A method and system for precise variable fertilization using drones based on artificial intelligence

Abstract

This invention relates to the field of fertilization technology, specifically to an artificial intelligence-based method and system for precise variable fertilization using unmanned aerial vehicles (UAVs). The method includes: obtaining real-time crop status data of a target crop area based on real-time multi-source sensor data acquired by the UAV during flight operations; inputting the real-time crop status data into a preset digital twin model for calculation to obtain dynamic variable fertilization control instructions for crop growth in future time periods; and controlling the UAV to perform fertilization operations on the target crop area according to the dynamic variable fertilization control instructions. This addresses the technical problem that existing variable fertilization methods rely on static analysis of the current or historical growth status of crops, failing to incorporate and optimize real-time predictions of future crop growth dynamics into the fertilization plan, resulting in a lack of foresight in variable fertilization control.

Description

Technical Field

[0001] This invention relates to the field of fertilization technology, specifically to an artificial intelligence-based method and system for precise variable fertilization using unmanned aerial vehicles (UAVs). Background Technology

[0002] Variable-rate fertilization is a key technology in modern agriculture. Its core lies in applying differentiated fertilizers to different locations in the field based on the actual needs of the crop during its growth process, aiming to increase yield and income, reduce waste, and protect the environment. Existing variable-rate fertilization methods typically rely on sensors mounted on drones to acquire real-time or historical crop spectral and image data. These data are then analyzed using onboard or cloud-based intelligent algorithms to generate a data map reflecting current nutrient differences in the field. Based on this map, a corresponding fertilization plan is generated, and the drone is controlled to execute the preset variable-rate fertilization operation. However, because this fertilization plan relies on diagnosing the crop's current or past growth status, it is essentially a reactive or in-process remedial fertilization approach. It fails to incorporate the crop's future growth dynamics and nutrient requirements into a real-time decision-making loop, resulting in a lack of predictability and optimality in the fertilization plan over time. By the time the fertilizer is absorbed and produces physiological effects, the crop's actual needs may have changed, thus affecting the final accuracy and yield increase effect of variable-rate fertilization. Summary of the Invention

[0003] To address the technical problem that existing variable fertilization methods rely on static analysis of the current or historical growth status of crops, and cannot incorporate and optimize predictions of future crop growth dynamics in real time into the fertilization plan, resulting in a lack of foresight in variable fertilization control, this application provides an artificial intelligence-based unmanned aerial vehicle (UAV) precision variable fertilization method and system.

[0004] The technical solution provided in this application for a method and system for precise variable fertilization using drones based on artificial intelligence is as follows: An AI-based method for precise variable fertilization using drones, comprising: Based on the real-time multi-source sensor data acquired by the UAV during flight operations, real-time crop status data of the target crop area is obtained. Real-time crop status data is input into a preset digital twin model for calculation to obtain dynamic variable fertilization control instructions for crop growth in future time periods. Based on dynamic variable fertilization control instructions, the drone is controlled to perform fertilization operations on the target crop area.

[0005] Furthermore, the UAV is equipped with a multispectral imaging unit, a hyperspectral imaging unit, and a millimeter-wave radar unit. The steps for obtaining real-time crop status data of the target crop area based on real-time multi-source sensor data acquired by the UAV during flight operations include: The system simultaneously acquires image data, spectral data, and three-dimensional point cloud data of the target crop area through a multispectral imaging unit, a hyperspectral imaging unit, and a millimeter-wave radar unit, respectively, to form real-time multi-source sensing data. Feature extraction and feature fusion are performed on real-time multi-source sensor data to obtain multi-dimensional feature vectors characterizing crop canopy nutrients, water stress, and canopy geometry. Based on the real-time flight condition parameters and real-time environmental parameters of the UAV, feature weighting and information compensation are performed on the multi-dimensional feature vector to obtain an enhanced feature vector; The enhanced feature vector is fused with the historical feature vector from the previous time step in the current job cycle using a sliding window in the time domain to generate a fused feature vector. The fused feature vectors are interpreted to obtain real-time crop status data.

[0006] Furthermore, the steps for interpreting the fused feature vectors to obtain real-time crop status data include: By using a pre-defined crop physiological model, dot product and nonlinear transformation calculations are performed on the fused feature vectors to obtain an initial set of crop physiological parameters. The pre-defined crop physiological model includes pre-defined crop state interpretation calculation rules and a pre-defined weight matrix associated with the pre-defined crop state interpretation calculation rules. The initial crop physiological parameter set is temporally correlated with the historical crop physiological parameter set of the previous time in the current operation cycle to obtain the target crop physiological parameter set; Based on the preset calibration mapping relationship, the target crop physiological parameter set is converted into real-time crop status data.

[0007] Furthermore, the steps of inputting real-time crop status data into a preset digital twin model for calculation to obtain dynamic variable fertilization control instructions for crop growth in future time periods include: By using a pre-set digital twin model, based on the input real-time crop status data, the predicted growth status of the target crop area under at least two different fertilization schemes is simulated in the future time period. Calculate the degree of deviation between each predicted growth state and the preset predicted growth state, and take the fertilization scheme corresponding to the degree of deviation being less than the preset deviation as the target fertilization scheme; Based on the drone's operational parameters and fertilizer container status, the fertilizer application rate and flight path of the target fertilization plan are adjusted to generate dynamic variable fertilization control commands.

[0008] Furthermore, the steps of simulating the predicted growth status of the target crop area under at least two different fertilization schemes in a future time period by using a pre-set digital twin model based on the input real-time crop status data include: Based on real-time crop status data, initialize the crop growth status and soil nutrient status of the target crop area in the preset digital twin model; Based on the spatiotemporal distribution of fertilizer application for each fertilization scheme, combined with meteorological forecast data for future periods, the predicted growth status is obtained by coupling the crop growth kinetic equation and soil nutrient transport equation in the preset digital twin model through time-step simulation.

[0009] Furthermore, the operational parameters include flight speed, operational swath width, and current hovering position. Based on the UAV's operational parameters and fertilizer container status, the steps for adjusting the fertilizer application rate and flight path of the target fertilization plan and generating dynamic variable fertilization control commands include: Based on the initial fertilizer requirements of each zone within the target crop area in the target fertilization plan, and the total amount of fertilizer remaining in the fertilizer tank, the total fertilizer constraint is calculated. Based on the total fertilizer application constraint, the initial fertilizer application requirements of each zone are adjusted to obtain an optimized fertilizer application requirement distribution. Based on the distribution of fertilizer demand, as well as the flight speed, operating width, and current hovering position, the flight operation path is adjusted to obtain the optimized flight operation path; Among them, the dynamic variable fertilization control instructions include optimizing the distribution of fertilization demand and optimizing flight operation paths.

[0010] Furthermore, based on the dynamic variable fertilization control instructions, the steps for controlling the drone to perform fertilization operations on the target crop area include: The optimized flight path is calculated into real-time control signals for the UAV in terms of horizontal position and flight speed. The UAV is then driven to perform the current flight operation along the optimized flight path using the real-time control signals. Based on the optimized distribution of fertilizer demand, as well as the real-time flight position and speed of the current flight operation, the real-time spraying flow rate of the fertilizer nozzle unit on the UAV is calculated. Based on the real-time spray flow rate, spray control signals are generated for the valve opening and liquid pump speed on the fertilizer nozzle unit, so as to drive the fertilizer nozzle unit to perform spraying operations based on the spray control signals; The measured spray flow rate of the spraying operation is obtained, and the measured spray flow rate is compared with the real-time spray flow rate. After obtaining the real-time flow error, the spraying control signal is adjusted in real time according to the real-time flow error to optimize the spraying operation.

