Unmanned aerial vehicle-based microclimate precision regulation system for understory cultivation area

Abstract

This invention belongs to the technical field of agricultural environmental control and UAV applications, and relates to a UAV-based microclimate precision control system for forest understory cultivation areas. The system includes: an environmental modeling and initialization module, a prediction and decision-making module, a command execution and observation module, a data verification and extraction module, an error calculation and correction source term generation module, a correction source term propagation and field update module, and a closed-loop iterative scheduling module. By constructing a three-dimensional voxel grid and generating an initial microclimate prediction field, identifying target voxels based on prediction trends and generating predictive control frames, the UAV executes intervention actions and collects observation data. After verification and error calculation, correction source terms are generated. The microclimate prediction field is updated based on predicted airflow propagation, achieving closed-loop iterative control. This invention solves the problems of lagging control and insufficient intervention targeting caused by the lack of dynamic prediction of three-dimensional microclimate fields and digital twin linkage capabilities in existing fixed sensor networks and traditional UAV operation methods.

Description

Technical Field

[0001] This invention belongs to the technical field of agricultural environmental control and drone applications, and relates to a drone-based microclimate precision control system for forest understory cultivation areas. Background Technology

[0002] In the cultivation and management of understory economic crops, maintaining suitable microclimate conditions plays a crucial role in the healthy growth and quality improvement of crops. Currently, relevant control methods mainly rely on the monitoring and intervention of key environmental parameters such as air humidity, temperature, and airflow.

[0003] Existing technological solutions commonly found in the industry mainly fall into two categories: one is fixed sensor networks deployed in forests, which trigger irrigation or spraying devices based on parameter thresholds at monitoring points, or rely on human experience for intervention; the other is using drones equipped with sensors for large-scale patrols or targeted spraying operations, improving operational flexibility and coverage. However, existing drone solutions are mostly based on preset flight paths and simple threshold triggers, and their control logic remains a monitoring-response mode, failing to deeply integrate with dynamic prediction models of the microclimate field. Operational decisions lack physical guidance and advanced judgment of three-dimensional spatial climate evolution trends.

[0004] Due to the shading and guiding effects of forest canopy structure on light and wind, the microclimate in forests exhibits significant spatial heterogeneity and temporal dynamics. Fixed sensor networks, limited by their discrete deployment, struggle to fully characterize the distribution features of the three-dimensional spatial field. Existing drone operations also lack a digital twin model linked to the physical field; their response mechanisms are based on logic triggered by detecting anomalies, which can lead to temporal lags in regulation, spatial blind spots, or a lack of targeted intervention. These methods may result in regulatory actions lagging behind actual crop needs, resource waste from whole-area spraying, insufficient localized regulation, and operational paths that cannot be dynamically adjusted based on real-time microclimate field predictions. Summary of the Invention

[0005] To address the aforementioned problems, this invention provides a precise microclimate control system for forest understory cultivation areas based on unmanned aerial vehicles (UAVs).

[0006] A drone-based microclimate control system for understory cultivation areas includes:

[0007] The environmental modeling and initialization module is configured to construct a three-dimensional voxel mesh covering the understory cultivation area and generate an initial microclimate prediction field based on environmental meteorological data input from outside the cultivation area. The environmental meteorological data input from outside the cultivation area includes at least ambient temperature and relative humidity.

[0008] The prediction and decision-making module is configured to infer the evolution trend of microclimate state based on the initial microclimate prediction field, identify target voxels that meet preset trigger conditions, and generate predictive control frames containing execution summary hash seeds.

[0009] The command execution and observation module is configured to allow the UAV to execute local fixed intervention actions based on predictive control frames, and to generate state observation packages based on measured microclimate data and execution summary hash seeds;

[0010] The data verification and extraction module is configured to use the execution digest hash seed to compare hash values, verify the data integrity in the state observation package, and extract the measured data of the causal tag.

[0011] The error calculation and correction source term generation module is configured to calculate the deviation between the measured data of the causal label and the predicted state of the corresponding spatiotemporal point, and generate correction source terms.

[0012] The correction source term propagation and field update module is configured to inject the correction source term into the corresponding voxel of the three-dimensional voxel grid and control its propagation path according to the predicted airflow vector to generate an updated microclimate prediction field.

[0013] The closed-loop iterative scheduling module is configured to take the updated microclimate prediction field as input and return it to the prediction and decision module, and then execute it cyclically for the next round of closed-loop regulation.

[0014] A further aspect of the present invention includes an environment modeling and initialization module, used to perform the following operations:

[0015] A 3D environment model was constructed based on forest understory 3D point cloud data;

[0016] The three-dimensional environment model is discretized into a three-dimensional voxel mesh, and each voxel is labeled with fluid or solid properties.

[0017] The calculation boundary conditions are set based on environmental meteorological data input from outside the cultivation area;

[0018] Iterative calculations are performed under the computational boundary conditions until the rate of change of field variables within the grid is lower than the preset stability threshold, thus generating the initial microclimate prediction field.

[0019] A further aspect of the present invention includes a prediction and decision-making module, used to perform the following steps:

[0020] Within a preset time window, predict the future state of each three-dimensional voxel in the three-dimensional voxel grid;

[0021] By traversing the simulation results, the target voxel whose predicted state value will exceed the preset microclimate threshold is located;

[0022] Extract the spatiotemporal coordinates of the target voxel, as well as the predicted airflow vector that causes its state change;

[0023] The flight vector of the UAV is calculated based on the spatiotemporal coordinates of the target voxel and the predicted airflow vector.

[0024] Select the corresponding behavior primitive activation ID from the preset behavior primitive library;

[0025] Call the random number generator to generate the execution digest hash seed;

[0026] The spatiotemporal coordinates, flight vector, behavioral primitive activation ID, and execution digest hash seed of the target voxel are encapsulated to generate a predictive control frame.

[0027] A further aspect of the present invention includes an instruction execution and observation module, used to perform the following operations:

[0028] The UAV analyzes predictive control frames to extract flight vectors and behavioral primitive activation IDs;

[0029] Maneuvering based on flight vectors, flying toward the target's spatial location defined by its spacetime coordinates;

[0030] At the target time, execute the corresponding localized, fixed intervention action based on the behavioral primitive activation ID;

[0031] After the intervention is completed, an instantaneous observation is conducted using an airborne multimodal sensor to obtain measured microclimate data;

[0032] Read the execution digest hash seed from the predictive control frame;

[0033] The actual intervention action parameters are concatenated into a parameter string, and the execution digest hash seed is used as the key to perform hash operation to generate the execution digest hash value;

[0034] The receipt frame ID, actual observation coordinates and time, measured microclimate data, and execution summary hash value are encapsulated to generate a state observation package.

[0035] A further aspect of the present invention includes a data verification and extraction module, which performs the following operations:

[0036] Extract the receipt frame ID from the status observation packet, and retrieve the original instruction parameters and the original issued execution digest hash seed corresponding to the ID;

[0037] Using the originally issued execution digest hash seed, perform local hashing on the original instruction parameters to generate locally computed hash values;

[0038] The locally computed hash value is compared with the execution digest hash value in the state observation package; if they match, the verification is considered successful.

[0039] Measured microclimate data are extracted from the validated state observation package as measured data for causal labeling.

[0040] A further embodiment of the present invention includes an error calculation and correction source term generation module, which performs the following operations:

[0041] Extract the measured state vector from the causally labeled measured data;

[0042] Query the predicted state vector of the corresponding spatiotemporal point from the spatiotemporal prediction field engine;

[0043] Calculate the state residual vector between the measured state vector and the predicted state vector;

[0044] The strength of the correction source term is determined by multiplying the magnitude of the state residual vector by a preset gain coefficient.

