A mobile intelligent agricultural unmanned swarm operation method and system

Abstract

This invention belongs to the field of smart agriculture technology and collaborative operation of unmanned systems. It discloses a method and system for mobile intelligent agricultural unmanned swarm operations. The method includes: generating a global verification path that minimizes the sum of the total travel distance and the soil compaction penalty; generating a soil property distribution map and fusing aerial imagery data to obtain a farmland fusion perception map; rasterizing the farmland areas in the fusion perception map, treating each abnormal grid as a task, and assigning bidding weights to agents for tasks based on the suitability of the agents in performing the corresponding tasks; allocating agent tasks with the objective of maximizing the total sum of global bidding weights; and predicting the trajectory of the aerial operation drones using a discrete linear model, with each aerial operation drone solving the optimization problem locally within a distributed model predictive control framework. This invention improves operational efficiency and accuracy.

Description

Technical Field

[0001] This invention relates to the field of smart agriculture technology and collaborative operation of unmanned systems, and particularly to a mobile smart agriculture unmanned cluster operation method and system. Background Technology

[0002] Current agricultural operations generally suffer from problems such as labor shortages, low efficiency, serious resource waste, and low levels of intelligence. While existing technologies include independent applications such as drone-based plant protection and ground robot inspections, these systems lack efficient collaboration and largely rely on fixed base stations or manual intervention, making it impossible to achieve large-scale, fully automated, and closed-loop operations.

[0003] Existing agricultural unmanned systems mostly adopt a single operation mode, such as using only drones for spraying or only ground robots for data collection. They lack an integrated "sky-ground" collaborative perception and execution mechanism and lack a mobile, self-sufficient energy integrated operation platform, resulting in limited operation range, slow response speed, and low resource utilization efficiency.

[0004] Therefore, how to provide a mobile intelligent agricultural unmanned cluster operation method and system is an urgent problem to be solved. Summary of the Invention

[0005] This invention provides a configuration method to solve the problems described above in the prior art.

[0006] According to a first aspect of the present invention, a mobile intelligent agricultural unmanned cluster operation method is provided.

[0007] In one embodiment, the mobile intelligent agricultural unmanned swarm operation method includes:

[0008] Using aerial imagery data, a set of key verification points is generated. Combined with a path planning algorithm, a global verification path is generated that minimizes the sum of the total travel distance and the soil compaction penalty. Based on the global verification path, ground truth data is obtained to generate a soil property distribution map. By fusing the aerial imagery data, a farmland fusion perception map is obtained.

[0009] The farmland areas in the integrated perception map are rasterized, and each abnormal grid is treated as a task. Based on the suitability of the agent to perform the corresponding task, the bidding weight of the agent for the task is assigned. The task allocation of the agent is carried out with the goal of maximizing the sum of the global bidding weights.

[0010] When the intelligent agent performs a task, it uses a discrete linear model to predict the trajectory of the aerial operation drone. Each aerial operation drone solves the optimization problem locally in the distributed model predictive control framework to maintain formation and avoid obstacles in real time during flight. The cost function includes tracking error, control cost and cooperative collision avoidance.

[0011] According to a second aspect of the present invention, a mobile intelligent agricultural unmanned cluster operation system is provided.

[0012] In one embodiment, the mobile intelligent agricultural unmanned swarm operation system includes:

[0013] The map fusion construction module is used to generate a set of key verification points using aerial imagery data, and combined with a path planning algorithm to generate a global verification path that minimizes the sum of the total driving distance and the soil compaction penalty. Based on the global verification path, ground truth data is obtained to generate a soil attribute distribution map, and the aerial imagery data is fused to obtain a farmland fusion perception map.

[0014] The agent task allocation module is used to rasterize the farmland areas in the farmland fusion perception map and treat each abnormal grid as a task. Based on the suitability of the agent to perform the corresponding task, the module allocates the agent's bidding weight for the task. The goal is to maximize the total sum of global bidding weights in the agent task allocation.

[0015] The UAV trajectory solving module is used to predict the motion trajectory of aerial operation UAVs using a discrete linear model when the intelligent agent performs a task. Each aerial operation UAV solves the optimization problem locally in a distributed model predictive control framework to maintain formation and avoid obstacles in real time during flight. The cost function includes tracking error, control cost, and cooperative collision avoidance.

[0016] According to a third aspect of the present invention, a computer device is provided.

[0017] In some embodiments, the computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the method described above.

[0018] According to a fourth aspect of the present invention, a computer-readable storage medium is provided.

[0019] In one embodiment, a computer program is stored on the computer-readable storage medium, which, when executed by a processor, implements the steps of the above method.

[0020] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects:

[0021] 1. Improved Operational Efficiency and Accuracy: By employing a space-ground collaborative sensing algorithm and Kriging interpolation fusion, large-area rapid scanning is combined with precise point verification to generate centimeter-level accuracy operation prescription maps, resolving the contradiction between the limitations of traditional remote sensing (wide coverage but low accuracy) and ground mapping (accurate but slow speed). Through multi-agent task allocation based on dynamic weights and spatiotemporal A* path planning, heterogeneous cluster parallel and conflict-free operations are achieved, resulting in an overall system operational efficiency improvement of more than 10 times.

[0022] 2. Achieve true Level 4 agricultural unmanned swarm intelligence: Through distributed model predictive control (DMPC), realize high-density, high-safety collaborative flight and real-time collision avoidance of unmanned swarms in unstructured environments; through edge-cloud collaborative predictive agronomic decision-making, upgrade plant protection and nutrient management from reactive to predictive, which is the core intelligent guarantee for achieving a reduction of more than 30% in pesticide and fertilizer use.

[0023] 3. Significantly enhanced system reliability, autonomy, and economy: The heterogeneous cluster real-time collision avoidance algorithm ensures the safety of dense formation operations; task allocation based on multi-objective utility enables the system to automatically optimize scheduling according to the real-time status of equipment, such as power and load; the hybrid energy and automated logistics system supports 24 / 7 uninterrupted operation. The platform-based design reduces the cost per operation and provides a clear return on investment period.

[0024] 4. Building a Trusted Data Foundation and Value-Added Service Entry Point for the Entire Chain: Based on a permissioned blockchain notarization mechanism, each automated agricultural operation is solidified into an immutable digital certificate, creating a brand-new value transfer model for agricultural products and providing a trusted data foundation for value-added services such as branded agriculture, green finance, and agricultural insurance.

[0025] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

[0026] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0027] Figure 1 This is a flowchart illustrating a mobile intelligent agricultural unmanned swarm operation method according to an exemplary embodiment;

[0028] Figure 2 This is a block diagram illustrating a mobile intelligent agricultural unmanned swarm operation system according to an exemplary embodiment;

[0029] Figure 3 This is a schematic diagram of the structure of a computer device according to an exemplary embodiment;

[0030] Figure 4 This is a schematic diagram illustrating the overall system architecture and collaborative control according to an exemplary embodiment;

[0031] Figure 5 This is a schematic diagram of a space-ground collaborative sensing and data fusion algorithm according to an exemplary embodiment;

[0032] Figure 6 This is a multi-agent dynamic task allocation and cooperative control diagram illustrated according to an exemplary embodiment;

[0033] Figure 7 This is a diagram illustrating an edge-cloud collaborative predictive agronomic decision-making architecture according to an exemplary embodiment;

[0034] Figure 8 This is a flowchart illustrating blockchain agricultural data storage and traceability according to an exemplary embodiment;

[0035] Figure 9 This is a schematic diagram illustrating the Merkle tree verification principle according to an exemplary embodiment. Detailed Implementation

[0036] Figure 1 An embodiment of a mobile intelligent agricultural unmanned cluster operation method according to the present invention is shown.

[0037] In this optional embodiment, the mobile intelligent agricultural unmanned swarm operation method includes:

[0038] S101. Using aerial imagery data, generate a set of key verification points. Combined with a path planning algorithm, generate a global verification path that minimizes the sum of the total driving distance and the soil compaction penalty. Based on the global verification path, obtain ground truth data to generate a soil attribute distribution map. Integrate the aerial imagery data to obtain a farmland fusion perception map.

[0039] S102. The farmland area in the farmland fusion perception map is rasterized, and each abnormal grid is treated as a task. Based on the suitability of the agent to perform the corresponding task, the bidding weight of the agent for the task is assigned. The task allocation of the agent is carried out with the goal of maximizing the sum of the global bidding weights.

[0040] S103. When the intelligent agent performs the task, it uses a discrete linear model to predict the trajectory of the aerial operation drone. Each aerial operation drone solves the optimization problem locally in the distributed model prediction control framework to maintain formation and avoid obstacles in real time during the flight of the aerial operation drone. The cost function includes tracking error, control cost and cooperative collision avoidance.