[0011] This application also provides an artificial intelligence-based drone precision variable fertilization system, comprising: The data acquisition module is used to obtain real-time crop status data of the target crop area based on real-time multi-source sensor data acquired by the UAV during flight operations. The data calculation module is used to input real-time crop status data into a preset digital twin model for calculation, and obtain dynamic variable fertilization control instructions for crop growth in future time periods. The operation control module is used to control the drone to perform fertilization operations on the target crop area based on dynamic variable fertilization control instructions.

[0012] Beneficial effects achieved: This application provides an artificial intelligence-based method for precise variable fertilization by unmanned aerial vehicles (UAVs), comprising: obtaining real-time crop status data of a target crop area based on real-time multi-source sensor data acquired by the UAV during flight operations; inputting the real-time crop status data into a preset digital twin model for calculation to obtain dynamic variable fertilization control instructions for crop growth in future time periods; and controlling the UAV to perform fertilization operations on the target crop area according to the dynamic variable fertilization control instructions.

[0013] In this application, real-time crop status data is obtained by acquiring multi-source sensor data in real time during UAV flight operations, ensuring the immediacy and authenticity of the data input and overcoming the lag of relying on historical data. This data is then fed into a preset digital twin model for calculation. This preset digital twin model is essentially a virtual simulation system that includes crop growth patterns, soil characteristics, and environmental responses. Starting from the input real-time crop status data, it can simulate the crop growth trajectory and results in future time periods, obtaining dynamic variable fertilization control instructions for crop growth in future time periods. These generated dynamic variable fertilization control instructions are not only based on the diagnosis of current needs, but also on a scheme that pursues overall growth in future time periods after simulating and optimizing future scenarios. Finally, the UAV is controlled to perform fertilization operations based on the dynamic variable fertilization control instructions, ensuring that the forward-looking instructions derived from future predictions are accurately implemented. Therefore, this application integrates the dynamic prediction of crop growth in the future into the current fertilization control instructions in real time through a closed loop of "real-time perception - digital twin simulation prediction - dynamic control", fundamentally giving variable fertilization control a forward-looking perspective and solving the shortcomings of existing methods that rely only on static analysis of the current or historical state and lack future considerations. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the steps of a drone-based precision variable fertilization method according to the present application; Figure 2 This is a flowchart illustrating the steps involved in calculating real-time crop status data in this application. Figure 3 A flowchart illustrating the steps involved in calculating dynamic variable fertilization control instructions for future time periods in this application; Figure 4 This is a flowchart illustrating the steps involved in controlling a drone to perform fertilization operations according to this application. Figure 5 This is a schematic diagram of the modules of the AI-based drone precision variable fertilization system of this application. Detailed Implementation

[0015] The following combination Figures 1-5 This application will be described in further detail.

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

[0017] Furthermore, if the embodiments of this invention involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the use of "and / or" or "and / or" throughout the text includes three parallel solutions. For example, "A and / or B" includes solution A, solution B, or a solution where both A and B are satisfied simultaneously. Furthermore, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

[0018] This application discloses an artificial intelligence-based method for precise variable fertilization using drones.

[0019] Please refer to Figure 1 The artificial intelligence-based drone precision variable fertilization method proposed in this embodiment includes steps S10~S30: Step S10: Obtain real-time crop status data of the target crop area based on the real-time multi-source sensor data acquired by the UAV during flight operations.

[0020] This step utilizes real-time multi-source sensor data acquired synchronously by the UAV during flight operations to directly and without delay capture the true state of the target crop area at the moment of operation. This transforms the existing input of lagging information, which relies on historical data or asynchronous monitoring, into high spatiotemporal resolution on-site perception synchronized with the flight operation. This ensures that the crop state information upon which all subsequent calculations and decisions are based has high timeliness and spatial accuracy, fundamentally avoiding decision-making biases caused by using outdated or location-mismatched data. At the same time, by fusing multi-source sensor information, it is possible to more comprehensively analyze the crop's physiological and ecological parameters in multiple dimensions, such as nutrients, water, and canopy structure. This provides rich and reliable input data for subsequent complex simulations and predictions, achieving a tight coupling of perception and decision-making execution in time and space. This enables the entire fertilization system to respond to changes in the state of the fertilization area at the moment of operation, thus providing a data foundation for forward-looking dynamic optimization decisions based on a pre-set digital twin model.

[0021] Step S20: Input the real-time crop status data into the preset digital twin model for calculation to obtain the dynamic variable fertilization control instructions for crop growth in the future time period.

[0022] Real-time crop status data is input into a pre-defined digital twin model. This model incorporates the mechanisms or laws governing the interaction between crop growth physiology, soil nutrient transport, and environmental factors. Through calculations, it simulates the possible crop growth paths and final states caused by different fertilization schemes in future time periods, thus establishing a virtual farmland test field. This allows for instantaneous and cost-free testing of the potential effects of various fertilization schemes in future time periods, and the selection of the fertilization scheme most conducive to achieving preset growth goals, such as maximizing yield, optimizing quality, or maximizing nutrient utilization efficiency. This generates dynamic variable fertilization control instructions, transforming traditional, real-time diagnostic-based fertilization schemes into predictive fertilization schemes. This significantly improves the foresight, scientific rigor, and overall effectiveness of dynamic variable fertilization control instructions, solving the fundamental problem of control lag and insufficient optimization caused by the inability of existing methods to incorporate future dynamics into decision-making.

[0023] Step S30: Based on the dynamic variable fertilization control command, control the drone to perform fertilization operations on the target crop area.

[0024] The drone's flight and nozzle movements are driven by dynamic variable fertilization control commands, ensuring that it strictly adheres to the requirements of these commands and applies corresponding spraying rates to each zone of the target crop area. This guarantees that all dynamic variable fertilization control commands based on real-time perception and future prediction can effectively act on the target crop area, thus constructing a precision variable fertilization system that integrates "perception-decision-execution." This allows dynamic variable fertilization strategies with future optimization characteristics to be accurately implemented in reality, achieving the goal of on-demand supply and forward-looking regulation of precision variable fertilization in the target crop area, effectively improving fertilizer utilization efficiency and overall agronomic benefits.

[0025] In one feasible implementation, refer to Figure 2 As shown, step S10 may specifically include steps S11 to S15: Step S11: Simultaneously acquire image data, spectral data, and three-dimensional point cloud data of the target crop area through the multispectral imaging unit, hyperspectral imaging unit, and millimeter-wave radar unit to form real-time multi-source sensing data.

[0026] During flight operations, the multispectral imaging unit, hyperspectral imaging unit, and millimeter-wave radar unit carried by the UAV are triggered in a unified time sequence by the flight control system to synchronously collect data on the same crop area directly below, i.e., the target crop area.

[0027] Specifically, the multispectral imaging unit uses multiple discrete spectral channel sensors built in. Each spectral channel sensor is configured with optical filters and photodetectors for specific bands, such as the red light band related to chlorophyll absorption, the red edge band sensitive to crop physiological changes, and the near-infrared band related to water absorption. When the drone flies over the target crop area, each spectral channel sensor is exposed synchronously to capture the intensity of reflected light from the crop canopy under the corresponding specific band, thereby generating a set of spatially aligned image data in which each pixel contains multiple specific band reflectances.