[0045] The positive or negative nature of the correction source term is determined based on the direction of the state residual vector;

[0046] The correction source term is added as an instantaneous source term to the transport equation of the corresponding three-dimensional voxel.

[0047] A further aspect of the present invention includes a source term propagation and field update module, which performs the following operations:

[0048] Obtain the predicted airflow vector of the injection point voxel and its neighborhood, and calculate the local mainstream direction;

[0049] Construct an anisotropic tensor and set the diffusion coefficients along the mainstream direction and the perpendicular direction;

[0050] The propagation of the correction source term is controlled in the heat and moisture transport equation by replacing the scalar diffusion coefficient with an anisotropic tensor.

[0051] An updated microclimate prediction field is generated through iterative calculations.

[0052] A further aspect of the present invention includes a closed-loop iterative scheduling module, used to perform the following steps:

[0053] The updated microclimate prediction field is fed back as a new input to the prediction and decision-making module;

[0054] The spatiotemporal prediction field engine is driven to perform the next round of trend extrapolation based on the updated microclimate prediction field;

[0055] By iteratively executing the prediction and decision-making module to the correction source term propagation and field update module, an adaptive closed-loop control process is formed, including prediction, intervention, observation, verification, and correction.

[0056] A further aspect of the present invention involves constructing a three-dimensional voxel mesh, comprising the following steps:

[0057] Obtain the bounding box of the 3D environment model in 3D space, and discretize the bounding box into multiple cubic units according to the preset resolution;

[0058] Traverse each cube unit and use ray casting to determine whether its center point is located inside the 3D environment model.

[0059] If the center point is located inside the model, the cube element is marked as a solid property; if it is located outside the model, it is marked as a fluid property.

[0060] All cubic cells labeled with fluid properties are combined to form a three-dimensional voxel mesh for carrying out microclimate calculations, and fluid boundary conditions are established in the physics engine.

[0061] In a further embodiment of the present invention, the behavior primitive activation ID corresponds to a pre-set standardized operating procedure in the local memory of the UAV. The standardized operating procedure includes specific modes of atomization spraying, local blowing or supplementary lighting operations, and the predictive control frame is broadcast through the downlink of the wireless protocol.

[0062] In summary, the present invention has the following beneficial technical effects:

[0063] 1. By constructing a three-dimensional voxel grid of the covered cultivation area and initializing the microclimate prediction field based on physical equations, the complex three-dimensional spatial structure of the forest canopy is transformed into a computable digital model, providing a computational basis for predicting the spatiotemporal evolution of microclimate based on heat and water transport equations, thereby supporting the early identification of potential anomaly areas.

[0064] 2. By extrapolating the evolution trend of microclimate conditions and locating target voxels that are about to exceed the threshold, predictive control frames containing target coordinates and intervention commands can be generated based on the predicted airflow vector. This enables UAVs to arrive at the predicted location and perform intervention before anomalies occur, realizing a shift from a passive response to an active prediction control method.

[0065] 3. By introducing a hash seed into the execution instructions and performing integrity verification in the returned data, a reliable data traceability and verification mechanism is established, which helps to reduce the impact of invalid data caused by communication errors or execution deviations on the model, thereby improving the credibility of the data source during the closed-loop calibration process.

[0066] 4. By converting the deviation between observed data and model predictions into correction source terms and controlling the anisotropic propagation of the predicted airflow field in the three-dimensional voxel grid, the digital model can be directionally corrected based on actual observations, causing the predicted field to gradually approach the real physical state, realizing continuous adaptive calibration of the model, and thus supporting more accurate long-term microclimate predictions. Attached Figure Description

[0067] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. The drawings are used to provide a further understanding of the present invention.

[0068] Figure 1 This discloses a schematic diagram of the framework in the embodiments of this application.

[0069] Figure 2 This discloses a flowchart of an embodiment of this application. Detailed Implementation

[0070] The following is in conjunction with the appendix Figure 1 - Figure 2 A preferred description of the present invention is provided below.

[0071] See attached document Figure 1 - Figure 2 This invention proposes a microclimate precision control system for forest understory cultivation areas based on unmanned aerial vehicles (UAVs), comprising the following modules:

[0072] The environmental modeling and initialization module is configured to construct a three-dimensional voxel mesh covering the understory cultivation area and generate an initial microclimate prediction field based on environmental meteorological data input from outside the cultivation area. The environmental meteorological data input from outside the cultivation area includes at least ambient temperature and relative humidity.

[0073] The prediction and decision-making module is configured to infer the evolution trend of microclimate state based on the initial microclimate prediction field, identify target voxels that meet preset trigger conditions, and generate predictive control frames containing execution summary hash seeds.

[0074] The command execution and observation module is configured to allow the UAV to execute local fixed intervention actions based on predictive control frames, and to generate state observation packages based on measured microclimate data and execution summary hash seeds;

[0075] The data verification and extraction module is configured to use the execution digest hash seed to compare hash values, verify the data integrity in the state observation package, and extract the measured data of the causal tag.

[0076] The error calculation and correction source term generation module is configured to calculate the deviation between the measured data of the causal label and the predicted state of the corresponding spatiotemporal point, and generate correction source terms.

[0077] The correction source term propagation and field update module is configured to inject the correction source term into the corresponding voxel of the three-dimensional voxel grid and control its propagation path according to the predicted airflow vector to generate an updated microclimate prediction field.

[0078] The closed-loop iterative scheduling module is configured to take the updated microclimate prediction field as input and return it to the prediction and decision module, and then execute it cyclically for the next round of closed-loop regulation.

[0079] In one embodiment of the present invention, the environment modeling and initialization module is used to perform the following steps:

[0080] A three-dimensional environmental model is constructed based on forest understory three-dimensional point cloud data; the three-dimensional environmental model is discretized into a three-dimensional voxel grid, and each voxel is labeled with fluid or solid properties; computational boundary conditions are set based on environmental meteorological data input from outside the cultivation area; iterative calculations are performed under the computational boundary conditions until the rate of change of field variables within the grid is lower than the preset stability threshold, thereby generating an initial microclimate prediction field.

[0081] Specifically, firstly, a drone equipped with a lidar sensor is deployed. The lidar sensor operates in the near-infrared band, for example, 905 nm. The drone flies along a pre-planned coverage route within a certain altitude range above the ground surface of the understory cultivation area, collecting high-density three-dimensional point cloud data of the ground surface, tree trunks, and tree canopies. For example, when flying at an altitude of 30 to 50 meters, it collects three-dimensional point cloud data at a density of no less than 500 points per square meter. This data, along with the high-precision centimeter-level position and attitude information obtained by its onboard real-time dynamic differential positioning module, is transmitted in real-time to a ground server via a 5G communication link.

[0082] After receiving the raw point cloud data stream, the point cloud processing module on the ground server first executes a statistical outlier removal algorithm to filter out noise points generated during flight. Then, it uses a random sample consensus algorithm to perform planar detection on the point cloud, thereby separating the ground point cloud. Next, Euclidean clustering is applied to the non-ground point cloud, clustering the discrete point cloud into clusters representing individual trees. Finally, an algorithm based on Poisson surface reconstruction is called to mesh the classified ground point cloud and tree cluster point cloud, generating a high-precision 3D environment model with continuous surfaces in PLY file format. It should be noted that the high-precision 3D environment model is a digital model describing the surface geometry of all physical entities within the understory cultivation area using vertices, edges, and faces as basic elements, with a spatial geometric accuracy better than 0.1 m.