[0041] In this optional embodiment, using aerial imagery data, a set of key verification points is generated. Combined with a path planning algorithm, a global verification path is generated that minimizes the sum of the total travel distance and the soil compaction penalty, including:

[0042] Aerial imagery data was collected using an aerial monitoring drone. Adaptive threshold segmentation was applied to the aerial imagery data to identify abnormal regions and obtain a binary mask. The centroids of connected regions in the binary mask were used as the set of key verification points. The objective function for solving the path was constructed by minimizing the sum of the total travel distance and the soil compaction penalty term, combined with the set of key verification points and the penalty weight coefficient. Based on the traveling salesman problem model and the objective function of the solution path, a global verification path was planned.

[0043] In this optional embodiment, ground truth data is acquired based on the global verification path to generate a soil attribute distribution map. This map is then fused with aerial imagery data to obtain a farmland fusion perception map, including:

[0044] The ground operation robot performs measurements according to the global verification path to obtain ground truth data; the ground truth data is combined with aerial imagery data through spatial correlation to obtain a soil attribute distribution map; the soil attribute distribution map and aerial imagery data are fused to obtain a farmland integrated perception map.

[0045] In this optional embodiment, assigning bidding weights to agents for tasks based on their suitability for performing the corresponding tasks includes:

[0046] The bidding factors are defined as the cost distance from the agent's current location to the task location, the agent's ability matching degree to the task, the agent's current load, the agent's current remaining energy, the agent's maximum energy, and the urgency of the task. Weight coefficients are assigned to each bidding factor. Based on the bidding factors and weight coefficients, a multi-objective utility function is constructed to calculate the agent's bidding weight for the task. The agents include aerial drones and ground robots.

[0047] In this optional embodiment, the allocation of agent tasks with the objective of maximizing the total sum of global bidding weights includes:

[0048] The agent outputs the optimal n bids, and uses an optimization algorithm to allocate tasks to the agent with the objective of maximizing the sum of global bid weights. The optimization objective and constraints are as follows:

[0049] ;

[0050] In the formula, Indicates the total number of intelligent agents; Indicates the total number of tasks; Represents intelligent agents For the task Bidding weight; Represents a binary allocation variable; Represents intelligent agents Maximum task capacity; This means that each task can be assigned to at most one agent; This indicates that the number of tasks assigned to each agent does not exceed its maximum capacity.

[0051] In this optional embodiment, the discrete linear model includes:

[0052] ;

[0053] In the formula, Indicates drone At any moment The state vector; Indicates drone At any moment The control input vector; Represents the state transition matrix; This represents the control input matrix.

[0054] In this optional embodiment, each aerial operation UAV locally solves the optimization problem within the distributed model predictive control framework, including:

[0055] In each control cycle, each aerial operation UAV independently solves the optimal control problem in the finite time domain within a distributed model predictive control framework, and the cost function is:

[0056] ;

[0057] ;

[0058] in: Indicates the importance of drones control sequence Optimize; Indicates the prediction time domain; Indicates drone At the predicted time The predicted state; Indicates drone At the predicted time Expected reference state; The weighted norm squared represents the state tracking error. This represents a diagonal weight matrix, where the diagonal elements weight the state components such as position error and velocity error, respectively, to penalize the state tracking error (i.e., increase the weight of the state tracking error). The corresponding element will force the drone to track the desired trajectory more accurately. This represents the squared weighted norm of the control input. Indicates and Different diagonal weight matrices have diagonal elements that weight each control component (such as acceleration command) to penalize the magnitude of the control quantity (i.e., increase it). The corresponding element will suppress excessive control actions, achieving energy saving and smooth flight. Indicates drone The set of neighbors; This represents the weighting coefficient of the exclusion term; This represents the distance-based repulsive potential function; Indicates drone At the predicted time The control input vector; Indicates drone At the predicted time The position vector; Indicates drone At the predicted time The position vector; Indicates the minimum safe distance between drones; Represents the state transition matrix; The control input matrix is ​​represented; the first control step of the aerial operation UAV to execute the optimization solution, and in the next cycle, based on the new state and the latest predicted trajectory of the neighbors, the optimization problem is solved again to realize distributed real-time collision avoidance and formation maintenance in the rolling time domain.

[0059] In this optional embodiment, a mobile intelligent agricultural unmanned cluster operation method further includes: achieving predictive agronomic decision-making through edge-cloud collaboration; specifically including: using an AI model deployed in the cloud to make long-term macro-predictive agronomic decisions; distilling the AI ​​model in the cloud to obtain a lightweight model; and sending the lightweight model and the latest prediction results to the vehicle-mounted edge center to make real-time local predictions of agronomic decisions.

[0060] A permissioned blockchain-based evidence storage mechanism solidifies each automated agricultural operation into a digital certificate.

[0061] Figure 2 An embodiment of a mobile intelligent agricultural unmanned swarm operation system of the present invention is shown.

[0062] In this optional embodiment, the mobile intelligent agricultural unmanned swarm operation system includes:

[0063] The map fusion construction module 101 is used to generate a set of key verification points using aerial imagery data, and combine it with a path planning algorithm to generate a global verification path that minimizes the sum of the total driving distance and the soil compaction penalty. Based on the global verification path, ground truth data is obtained to generate a soil attribute distribution map, and the aerial imagery data is fused to obtain a farmland fusion perception map.

[0064] The agent task allocation module 102 is used to rasterize the farmland area in the farmland fusion perception map, and to treat each abnormal grid as a task. Based on the suitability of the agent to perform the corresponding task, the module allocates the bidding weight of the agent for the task. The goal is to maximize the total sum of global bidding weights in the agent task allocation.

[0065] The UAV trajectory solving module 103 is used to predict the motion trajectory of the aerial operation UAV using a discrete linear model when the intelligent agent performs the task. Each aerial operation UAV solves the optimization problem locally in the distributed model prediction control framework to maintain formation and avoid obstacles in real time during the flight of the aerial operation UAV. The cost function includes tracking error, control cost and cooperative collision avoidance.

[0066] To facilitate understanding of the above technical solutions of the present invention, the following further describes the above technical solutions of the present invention from the perspectives of architecture and principle, as follows:

[0067] This invention integrates a mobile carrier platform, an aerial drone swarm, a ground robot, intelligent logistics support, and cloud-edge-device collaborative control into an unmanned agricultural operation method and system, in order to solve the technical problems of poor coordination, discontinuous operation, low level of intelligence, and low resource utilization in existing agricultural operations.

[0068] This invention includes: a mobile, high-performance carrier platform—a mothership—utilizing an off-road capable vehicle chassis and integrating an intelligent cockpit, a drone takeoff and landing deck, equipment bay, and ground robot deployment and retrieval devices; an unmanned system cluster, including aerial agricultural drones, monitoring drones, and ground robots, capable of collaborative operations; an intelligent logistics support system, including an automatic pesticide dispensing system, a hybrid energy system, and automated charging / battery swapping devices; an information control and cloud-edge-device collaborative system, including a vehicle-mounted edge computing center, a self-organizing network communication module, a remote command platform, and an AI agronomic decision-making model; and a blockchain traceability module for recording agricultural operation data and constructing reliable agricultural product traceability archives.

[0069] A three-dimensional perception and closed-loop operation mechanism integrating air and ground: Aerial drones handle large-scale rapid monitoring and spraying, while ground robots perform detailed monitoring of the soil and root zone. Data fusion and real-time decision-making are achieved through a vehicle-mounted edge computing center, forming a closed loop of monitoring-decision-execution-evaluation. Multi-machine collaborative path planning and dynamic task allocation are supported, achieving fully automated, unmanned operation, encompassing three-dimensional perception, AI decision-making, precise execution, and evaluation. A mobile integrated logistics support system integrates an automatic pesticide dispensing system, supporting raw material identification, precise proportioning, and unmanned filling. A hybrid energy system, such as diesel generators combined with photovoltaic and energy storage, achieves energy self-sufficiency and intelligent scheduling. Automatic battery swapping for drones and wireless charging for ground robots are provided, supporting 24 / 7 continuous operation. A blockchain-based trusted traceability method for agricultural operations: Information such as the time, location, pesticide type, dosage, and operator for each operation is recorded on the blockchain; an immutable digital archive of agricultural products is constructed, supporting full-process traceability from field to table. A cloud-edge-device collaborative intelligent decision-making architecture: The vehicle-mounted edge computing center is responsible for real-time data processing and cluster control; the cloud-based AI model provides agronomic decision support such as pest and disease prediction and growth simulation; and remote collaborative operations in areas without network coverage are achieved through 5G / satellite communication.