[0028] The hyperspectral imaging unit uses its built-in spectrometer and beam-splitting elements such as gratings or prisms to disperse the incident crop canopy reflected light into hundreds of continuous narrow bands in space. A two-dimensional detector array synchronously records the spectral intensity of each spatial pixel in each narrow band, thereby obtaining high-resolution continuous spectral data covering the visible to near-infrared range in a single scan. Each spatial pixel corresponds to a fine spectral curve.

[0029] The millimeter-wave radar unit actively transmits frequency-modulated continuous waves or pulsed millimeter-wave signals into the crop canopy of the target crop area through its transmitting antenna. It then uses a receiving antenna to capture the echo signals reflected from different depths and layers of the crop canopy, such as leaves and stems. By measuring the time difference or phase difference between the transmitted millimeter-wave signal and the echo signal, the straight-line distance between the transmitted millimeter-wave signal and the transmitting antenna is calculated. The millimeter-wave radar unit simultaneously records the scanning angle when acquiring the echo signal. This scanning angle typically includes the azimuth angle in the horizontal direction and the elevation angle in the vertical direction. The azimuth angle and the elevation angle together define the direction vector of the echo signal source. After obtaining the slant range and direction vector, the millimeter-wave radar unit itself can determine the relative three-dimensional position of the reflection point corresponding to the echo signal in a polar coordinate system with the millimeter-wave radar unit as the origin. To unify all reflection points into a global coordinate system with actual geographical significance (i.e., three-dimensional spatial coordinates) to generate a three-dimensional point cloud describing the entire canopy structure, the flight position information (usually provided by the Global Navigation Satellite System (GNSS) onboard the UAV) and flight attitude information (usually provided by the Inertial Measurement Unit (IMU) for pitch, roll, and yaw angles) acquired by the millimeter-wave radar unit at the moment of measurement are incorporated into the calculation. Specifically... In other words, through coordinate transformation, the relative position coordinates based on the radar polar coordinate system are first transformed into the carrier coordinate system with the UAV's center of mass as the reference, using the installation geometry of the millimeter-wave radar unit on the UAV platform. Then, combined with the UAV's real-time position and attitude matrix, it is further transformed into a fixed Earth coordinate system, such as the Northeast-Eastern Sky coordinate system. Thus, each reflection point is assigned a set of three-dimensional spatial coordinates (X, Y, Z) with global geographic reference significance. Finally, all reflection points collected in a single flight operation are gathered together to generate three-dimensional point cloud data for describing the three-dimensional geometric morphology and internal spatial distribution of the crop canopy surface.

[0030] The three types of heterogeneous raw data that are strictly synchronized in time and space and complementary in information dimensions constitute real-time multi-source sensing data. This enables the simultaneous acquisition of two-dimensional spectral reflectance information (i.e., image data), continuous and fine spectral information (i.e., spectral data), and three-dimensional structural information (i.e., three-dimensional point cloud data) of the target crop area through a single synchronous scan. This provides a data foundation with rich spectral features, accurate material identification capabilities, and three-dimensional spatial context for subsequent analysis. It overcomes the limitation of the information dimension of a single sensor and lays a data foundation for the subsequent high-precision and high-reliability interpretation of crop physiological state.

[0031] Step S12 involves extracting and fusing features from real-time multi-source sensor data to obtain multi-dimensional feature vectors characterizing crop canopy nutrients, water stress, and canopy geometry.

[0032] For image data, spectral features are extracted using a predefined vegetation index formula. For example, the reflectance values ​​of the near-infrared and red bands are substituted into the vegetation index formula (near-infrared reflectance - red reflectance) / (near-infrared reflectance + red reflectance) to calculate the normalized difference vegetation index value for each pixel, which is used to characterize chlorophyll content and photosynthetic activity. Specific band combinations are then substituted into the photochemical reflectance index formula, such as... This yields an index value reflecting photosynthetic efficiency and water stress status. These calculated exponential plots constitute the first spectral feature. Among them, This indicates a specific wavelength band of 531 nanometers. This indicates a specific wavelength band of 570 nanometers.

[0033] For spectral data, a continuous projection algorithm is used. This algorithm iteratively selects the bands from the full-band spectrum that are most correlated with the target agronomic parameters, such as leaf nitrogen content and water content, and have the lowest redundancy among them. Alternatively, it directly calculates the band ratio of sensitive bands as features and calculates the first or second derivative of the spectral curves of these bands to enhance the sensitivity to early stresses, thereby extracting the second spectral features.

[0034] For 3D point cloud data, it is first layered vertically along the Z-axis by height, and the density of 3D point cloud data within each height layer is statistically analyzed to obtain the characteristics representing the vertical distribution of leaf area density. The surface roughness of the canopy top is quantified by calculating the elevation standard deviation of the canopy top surface model fitted by the 3D point cloud data. The space containing the 3D point cloud data is divided into regular 3D voxel grids, and the binary state of whether there are points in each voxel grid is calculated. Then, the proportion of empty voxel grids in the entire canopy space is statistically analyzed as the porosity. These statistics together constitute the 3D geometric structural features describing the canopy height, vertical structural complexity, and permeability.

[0035] After completing the feature extraction, feature fusion is performed: First, using the geographic location information and timestamps recorded synchronously during the acquisition of image data, spectral data, and 3D point cloud data, the first spectral feature, the second spectral feature, and the 3D geometric structure feature are mapped to a unified geographic coordinate system, such as the same spatial grid under WGS-84, to achieve strict spatial registration and alignment at the pixel level and the point cloud level. Then, for the first spectral feature, the second spectral feature, and the 3D geometric structure feature from different sensors, with different dimensions and numerical ranges, methods such as max-min normalization are used to transform them to a unified numerical range, such as [0, 1]. Finally, the first spectral feature, the second spectral feature, and the three-dimensional geometric structure feature corresponding to a spatial grid, which have been spatially aligned and numerically normalized, are connected in order of their physical meaning to form a multimodal feature vector. This multimodal feature vector simultaneously contains nutrient information, water stress information, and three-dimensional geometric structure information of the crop canopy. It overcomes the limitations and ambiguities of a single data source, such as relying solely on the spectrum, which is easily affected by canopy structure interference. It generates a multidimensional feature expression that characterizes the comprehensive physiological and ecological state of the crop, providing a reliable and information-rich input for subsequent accurate state interpretation and decision-making.

[0036] Step S13: Based on the real-time flight condition parameters and real-time environmental parameters of the UAV, perform feature weighting and information compensation on the multi-dimensional feature vector to obtain the enhanced feature vector.

[0037] First, two key parameter sets are acquired and input in real time: one is the real-time flight condition parameters of the UAV, mainly including the current flight altitude, flight speed, pitch angle, and roll angle; the other is the real-time environmental parameters, mainly the current solar altitude angle and light intensity. Next, the multi-dimensional feature vector is processed based on the real-time flight condition parameters and the real-time environmental parameters.

[0038] For feature weighting, the sensitivity of the first spectral feature, the second spectral feature, and the three-dimensional geometric feature to the current flight conditions and environmental interference is evaluated according to a set of preset rules. A dynamic weight is calculated for each of the first spectral feature, the second spectral feature, and the three-dimensional geometric feature. For example, when the light intensity is significantly reduced, the reliability of the first spectral feature and the second spectral feature based on optical reflection will decrease, and their weights will be adjusted accordingly. When the UAV attitude angle changes significantly, the reliability of the three-dimensional geometric feature may be reduced due to the measurement of geometric deformation, and its weight will also be reduced.

