone according to claim 2, characterized in that, The multidimensional field information is iteratively compared and mapped with a pre-constructed crop pathology knowledge graph to dynamically identify potential target patches for pesticide intervention, specifically including: The crop pathology knowledge graph is stored in a graph structure, where nodes represent specific disease or pest states and edges represent the causal relationships and probability weights between different environmental factors and the occurrence and development of diseases. The iterative comparison refers to inputting the real-time acquired multidimensional field information into the crop pathology knowledge graph, activating the matching nodes, and performing reasoning propagation along the edges of the graph to calculate the confidence value of each field location in different pest and disease states. The association mapping refers to binding the inferred pest and disease status confidence value with the corresponding geographic coordinates to generate a dynamically updated pest and disease risk heat distribution layer on the field digital base map. The potential target patches are automatically extracted by setting a confidence threshold on the pest and disease risk heat distribution layer, and their boundaries are composed of continuous pixel regions exceeding the threshold. The dynamic identification process is repeated after each information collection cycle, so that the shape, area and location of the potential target patches can evolve and change over time.

4. The method for precise pesticide application using a drone according to claim 3, characterized in that, Based on the boundary characteristics and internal heterogeneity of the potential target patches, a biomimetic fractal segmentation rule is used to adaptively divide the potential target patches into fine-grained regions, generating a hierarchical drug delivery management unit, specifically including: The boundary features are quantified by calculating the tortuosity and concavity of the contour of the potential target patch; The internal heterogeneity is assessed by analyzing the statistical variance and spatial autocorrelation of the confidence values ​​on the heat map of the pest and disease risk distribution within the potential target patches. The biomimetic fractal segmentation rule simulates the branching growth pattern of veins or roots in nature. Its core is to recursively divide an irregular region into unequal areas according to the degree of heterogeneity until the internal homogeneity of the sub-region meets the preset standard. The adaptive fine-grained region segmentation process is as follows: taking the entire potential target patch as the initial region, and determining whether to trigger segmentation based on the internal heterogeneity evaluation results; if triggered, determining the direction of the initial segmentation line based on the boundary tortuosity, and dividing the region into two sub-regions; recursively repeating the evaluation and segmentation process for each sub-region. The hierarchical drug application management unit ultimately manifests as a tree structure, with the root node corresponding to the original potential target patch, the leaf nodes corresponding to the final drug application management unit that meets the homogeneity standard, and the intermediate nodes representing the transition management areas at different scales.

5. A method for precise pesticide application using a drone according to claim 4, characterized in that, For each of the aforementioned application management units, parallel execution of pesticide demand reasoning and nozzle motion planning generates a machine-executable application instruction set containing spatial coordinates, pesticide formulation, and spraying parameters, specifically including: The pesticide demand reasoning is performed independently for each pesticide application management unit. Its input is the aggregated value of the multidimensional field information from all collection points within the pesticide application management unit, as well as the dominant pest and disease types and degrees obtained from the crop pathology knowledge graph. The drug demand reasoning is accomplished through a trained neural network model, which directly outputs recommended drug type, concentration, and application rate per unit area. The nozzle motion planning is carried out synchronously with the liquid demand reasoning, based on the geographical boundary of the application management unit, the current flight speed and altitude of the UAV, and the mechanical constraints of the variant spray boom system. The nozzle motion planning calculation yields the nozzle's switching sequence, lateral swing angle, and flow modulation curve within each swing cycle; The machine can execute a set of application instructions that encodes the spatial coordinates, drug formulation, and spraying parameters into a sequence of instructions ordered by timestamps, wherein each instruction precisely corresponds to the action of the UAV when it flies over a specific application management unit.

6. The method for precise pesticide application using a drone according to claim 5, characterized in that, The variable-propeller spray system on the drone adjusts its attitude and flow rate according to the machine-executable drug delivery command set, specifically including: The variant spray bar system is composed of multiple independently telescopic segments hinged together, with an atomizing nozzle controlled by a solenoid valve installed at the end of each segment; Attitude adjustment refers to the variant spray boom system dynamically changing the extension length and hinge angle of each segment according to the boundary of the application management unit specified in the machine's executable application instruction set, so that the spray profile formed by the nozzle array matches the shape of the application management unit; The flow adjustment refers to the control of the on / off state and flow rate of the liquid medicine by each atomizing nozzle controlled by the solenoid valve according to the flow modulation curve specified in the set of executable drug delivery instructions of the machine, with millisecond-level precision. The variant spray boom system has a built-in inertial measurement unit that provides real-time feedback on the actual attitude of the spray boom, forming a closed-loop control with the commanded expected attitude.

7. A method for precise pesticide application using a drone according to claim 6, characterized in that, Simultaneously, during the spraying process, real-time environmental perception data is integrated to continuously optimize subsequent instructions, specifically including: The real-time environmental perception data is provided by the UAV's onboard downward-looking wind field sensor and real-time imaging spectrometer, including real-time wind speed and direction below the operational flight path and instantaneous changes in canopy spectral reflectance. The rolling optimization is performed in units of sliding time windows. When the UAV executes the machine-executable pesticide application instruction set within the current time window, the real-time environmental perception data of the field area to be covered in the next time window is analyzed simultaneously. The rolling optimization first predicts the trajectory of the liquid drift based on real-time wind speed and direction data, and then corrects the nozzle switching timing and swing angle in the nozzle action planning in advance accordingly. The rolling optimization then compares the real-time canopy spectral reflectance data with the baseline spectral data before the operation. If a significant difference is detected, a rapid recalculation of the inference of the drug demand in the affected area is triggered. Based on the results of predictive drift correction and fast recalculation, the set of machine-executable drug delivery instructions that have not yet been executed is dynamically updated.

8. A method for precise pesticide application using a drone according to claim 7, characterized in that, The dynamic relocation strategy of the mobile ground sensor node network specifically includes: The dynamic relocation strategy is centrally scheduled by the ground monitoring platform, and its decision-making is based on the dynamic changes of the pest and disease risk heat distribution layer and the operation route that the UAV is about to execute. When the pest and disease risk heat distribution layer shows a sharp increase in the confidence value of a certain area, the dynamic relocation strategy will instruct the nearest mobile ground sensor node to move to the area and collect data more intensively. Before the drone performs spraying operations, the dynamic relocation strategy instructs the mobile ground sensor node network to perform supplementary intensive sampling in the area below the planned flight path to provide the latest field information for final command verification.

9. A method for precise pesticide application using a drone according to claim 8, characterized in that, The updating and expansion mechanism of the crop pathology knowledge graph specifically includes: The initial structure of the crop pathology knowledge graph was constructed from domain expert knowledge; After each operation cycle is completed, the system will archive the complete multi-dimensional field information collected this time, the final set of machine-executable pesticide application instructions, and the pesticide application effect data evaluated by the mobile ground sensor node network as a complete case. By analyzing historical case databases using offline machine learning algorithms, potential correlation patterns between new combinations of environmental factors and the occurrence of pests and diseases can be discovered. The newly discovered association patterns are integrated into the crop pathology knowledge graph by adding new nodes or edges, or by adjusting the probability weights of existing edges, so as to achieve the autonomous evolution of the graph.

10. A precision pesticide spraying drone system, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the precise pesticide application method using a pesticide spraying drone as described in any one of claims 1 to 9.

Images