[0070] This invention is physically composed of the following units:

[0071] 1. Mobile Intelligent Support Platform - Mothership:

[0072] It is modified from a heavy-duty vehicle chassis with high off-road capability. The driver's cab is equipped with a dual-redundant drive-by-wire system and an operator monitoring interface. The middle of the vehicle is the work compartment, which integrates a hydraulic lift-type UAV take-off and landing platform and an automated drone pod that can accommodate 4 operational UAVs and 2 monitoring UAVs. The rear of the vehicle is the equipment compartment, which houses an intelligent drug dispensing unit, a hybrid energy system such as a diesel generator, photovoltaic panels, lithium battery energy storage, and an onboard edge computing center. The rear of the vehicle is equipped with a hydraulic lift tailgate for deploying and retrieving ground robots.

[0073] 2. Unmanned system cluster:

[0074] Aerial Operation Drones (UAV-W): 4 units, payload ≥ 50kg, equipped with millimeter-wave radar omnidirectional obstacle avoidance, RTK / PPK high-precision positioning module, and variable spraying system. Aerial Monitoring Drones (UAV-S): 2 units, equipped with a five-channel multispectral camera and a visible light camera. Ground Operation Robots (UGV): 1 unit, articulated tracked chassis, integrating multi-parameter soil sensors, robotic arm sampling device, lidar, and visual navigation system.

[0075] 3. Remote cloud control center:

[0076] A cloud server cluster is deployed, equipped with AI agronomic models, a digital twin engine, and blockchain-based evidence storage services. Each unit is intelligently connected and scheduled through a collaborative control system, forming a four-in-one collaborative operation system comprising mobile base stations, airborne clusters, ground units, and a cloud-based central processing unit. The collaborative control system serves as the core software, running in the vehicle-mounted edge computing center and some airborne computing units.

[0077] The core algorithm for space-ground coordinated control is as follows:

[0078] A space-ground collaborative sensing algorithm based on multi-source heterogeneous data fusion: This algorithm aims to integrate large-scale aerial remote sensing data with fine ground-based point measurement data to construct accurate farmland situation maps and provide reliable input for decision-making.

[0079] Input: Multispectral image data from UAV-S, preprocessed (e.g., radiometric correction, geometric correction, and mosaicking) to generate a Normalized Difference Vegetation Index (NDVI) map, denoted as a matrix. ,in The geographic coordinates are represented by point measurements of soil moisture (s), electrical conductivity (e), etc., from UGV, denoted as set G. ground ={(x i ,y i ,s i ,e i ,...)}, i=1...N, where N is the number of measurement points. Algorithm flow:

[0080] 1. Initial screening of abnormal areas: for Adaptive threshold segmentation is applied to identify regions with significantly abnormal growth, such as suspected areas of pests and diseases, and a binary mask is obtained. .

[0081] ;

[0082] In the formula, A binary image mask, a matrix, in coordinates At this point, a value of 1 indicates that the area is identified as abnormal, and 0 indicates that it is normal. Vegetation indices, such as NDVI, calculated from aerial UAV multispectral imagery, are displayed in coordinates. The value at that location. : Dynamically set lower and upper thresholds. Values ​​below This may indicate crop wilting or disease, and is higher than [the specified value]. This could indicate weeds or abnormal overgrowth. This formula, through a simple threshold segmentation method, quickly identifies potential problem areas requiring further attention from large-scale remote sensing imagery. Among these, and The threshold is dynamically determined based on historical data and crop type.

[0083] 2. UGV verification path planning: based on The centroids of the connected regions are used as the set of key verification points. An improved Traveling Salesman Problem (TSP) model is used to plan the global verification path for UGVs. The objective function is to minimize the sum of the total travel distance and the soil compaction penalty term.

[0084] ;

[0085] : indicates that the optimization objective is to minimize the subsequent expression. : The set of key verification points, i.e. the set of centroids of the abnormal region. From point Time The passable distance is usually determined by taking into account terrain and obstacles. : Binary decision variable. If the path contains from arrive If the edge is not a valid path length, the value is 1; otherwise, it is 0. λ: Penalty weight coefficient, used to balance the influence of path length and soil compaction. At point The soil compaction index is used to quantify the degree of soil structure damage caused by ground robot movement. This is an improved Traveling Salesman Problem (TSP) model. The objective is not only to plan a shortest path to all critical points—the first term—but also to introduce a penalty for soil compaction—the second term, reflecting soil conservation considerations in agricultural operations.

[0086] The improved Traveling Salesman Problem (TSP) model is a mixed-integer linear programming model constructed by introducing a soil compaction penalty term strongly correlated with agricultural operations, based on the classic symmetric TSP. Its complete structure and key parameters are described below:

[0087] 1. Model Structure. The model aims to plan a route for a Ground Working Robot (UGV) to access all key verification points. The optimal sequence, whose mathematical expression contains the following core parts:

[0088] Decision variable: binary variable x ab If the planned path includes an arc segment that travels directly from point a to point b, then x ab =1; otherwise 0. Where a, b∈ .

[0089] Objective function (minimize total cost):

[0090] ;

[0091] Part One: This represents the total distance traveled along the route. Where dab This is the Euclidean distance from point a to point b, or the traversable distance considering the terrain.

[0092] Part Two: This is a newly added penalty item for soil compaction. Among them, CPI (p a Let p be a point. a The soil compaction index at this location is obtained through a pre-set soil type map or real-time sensor data. The looser the soil and the higher the moisture content, the higher the value, indicating a greater potential for damage to the soil structure caused by driving at this location. a This is a binary auxiliary variable indicating whether point a is visited by the path (usually y). a =1 holds true for any access point. λ is the penalty weighting coefficient, used to balance the two objectives of travel distance and soil conservation. Its value is preset according to agronomic requirements; for example, setting λ=0.5 means that the focus on soil compaction is equivalent to half the focus on distance.

[0093] Constraints (based on classic TSP constraints):

[0094] Each point is left only once: Each point is visited only once. Eliminating sub-loop constraints: Miller-Tucker-Zemlin (MTZ) constraints or sub-loop elimination constraints are used to ensure the formation of a single Hamiltonian cycle. Soil compaction-related constraints: This ensures that soil compaction penalties are only calculated for points accessed by the path.

[0095] 2. Parameter Determination and Training Instructions:

[0096] It should be clarified that the improved TSP model here is an optimization calculation model. The methods for determining its core parameters λ and CPI(pa) are as follows:

[0097] The penalty weight coefficient λ is determined as follows: This coefficient does not require iterative training but is pre-configured based on agronomic knowledge and field management policies. The configuration method involves providing a pre-configured table, which operators can select according to the soil conservation priority of the current operation. For example, Table 12:

[0098] Table 12 Configuration Table

[0099] Soil protection priority λ Recommended Value illustrate Minimum (efficiency first) 0.1 Almost only path length is optimized Medium (Balanced Mode) 0.5 Balancing path length and soil conservation Highest priority (protection first) 2.0 Significantly inclined to avoid compaction-sensitive areas

[0100] Soil compaction index CPI (pa). Data source: This index is not obtained through training, but is calculated based on a known model of soil physical properties (such as texture, bulk density, and moisture), or directly retrieved from a pre-generated digital soil map of farmland. Calculation / query method: Before path planning, a soil database or real-time sensor data is called to assign a value to pa for each grid or key point. For example, it can be simplified as follows: ,in, The fixed weights are determined based on the soil mechanics model. Model solution: This optimization model is solved using standard solvers (such as Gurobi, CPLEX) or efficient heuristic algorithms (such as the LKH algorithm) to obtain a globally validated path. The solution process is a deterministic mathematical optimization computation and does not involve parameter training based on the dataset.

[0101] 3. Data Fusion and Interpolation: UGV performs measurements along the planned path to obtain ground truth data. Kriging interpolation is used to combine the spatial correlation of discrete point data Gground with aerial imagery data Vaerial to generate a high-precision soil property distribution map S. map (x,y) is the final generated continuous soil property distribution map, which is a continuous function of two-dimensional geographic coordinates (x,y), and each (x,y) location has a soil property value (such as moisture).