[0039] For information compensation, a preset compensation model is invoked to directly perform numerical corrections on the first spectral feature, the second spectral feature, and the three-dimensional geometric structure feature. For example, based on the real-time solar altitude angle and sensor viewpoint, the reflectance values ​​of the first and second spectral features are corrected using a bidirectional reflectance distribution function to compensate for the impact of changes in illumination angle; or, based on the current flight altitude and speed, the three-dimensional geometric structure feature is normalized to eliminate data acquisition differences caused by different flight parameters. Finally, all features after dynamic weight adjustment and numerical compensation are combined to form an enhanced feature vector. This enables the enhanced feature vector to dynamically and adaptively adjust the reliability of different features and correct errors, thereby improving its robustness to complex and changing operating environments and flight states.

[0040] It should be noted that the preset rule evaluation refers to a set of if-then logic or fuzzy rules based on physical mechanisms and experimental calibration, used to dynamically determine the credibility level of each feature and assign weights to it according to real-time flight condition parameters and real-time environmental parameters; the preset compensation model refers to a series of mathematical models used to correct specific physical disturbances, such as the BRDF model and the point cloud density height normalization model, which quantitatively compensates the disturbed features through calculation.

[0041] Step S14: The enhanced feature vector is fused with the historical feature vector of the previous time step in the current job cycle through a sliding window in the time domain to generate a fused feature vector.

[0042] A fixed-length first-in-first-out queue is maintained as a sliding window to cache the historical feature vectors of adjacent and preceding consecutive times of the enhanced feature in the current job cycle. When the enhanced feature vector is generated, it is added to the end of the sliding window, while the oldest historical feature vector in the sliding window is removed, so that the sliding window always contains the latest feature data of several consecutive times. Next, a temporal fusion calculation is performed on the enhanced feature vectors and historical feature vectors arranged in chronological order within the sliding window. This is achieved by weighted averaging of the corresponding dimensions of each feature vector within the sliding window. The weights can be allocated based on temporal proximity, such as the highest weight at the current moment, decreasing as time goes back. Alternatively, temporal convolution or recurrent neural networks can be used to encode the window sequence to extract the evolution patterns and dependencies of the feature vectors in the temporal dimension. Through this calculation, the sequence information of each feature dimension within the sliding window is aggregated into a new feature value. The new feature vector formed by aggregating all dimensions is the fused feature vector, which simultaneously contains spatial features and their dynamic information changing over time, i.e., it includes spatiotemporal correlation information. This allows the feature expression of the fused feature vector to not only include current spatial information but also integrate short-term historical change trends. This effectively suppresses random noise or transient interference that may exist in single-frame data, enhances the stability and continuity of the fused feature vector in the temporal dimension, and enables the fused feature vector to reflect the dynamic changes in crop status.

[0043] Step S15: Interpret the crop physiological state of the fused feature vector to obtain real-time crop state data.

[0044] The fused feature vectors are interpreted into crop physiological state parameters, providing direct and accurate input for subsequent dynamic optimization decisions based on a pre-defined digital twin model. Specifically, by "decoding" and "translating" the comprehensive crop canopy state represented by the fused feature vectors, real-time crop state data reflecting real-time nutrient and water conditions, such as relative chlorophyll content, nitrogen content index, and water stress index, are extracted from abstract features containing spatiotemporal variation patterns. This completes the conversion from real-time multi-source sensor data to real-time crop state data usable for final decision-making, providing a scientific and reliable data foundation for subsequent dynamic variable fertilization control based on future growth simulations.

[0045] Furthermore, step S15 may include steps S151 to S153: Step S151: Using a preset crop physiological model, the fused feature vector is subjected to dot product and nonlinear transformation calculations to obtain an initial crop physiological parameter set. The preset crop physiological model includes preset crop state interpretation calculation rules and preset weight matrices associated with the preset crop state interpretation calculation rules.

[0046] The preset crop physiological model in this embodiment is essentially a lightweight feedforward neural network model trained on a large amount of labeled data. The preset crop state interpretation calculation rules define the specific computational architecture of this feedforward neural network model, namely the connections and data flow order between the input layer, several hidden layers, and the output layer. The preset weight matrix corresponds to the set of connection weight parameters between neurons in each layer of the network. These weight parameters are core knowledge learned by the model through training, capable of characterizing the complex mapping relationship between fused features and crop physiological parameters. During computation, the fused feature vector containing spatiotemporal correlation information is used as input. It is first multiplied by the preset weight matrix of the first hidden layer to obtain a linear combination result. Then, a nonlinear transformation is immediately applied to this linear combination result, for example, through the ReLU activation function, introducing nonlinearity so that the preset crop physiological model can fit complex nonlinear relationships.

[0047] The aforementioned calculation process of the dot product followed by a nonlinear transformation is performed sequentially in multiple hidden layers. Each layer uses its corresponding preset weight matrix and activation function to progressively abstract and transform the fused feature vector at a higher level. Finally, the output layer maps the final calculation result to an initial set of crop physiological parameters, which may directly correspond to the output values ​​of physiological indicators such as chlorophyll content and nitrogen content.

[0048] Step S152: Perform time-series correlation correction between the initial crop physiological parameter set and the historical crop physiological parameter set of the previous time in the current operation cycle to obtain the target crop physiological parameter set.

[0049] Maintain a cache that stores several historical crop physiological parameter sets from previous time points in the current job cycle. When the initial crop physiological parameter set is generated, it is combined with the historical crop physiological parameter sets in the cache in chronological order to form a short time series. Then, a time series filtering or smoothing algorithm is applied to process this short time series, such as using a first-order low-pass filter or a Kalman filter. The time series filtering or smoothing algorithm will perform optimal estimation of the initial crop physiological parameter set at the current time based on a preset dynamic model and the estimation of observation noise to obtain an estimated value. This estimated value is then fused with the predicted value based on the historical crop physiological parameter sets to output a corrected value.

[0050] Specifically, the preset dynamic model used is a mathematical description of the variation law of crop physiological parameters in adjacent short time intervals. It can usually be simplified into a state transition equation. For example, it is assumed that in a very short operating cycle, the physiological parameters of crops such as chlorophyll content and nitrogen content mainly show slow continuous changes rather than abrupt changes. Its mathematical form can be expressed as the parameter state at the current moment is roughly equal to the parameter state at the previous moment plus a small perturbation (process noise) that conforms to a Gaussian distribution. The observation noise is a statistical estimate of the random error contained in the output of the initial crop physiological parameter set of the preset crop physiological model in step S151. It quantifies the uncertainty of the output estimate. A specific implementation example is as follows: Using a Kalman filter, based on the corrected historical target crop physiological parameter set from the previous time step (i.e., the historical crop physiological parameter set), a pre-defined kinetic model (defined by the state transition matrix and process noise covariance matrix) is used to predict the current value of the parameter set. Subsequently, a time-series filtering or smoothing algorithm compares the crop physiological parameter values ​​actually observed in the initial crop physiological parameter set at the current time with the predicted values, and calculates a Kalman gain based on the process noise covariance of the pre-defined kinetic model and the pre-defined observation noise covariance. Finally, this Kalman gain is used to weightedly fuse the predicted value and the crop physiological parameter value, i.e., the corrected value = predicted value + Kalman gain × (crop physiological parameter value - predicted value), thus outputting the optimal estimate of the parameter set at the current time step, i.e., the target crop physiological parameter set. The core of this process is that the algorithm trusts the trend predicted by the kinetic model, but corrects it with actual crop physiological parameter values, and the strength of the correction (Kalman gain) is dynamically determined by the uncertainty of the model prediction (process noise) and the unreliability of the observation (observation noise). The main effect of this approach is that, through the Bayesian optimal estimation principle, random noise and gross errors that do not conform to the physiological change pattern in the initial set of crop physiological parameters are effectively filtered out, making the output set of target crop physiological parameters smoother, more continuous, and more reliable in the time series.