[0083] The spatial partitioning module reads the high-precision 3D environment model, obtains its minimum bounding box in 3D space, and discretizes the space defined by the bounding box into a series of cubic units, i.e., 3D voxels, according to a preset resolution of 0.5 m. It should be noted that a 3D voxel is a cubic spatial unit with a unique index (i, j, k) in a 3D Cartesian coordinate system, and its typical side length is set to 0.3 m to 1.0 m (e.g., 0.5 m). This setting is based on an engineering trade-off of obtaining an effective characterization of the turbulent effects under the canopy within an acceptable computation time: too high a resolution (e.g., <0.3 m) will lead to a surge in the number of voxels, exceeding real-time computing capabilities; too low a resolution (e.g., >1.0 m) will fail to capture the details of the turbulent structure under the canopy. Based on the typical forest understory spatial scale (canopy gap of about 1–3 m) and computational fluid dynamics simulation experience, the preferred resolution is 0.3 m to 1.0 m, and more preferably 0.5 m. This value can be calibrated by analyzing the errors between simulation results and measured data at different resolutions. For example, at a resolution of 0.5 m, the root mean square error between the simulated wind speed and the measured value can be controlled within 0.2 m / s. Each 3D voxel is traversed, and its center point is determined using ray casting to determine if it is located inside the high-precision 3D environment model. If it is inside the model, the 3D voxel is marked as a solid property; otherwise, it is marked as a fluid property. The set of all 3D voxels marked as fluid properties constitutes the 3D voxel mesh used for subsequent calculations.

[0084] Finally, the spatiotemporal prediction field engine is initialized. This engine is typically a computational fluid dynamics software developed in C++ and based on the finite volume method for solving physical equations, running on a server configured with at least 64 GB of memory and more than 24 physical cores. During initialization, the engine initiates a network request to an external public meteorological service through an application programming interface (API) to obtain environmental meteorological data from outside the cultivation area, representing its geographical location. It should be understood that this external environmental meteorological data is a set of parameters describing regional-scale weather conditions, specifically including real-time wind speed, wind direction, ambient temperature, relative humidity, and total solar irradiance. Typical values ​​include, for example, temperature from 5.0 ℃ to 40.0 ℃, relative humidity from 20% to 95%, wind speed from 0.1 m / s to 10.0 m / s, and total solar irradiance from 100 W / m² to 1200 W / m². The engine parses the external environmental meteorological data and sets it as the computational boundary conditions for a three-dimensional voxel mesh. For example, the total solar irradiance is applied to the top boundary of the grid, and wind speed and direction are converted into velocity vectors and applied to the inflow side boundary of the grid.

[0085] After setup, the engine performs a brief startup calculation. Without considering internal heat and humidity sources, it iterative calculations are performed based solely on boundary conditions until the rate of change of field variables at each point within the grid falls below a stability threshold. This stability threshold is set according to the convergence criteria of numerical calculations, typically defined as the rate of change of the state variable (e.g., temperature, humidity) of the same voxel between adjacent time steps being less than 0.1% / s, or the absolute change being less than 0.01℃ or 0.1% per second. This threshold can be determined through grid independence verification and convergence testing to ensure that the startup calculation results are not affected by initial transients. After this process, each three-dimensional voxel of the fluid property obtains an initial state vector, and the initial state vectors of all voxels together constitute the initial microclimate prediction field.

[0086] It should be noted that the initial microclimate prediction field is a multi-dimensional floating-point array. The dimensions of the array correspond to the dimensions of the three-dimensional voxel grid. Each array element stores the initial state vector of the corresponding voxel. This vector includes at least five physical quantities: temperature, humidity, and three-dimensional velocity components (u, v, w). The prediction field is stored on the server hard drive in binary file format, waiting to be called in subsequent steps.

[0087] Among them, the three-dimensional voxel mesh is a logical set of all three-dimensional voxels that are determined to be fluid properties, which constitutes the discretized spatial domain for microclimate simulation calculation.

[0088] For example, in this embodiment, a drone is first dispatched to collect point cloud data of a 100 m × 100 m × 20 m area. After the aforementioned processing flow, a high-precision 3D environment model including 800,000 triangular facets is generated. Next, the space is divided with a voxel side length of 0.5 m, generating a mesh of 200 × 200 × 40, totaling 1.6 million 3D voxels. By spatially determining the location against the high-precision 3D environment model, 700,000 voxels are marked as solids, and the remaining 900,000 voxels constitute the 3D voxel mesh. Subsequently, the spatiotemporal prediction field engine is activated, and a set of environmental meteorological data input from outside the cultivation area is obtained via API. The specific values ​​are: {Ambient temperature: 28.2℃, relative humidity: 75%, wind speed: 1.5 m / s, wind direction: southwest (225°), total solar irradiance: 850 W / m²}. The engine applies 850 W / m² to the upper boundary of the grid at z=20 m, and applies the inlet velocity vector (u=-1.06 m / s, v=-1.06 m / s, w=0 m / s) calculated from a wind speed of 1.5 m / s and a wind direction of 225° to the south and west boundaries of the grid. After a startup calculation lasting 300 time steps, an initial microclimate prediction field is generated. For example, for a voxel located in an open area with coordinate indices (30, 45, 12), the initial state vector obtained is {temperature: 28.8 ℃, humidity: 72%, velocity components: (-0.95, -0.98, -0.1) m / s}. For the voxel with coordinate index (150, 160, 5) located under the dense tree canopy, its initial state vector is {Temperature: 28.0 ℃, Humidity: 78%, Velocity component: (-0.15, -0.2, -0.05) m / s}. This dataset, including 900,000 state vectors, is written to the hard disk as the output of the environment modeling and initialization module.

[0089] In one embodiment of the present invention, the prediction and decision-making module is configured to perform the following steps:

[0090] Within a preset time window, the future state of each 3D voxel in the 3D voxel grid is extrapolated; the extrapolation results are traversed to locate the first target voxel whose predicted state value will exceed the preset microclimate threshold; the spatiotemporal coordinates of the target voxel and the predicted airflow vector that causes its state change are extracted; the flight vector of the UAV is calculated based on the spatiotemporal coordinates of the target voxel and the predicted airflow vector; the corresponding behavior primitive activation ID is selected from the preset behavior primitive library; a random number generator is called to generate an execution digest hash seed; the spatiotemporal coordinates, flight vector, behavior primitive activation ID, and execution digest hash seed of the target voxel are encapsulated to generate a predictive control frame.

[0091] Specifically, firstly, the spatiotemporal prediction field engine loads the initial microclimate prediction field from the server's hard drive and sets an appropriate prediction duration as the preset time window and calculation time step, such as a preset time window of 300 s and a calculation time step of 1 s. The preset time window is set based on the balance between system response time and prediction effectiveness: a time window that is too short, such as <120 s, may lead to insufficient intervention preparation time; a time window that is too long, such as >600 s, will reduce prediction accuracy due to increased meteorological uncertainty. The preferred range is 180 s to 480 s, and more preferably 300 s. This value can be optimized based on historical meteorological data statistics and prediction model error analysis. For example, by analyzing the meteorological change amplitude every 5 minutes over the past 72 hours, it is determined that the predictability of wind speed and temperature changes within 300 s remains above 85%. The preset time window is a parameter representing the duration of the prediction period, and its typical value range is 120 s to 600 s. Subsequently, the spatiotemporal prediction field engine drives its built-in solver to iteratively calculate the state of each three-dimensional voxel in the three-dimensional voxel grid over multiple time steps, such as 300 time steps, based on the initial values ​​of the entire field provided by the initial microclimate prediction field and by performing explicit Euler integrals over time on the heat and moisture transport equations.