[0102] ;

[0103] At the point to be estimated The estimated values ​​of soil properties (such as moisture) at a specific location are used in the Kriging interpolation formula. The estimated value, Here, we have the coordinates of a specific point to be estimated. Although it is two-dimensional, it is represented by a single symbol. express, It is to calculate S map (x, y) at a certain point The method for calculating the value of S is as follows: when this formula is applied to all (x,y) points on the map, the complete S is obtained. map (x,y); The number of measurement points is known. : in the Measured values ​​of soil properties at known measurement points. : Assign the first The weights of each known point. The sum of all weights is 1. This is the core expression of Kriging interpolation. It utilizes known spatial sample points. The unknown location is estimated by weighted summation. The value of the weight. The values ​​are not arbitrarily set, but determined by solving a semi-variogram model to ensure that the estimate is unbiased and has the minimum variance. This formula combines ground discrete point data with aerial imagery trends to generate a continuous, high-precision soil property distribution map. Weight The estimation variance was minimized by solving the semi-variogram model. Finally, the fused sensing result is a multi-layered digital map containing both crop canopy and root zone soil conditions. .

[0104] Table 1 Algorithm Performance Metrics

[0105] index Traditional methods This invention Increase Perception accuracy 70-80% ≥95% +20% Data coverage speed 10 mu / hour 200 mu / hour 20 times Spatial resolution meter level centimeter level 100 times Soil compaction reduced - ≥60%

[0106] Adaptive threshold segmentation: Dynamically adjusts anomaly detection thresholds based on historical data and crop type to improve adaptability. Soil conservation pathway planning: Introduces a soil compaction penalty term into pathway optimization to achieve sustainable agricultural development. Kremlin interpolation: Combines aerial imagery trends with ground-based ground-based measurements to generate high-precision farmland status maps.

[0107] A multi-agent task allocation and path planning algorithm based on dynamic weights:

[0108] In obtaining a fusion perception map Subsequently, the system needs to dynamically allocate tasks to heterogeneous unmanned aerial vehicle (UAV-W, UGV) clusters. This invention proposes a competitive task allocation algorithm based on dynamic weights.

[0109] Problem modeling: The farmland area requiring the operation is rasterized, and each abnormal raster is regarded as a subtask. Let the task set be... Each task Based on its location, type, and urgency Definition. Types include spraying and sampling.

[0110] Let the set of intelligent agents be... It includes UAV-W and UGV. Each intelligent agent It has its current position pos k Capability attributes, current load k and surplus energy E k .

[0111] Dynamic weight calculation: agent For the task Bidding weight Calculated from the following multi-objective utility function: ;

[0112] in, Intelligent agent (Such as a drone) for the mission The bidding weight. The higher the value, the more suitable the agent is to perform this task. Current position of the agent To the mission The cost of location: distance. Capability matching degree: 1 indicates a perfect match (e.g., agricultural drones matching spraying tasks), and 0 indicates a mismatch. Intelligent agent The current load (e.g., the number of tasks already undertaken). Intelligent agent The current remaining energy. Intelligent agent The maximum energy. :Task The level of urgency. Different adjustable weighting coefficients control the contribution ratio of factors such as distance, capability, load, energy, and urgency in the bidding process. This is a multi-objective comprehensive utility function. It quantifies an agent's willingness or suitability to perform a task. The function design encourages agents with close proximity, matching capabilities, light load, and sufficient power to perform urgent tasks. (Exponential decay term...) This makes the effect of distance non-linear, and the advantages of short-range tasks are more obvious.

[0113] Algorithm Flow (Centralized Auction):

[0114] Collaborative Control Center Generate a task list T and broadcast it. Each agent... The above formula is used to calculate the results for all reachable tasks. The agent submits its optimal number of bids to the center. The center then uses either the Hungarian algorithm or a greedy allocation algorithm to allocate tasks, aiming to maximize the total global bid weights. ;

[0115] max: The optimization goal is to maximize the subsequent expression. The total number of intelligent agents. Total number of tasks. Intelligent agent For the task The bidding weight. : Binary allocation variable. If the task Assigned to intelligent agents If the result is positive, then it is 1; otherwise, it is 0. Intelligent agent Maximum task capacity (the maximum number of tasks that can be handled at once). Constraints: Each task can be assigned to at most one agent. The constraint is that each agent cannot be assigned more tasks than its maximum capacity. This is a standard 0-1 integer programming problem. The goal is to optimize the task allocation scheme while satisfying the constraints. This maximizes the sum of the bidding weights of all assigned tasks, thereby achieving optimal overall system performance. The Hungarian algorithm or a greedy algorithm can typically be used to solve this problem.

[0116] Real-time collision avoidance and formation maintenance control algorithms for heterogeneous clusters:

[0117] For UAV-W swarms performing large-area spraying, maintaining formation and real-time obstacle avoidance are crucial during high-density flight. This invention employs a distributed model predictive control (DMPC) framework. The predictive model approximates the dynamics of each UAV i using a discrete linear model: ;

[0118] Drones At any moment The state vector typically includes three-dimensional position and velocity, i.e. . Drones At any moment The control input vector is usually an acceleration command. : State transition matrix, which describes how the system state evolves on its own. The control input matrix describes how the control inputs affect the state changes. This is the state-space equation of a linear discrete-time system. It is used in distributed model predictive control (DMPC) to predict the trajectory of a UAV over a future period. This model is a simplification of complex nonlinear UAV dynamics, facilitating online optimization calculations.

[0119] Distributed optimization problem: In each control cycle, each drone... Solve an optimal control problem in a finite time domain independently, but its cost function includes the cost of its neighbors. Collaborative items:

[0120] ;

[0121] ;

[0122] in, For drones control sequence Optimize to minimize costs. Prediction time domain, i.e., the number of steps to predict forward. Drones At the predicted time The predicted state. Drones At the predicted time The expected reference state, that is, the position it should reach in the formation. The weighted norm square of the state tracking error. This represents a diagonal weight matrix, where the diagonal elements weight the state components such as position error and velocity error, respectively, to penalize the state tracking error (i.e., increase the weight of the state tracking error). The corresponding element will force the drone to track the desired trajectory more accurately. : Control the squared norm of the input. Indicates and Different diagonal weight matrices have diagonal elements that weight each control component (such as acceleration command) to penalize the magnitude of the control quantity (i.e., increase it). The corresponding element in the middle will suppress excessive control actions, achieving energy saving and smooth flight. Drones The neighbor set, other drones within communication range. Distance-based repulsive potential function; Drones At the predicted time The control input vector; Drones At the predicted time The position vector; Drones At the predicted time The position vector; Minimum safe distance between drones; : State transition matrix; : Control input matrix. This function increases when the predicted positions of two drones are too close, thus driving them to maintain distance during optimization. : Weighting coefficients of the exclusion term. This is an optimization problem that each UAV solves locally within the DMPC framework. The cost function consists of three parts: 1) tracking error, keeping up with the formation; 2) control cost, saving energy / smoothing actions; 3) cooperative collision avoidance, maintaining a safe distance from neighbors. By periodically solving this problem, the UAV can autonomously plan a safe flight path that conforms to the formation, taking into account the intentions of its neighbors.

[0123] Drones share their currently predicted state trajectories via a communication network to achieve collaborative optimization. Execution and iteration: Each drone executes only the first control step of its optimized solution. Then, in the next cycle, based on the new state and the latest predicted trajectories of the neighbors, the optimization problem is solved again to achieve distributed real-time collision avoidance and formation maintenance in the rolling time domain.

[0124] Table 2 Agent Configuration

[0125]

[0126] Table 3 Control Parameters

[0127] parameter value illustrate Prediction Time Domain 10 Step length safe distance 5.0 rice Control cycle 100 millisecond Convergence accuracy 0.01 -

[0128] Dynamic weighted bidding mechanism: Intelligent scheduling is achieved by considering multiple dimensions such as distance, capability, load, energy, and urgency. Distributed model predictive control: Multi-machine collaborative collision avoidance and formation maintenance are implemented while ensuring real-time performance. Mixed-integer programming solution: The Hungarian algorithm is used to efficiently solve the task allocation problem with a computational complexity of O(n³).

[0129] Edge-cloud collaborative predictive agronomic decision-making algorithm:

[0130] The vehicle-mounted edge computing center collaborates with the remote cloud control center to achieve a leap from on-demand operations to predictive operations. Cloud-based AI large-scale model (long-term, macroscopic): Deploying a large-scale crop growth and pest / disease prediction model based on the Transformer architecture in the cloud. .

[0131] Input: Historical multispectral sequences, high-precision meteorological data, soil data, and agricultural operation records constitute a multimodal tensor X. Processing: The model captures spatiotemporal dependencies through a self-attention mechanism.

[0132] ;

[0133] in, (Query), (Key), (Value): Three different matrices, each composed of input data. Data such as historical multispectral sequences and meteorological data are obtained through linear transformation. In crop prediction models, these represent different projections of the input features.