[0051] It should be noted that the specific operation of predicting the current value of the parameter set using the preset dynamic model is as follows: The core of the preset dynamic model is defined by the state transition matrix F and the process noise covariance matrix Q. The state transition matrix F (usually set as an identity matrix or a matrix close to an identity matrix) describes the ideal evolution relationship of the crop physiological parameter set from the previous time to the current time. Its mathematical form is x_pred=F×x_est, where x_est represents the historical crop physiological parameter set. The predicted value x_pred of the current state is obtained through this matrix multiplication. At the same time, the process noise covariance matrix Q is used to model the uncertainty in this prediction process. It represents the error introduced by unmodeled factors (such as short-term microclimate fluctuations, small random changes in crop physiology, etc.) in addition to the known state transition relationship. In the prediction step, the uncertainty of the system state estimate (represented by the error covariance matrix P) is also updated according to this model. Its update formula is P_pred=F×P_est×F^T+Q, where P_est is the error covariance of the state estimate at the previous time, and P_pred is the uncertainty of the predicted value at the current time.

[0052] Step S153: Based on the preset calibration mapping relationship, the target crop physiological parameter set is converted into real-time crop status data.

[0053] It should be noted that the preset calibration mapping relationship is a set of mathematical transformation functions or lookup tables established in advance through a large amount of field sampling and analysis. This preset calibration mapping relationship establishes a definite correspondence between the target crop physiological parameter set and real-time crop status data with clear agronomic and physical significance, such as the relative chlorophyll content SPAD value, nitrogen content index NNI, and water stress index CWSI.

[0054] Each parameter value in the target crop physiological parameter set is input into its corresponding pre-calibrated mathematical transformation function, such as a linear or nonlinear regression equation or a pre-defined lookup table, through regression analysis. This directly maps the values ​​to obtain the final physical quantities that can be directly used for agronomic interpretation. For example, a parameter value can be converted into relative chlorophyll content through a linear equation y=ax+b. The set of specific indicators such as relative chlorophyll content, nitrogen content index, and water stress index obtained after converting all parameter values ​​constitutes the final output real-time crop status data.

[0055] In one feasible implementation, refer to Figure 3 As shown, step S20 may specifically include steps S21 to S23: Step S21: Using a preset digital twin model, based on the input real-time crop status data, simulate the predicted growth status of the target crop area under at least two different fertilization schemes in the future time period.

[0056] By invoking a pre-defined digital twin model, the real-time crop status data depicting the actual regional conditions is used as the starting point for the simulation. Then, at least two fertilization schemes differing in fertilization amount or mode are loaded in parallel into this pre-defined digital twin model. Combined with future weather forecasts, the embedded crop growth kinetics equation and soil nutrient transport equation are coupled and simulated in a forward-looking, time-step manner. This calculates the predicted growth status of the crop population in the future, resulting from each fertilization scheme. This transforms the fertilization scheme from a static analysis of the current situation to a quantitative projection and comparison of the future dynamic succession trajectory under different fertilization schemes. Thus, before the actual fertilization scheme is implemented, the pre-defined digital twin model can scientifically predict and select the fertilization scheme most conducive to achieving long-term agronomic goals, fundamentally endowing variable fertilization control with forward-looking and global optimization capabilities.

[0057] Furthermore, step S21 may include steps S211 to S212: Step S211: Based on real-time crop status data, initialize the crop growth status and soil nutrient status of the target crop area in the preset digital twin model.

[0058] The core of the initialization operation is to establish a precise mapping between the virtual farmland test field in the preset digital twin model and the real physical world, analyze the input real-time crop status data, and map the real-time crop status data to the initial values ​​defined by the state variables inside the preset digital twin model based on the preset data conversion relationship established through agronomic experiments. For example, the relative chlorophyll content and nitrogen content index are converted into the initial growth status parameters such as leaf nitrogen concentration and biomass defined by the preset digital twin model through empirical formulas or lookup tables. At the same time, the initial water stress level of the crop is set by combining the water stress index. For the initialization of soil nutrient status, the preset digital twin model can back-calculate or assimilate the initial concentration level of effective nutrients such as nitrogen in the current root zone soil based on the crop's nitrogen content index and historical fertilization records, which serves as the starting point for soil nutrient dynamics simulation.

[0059] Through this series of transformations and assignments, the growth status parameters of the virtual crops representing the target crop area and the nutrient status parameters of the virtual soil in the preset digital twin model are set to values ​​consistent with the current state in the real world.

[0060] Step S212: Based on the spatiotemporal distribution of fertilizer application for each fertilization scheme, combined with meteorological forecast data for future periods, a time-step simulation is performed by coupling the crop growth kinetic equation and soil nutrient transport equation in the preset digital twin model to obtain the predicted growth status.

[0061] Starting with the initialized crop growth and soil nutrient states, the simulation iterates forward at fixed time steps (e.g., 1 hour) within a set future time period (e.g., the next 7 days). Within each time step, the simulation first determines whether a fertilization event occurs at a specific location in the target crop area based on the spatiotemporal distribution data of the current fertilization plan. If so, the fertilization amount is added as a source term to the soil nutrient transport equation, updating the available nutrient content in the soil in real time. Simultaneously, meteorological forecast data for the future time period, including solar radiation, temperature, precipitation, and humidity at the corresponding time points, is input into the crop growth kinetics equation to calculate the potential photosynthetic rate of the crop. The model first calculates the crop growth rate, respiration consumption, and transpiration demand. Then, it couples and solves the crop growth kinetics equation with the soil nutrient transport equation. The soil nutrient transport equation calculates the transformation, fixation, leaching, and transport of nutrients to the root surface in the soil, thereby determining the effective nutrient concentration that the crop roots can absorb. The crop growth kinetics equation, based on the meteorological conditions at the corresponding time point, the crop's own state, and the nutrient availability calculated from the soil equation, calculates the crop's dry matter growth, nitrogen uptake, and distribution in various organs at that time point using a dry matter accumulation model based on light energy utilization efficiency and an absorption and distribution model based on nutrient concentration. It also updates crop growth state variables such as leaf area index, aboveground biomass, and plant nitrogen concentration. After completing the calculation for one time point, all updated crop state data are used as the initial conditions for the next time point, and the above coupled calculation process is repeated until the entire future period is covered. Finally, the model outputs the crop growth state at each time point in the future period under the corresponding fertilization scheme; this state sequence is the predicted growth state.

[0062] It should be noted that the crop growth kinetic equation is a mathematical model describing the mechanism of physiological processes such as crop dry matter accumulation, organ formation and yield formation as they change with time and environmental conditions; the soil nutrient transport equation is a mathematical model describing the mechanism of physicochemical processes such as the transformation, adsorption, desorption, diffusion and mass flow of nutrients in the soil as they change with time and space.