[0092] This process generates a microclimate state evolution trend, which is a four-dimensional array with dimensions (i, j, k, t), recording the state vector of each three-dimensional voxel at each time step within a preset time window. Next, a parallel processing task is initiated to traverse this four-dimensional dataset. For each three-dimensional voxel's predicted state vector at each future time step, its humidity component is extracted and compared with a preset microclimate threshold, such as 60%.

[0093] It should be understood that the preset microclimate threshold is a critical value used to determine whether the microclimate state is abnormal. It is set according to the physiological characteristics of the target crop. Specifically, for different understory economic crops (such as ginseng, Panax notoginseng, edible fungi, etc.), the suitable humidity range for their growth is determined by consulting agricultural meteorology manuals, crop physiology literature, or conducting field experiments. For example, for ginseng, which prefers shade and moisture, the suitable relative humidity is 70%-85%, so the lower limit of the threshold can be set to 70%, and the upper limit to 85%. The system can have a built-in crop-threshold comparison table, and users can select the corresponding threshold according to the actual cultivated crop. Alternatively, the threshold can be dynamically optimized based on the relationship between historical yield and microclimate data through machine learning models, such as using support vector machines or decision tree models to back-deduce the optimal humidity threshold range with the goal of maximizing yield. When the predicted humidity value of a three-dimensional voxel is detected to be lower than the threshold for the first time, the traversal is immediately stopped, and the three-dimensional voxel is identified as the target voxel. At the same time, the three-dimensional spatial index (i, j, k) at which the breakthrough occurred and the corresponding time step t are recorded. Subsequently, the complete predicted state vector of the target voxel at time step t is retrieved and extracted from the four-dimensional dataset, and the three-dimensional velocity components (u, v, w) included therein are identified as the predicted airflow vector. These two together constitute the control target, namely, a data structure that includes the spatiotemporal information of the target voxel and the dynamic information that causes its state changes.

[0094] Finally, predictive control frames are generated based on the control target. The spatial index (i, j, k) of the target voxel is multiplied by its physical size to obtain the physical coordinates (x, y, z) of its center point, and combined with the time step t to construct the target spatiotemporal coordinates. The flight vector is set as a speed command that guides the UAV from its current hovering position to an interception point a certain distance upwind of the target spatiotemporal coordinates. The certain distance upwind is typically set to 2-5 times the voxel side length in the opposite direction of the predicted airflow vector. For example, if the voxel side length is 0.5 m, the interception point is 1.0 m to 2.5 m from the center of the target voxel. This distance is set to ensure that the droplets or airflow sprayed by the UAV have sufficient diffusion and mixing time before reaching the target voxel, while avoiding direct interference to sensor readings due to excessive proximity. This value can be optimized and determined through droplet trajectory simulation in computational fluid dynamics simulation.

[0095] Simultaneously, a specific ID representing the standard atomized spraying behavior primitive is queried and selected from the behavior primitive library. A pseudo-random number generator is then used to generate an integer of a specific length, such as 256 bits, as the execution digest hash seed. This seed is used for subsequent instruction execution verification to ensure data integrity and source reliability. Finally, the frame ID, target spatiotemporal coordinates, flight vector, behavior primitive activation ID, and execution digest hash seed are encapsulated together into a fixed-length data packet, i.e., a predictive control frame, and broadcast to the dynamic probes and intervention units within the cultivation area via a UDP-based downlink.

[0096] Predictive control frames are structured data packets used to issue specific action commands to drones.

[0097] In this embodiment, the heat and moisture transport equations are simplified into scalar transport equations that include advection and diffusion terms, as follows:

[0098] Heat transport equation:

[0099]

[0100] Water transport equation:

[0101]

[0102] in, It represents a scalar quantity of temperature field, and its unit is °C. This represents a scalar quantity representing the water vapor concentration field, with units of kg / m³. This value can be correlated with temperature. The relative humidity is converted using the equation of state. It is a time variable. It is a three-dimensional velocity vector field obtained from the initial microclimate prediction field or the subsequently updated microclimate prediction field, representing the predicted airflow. and These are the turbulent diffusion coefficients for heat and moisture, respectively, and their values ​​are estimated at each time step using a simplified k-epsilon turbulence model. and It is the source term, whose value is set to zero in the prediction phase of this step, and is used to inject correction amount in subsequent correction steps.

[0103] For example, in this embodiment, the spatiotemporal prediction field engine loads an initial microclimate prediction field, including 900,000 initial state vectors, generated by the environmental modeling and initialization module. The engine sets a preset time window of 300 s and begins solving the heat and moisture transport equations. At the 182nd time step, a three-dimensional voxel with index (150, 160, 5) is detected, whose predicted relative humidity drops for the first time from 60.1% in the previous second to 59.8%, exceeding the preset microclimate threshold of 60%. This voxel is immediately identified as the target voxel, and the time t=182 s is recorded. The predicted airflow vectors of this target voxel at t=182 s are then retrieved as {-0.1 m / s, -0.15 m / s, -0.02 m / s}. Thus, the control target is determined. Subsequently, the voxel index (150, 160, 5) is multiplied by the voxel side length of 0.5 m to calculate the target physical coordinates as {x: 75.0 m, y: 80.0 m, z: 2.5 m}. Combined with the time t = 182 s, the target spatiotemporal coordinates are constructed. A behavior primitive activation ID with ID = 0x01 is generated, along with a 256-bit execution digest hash seed, such as "0xA1B2C3D4...E5F6". Finally, the frame ID "0x0001", the target spatiotemporal coordinates, the calculated flight vector, the behavior primitive activation ID "0x01", and the hash seed are encapsulated into a predictive control frame and transmitted via the downlink.

[0104] In one embodiment of the present invention, the instruction execution and observation module is configured to perform the following operations:

[0105] The UAV parses the predictive control frame, extracts the flight vector and behavioral primitive activation ID; it maneuvers according to the flight vector, flying to the spatial position defined by the target's spatiotemporal coordinates; at the target time, it executes the corresponding local fixed intervention action according to the behavioral primitive activation ID; after the intervention action is completed, it uses an onboard multimodal sensor to perform an instantaneous observation to obtain measured microclimate data; it reads the execution digest hash seed from the predictive control frame; it concatenates the parameters of the actually executed intervention action into a parameter string, uses the execution digest hash seed as a key to perform hash operation, and generates an execution digest hash value; it encapsulates the receipt frame ID, actual observation coordinates and time, measured microclimate data, and execution digest hash value to generate a state observation package.

[0106] Specifically, first, the onboard computer of the dynamic probe and intervention unit located on the UAV receives the predictive control frame via its wireless communication module. The built-in parsing program immediately reads the data packet content and verifies and checks data fields such as frame ID and target coordinates. After confirming the validity of the data packet, the parsing program extracts the flight vector and converts it into a series of pulse-width modulated signals for the UAV's flight controller.