[0134] Key vector The dimension is used to scale the dot product and prevent the softmax function from entering regions with minimal gradients. Normalized exponential function: Converts a vector into a probability distribution. This involves calculating the similarity (dot product) between the query and the key to obtain an attention weight matrix. The formula's purpose is to be the core of the self-attention mechanism in the Transformer architecture. It allows the model to dynamically calculate the association weights between different parts when processing sequential data. For example, the model can learn a strong association between the drought situation of a field last month and the risk of pests and diseases this month. Through this mechanism, large AI models can capture the complex spatiotemporal dependencies of farmland conditions, thereby making more accurate macro-level predictions. Sequential data includes data from different times and different fields.

[0135] Lightweight edge inference and decision-making (real-time, local): Model delivery. The cloud will... Lightweight model obtained by distillation And the latest forecast results, risk forecast chart Nutritional requirements chart Data is sent to the vehicle's edge center. Real-time diagnostics and prescription generation: edge-side combined with real-time sensing. Execute the decision function Φ:

[0136] ;

[0137] Where Θ is the set of decision rule parameters. For It shows that it has occurred and For the identified area, generate a therapeutic prescription, such as bactericide A, dose D1. The display is normal, but... High-risk areas are identified, and preventative prescriptions are generated, such as immune inducer B, dose D2. The prescription map is a structured data layer containing the job type, drug ID, dosage, and priority for each raster.

[0138] The large-scale crop growth and pest prediction model deployed in the cloud It is a multi-task spatiotemporal prediction model based on the Transformer architecture. It is trained on a large-scale, multi-source agricultural dataset using supervised learning. The specific training method is as follows:

[0139] 1. Training Data Preparation. Data Sources and Composition: Training Dataset The dataset consists of multimodal spatiotemporal sequence data, primarily including: Remote sensing image sequences: historical multispectral / hyperspectral satellite or UAV imagery, used to extract crop growth indicators such as Normalized Difference Vegetation Index (NDVI) and chlorophyll content. Meteorological and environmental data: historical daily or hourly values ​​of temperature, precipitation, humidity, wind speed, and sunshine duration. Soil property data: static or periodic measurements of soil type, pH value, organic matter content, and nitrogen, phosphorus, and potassium nutrient content. Agricultural operation records: historical records of irrigation, fertilization, and pesticide application, used as input conditions.

[0140] True values ​​should be labeled. For growth forecasts: measurements of crop leaf area index (LAI), biomass, or final yield at the corresponding period. For pest and disease forecasts: pest and disease occurrence labels (such as disease type, occurrence level, and location) confirmed by on-site inspections by agronomic experts.

[0141] Data preprocessing. Spatiotemporal alignment: Unify all data onto the same spatiotemporal grid (e.g., a specific field's grid or the entire field) and interpolate to the same time frequency (e.g., daily). Feature engineering: Construct derived features from the raw data, such as calculating consecutive drought days and accumulated temperature. Sequence construction: Using a sliding window approach, construct a fixed-length historical sequence (e.g., data from the past 60 days) as the model input and the corresponding future target sequence (e.g., the state for the next 7 or 14 days). Dataset partitioning: Divide the data into training, validation, and test sets in chronological order to ensure temporal causality.

[0142] 2. Model Architecture and Training Objectives. Model Architecture: The core architecture employs an Encoder-Decoder Transformer variant. The Encoder processes historical multimodal input sequences. Data from each modality is first projected onto a unified feature space through independent embedding layers, and then spatiotemporal location encoding is added. The encoder captures the complex dependencies between different time points and features in the historical sequence through a multi-layer self-attention mechanism. The Decoder, during training, uses historical encoding and some known future conditions (such as future weather forecasts) as input to autoregressively predict future crop state sequences. During inference, it performs multi-step predictions entirely based on historical encoding and conditional inputs. The training objective (loss function) is multi-task learning, jointly optimizing growth prediction and pest and disease prediction.

[0143] ;

[0144] in: The loss function for growth prediction is usually the mean squared error (MSE) or smoothed L1 loss, which is used to regress and predict continuous growth indicators (such as leaf area index LAI). The loss function for predicting pests and diseases is usually the cross-entropy loss, which is used to classify and predict the probability or level of pest and disease occurrence. and To balance the different hyperparameters of the two tasks and satisfy ; This represents the true leaf area index (LAI) value (i.e., the measured value in the training samples). This represents the leaf area index value predicted by the model; Represents actual pest and disease labels (e.g., discrete categories such as disease type and occurrence level). This represents the probability distribution or category of pest and disease labels predicted by the model.

[0145] 3. Training Process. Optimizer and Hyperparameters: The AdamW optimizer is used for training, with an initial learning rate set to... A cosine annealing learning rate scheduling strategy with hot restart is employed. The batch size is set to 32 or 64 based on GPU memory. Training epochs are typically 100-200, and training is stopped early when the loss on the validation set no longer decreases. Data augmentation: Random time shifts and slight noise are added to the training data to improve model robustness. Regularization: Dropout and weight decay are used to prevent overfitting. Cross-validation: Time-series cross-validation can be used on large datasets to more robustly evaluate model performance. Distillation yields a lightweight model. :exist After training converges, knowledge distillation is used to transfer the knowledge from the large model to a student model with a simpler structure and fewer parameters. In the middle. Using the same training data, let the student model... Simultaneously learn real labels and teacher models The output soft labels, which contain inter-category association information, help the small model achieve better generalization ability. The final model performance is evaluated on an independent test set. Growth prediction uses metrics such as root mean square error (RMSE) and coefficient of determination (R²); pest and disease prediction uses metrics such as precision, recall, F1 score, and area under the AUC-ROC curve. (The training is complete.) Deployed on a cloud server, it periodically (e.g., weekly) uses the latest data to perform rolling forecasts for the entire region, generating risk maps and demand maps. The Medge obtained through the above distillation process is then sent to the vehicle-mounted edge computing center for real-time local diagnosis and prescription generation.

[0146] Table 4 Decision Rule Table

[0147] Real-time status Predicting risks Decision-making actions Prescription type Severe stress / suspected pests and diseases high Therapeutic occupations <![CDATA[Fungicide A, dose D1]]> Healthy / Mild Stress high Preventive work <![CDATA[Immunopotentiator B, dose D2]]> healthy Low / Medium Nutritional management According to fertilization Mild stress Low monitor Increase monitoring frequency

[0148] Table 5 Cloud Model Configuration

[0149] Model Architecture Parameters Forecast cycle Pest and disease prediction Transformer 1.2B 7 days Growth simulation LSTM+Attention 850M 14 days Production forecast CNN+GRU 650M 30 days

[0150] Table 6 Edge Computing Configuration

[0151] hardware Specification Reasoning speed Power consumption GPU Server A100 or other similar specifications 500 FPS 400W Memory 64GB DDR5 - - storage 2TB NVMe 7GB / s -

[0152] Edge-cloud collaborative inference: Large cloud models handle long-term macro-level forecasting, while lightweight edge models handle real-time local decision-making. Predictive agronomic management: Upgrading from on-demand to predictive operations, moving the plant protection checkpoint forward. Formalized decision rules: Transforming agronomic expert experience into actionable decision rule tables, achieving transparent decision-making.

[0153] Blockchain-based evidence storage implementation process:

[0154] To ensure the reliability of agricultural data, every collaborative operation is recorded and documented on the blockchain. Data encapsulation: The collaborative control system generates a record (RR) the moment an instruction is executed. On-chain storage: Record R is sent to the blockchain network of the cloud control center via a secure link. This network adopts a permissioned blockchain architecture and is jointly maintained by nodes such as agricultural regulatory departments, certification bodies, and farmers. The efficient RAFT consensus algorithm is used to quickly reach consensus, appending a new block containing R to the chain. Traceability verification: The QR code on the final agricultural product packaging is linked to the Merkle root hash of all RR records for its production batch. After scanning the code, consumers can query and verify the integrity and immutability of these records on the blockchain, presenting a reliable end-to-end archive from farm to table.

[0155] Table 7 Blockchain Configuration

[0156] parameter value illustrate Blockchain type Permissioned Chain Consortium blockchain architecture consensus algorithm RAFT High-efficiency consensus Block time 5 seconds Average block interval Number of nodes 5-20 Scalable Hash Algorithm SHA-256 Collision resistance

[0157] Table 8 Performance Indicators

[0158] index value illustrate TPS 1000+ transactions / second Delay in evidence preservation <3 seconds Recorded on the blockchain Query delay <1 second Source tracing query Data integrity 100% Unalterable

[0159] Agricultural-Specific Licensing Chain: A consortium blockchain architecture designed for agricultural scenarios, jointly maintained by farmers, regulatory agencies, and certification centers. End-to-End Trusted Traceability: Data is stored and verified throughout the entire chain, from field operations to table consumption, constructing a digital identity for agricultural products. Smart Contract Automation: Business logic such as insurance payouts and subsidy disbursements is automatically executed based on the stored data.