[0063] Step S22: Calculate the degree of deviation between each predicted growth state and the preset predicted growth state, and take the fertilization scheme corresponding to the degree of deviation being less than the preset degree of deviation as the target fertilization scheme.

[0064] It should be noted that the preset predicted growth state is a standard state sequence at multiple time points within the same future period, representing the ideal or expected crop growth trajectory.

[0065] For each fertilization scheme obtained from the simulation, the predicted growth state is calculated using a distance metric algorithm to determine the degree of deviation from the preset predicted growth state. Specifically, a dynamic time warping algorithm is used. This dynamic time rule algorithm can effectively align and compare two states that may have local scaling on the time axis. The overall difference is quantified by calculating the sum of the Euclidean distances under the optimal matching path between the two states. The sum of the Euclidean distances is the degree of deviation.

[0066] After calculating the deviation of all fertilization schemes, the fertilization scheme with the smallest deviation is selected as the target fertilization scheme. The preset deviation in this step is the second smallest deviation among all deviations.

[0067] Step S23: Based on the UAV's operating parameters and fertilizer container status, adjust the fertilizer application amount and flight path of the target fertilization plan to generate dynamic variable fertilization control instructions.

[0068] Based on the real-time operational parameters of the drone, such as flight speed, operational width, current hovering position, and fertilizer tank status (i.e., the total amount of fertilizer remaining), the target fertilization plan is adjusted.

[0069] First, a hard constraint is set based on the total amount of remaining fertilizer. The theoretical fertilizer demand is then globally optimized and adjusted to approximate the optimal distribution as closely as possible without exceeding the total amount of remaining fertilizer, resulting in an optimized fertilizer demand distribution. Then, based on this optimized fertilizer demand distribution and the current hovering position and performance of the UAV, an optimized flight path is planned that covers the entire area, can be executed precisely according to the adjusted demand, and has the shortest total operation time. This ensures that the target fertilization plan will not be interrupted or distorted by the fertilizer quantity limitations and UAV performance limitations in actual operations, but can be transformed into a series of dynamic variable fertilization control commands that can be directly executed by the UAV.

[0070] Furthermore, step S23 may include steps S231 to S233: Step S231: Based on the initial fertilization requirements of each zone within the target crop area in the target fertilization plan and the total amount of remaining fertilizer in the fertilizer tank, calculate the total fertilization constraint.

[0071] Obtain the initial fertilizer requirements for each zone within the target area as defined in the target fertilization plan. Then, sum the initial requirements for all zones to obtain the total theoretical fertilizer requirement. Compare the remaining fertilizer quantity with the total theoretical fertilizer requirement and set the total fertilizer requirement constraint for the current work cycle, which is the upper limit of the total fertilizer quantity allowed by the system.

[0072] It should be noted that comparing the remaining total fertilizer with the theoretical total fertilizer requirement, if the comparison shows that the remaining total fertilizer is greater than or equal to the theoretical total fertilizer requirement, it means that fertilizer resources are sufficient. In this case, the theoretical total fertilizer requirement can be directly used as the actual executable total fertilizer requirement constraint, and no further reduction or adjustment of the requirement is required. Conversely, if the remaining total fertilizer is less than the theoretical total fertilizer requirement, it means that fertilizer resources are insufficient. In this case, the remaining total fertilizer must be set as the total fertilizer requirement constraint, and the initial fertilizer requirement of each zone must be reduced and redistributed in subsequent steps.

[0073] Step S232: Based on the total fertilizer application constraint, adjust the initial fertilizer application requirements of each zone to obtain the optimized fertilizer application requirement distribution.

[0074] The initial fertilizer demand of each partition is used as the baseline distribution, i.e., the initial fertilizer demand distribution, and the total fertilizer application constraint is defined as the upper limit of the sum of the adjusted fertilizer demands of each partition. To achieve the adjustment, a constrained optimization problem is constructed, with the decision variable being the adjusted first fertilizer demand distribution of each partition, and the objective function being to minimize the overall difference between the first fertilizer demand distribution and the initial fertilizer demand distribution. For example, the least squares criterion can be used, i.e., minimizing the sum of squares of the differences in fertilizer demands before and after adjustment for each partition. The constraint is that the adjusted fertilizer demands of each partition are all non-negative, and their sum does not exceed the total fertilizer application constraint.

[0075] Solving the above optimization problem involves the following steps: Suppose the target crop region is divided into n partitions, and the initial fertilizer requirement distribution for each partition is a known positive number di (i = 1, 2, ..., n). The decision variable is the adjusted first fertilizer requirement distribution xi for each partition. The objective function is to make the first fertilizer requirement distribution xi as close as possible to the initial fertilizer requirement distribution di. To achieve this, the least squares criterion is used, i.e., minimizing the objective function. There are two constraints: first, the sum of all first-order fertilizer demand distributions xi does not exceed the total fertilizer demand constraint C; second, all first-order fertilizer demand distributions xi must be non-negative. Next, the Lagrange multiplier method is used to construct the Lagrange function: in, It is a Lagrange multiplier for total fertilizer application constraints, requiring ≥0; It is the Lagrange multiplier of the non-negative constraint of the i-th partition, requiring ≥0. Then, by taking the partial derivatives with respect to each variable and setting them to zero, and combining this with the Kuhn-Tucker conditions (KKT conditions), the solution is obtained from... And combined with complementary relaxation conditions and The structure of the optimal solution can be derived. In the optimal case, if the total fertilizer application constraint has no effect (i.e., ∑di itself is ≤ C), then λ = 0, and the optimal solution is xi = di; if the total fertilizer application constraint has an effect (i.e., ∑di > C), then λ > 0, and the optimal solution xi is the initial fertilizer demand distribution di minus a constant term related to λ ( However, it is necessary to simultaneously satisfy xi ≥ 0. This requires searching for a suitable λ value using a bisection method, such that all calculated first fertilizer demand distributions xi are non-negative and their sum is exactly equal to the total fertilizer constraint C. For partitions where the calculated value may be negative, their first fertilizer demand distributions xi will be set to 0 and removed from subsequent calculations accordingly, resulting in an analytical trial solution. Then, it is checked whether the trial solution satisfies the constraints. If there is a violation, the working set is updated, for example, by adding a variable with a negative value to the set that is fixed at zero, or by relaxing a constraint that should be strictly less than the total fertilizer constraint C. Then, the solution is resolved under the new working set. This process is iterated until an optimal solution that satisfies all constraints and meets the KKT conditions is found.

[0076] Finally, the set of the first fertilizer demand distributions xi obtained through the above solution is the optimal fertilizer demand distribution that satisfies the total fertilizer application constraint and has the smallest difference from the initial distribution.

[0077] It should be noted that, This means that the total amount of fertilizer applied must either be used up exactly or have no effect. This indicates that for each partition, its fertilizer requirement is either positive or reduced to zero.

[0078] Step S233: Based on the optimized fertilizer demand distribution, as well as the flight speed, operating width, and current hovering position, adjust the flight operation path to obtain the optimized flight operation path. The dynamic variable fertilizer control command includes the optimized fertilizer demand distribution and the optimized flight operation path.

[0079] The target crop area is divided into a series of zones that require differentiated fertilization based on the distribution of optimized fertilization demand. Each zone corresponds to a specific fertilization demand. A sequence and trajectory are planned that starts from the current hovering position of the drone, flies over and covers all zones in turn.