[0107] The flight controller adjusts the rotational speed of the UAV's four or more rotors accordingly, driving the UAV to maneuver from its current position according to the heading and speed provided by the flight vector, flying towards the spatial location defined in the target's spatiotemporal coordinates. Simultaneously, the onboard computer starts a countdown timer set to the time t defined in the target's spatiotemporal coordinates. When the UAV confirms, via its positioning system (such as GPS and barometric altimeter), that the error between its own position and the target position is less than a certain tolerance (e.g., 0.5 m), and the countdown timer reaches zero, the onboard computer immediately activates the program associated with the behavior primitive activation ID "0x01" carried in the predictive control frame. This program sends a working signal to a miniature water pump connected to the atomizing nozzle via a GPIO interface, causing it to operate continuously at a set flow rate for a set duration, such as 50 mL / s for 2.0 s, performing a standardized atomizing spraying action, i.e., a local solidification intervention action. The spray flow rate and spraying duration are determined based on the predicted humidity deficit of the target voxel, the UAV's liquid load, and droplet settling and evaporation models. The specific water volume for spraying can be initially set using the following empirical formula:

[0108]

[0109] in To predict the difference (%) between humidity and the threshold. Voxel volume (m 3 ), This is an empirical coefficient, for example, 0.1-0.3 mL / (%·m 3 The flow rate and duration are allocated based on nozzle characteristics (such as atomization angle and droplet diameter) and the drone's hovering stability. Typically, a single spraying session is controlled to last 1-5 seconds to avoid localized over-wetting and excessive drone battery depletion. After spraying is complete, the water pump stops operating. The onboard computer immediately triggers its connected onboard multimodal sensors to perform a momentary observation.

[0110] Specifically, temperature and humidity sensors record the current air temperature and relative humidity, while laser wind radar measures the three-dimensional wind field at the current location. Simultaneously, the onboard computer reads the execution digest hash seed issued in the predictive control frame. The actual action parameters, namely the spray flow rate and spray duration, are concatenated into a string. Using the execution digest hash seed as the key, a specific hash algorithm is called to process this string, generating an execution digest hash value. For example, a 256-bit execution digest hash seed with a spray flow rate of 50 mL / s and a spray duration of 2.0 s would be concatenated into the string "50.0_2.0". The SHA-256 hash algorithm is then used to process this string, generating a 256-bit execution digest hash value.

[0111] Finally, the onboard computer sequentially encapsulates the newly generated receipt frame ID, the actual observation coordinates and time obtained through GPS and barometer, the measured microclimate data obtained by temperature and humidity sensors and laser wind radar, and the just calculated execution digest hash value into a structured state observation package, and sends it back to the spatiotemporal prediction field engine of the ground server through the uplink. The state observation package encapsulates the execution results of the intervention mission and the real observation data on site, which is a complete feedback of the digital twin command.

[0112] It should be noted that the dynamic probe and intervention unit are functional units integrated into the UAV, including hardware such as various sensors and actuators, and embedded software for controlling this hardware and communicating. The flight vector is a vector containing three-dimensional velocity and angular velocity commands, used to guide the UAV's spatial displacement. The behavior primitive activation ID is a 16-bit integer, such as "0x01," corresponding to a predefined, unchangeable short program in the UAV firmware that directly controls the specific operation of the actuator. Locally fixed intervention actions refer to a series of standardized operations pre-programmed and stored in the UAV's local non-volatile memory, such as specific patterns of spraying, blowing, or localized lighting, with specific parameters set according to commands. Instantaneous observation refers to the simultaneous sampling of multiple microclimate parameters at the current location within a short time window, typically 500 ms, after the intervention action is completed. The 500 ms time window is set to capture the immediate effect of the intervention after the microclimate state changes rapidly due to the intervention but before it has fully mixed with the surrounding environment, while avoiding errors caused by sensor response delays. This window value is determined based on the response time of a typical microclimate sensor and the decay time of airflow disturbances on the UAV rotor. The execution digest hash seed, a random number from the predictive control frame, is used as input to the hash operation in this step to ensure closed-loop verification logic. The execution digest hash value is a fixed-length string generated by hashing the key parameters of the actual executed action; it is used to verify whether the action was executed as instructed.

[0113] For example, in this embodiment, the UAV receives a predictive control frame with frame ID "0x0001". It resolves the flight vector and maneuvers to a position with physical coordinates {x: 75.0 m, y: 80.0 m, z: 2.5 m}. At t=182 s indicated by the countdown timer, the UAV activates ID "0x01" based on the behavior primitive and performs a 2.0 s atomization spray at a flow rate of 50 mL / s. After spraying, the UAV immediately performs instantaneous observation, and its onboard multimodal sensor measures the measured microclimate data as {temperature: 27.5 ℃, relative humidity: 85.2%, wind speed components: (-0.12, -0.16, -0.03) m / s}. Simultaneously, the onboard computer reads the execution digest hash seed "0xA1B2C3D4...E5F6" from the predictive control frame and uses the actual action parameter "50.0_2.0" as input to calculate the execution digest hash value using the SHA-256 algorithm, for example, "0x9F86D081...A4C5D6". Finally, the UAV encapsulates the new receipt frame ID "0x0001_Resp", the actual observed coordinates {x: 75.1 m, y: 79.9 m, z: 2.6 m} and time t=182.5 s, the measured microclimate data such as {temperature: 27.5 ℃, humidity: 85.2%, ...,}, and the calculated execution digest hash value "0x9F86D081...A4C5D6" into a state observation packet and sends it back via the uplink.

[0114] In one embodiment of the present invention, the data verification and extraction module is used to perform the following steps:

[0115] Extract the receipt frame ID from the state observation package, retrieve the original instruction parameters and the original issued execution digest hash seed corresponding to the ID; use the original issued execution digest hash seed to perform a local hash operation on the original instruction parameters to generate a locally calculated hash value; compare the locally calculated hash value with the execution digest hash value in the state observation package; when the two match, the verification is deemed successful; extract the measured microclimate data from the verified state observation package as measured data for causal labeling.

[0116] Specifically, firstly, the spatiotemporal prediction field engine receives state observation packets sent by the dynamic probe and intervention unit via TCP protocol at its uplink data receiving port. Upon arrival of the data packets, a mandatory verification process built into the engine's data receiving logic, which cannot be skipped, is initiated to ensure that only feedback data from reliable sources and executed correctly can enter the subsequent core algorithms.

[0117] The first step of the process is to extract the acknowledgment frame ID from the state observation packet. The identifier portion of this acknowledgment frame ID, "0x0001", is compared with an internally maintained list that records all sent predictive control frame IDs. The process proceeds only to the next step if a matching ID is found in the list; otherwise, the state observation packet is discarded, and a communication anomaly is recorded.

[0118] The second step in the process is hash verification. Based on the matching frame ID "0x0001", the original instruction parameters used to generate the frame are retrieved from its local storage, namely the action parameter "50.0_2.0" corresponding to the behavior primitive activation ID "0x01", and the execution digest hash seed "0xA1B2C3D4…E5F6" issued at that time. Next, the spatiotemporal prediction field engine calls its own SHA-256 hash calculation module, using the same hash seed as the key, to perform a hash operation on the retrieved original instruction parameter "50.0_2.0". This operation generates a locally calculated hash value. Then, this locally calculated hash value is compared bit by bit with the execution digest hash value included in the state observation packet. The hash verification passes only when the two 256-bit hash values ​​are completely identical.

[0119] After both the receipt frame ID and the execution digest hash value are verified, the state observation packet is determined to be a legitimate and valid feedback, and the data packet is formally received. Finally, the measured microclimate data encapsulated within this verified state observation packet is extracted, including temperature, humidity, and wind speed components. Since this data was obtained through a rigorously validated closed-loop process, it carries explicit contextual information; that is, it is the direct result of a specific action performed at a specific point in time by a specific instruction with frame ID "0x0001". Therefore, this data is marked as causal-labeled measured data, which is an observation record that includes not only numerical values ​​but also its causes and background metadata. This data is stored in a dedicated memory buffer, awaiting invocation by the error calculation and correction source term generation module.