[0160] Table 9 Advantages of the prior art compared to the present invention

[0161] Comparison items Existing technology This invention Work mode Single equipment operating independently Space-ground coordination and multi-machine linkage Operational continuity Relying on human intervention and supplies Fully automated logistics support, supporting long-term operation Decision intelligence Relying on human experience or simple models Predictive agronomic decisions based on AI large models Data credibility Records are easily altered and difficult to trace. Blockchain-based evidence storage ensures full traceability. System scalability Fixed architecture, difficult to scale Modular design, supports plug and play

[0162] like Figure 4 As shown, the system employs a four-in-one collaborative operation architecture comprising a mobile mothership, an aerial swarm, ground units, and a cloud-based central control unit. The mobile platform, serving as the system's mobile base station and logistics center, integrates an intelligent dispensing system, a hybrid energy system, and an edge computing center, including ground units, equipment bays, operational bays, and a cockpit. The aerial drone swarm—monitoring and operations—achieves integrated air-ground operations with the ground robots through a collaborative control system.

[0163] like Figure 5As shown, the air-ground collaborative sensing algorithm integrates large-scale aerial remote sensing data with fine-grained ground-based point measurement data. First, multispectral images are acquired by aerial drones and vegetation indices are calculated to identify abnormal areas. Then, ground robots are dispatched to perform fine-grained verification of the abnormal areas. Finally, the discrete point data and aerial image data are fused using the Kriging interpolation algorithm to generate a high-precision, multi-layered digital map.

[0164] like Figure 6 As shown, the multi-agent dynamic task allocation algorithm assigns tasks to heterogeneous unmanned aerial vehicle (UAV) swarms based on bidding weights. The system considers multiple factors such as agent location, capability, load, energy status, and task urgency, and determines the globally optimal allocation scheme through an optimization algorithm. Simultaneously, a distributed model predictive control (DMPC) framework is employed to achieve real-time collision avoidance and formation maintenance for the UAV swarm.

[0165] like Figure 7 The cloud-based AI large model layer (Transformer architecture) includes historical multispectral data, meteorological and environmental data, soil property database, agronomic knowledge graph, and the specific components of each data set. The edge computing layer (vehicle-mounted edge nodes) includes real-time perception data, lightweight AI models, a real-time decision engine, prescription generation module, and their specific components. The farmland execution layer includes therapeutic prescription areas, preventative prescription areas, detection areas, and nutrient supplementation areas, explaining the function of each area.

[0166] like Figure 8 As shown, the blockchain-based evidence storage system generates an immutable record for each agricultural operation. This record contains key data such as timestamps, agent IDs, task IDs, geographical information, and operational parameters, and its integrity is ensured through the SHA256 hash algorithm. The records are organized into a Merkle tree structure and stored in blocks. Consensus is reached through the RAFT consensus mechanism within a permissioned blockchain network composed of farmers, regulatory agencies, and certification centers, thus constructing a trusted traceability archive for agricultural products. Figure 9 The diagram shown illustrates the principle of Merkle tree verification.

[0167] Table 10 System Composition

[0168]

[0169] Table 11 Technical Parameters

[0170] parameter index unit Work efficiency ≥2000 mu / day Positioning accuracy ≤0.02 m Communication delay ≤50 ms Energy self-sufficiency ≥72 Hour System reliability ≥300 Hours (MTBF)

[0171] Mobile Integrated Platform: A pioneering mobile base station design integrating energy, communication, computing, and logistics, enabling unlimited expansion of operational range. Heterogeneous Cluster Collaboration: Enables heterogeneous collaborative control of aerial drones and ground robots, overcoming the limitations of single-device operation. Cloud-Edge-Device Collaborative Architecture: Constructs a three-tiered intelligent decision-making system: cloud training, edge inference, and terminal execution.

[0172] This invention creates a system-level solution for unmanned agriculture that integrates energy self-sufficiency, mobile mothership, space-ground integration, cloud-edge collaboration, and intelligent decision-making. It enables two operators to complete a closed-loop operation of monitoring, analysis, decision-making, execution, and evaluation across at least 2,000 mu (approximately 133 hectares) of farmland within a single day and single-vehicle work cycle. This increases agricultural production efficiency to more than 10 times that of traditional methods, while simultaneously achieving a reduction of pesticide and fertilizer use by more than 30% and a 50% increase in water resource utilization.

[0173] Advanced Features: The system's integration, intelligence, and operational efficiency reach internationally leading levels. Reliability: Capable of stable operation in harsh agricultural environments with high temperature, high humidity, and high dust levels, with a mean time between failures (MTBF) exceeding 300 hours. Safety: Equipped with comprehensive physical, network, and data security protection mechanisms. Scalability: Adopts a modular and platform-based design, supporting plug-and-play compatibility with future new agricultural robot equipment. Economic Efficiency: While improving efficiency, it controls the total cost of ownership (TCO), demonstrating excellent commercialization prospects.

[0174] Core functions:

[0175] Integrated Space-Ground Perception: Combining the wide-area, high-efficiency advantages of aerial remote sensing with the precision and depth of ground-based detection, a three-dimensional cognitive system for farmland, from soil to canopy, is constructed. Closed-Loop Operation: Monitoring from the air and ground; AI decision-making; precise execution via aerial spraying; effect evaluation via re-monitoring, forming a complete, automated intelligent closed loop. Unmanned Full-Process Operation: From data collection to pesticide and fertilizer spraying, personnel do not need to go to the fields, greatly improving the working environment and ensuring safety. Ultra-Efficient Cluster Collaboration: Four aerial sprayers work together for extremely high efficiency; ground and air equipment can operate in parallel. For example, while spraying from the air, ground robots can simultaneously conduct soil surveys in the next field, maximizing the use of platform resources. Energy and Logistics Self-Sustainability: An independent microgrid and automatic replenishment system support the platform's long-term independent operation in the field. Remote Expert Collaboration: Breaking geographical limitations, it gathers top agronomic wisdom to serve vast farmlands. This invention comprises four main entities: a mobile platform, an unmanned cluster, a command center, and a collaborative control system, as well as space-ground collaborative control and four collaborative scenarios: platform-cluster, space-ground, air-to-air, and edge-cloud.

[0176] Detailed system construction content:

[0177] 1. Mobile high-performance carrier platform - "Mothership":

[0178] Chassis Selection: Utilizing a heavy-duty off-road truck chassis that meets China VI emission standards or is purely electric, possessing excellent off-road capability and load-bearing capacity. Cabin Modifications: Driver's Cab: Upgraded to an intelligent cockpit, equipped with a dual-redundant drive-by-wire system, reserving interfaces for future autonomous truck deployment, and integrating a dual-operator workstation. Work Cabin: Employs a hydraulically lifting drone take-off and landing deck, integrating a drone nesting system, featuring windproof, rainproof, automatic centering, and locking functions. Equipment Cabin: Integrates a dispensing system, energy system, and computing unit in separate zones, featuring temperature control, shock absorption, and corrosion resistance design. Ground Unit Cabin: Equipped with a hydraulically lifting tailgate at the rear, serving as a deployment and retrieval channel for ground robots.

[0179] 2. Unmanned system cluster:

[0180] (1) Aerial Unmanned Aerial Vehicle (UAV) Cluster. 4 Plant Protection Clusters: Utilizing industry-leading payload platforms of 50kg or more, supporting RTK / PPK high-precision navigation, terrain-following flight, two-way communication, and possessing millimeter-wave radar omnidirectional obstacle avoidance capabilities. 2 Monitoring Clusters: Equipped with high-precision visible light, multispectral (5 channels or more), and hyperspectral imagers, possessing rapid stitching and real-time transmission capabilities.

[0181] (2) Ground robot group, 1 unit. Platform: Lightweight composite material, articulated tracked chassis, ground pressure <25Kpa, ultra-low soil compaction. Sensing system: Integrated needle-type soil multi-parameter sensor, including humidity, temperature, EC, pH; robotic arm-assisted sampling system; lidar and visual SLAM navigation system; near-ground chlorophyll fluorescence sensor. Function: To realize in-situ measurement of soil profile parameters, monitoring of plant root microenvironment, and acquisition of insect spore capture images, serving as an important verification and supplement to aerial remote sensing data.