[0080] Specifically, the process is as follows: In the first stage, starting from the current hovering position, the center points or entry points of all partitions are designated as essential nodes. At the cost of the time the UAV spends moving between nodes (calculated by dividing the distance between nodes by the flight speed), a minimum insertion method is used. Starting with an initial sub-path consisting only of the starting point and the node closest to it, the process iteratively selects a node from the remaining unvisited nodes and attempts to insert it into all possible positions within the current sub-path. Specifically, for each pair of adjacent nodes, the total path time increment caused by each insertion is calculated, and the insertion position with the smallest increment is selected. This process is repeated until all nodes are inserted, ultimately forming a shortest Hamiltonian path that visits all nodes, thus determining the globally optimal access sequence for each partition. In the second stage, within each partition, a bow-shaped reciprocating filling path is generated based on the partition's shape and width. Finally, the globally optimal access sequence obtained in the first stage is seamlessly connected to the filling paths within each partition in the second stage, generating smooth turning and transfer trajectories between the connection points. This results in an optimized flight path that starts from the current hovering position, covers all partitions, and has the shortest overall time.

[0081] At this point, the optimized flight path and the optimized fertilizer demand distribution are combined to obtain dynamic variable fertilizer control instructions.

[0082] In one feasible implementation, refer to Figure 4 As shown, step S30 may specifically include steps S31 to S34: Step S31: The optimized flight operation path is calculated into real-time control signals for the UAV in terms of horizontal position and flight speed. The UAV is then driven to perform the current flight operation along the optimized flight operation path using the real-time control signals.

[0083] The optimized flight path is input into the UAV's flight controller. During flight, the flight controller acquires the UAV's GPS position and inertial measurement unit attitude data in real time to calculate the lateral and heading deviations between the UAV's current flight position and the optimized flight path. Based on these deviations, the flight controller calculates the lateral and velocity control quantities required to eliminate the deviations in real time.

[0084] Specifically, lateral control aims to bring the UAV closer to the target path, while speed control ensures that the UAV flies at the speed set at the waypoint. These continuously calculated desired roll angles and desired speeds, along with flight altitude commands, constitute a set of real-time control signals, which are sent to the UAV's underlying flight control system. The underlying flight control system dynamically adjusts the UAV's attitude and thrust by regulating the speed of each motor, thereby precisely driving the UAV to fly along the optimized flight path.

[0085] Step S32: Based on the optimized distribution of fertilizer demand and the real-time flight position and speed of the current flight operation, calculate the real-time spraying flow rate of the fertilizer nozzle unit on the UAV.

[0086] As the drone flies along its optimized flight path, its global navigation satellite system continuously provides real-time flight position with centimeter-level accuracy, while its inertial measurement unit provides real-time flight speed. Based on the real-time flight position, the fertilizer demand corresponding to the zone directly below the drone is obtained by querying the optimized fertilizer demand distribution. Subsequently, the formula is used: Real-time spray flow rate = (Fertilizer demand × Real-time flight speed × Spraying operation width) / Fertilizer solution concentration. The product of the fertilizer demand and the real-time flight speed, multiplied by the spraying width, yields the amount of fertilizer that the drone needs to cover per unit area of ​​ground per time at the current zone and flight speed. Finally, dividing the fertilizer mass by the pre-configured fertilizer stock solution concentration converts it into the real-time spray flow rate that the fertilizer nozzle unit needs to output.

[0087] In this embodiment, the above calculation process is repeated in each control cycle to generate a series of continuously changing spray flow commands that are matched in real time with spatial position and flight speed.

[0088] Step S33: Based on the real-time spray flow rate, generate spray control signals for the valve opening and liquid pump speed on the fertilizer nozzle unit, so as to drive the fertilizer nozzle unit to perform spraying operation based on the spray control signals.

[0089] Using the real-time spray flow rate as the setpoint for the current control cycle, and based on a mapping data table describing the correspondence between "flow rate - valve opening - pump speed" established through pre-experimental calibration, this setpoint is simultaneously converted into two parallel control commands: one for the valve opening percentage of the proportional control valve, and the other for the liquid pump speed driving the liquid pump. Then, the spray controller generates corresponding spray control signals. For the valve, this typically involves generating a pulse width modulation signal to control its opening; for the liquid pump, it generates an analog voltage signal or a specific speed control signal to set the liquid pump speed.

[0090] The spraying control signal generated above is sent to the valve actuator and liquid pump motor on the fertilizer nozzle unit through the electrical circuit, driving them to act immediately, so that the actual liquid flow rate output by the fertilizer nozzle unit is close to the set real-time spraying flow rate value.

[0091] Step S34: Obtain the measured spray flow rate of the spraying operation, compare the measured spray flow rate with the real-time spray flow rate, obtain the real-time flow error, and adjust the spraying control signal in real time according to the real-time flow error to optimize the spraying operation.

[0092] By installing electromagnetic flow meters or turbine flow meters on the fertilizer pipeline, the actual volume or mass flow rate of the fertilizer nozzle unit is collected in real time, i.e., the measured spray flow rate. The measured spray flow rate is fed back to the spray controller. In each control cycle, the spray controller algebraically subtracts the measured spray flow rate from the real-time spray flow rate to obtain a real-time flow error with a positive or negative sign.

[0093] Next, the real-time flow error is read by the PID controller, and based on the preset proportional coefficient Kp, integral coefficient Ki, and derivative coefficient Kd, the proportional term of the real-time flow error, the integral term of the accumulated error, and the derivative term of the error change are calculated respectively. The sum of the three terms is used as the control output increment. Specifically: the proportional term is calculated by directly multiplying the real-time flow error e(t) by the proportional coefficient Kp, i.e., proportional term P_out = Kp × e(t); the integral term is calculated by first accumulating the historical flow error, obtaining the accumulated historical flow error, and then adding the real-time flow error e(t) to the accumulated historical flow error Sum_e_prev to obtain the new error accumulation. The cumulative error Sum_e_new = Sum_e_prev + e(t) is calculated. This new cumulative error Sum_e_new is then multiplied by the integral coefficient Ki, resulting in the integral term I_out = Ki × Sum_e_new. The differential term is calculated by first determining the rate of change of the flow error: subtracting the flow error e(t-1) from the real-time flow error e(t) from the flow error e(t-1) of the previous control cycle, and then dividing by the control cycle length Δt to obtain the rate of change of error [e(t) - e(t-1)] / Δt. This rate of change of error is then multiplied by the differential coefficient Kd, resulting in the differential term D_out = Kd × [e(t) - e(t-1)] / Δt. Finally, the proportional term P_out, the integral term I_out, and the differential term D_out are added together to output the control output increment.

[0094] The calculated control output increment is superimposed on the spraying control signal in step S33 to generate the adjusted spraying control signal. For example, when the measured spraying flow rate is lower than the real-time spraying flow rate, the real-time flow error is positive. The control output increment of the PID controller will increase the spraying control signal, thereby instructing the valve opening to increase or the liquid pump speed to increase the flow rate, forming a real-time closed loop of "measurement-comparison-calculation-adjustment-remeasurement". This achieves continuous monitoring and dynamic compensation of the measured spraying flow rate, thereby solving the deviation of the measured spraying flow rate caused by factors such as pipeline pressure fluctuations, slight nozzle blockage, changes in liquid viscosity, or actuator response lag, and ensuring the accuracy and stability of the fertilizer flow rate actually output by the fertilizer spraying head unit.