[0120] Among them, the causal relationship of the measured data with causal labels is clear and traceable, which is fundamentally different from random sensor readings without background.

[0121] For example, in this embodiment, the spatiotemporal prediction field engine receives a data packet. First, it parses the acknowledgment frame ID from the packet, which is “0x0001_Resp”. It queries the list of sent instructions and confirms that “0x0001” is a valid sent ID, thus the first verification step passes. Next, based on the ID “0x0001”, it retrieves the original instruction parameters “50.0_2.0” and the hash seed “0xA1B2C3D4...E5F6” from the local database. It performs a SHA-256 operation locally (“50.0_2.0”, key="0xA1B2C3D4...E5F6") to obtain the locally calculated hash value, which is “0x9F86D081…A4C5D6”. Then, it compares this locally calculated hash value with the execution digest hash value “0x9F86D081...A4C5D6” carried in the data packet, finding that they are completely identical, thus the second verification step passes. Since both verifications passed, the state observation packet was accepted. Subsequently, the measured microclimate data {temperature: 27.5 ℃, humidity: 85.2%, wind speed components: (-0.12, -0.16, -0.03) m / s} was extracted from the packet and marked as strongly correlated with the instruction "0x0001", thus obtaining the causal label. This measured data with the causal label was sent to a first-in-first-out queue as the final output of this step. If the hash value in the data packet is "0x9F86D081...A4C5D7", the entire data packet will be discarded because it does not match the locally calculated value.

[0122] In one embodiment of the present invention, the error calculation and correction source term generation module is used to perform the following steps:

[0123] Extract the measured state vector from the causally labeled measured data; query the predicted state vector of the corresponding spatiotemporal point from the spatiotemporal prediction field engine; calculate the state residual vector between the measured state vector and the predicted state vector; determine the strength of the correction source term by multiplying the magnitude of the state residual vector by a preset gain coefficient; determine the positive or negative nature of the correction source term based on the direction of the state residual vector; add the correction source term as an instantaneous source term to the transport equation of the corresponding three-dimensional voxel.

[0124] Specifically, firstly, the measured data with causal labels are extracted from the output of the data verification and extraction module. This data includes measured temperature, humidity, and three-dimensional velocity components, which together constitute the measured state vector. Simultaneously, based on the causal label information carried by this data—namely, the corresponding observation coordinate system and time step—the predicted state vector at the same spatiotemporal point is queried and located from the four-dimensional prediction dataset generated by the prediction and decision module. Next, component-wise subtraction is performed on these two vectors to calculate the state residual vector, i.e., It quantifies the discrepancy between predictions and reality.

[0125] The measured temperature is subtracted from the predicted temperature to obtain the temperature residual; the measured humidity is converted to water vapor concentration and then subtracted from the predicted water vapor concentration to obtain the concentration residual; and the measured velocity component is subtracted from the predicted velocity component to obtain three velocity residuals. These residual values ​​together constitute the state residual vector. Subsequently, based on the magnitude and direction of the state residual vector, it is transformed into a correction source term. It should be noted that the correction source term is a scalar or vector representing the magnitude and direction of the virtual energy required to correct the state residual vector. In this embodiment, it is simplified to a scalar, whose physical meaning is the mass of water vapor that needs to be injected into or removed from the voxel per unit time. First, the absolute value of the concentration residual component representing the core control target (water vapor concentration in this example) in the state residual vector is calculated, and then multiplied by a preset gain coefficient, such as 0.1. The gain coefficient is set according to the stability and response speed requirements of the control system. If the value is too large, it will cause the digital twin model to overcorrect and produce oscillations; if it is too small, convergence will be slow. The value can be determined through system identification and PID parameter tuning methods, such as the Ziegler-Nichols method or gradient descent based on historical error data. Typical values ​​range from 0.05 to 0.2. In this embodiment, 0.1 is preferred. This coefficient is experimentally calibrated based on the system response characteristics to determine the strength or magnitude of the correction source term. The sign (positive or negative) of the concentration residual component in the state residual vector determines the nature of the correction source term; a positive residual represents the positive potential energy required to increase the concentration, while a negative residual represents the opposite. The setting of this correction source term is based on the proportional control concept in control theory, i.e., the larger the deviation, the larger the correction applied.

[0126] Finally, the spatiotemporal prediction field engine injects this calculated correction source term as an instantaneous, local energy source into the 3D voxel mesh. Specifically, this injection operation involves using the value of this correction source term as the source term when solving the heat and moisture transport equations for the next time step. The injection is added to the calculation formula for the voxel corresponding to the causally labeled measured data observation point, i.e., the voxel with index (150, 160, 5). This injection only occurs in the next computation time step, after which the source term is restored to zero.

[0127] The measured state vector is a vector comprising multiple observed physical quantities, and its dimensions match those of the predicted state vector. For example... The predicted state vector is the state vector predicted by the spatiotemporal prediction field engine at the same time and space point before intervention, for example... .

[0128] For example, in this embodiment, measured data with causal tags are retrieved from the queue. This includes a measured state vector of {T: 27.5 ℃, C: 0.0216 kg / m³, u: -0.12 m / s, v: -0.16 m / s, w: -0.03 m / s}, where the water vapor concentration C is calculated based on a relative humidity of 85.2% and a temperature of 27.5 ℃. Assume that at t=182 s, the predicted state vector for voxels (150, 160, 5) is {T: 28.0 ℃, C: 0.0163 kg / m³, u: -0.1 m / s, v: -0.15 m / s, w: -0.02 m / s}, where the water vapor concentration is calculated based on a predicted humidity of 59.8% and a predicted temperature of 28.0 ℃. The state residual vector was calculated as {ΔT: -0.5 ℃, ΔC: +0.0053 kg / m³, Δu: -0.02 m / s, Δv: -0.01 m / s, Δw: -0.01 m / s}. Next, the strength of the correction source term was calculated based on the concentration residual ΔC. Assuming a preset gain coefficient k of 0.1, the magnitude of the correction source term is 0.0053 kg / m³ × 0.1 = 0.00053. Since ΔC is positive, it indicates that the actual concentration is higher than the prediction; therefore, this represents a positive potential energy for increasing the concentration. Finally, the value 0.00053 was taken as the source term. The instantaneous value is injected only for the equation of voxel (150, 160, 5) when calculating the next time step (t=183).

[0129] In one embodiment of the present invention, the source term propagation and field update correction module is used to perform the following steps:

[0130] Obtain the predicted airflow vector of the injection point voxel and its neighborhood, and calculate the local mainstream direction; construct an anisotropic tensor and set the diffusion coefficients along the mainstream direction and the vertical direction; use the anisotropic tensor to replace the scalar diffusion coefficient to control the propagation of the correction source term in the heat and moisture transport equations; generate the updated microclimate prediction field through iterative calculation.

[0131] Specifically, before the spatiotemporal prediction field engine prepares to solve the heat and moisture transport equations for the next time step, it uses the local airflow vector field at the injection point to define a mathematical object, namely an anisotropic tensor, such as a second-order tensor, to describe the characteristics of the medium having different diffusion properties in different directions. Specifically, the engine queries the predicted airflow vectors of the voxel and its surrounding neighboring voxels for the current time step, and calculates a smoothed local mainstream direction vector through a weighted average.