[0182] 3. Intelligent logistics support system:

[0183] (1) Automated Agricultural Input Distribution System. Intelligent Liquid Storage Tank: RFID technology is used to manage the liquid, enabling pesticide identification, inventory monitoring, and safety control. Precision Dispensing Unit: Based on a high-precision metering pump and flow sensor, the system automatically completes the extraction, mixing, and stirring of the mother liquor according to the prescription map sent from the cloud or edge. Unmanned Filling: Through robotic arms and vision guidance, the system achieves automatic alignment, sealed docking, and precise filling of the pesticide tank of the plant protection drone, with zero contact throughout the entire process.

[0184] (2) Intelligent Hybrid Energy System. Multi-source power generation: Integrating diesel generator set as the main power source, rooftop flexible photovoltaic system as the auxiliary power source, and large-capacity lithium battery energy storage system as the peak-shaving and silent operation power source. Intelligent Energy Management System (EMS): Dynamically scheduling energy based on AI algorithms, prioritizing photovoltaic power supply to achieve optimal energy consumption. Automated charging / battery swapping: The drone adopts an automated battery swapping mode with a robotic arm, and the ground robot adopts a wireless charging mode, achieving 24 / 7 uninterrupted operation capability.

[0185] 4. Information control and cloud-edge-device collaborative system:

[0186] (1) Vehicle-mounted edge computing center: Equipped with a high-performance GPU computing server, responsible for: (2) Real-time data processing: Rapidly stitching multispectral images, calculating vegetation index, and inferring AI pest and disease identification models. (3) Real-time cluster control: Running the collaborative control algorithm of the heterogeneous unmanned system between land and space to realize dynamic task allocation and collision avoidance. (4) Local digital twin: Constructing a high-precision three-dimensional model of local farmland for operation simulation and simulation. (5) Local self-organizing network: Using 5G LAN / Wi-Fi 6 / long-distance image transmission technology to build a low-latency, high-reliability data link between vehicles and unmanned equipment. (6) Remote backhaul network: Standard configuration of 4G / 5G, and integration of low-orbit satellite communication terminals, such as Starlink, to ensure remote monitoring and data synchronization capabilities in areas without public network coverage. (7) Global command dashboard: Supporting remote command center to conduct global monitoring, task distribution, and resource scheduling for multiple vehicles. (8) AI agronomic model: Deploying crop growth models, pest and disease prediction models, and yield prediction models based on massive data training. (9) Blockchain traceability system: Generates an immutable record for every agricultural operation, and builds a trustworthy digital archive of agricultural products from the field to the table.

[0187] Key technological highlights and advancements of this invention: **Integrated Space-Ground-Cloud Architecture:** This architecture achieves cross-domain collaboration in perception, decision-making, and execution, representing an industry first in terms of technological complexity and integration. **AI-Based Predictive Agronomic Management:** The system not only applies fertilizer on demand but also predictively applies pesticides through multi-source data fusion analysis, moving the plant protection checkpoint forward. **Level 4 Agricultural Unmanned System Cluster Control:** Overcoming core technologies such as heterogeneous platform collaboration, dynamic obstacle avoidance, and task orchestration, this system achieves truly unmanned cluster operations. **Blockchain-Based Trusted Data Traceability:** A new highlight, this system blockchainizes agricultural production data, significantly enhancing the brand value of agricultural products and consumer trust, providing core support for value-added services. **Platform-Based and Modular Design:** The entire vehicle serves as an open platform with standardized energy, communication, and control interfaces, enabling rapid integration of third-party agricultural robots and strong ecosystem scalability.

[0188] This invention focuses on the following key research and development areas: **Agricultural AI Large-Scale Model Training:** Researching pre-trained large-scale models based on multimodal agricultural data, including remote sensing, meteorological, soil, and agronomic data, to improve model generalization ability and decision-making accuracy. **High-Precision Navigation and Collaborative Control of Unmanned Systems in Unstructured Farmland Environments:** Solving problems such as GPS signal obstruction and visual navigation interference in complex crop canopies. **System-Level Reliability Design and Testing Verification:** Establishing a rigorous V-shaped development and testing process, conducting HIL (Hardware-in-the-Loop) and real-vehicle durability tests to ensure long-term system stability and reliability. **Data Security and Network Security Protection System Construction:** Designing end-to-end encrypted communication protocols to prevent data leakage and malicious network attacks, safeguarding agricultural data sovereignty.

[0189] Social and Economic Benefits Analysis. Social Benefits: First, it represents a strategic technological equipment solution to address the structural shortage of rural labor and ensure national food security. Second, it promotes the transformation of agricultural production methods towards green, low-carbon, and efficient practices. Third, the resulting technical standards and application paradigms can be replicated and promoted throughout the region and the country, leading to the leapfrog development of my country's smart agriculture industry. Economic Benefits: For operation service providers: Annual revenue from single-machine operation services is expected to reach several million yuan, with a controllable investment return cycle and a clear business model. For large-scale farms: It significantly reduces labor and agricultural input costs, increases crop yield and quality, and improves overall economic benefits by more than 20%. For the equipment manufacturing industry: It drives the development of related industrial chains such as high-end agricultural machinery, sensors, robots, and artificial intelligence in China, creating new economic growth points.

[0190] The core functions, technical support, unmanned system equipment, and mobile carrier platform of the mobile intelligent agricultural operation system are as follows:

[0191] The core functions consist of comprehensive perception, intelligent decision-making, precise execution, automated logistics, and full-cycle management. Comprehensive perception includes multispectral remote sensing monitoring (aerial) to monitor crop growth and analyze soil nutrient distribution; in-situ soil profile detection (ground) to detect soil physical properties and monitor soil chemical indicators; and near-ground canopy scanning (ground) to acquire canopy structure information and analyze vegetation health. Intelligent decision-making includes: AI data fusion and prescription map generation, fusing multi-source data to generate precision agriculture prescription maps; predictive early warning of pests and diseases, analyzing pest and disease occurrence patterns and issuing early warning information; and agronomic model analysis and decision support, constructing agronomic models and providing decision support. Precise execution includes cluster-coordinated variable spraying (aerial), with drone clusters working collaboratively to perform variable spraying based on prescription maps; and autonomous mobile reconnaissance (ground), with robots autonomously moving to reconnoiter farmland. Automated logistics includes fully automated pesticide dispensing and filling, precisely dispensing pesticides and automatically filling them; automated battery swapping / charging, with drones / robots automatically swapping and charging batteries; and intelligent hybrid energy scheduling, rationally allocating solar, electrical, and other energy sources to achieve efficient energy utilization. Full-cycle management includes: digital archives of crop growth, recording data throughout the entire crop growth process to provide data support for precision agriculture; blockchain-based agricultural traceability, using blockchain technology to record agricultural operations and achieve traceability of agricultural product quality; and effect evaluation and report generation, assessing the effects of agricultural production and generating evaluation reports.

[0192] The technical support consists of a cloud-edge-device architecture, communication networks, and key technologies. The cloud-edge-device architecture includes: a cloud platform (cloud) for data storage and management, and data analysis and processing; a vehicle-mounted edge computing center (edge) for real-time data processing and intelligent decision support; and drones / robots (devices) for data collection and transmission, executing agricultural production tasks. The communication network includes: 5G / 4G networks for high-speed data transmission and remote control; low-Earth orbit satellite communication (such as Starlink) for wide coverage and high communication stability; and local self-organizing networks (Wi, Fi6 / image transmission) for interconnectivity between devices and ensuring reliable data transmission. Key technologies include heterogeneous unmanned system cluster control for collaborative control of drones, robots, and other heterogeneous unmanned systems, improving operational efficiency; agricultural AI big data models for analyzing agricultural data using artificial intelligence to provide precision agriculture decision support; multi-source data fusion for integrating data from satellite remote sensing, drones, robots, and other sources to improve data accuracy and reliability; and network security and data encryption to ensure the network security of the agricultural system and protect data privacy and security.

[0193] The unmanned system equipment consists of an air unit and a ground unit. The air unit includes: a swarm of 4 agricultural drones with a large payload capacity, capable of carrying a large amount of pesticides; RTK positioning to improve operational accuracy; omnidirectional obstacle avoidance to ensure flight safety; 2 detection drones equipped with a multispectral camera to acquire multispectral information of crops and a visible light camera to capture visible light images of crops; and 1 all-terrain reconnaissance robot with a tracked design to adapt to various terrains, soil sensors, and a robotic arm.