[0132] Then, a 3×3 anisotropic tensor is constructed, with its principal axis aligned with the local mainstream direction vector. The component values ​​along this principal axis are set to larger values, for example, 5 times the base diffusion coefficient, while the component values ​​along the other two directions orthogonal to the principal axis are set to smaller values, for example, 0.5 times the base diffusion coefficient. The "base diffusion coefficient" refers to the scalar diffusion coefficient in the standard isotropic diffusion model. or Its typical value is based on the thermal diffusivity and turbulent diffusivity of air. For heat, the molecular thermal diffusivity is approximately 2.0 × 10⁻⁶. 5 m 2 / s, the turbulent thermal diffusivity is typically 1-3 orders of magnitude greater; for water vapor, the molecular diffusivity is approximately 2.5 × 10⁻⁶. 5 m 2 / s. In actual forest underflow turbulent environments, turbulent diffusion dominates, and the basic diffusion coefficient can be taken as 1.0 × 10⁻⁶. 3 m 2 / s to 5.0×10 2  m 2 The specific value can be calibrated by fitting measured concentration decay curves or using large eddy simulation. The anisotropy ratio (e.g., 5x and 0.5x) is set based on the observation that understory airflow is generally strongly directional along canopy gaps or channels, while lateral diffusion is strongly hindered by vegetation. This ratio can be obtained by analyzing multiple sets of tracer gas diffusion experimental data under different wind directions. It should be noted that this tensor is used to specify that the rate and direction of the source term propagation in space are not uniform, but depend on the local flow field direction.

[0133] In the calculation of the next time step in solving the heat and moisture transport equations, the engine will use the scalar diffusion coefficient in the equations. Replace it with the previously defined anisotropic tensor. This means that when calculating the diffusion term... At this point, the water vapor concentration increment represented by the correction source term will propagate at a higher rate along the predicted breeze path, while propagation in other directions will be suppressed. In subsequent time steps, this injected correction source term will physically diffuse and attenuate throughout the entire 3D voxel grid, based on the anisotropic tensor of each local voxel. Specifically, the initial concentration increment concentrated at the injection point will drift along the predicted airflow path like a wisp of smoke, its influence gradually spreading to downstream voxels, while its peak intensity decreases with increasing propagation distance. This process smoothly adjusts the state variables of all affected voxels, primarily water vapor concentration and temperature values ​​in this embodiment, ensuring a smooth and overall correction of the simulated state of the entire prediction field towards a direction validated by real-world data. After the computation is complete, the state vectors of all 3D voxels are updated, collectively forming an updated microclimate prediction field that more closely resembles reality. Compared to the initial microclimate prediction field, it better matches the real-world state in local areas. The microclimate prediction field is stored in memory as the final output of this step.

[0134] The heat and moisture transport equations are physical equations defined in the prediction and decision-making module. The correction source term is calculated in the error calculation and correction source term generation module, representing the correction amount for the deviation between actual observations and model predictions.

[0135] For example, in this embodiment, the spatiotemporal prediction field engine locates the injection point voxel (150, 160, 5). The predicted airflow vectors of this voxel and its neighborhood are averaged to obtain a local mainstream direction vector, with a direction approximately -135°. Based on this, an anisotropic tensor is constructed, and the diffusion coefficient of this tensor in the -135° direction is temporarily set to 5.0e. -5 The diffusion coefficient is set to m² / s, while the diffusion coefficient in the direction perpendicular to it at 45° is set to 0.5e. -5The thermal density is m² / s, while the basic diffusion coefficient is 1.0e⁻⁵ m² / s. When solving the heat and moisture transport equations at time step t=183, the engine uses this tensor to calculate the diffusion term. The correction source term 0.00053 injected in the previous step serves as the source term, causing an instantaneous increase in the concentration of voxels (150, 160, 5). In subsequent time steps t=184, 185, ..., this concentration increment mainly propagates along the -135° direction, i.e., downstream voxels (149, 159, 5), (148, 158, 5), etc., with its influence gradually expanding. Simultaneously, the peak concentration shifts away from voxel (150, 160, 5) and gradually decreases. After approximately 50 time steps of propagation and decay, the injected energy is completely absorbed, and the entire field reaches a new equilibrium state. At this point, the set of state vectors of all 900,000 three-dimensional voxels constitutes the updated microclimate prediction field. For example, the humidity of voxel (150, 160, 5) is corrected to 79%, and the humidity of its downstream voxel (149, 159, 5) is also slightly increased from the predicted 61% to 63%.

[0136] In one embodiment of the present invention, the closed-loop iterative scheduling module is configured to perform the following steps:

[0137] The updated microclimate prediction field is fed back to the prediction and decision-making module as a new input; the spatiotemporal prediction field engine is driven to perform the next round of trend extrapolation based on the updated microclimate prediction field; by cyclically executing the prediction and decision-making module to the correction source term propagation and field update module, an adaptive closed-loop control process including prediction, intervention, observation, verification, and correction is formed.

[0138] Specifically, firstly, the updated microclimate prediction field generated by the source term propagation and field update module is directly used as the new input and returned to the starting position of the prediction and decision module. The microclimate prediction field includes full-field state information corrected by real-world observation data. At this point, the spatiotemporal prediction field engine no longer uses the initial microclimate prediction field, but instead uses this newly updated microclimate prediction field, which is more consistent with physical reality, as the initial condition for its next round of prediction and inference.

[0139] The spatiotemporal prediction field engine is driven again, repeating the complete process of the prediction and decision-making module. Based on new initial conditions, it deduces the microclimate state evolution trend of each three-dimensional voxel within a preset time window by solving the heat and moisture transport equations. Since the input data has been updated, the engine may predict new anomalous trends completely different from the previous round, or confirm that previous interventions have successfully eliminated the anomalies. If a new target voxel is predicted, a new predictive control frame is generated and issued. Subsequently, the dynamic probe and intervention unit on the UAV receive the new instructions and execute the operation of the instruction execution and observation module, completing maneuvering, intervention, and observation, and transmitting a new state observation package. Next, the data verification and extraction module performs forced verification of this new observation package. After verification, the error calculation and correction source term generation module calculates the new state residual vector and generates a new correction source term. Immediately afterwards, the correction source term propagation and field update module injects and propagates the new correction source term, generating a newly updated microclimate prediction field.

[0140] This complete loop, encompassing the prediction and decision-making module to the correction source term propagation and field update module, constitutes a closed-loop regulation. By continuously executing these complete steps in an uninterrupted, event-triggered manner, an adaptive closed-loop regulation process is formed. This process dynamically adjusts its predictive and control behaviors based on actual environmental changes and the actual effects of interventions. This process does not rely on fixed time periods but is driven by each successful prediction and verification, thereby achieving proactive, continuous, and predictive shaping of the microclimate of the entire understory cultivation area. The core characteristic of this process is its predictability; all intervention actions are initiated based on predictions of the future, aiming to resolve potential anomalies in their nascent stage.

[0141] The adaptive closed-loop control process refers to the cyclical process of continuously repeating the prediction and decision-making module to the correction source term propagation and field update module, rather than executing according to a fixed script.

[0142] For example, in this embodiment, the updated microclimate prediction field is immediately transmitted to the prediction and decision-making module. In this prediction field, the humidity of voxels (150, 160, 5) and their downstream areas has been corrected. The spatiotemporal prediction field engine extrapolates from this new starting point and finds that the previously predicted humidity drop below the threshold no longer occurs within the next 300 seconds. However, at the 250th second of the extrapolation, the engine predicts that the temperature of voxels (80, 85, 15) located at the edge of the forest window will exceed the temperature threshold of 35°C due to direct afternoon sunlight. This temperature threshold of 35°C is set based on the high-temperature stress critical point of common understory crops. For example, most shade-loving crops will experience physiological stress such as decreased photosynthesis and transpiration disorder when the temperature is consistently above 32-35°C. The specific threshold can be adjusted according to crop variety, growth stage, and historical meteorological disaster records, or it can be integrated with crop growth models to dynamically calculate the suitable temperature range for different growth stages.

[0143] A new predictive control frame with frame ID "0x0002" is generated, instructing the UAV to proceed to the new location to perform a misting spray aimed at cooling. The UAV receives and executes the instruction, subsequently transmitting a state observation packet including the execution result and measured data. The temperature field is then corrected through a process of verification, residual calculation, injection, and propagation of correction source terms. This process is continuously iterated; whenever the model's predictions indicate a risk of exceeding a threshold, an intervention is proactively initiated, and the digital twin model is continuously calibrated based on real-world feedback, thereby maintaining the microclimate of the entire cultivation area within the range most suitable for crop growth.

[0144] Each of the modules can be implemented in whole or in part through software, hardware, or a combination thereof. It supports hardware embedded in or independent of the processor in the computer device, and also supports software stored in the memory of the computer device, so that the processor can call and execute the operations corresponding to each of the above modules.

[0145] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A microclimate precision control system for forest understory cultivation areas based on unmanned aerial vehicles (UAVs), characterized in that, include: The environmental modeling and initialization module is configured to construct a three-dimensional voxel mesh covering the understory cultivation area and generate an initial microclimate prediction field based on environmental meteorological data input from outside the cultivation area. The environmental meteorological data input from outside the cultivation area includes at least ambient temperature and relative humidity. The prediction and decision-making module is configured to extrapolate the evolution trend of microclimate state based on the initial microclimate prediction field, traverse the extrapolation results, locate the first target voxel whose predicted state value will exceed the preset microclimate threshold; extract the spatiotemporal coordinates of the target voxel and the predicted airflow vector that causes its state change; calculate the UAV's flight vector based on the target voxel's spatiotemporal coordinates and the predicted airflow vector; select the corresponding behavior primitive activation ID from the preset behavior primitive library; call the random number generator to generate the execution digest hash seed; and encapsulate the target voxel's spatiotemporal coordinates, flight vector, behavior primitive activation ID, and execution digest hash seed to generate a predictive control frame. The command execution and observation module is configured to allow the UAV to execute local fixed intervention actions based on predictive control frames, and to generate state observation packages based on measured microclimate data and execution summary hash seeds; The data verification and extraction module is configured to use the execution digest hash seed to compare hash values, verify the data integrity in the state observation package, and extract the measured data of the causal tag. The error calculation and correction source term generation module is configured to calculate the deviation between the measured data of the causal label and the predicted state of the corresponding spatiotemporal point, and generate correction source terms. The correction source term propagation and field update module is configured to inject the correction source term into the corresponding voxel of the three-dimensional voxel grid and control its propagation path according to the predicted airflow vector to generate an updated microclimate prediction field. The closed-loop iterative scheduling module is configured to take the updated microclimate prediction field as input and return it to the prediction and decision module, and then execute it cyclically for the next round of closed-loop regulation.

2. The microclimate precision control system for forest understory cultivation areas based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, The environment modeling and initialization module is used to perform the following operations: A 3D environment model was constructed based on forest understory 3D point cloud data; The three-dimensional environment model is discretized into a three-dimensional voxel mesh, and each voxel is labeled with fluid or solid properties. The calculation boundary conditions are set based on environmental meteorological data input from outside the cultivation area; Iterative calculations are performed under the computational boundary conditions until the rate of change of field variables within the grid is lower than the preset stability threshold, thus generating the initial microclimate prediction field.

3. The microclimate precision control system for forest understory cultivation areas based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, The instruction execution and observation module is used to perform the following operations: The UAV analyzes predictive control frames to extract flight vectors and behavioral primitive activation IDs; Maneuvering based on flight vectors, flying toward the target's spatial location defined by its spacetime coordinates; At the target time, execute the corresponding localized, fixed intervention action based on the behavioral primitive activation ID; After the intervention is completed, an instantaneous observation is conducted using an airborne multimodal sensor to obtain measured microclimate data; Read the execution digest hash seed from the predictive control frame; The actual intervention action parameters are concatenated into a parameter string, and the execution digest hash seed is used as the key to perform hash operation to generate the execution digest hash value; The receipt frame ID, actual observation coordinates and time, measured microclimate data, and execution summary hash value are encapsulated to generate a state observation package.

4. The microclimate precision control system for forest understory cultivation areas based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, The data validation and extraction module is used to perform the following operations: Extract the receipt frame ID from the status observation packet, and retrieve the original instruction parameters and the original issued execution digest hash seed corresponding to the ID; Using the originally issued execution digest hash seed, perform local hashing on the original instruction parameters to generate locally computed hash values; The locally computed hash value is compared with the execution digest hash value in the state observation package. If they match, the verification is considered successful. Measured microclimate data are extracted from the validated state observation package as measured data for causal labeling.

5. The microclimate precision control system for forest understory cultivation areas based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, The error calculation and correction source term generation module is used to perform the following steps: Extract the measured state vector from the causally labeled measured data; Query the predicted state vector of the corresponding spatiotemporal point from the spatiotemporal prediction field engine; Calculate the state residual vector between the measured state vector and the predicted state vector; The strength of the correction source term is determined by multiplying the magnitude of the state residual vector by a preset gain coefficient. The positive or negative nature of the correction source term is determined based on the direction of the state residual vector; The correction source term is added as an instantaneous source term to the transport equation of the corresponding three-dimensional voxel.

6. The microclimate precision control system for forest understory cultivation areas based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, The source term propagation and field update correction module is used to perform the following steps: Obtain the predicted airflow vector of the injection point voxel and its neighborhood, and calculate the local mainstream direction; Construct an anisotropic tensor and set the diffusion coefficients along the mainstream direction and the perpendicular direction; The propagation of the correction source term is controlled in the heat and moisture transport equation by replacing the scalar diffusion coefficient with an anisotropic tensor. An updated microclimate prediction field is generated through iterative calculations.

7. The microclimate precision control system for forest understory cultivation areas based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, The closed-loop iterative scheduling module is used to perform the following operations: The updated microclimate prediction field is fed back as a new input to the prediction and decision-making module; The spatiotemporal prediction field engine is driven to perform the next round of trend extrapolation based on the updated microclimate prediction field; By iteratively executing the prediction and decision-making module to the correction source term propagation and field update module, an adaptive closed-loop control process is formed, including prediction, intervention, observation, verification, and correction.

8. The microclimate precision control system for forest understory cultivation areas based on unmanned aerial vehicles (UAVs) according to claim 2, characterized in that, Constructing a 3D voxel mesh includes the following steps: Obtain the bounding box of the 3D environment model in 3D space, and discretize the bounding box into multiple cubic units according to the preset resolution; Traverse each cube unit and use ray casting to determine whether its center point is located inside the 3D environment model. If the center point is located inside the model, the cube element is marked as a solid property; if it is located outside the model, it is marked as a fluid property. All cubic cells labeled with fluid properties are combined to form a three-dimensional voxel mesh for carrying out microclimate calculations, and fluid boundary conditions are established in the physics engine.

9. The microclimate precision control system for forest understory cultivation areas based on unmanned aerial vehicles (UAVs) according to claim 3, characterized in that, The behavior primitive activation ID corresponds to the standardized operating procedure preset in the UAV's local memory. The standardized operating procedure includes specific modes of atomization spraying, local blowing, or supplemental lighting operations, and the predictive control frame is broadcast through the downlink of the wireless protocol.

Images