[0194] The mobile carrier platform is a mothership-style truck, including: a highly mobile heavy-duty chassis with excellent off-road performance, capable of driving in complex terrain; an electric lifting and lowering platform for rapid take-off and landing to improve work efficiency; an intelligent hybrid microgrid that allocates energy sources such as solar and electrical power to ensure stable equipment operation; and a modular equipment compartment that allows for module replacement according to different operational needs, improving the equipment's versatility and flexibility.

[0195] In this invention, the intelligent application scenario layer includes: intelligent field inspection, precise variable application of pesticides / fertilizers, intelligent sowing and animal feeding, large-scale multispectral remote sensing and mapping of farmland, pest and disease monitoring, post-disaster assessment, precise loss assessment for agricultural insurance, and crop growth monitoring and yield prediction.

[0196] The core capability layer includes: digital agriculture brain, autonomous flight control, intelligent scheduling of drone swarms, crop growth and disaster early warning algorithm models, farmland digital twins, visualization management system, data analysis and display, and intelligent planning of agricultural operations.

[0197] The cloud-integrated computing power support layer includes: ground mobile platform, cloud infrastructure, 3D modeling, cloud micro-applications, and data storage / analysis / retrieval.

[0198] The ubiquitous reliable connectivity layer includes: 5G / 6G low-altitude continuous coverage network, air-ground integrated communication collaboration, BeiDou satellite navigation and ground augmentation network, and end-to-end secure transmission guarantee system.

[0199] The edge-to-all intelligent sensing layer includes: agricultural drones (plant protection, seeding), monitoring drones (visible light, infrared light, multispectral), agricultural robots, farmland IoT terminals, vehicle / airborne sensors (multispectral / hyperspectral cameras, lidar, visible light sensors), and Beidou high-precision positioning terminals.

[0200] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 3As shown, the computer device includes a processor, memory, and a network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores static and dynamic information data. The network interface communicates with external terminals via a network connection. When the computer program is executed by the processor, it implements the steps in the above method embodiments.

[0201] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the computer device to which the present invention is applied. A specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0202] In addition, the present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0203] In addition, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0204] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0205] This invention is not limited to the structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this invention is limited only by the appended claims.

Claims

1. A mobile intelligent agricultural unmanned swarm operation method, characterized in that, include: Using aerial imagery data, a set of key verification points is generated. Combined with a path planning algorithm, a global verification path is generated that minimizes the sum of the total travel distance and the soil compaction penalty. Based on the global verification path, ground truth data is obtained to generate a soil attribute distribution map, which is then fused with aerial imagery data to obtain a farmland fusion perception map. The farmland areas in the integrated perception map are rasterized, and each abnormal grid is treated as a task. Based on the suitability of the agent to perform the corresponding task, the bidding weight of the agent for the task is assigned. The task allocation of the agent is carried out with the goal of maximizing the sum of the global bidding weights. When the intelligent agent performs a task, it uses a discrete linear model to predict the trajectory of the aerial operation drone. Each aerial operation drone solves the optimization problem locally in the distributed model predictive control framework to maintain formation and avoid obstacles in real time during flight. The cost function includes tracking error, control cost and cooperative collision avoidance.

2. The mobile intelligent agricultural unmanned cluster operation method according to claim 1, characterized in that, The process of generating a set of key verification points using aerial imagery data, and then combining this with a path planning algorithm to generate a global verification path that minimizes the sum of the total travel distance and the soil compaction penalty, includes: Aerial imagery data is collected using aerial monitoring drones. Adaptive threshold segmentation is applied to the aerial imagery data to identify abnormal regions and obtain a binary mask. Use the centroids of connected regions in the binary mask as the set of key verification points; To minimize the sum of the total driving distance and the soil compaction penalty, and combining the set of key verification points and the penalty weight coefficient, an objective function for solving the path is constructed. Based on the Traveling Salesman Problem model and the objective function of the solution path, a global verification path is planned.

3. The mobile intelligent agricultural unmanned cluster operation method according to claim 1, characterized in that, The process of acquiring ground truth data based on a global verification path to generate a soil attribute distribution map, and fusing it with aerial imagery data to obtain a farmland fusion perception map includes: The ground operation robot performs measurements according to the global verification path to obtain ground truth data; By combining ground truth data with aerial imagery data through spatial correlation analysis, a soil property distribution map is obtained. By fusing soil property distribution maps and aerial imagery data, a farmland integrated perception map is obtained.

4. The mobile intelligent agricultural unmanned cluster operation method according to claim 1, characterized in that, The allocation of bidding weights for tasks by agents based on their suitability for performing the corresponding tasks includes: The bidding factors are the cost distance from the agent's current position to the task position, the agent's ability to match the task, the agent's current load, the agent's current remaining energy, the agent's maximum energy, and the urgency of the task. Weight coefficients are assigned to each bidding factor. Based on bidding factors and weighting coefficients, a multi-objective utility function is constructed to calculate the agent's bidding weight for the task; Intelligent agents include aerial drones and ground robots.

5. A mobile intelligent agricultural unmanned cluster operation method according to claim 1, characterized in that, The allocation of agent tasks with the goal of maximizing the total sum of global bidding weights includes: The agent outputs the optimal n bids, and uses an optimization algorithm to allocate tasks to the agent with the objective of maximizing the sum of global bid weights. The optimization objective and constraints are as follows: ； In the formula, Indicates the total number of intelligent agents; Indicates the total number of tasks; Represents intelligent agents For the task Bidding weight; Represents a binary allocation variable; Represents intelligent agents Maximum task capacity; This means that each task can be assigned to at most one agent; This indicates that the number of tasks assigned to each agent does not exceed its maximum capacity.

6. A mobile intelligent agricultural unmanned cluster operation method according to claim 1, characterized in that, The discrete linear model includes: ； In the formula, Indicates drone At any moment The state vector; Indicates drone At any moment The control input vector; Represents the state transition matrix; This represents the control input matrix.

7. A mobile intelligent agricultural unmanned cluster operation method according to claim 1, characterized in that, The optimization problems that each aerial operation UAV solves locally in the distributed model predictive control framework include: In each control cycle, each aerial operation UAV independently solves the optimal control problem in the finite time domain within a distributed model predictive control framework, and the cost function is: ； ； in: Indicates the importance of drones control sequence Optimize; Indicates the prediction time domain; Indicates drone At the predicted time The predicted state; Indicates drone At the predicted time Expected reference state; The weighted norm squared represents the state tracking error. This represents the squared weighted norm of the control input. and Representing different diagonal weight matrices; Indicates drone The set of neighbors; This represents the weighting coefficient of the exclusion term; This represents the distance-based repulsive potential function; Indicates drone At the predicted time The control input vector; Indicates drone At the predicted time The position vector; Indicates drone At the predicted time The position vector; Indicates the minimum safe distance between drones; Represents the state transition matrix; Represents the control input matrix; The first control step of the aerial operation UAV executes the optimization solution. In the next cycle, based on the new state and the latest predicted trajectory of the neighbors, the optimization problem is solved again to achieve distributed real-time collision avoidance and formation maintenance in the rolling time domain.

8. A mobile intelligent agricultural unmanned cluster operation method according to claim 1, characterized in that, Also includes: Predictive agronomic decisions are made through edge-cloud collaboration; specifically including: Long-term macro-prediction of agronomic decisions through AI models deployed in the cloud; The AI ​​model is distilled in the cloud to obtain a lightweight model; The lightweight model and the latest prediction results are sent to the vehicle edge center for real-time local prediction of agronomic decisions.

9. A mobile intelligent agricultural unmanned cluster operation method according to claim 1, characterized in that, Also includes: A permissioned blockchain-based evidence storage mechanism solidifies each automated agricultural operation into a digital certificate.

10. A mobile intelligent agricultural unmanned swarm operation system, characterized in that, include: The map fusion building module is used to generate a set of key verification points using aerial imagery data, and combined with the path planning algorithm, to generate a global verification path that minimizes the sum of the total driving distance and the soil compaction penalty. Based on the global verification path, ground truth data is obtained to generate a soil attribute distribution map, which is then fused with aerial imagery data to obtain a farmland fusion perception map. The agent task allocation module is used to rasterize the farmland areas in the farmland fusion perception map and treat each abnormal grid as a task. Based on the suitability of the agent to perform the corresponding task, the module allocates the agent's bidding weight for the task. The goal is to maximize the total sum of global bidding weights in the agent task allocation. The UAV trajectory solving module is used to predict the motion trajectory of aerial operation UAVs using a discrete linear model when the intelligent agent performs a task. Each aerial operation UAV solves the optimization problem locally in a distributed model predictive control framework to maintain formation and avoid obstacles in real time during flight. The cost function includes tracking error, control cost, and cooperative collision avoidance.

Images