[0095] This application also provides an AI-based drone precision variable fertilization system. Referring to FIG5, the AI-based drone precision variable fertilization system includes: The data acquisition module is used to obtain real-time crop status data of the target crop area based on real-time multi-source sensor data acquired by the UAV during flight operations. The data calculation module is used to input real-time crop status data into a preset digital twin model for calculation, and obtain dynamic variable fertilization control instructions for crop growth in future time periods. The operation control module is used to control the drone to perform fertilization operations on the target crop area based on dynamic variable fertilization control instructions.

[0096] Optionally, the data acquisition module is also used for: The system simultaneously acquires image data, spectral data, and three-dimensional point cloud data of the target crop area through a multispectral imaging unit, a hyperspectral imaging unit, and a millimeter-wave radar unit, respectively, to form real-time multi-source sensing data. Feature extraction and feature fusion are performed on real-time multi-source sensor data to obtain multi-dimensional feature vectors characterizing crop canopy nutrients, water stress, and canopy geometry. Based on the real-time flight condition parameters and real-time environmental parameters of the UAV, feature weighting and information compensation are performed on the multi-dimensional feature vector to obtain an enhanced feature vector; The enhanced feature vector is fused with the historical feature vector from the previous time step in the current job cycle using a sliding window in the time domain to generate a fused feature vector. The fused feature vectors are interpreted to obtain real-time crop status data.

[0097] Optionally, the data acquisition module is also used for: By using a pre-defined crop physiological model, dot product and nonlinear transformation calculations are performed on the fused feature vectors to obtain an initial set of crop physiological parameters. The pre-defined crop physiological model includes pre-defined crop state interpretation calculation rules and a pre-defined weight matrix associated with the pre-defined crop state interpretation calculation rules. The initial crop physiological parameter set is temporally correlated with the historical crop physiological parameter set of the previous time in the current operation cycle to obtain the target crop physiological parameter set; Based on the preset calibration mapping relationship, the target crop physiological parameter set is converted into real-time crop status data.

[0098] Optionally, the data calculation module is also used for: By using a pre-set digital twin model, based on the input real-time crop status data, the predicted growth status of the target crop area under at least two different fertilization schemes is simulated in the future time period. Calculate the degree of deviation between each predicted growth state and the preset predicted growth state, and take the fertilization scheme corresponding to the degree of deviation being less than the preset deviation as the target fertilization scheme; Based on the drone's operational parameters and fertilizer container status, the fertilizer application rate and flight path of the target fertilization plan are adjusted to generate dynamic variable fertilization control commands.

[0099] Optionally, the data calculation module is also used for: Based on real-time crop status data, initialize the crop growth status and soil nutrient status of the target crop area in the preset digital twin model; Based on the spatiotemporal distribution of fertilizer application for each fertilization scheme, combined with meteorological forecast data for future periods, the predicted growth status is obtained by coupling the crop growth kinetic equation and soil nutrient transport equation in the preset digital twin model through time-step simulation.

[0100] Optionally, the data calculation module is also used for: Based on the initial fertilizer requirements of each zone within the target crop area in the target fertilization plan, and the total amount of fertilizer remaining in the fertilizer tank, the total fertilizer constraint is calculated. Based on the total fertilizer application constraint, the initial fertilizer application requirements of each zone are adjusted to obtain an optimized fertilizer application requirement distribution. Based on the distribution of fertilizer demand, as well as the flight speed, operating width, and current hovering position, the flight operation path is adjusted to obtain the optimized flight operation path; Among them, the dynamic variable fertilization control instructions include optimizing the distribution of fertilization demand and optimizing flight operation paths.

[0101] Optionally, the operation control module is also used for: The optimized flight path is calculated into real-time control signals for the UAV in terms of horizontal position and flight speed. The UAV is then driven to perform the current flight operation along the optimized flight path using the real-time control signals. Based on the optimized distribution of fertilizer demand, as well as the real-time flight position and speed of the current flight operation, the real-time spraying flow rate of the fertilizer nozzle unit on the UAV is calculated. Based on the real-time spray flow rate, spray control signals are generated for the valve opening and liquid pump speed on the fertilizer nozzle unit, so as to drive the fertilizer nozzle unit to perform spraying operations based on the spray control signals; The measured spray flow rate of the spraying operation is obtained, and the measured spray flow rate is compared with the real-time spray flow rate. After obtaining the real-time flow error, the spraying control signal is adjusted in real time according to the real-time flow error to optimize the spraying operation.

[0102] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A method for precise variable fertilization using drones based on artificial intelligence, characterized in that, include: Based on the real-time multi-source sensor data acquired by the UAV during flight operations, real-time crop status data of the target crop area is obtained, wherein the real-time multi-source sensor data includes image data, spectral data and three-dimensional point cloud data of the target crop area. The real-time crop status data is input into a preset digital twin model for calculation to obtain dynamic variable fertilization control instructions for crop growth in future time periods. Based on the dynamic variable fertilization control command, the drone is controlled to perform fertilization operations on the target crop area; The step of inputting the real-time crop status data into a preset digital twin model for calculation to obtain dynamic variable fertilization control instructions for crop growth in future time periods includes: Using the preset digital twin model, based on the input real-time crop status data, the predicted growth status of the target crop area under at least two different fertilization schemes is simulated in the future time period; Calculate the degree of deviation between each predicted growth state and the preset predicted growth state, and take the fertilization scheme corresponding to the degree of deviation being less than the preset degree of deviation as the target fertilization scheme; Based on the UAV's operational parameters and fertilizer tank status, the fertilizer application rate and flight path of the target fertilization plan are adjusted to generate the dynamic variable fertilization control command. The operational parameters include flight speed, operational width, and current hovering position. The total fertilizer application constraint is calculated based on the initial fertilizer requirement of each zone within the target crop area in the target fertilization plan and the total remaining fertilizer in the fertilizer tank. Based on this constraint, the initial fertilizer requirement of each zone is adjusted to obtain an optimized fertilizer requirement distribution. The flight path is then adjusted according to the optimized fertilizer requirement distribution, the flight speed, operational width, and current hovering position to obtain an optimized flight path. The dynamic variable fertilization control command includes the optimized fertilizer requirement distribution and the optimized flight path. The step of controlling the drone to perform fertilization operations on the target crop area according to the dynamic variable fertilization control command includes: The optimized flight path is calculated into a real-time control signal for the UAV in terms of horizontal position and flight speed. The UAV is then driven to perform the current flight operation along the optimized flight path using the real-time control signal. Based on the optimized fertilizer demand distribution, and the real-time flight position and real-time flight speed of the current flight operation, the real-time spraying flow rate of the fertilizer nozzle unit on the UAV is calculated. Based on the real-time spray flow rate, a spray control signal is generated for the valve opening and liquid pump speed on the fertilizer nozzle unit, so as to drive the fertilizer nozzle unit to perform spraying operation based on the spray control signal; The measured spray flow rate of the spraying operation is obtained, and the measured spray flow rate is compared with the real-time spray flow rate to obtain the real-time flow error. The spraying control signal is then adjusted in real time based on the real-time flow error to optimize the spraying operation.

2. The artificial intelligence-based drone-based precision variable fertilization method according to claim 1, characterized in that, The UAV is equipped with a multispectral imaging unit, a hyperspectral imaging unit, and a millimeter-wave radar unit. The step of obtaining real-time crop status data of the target crop area based on real-time multi-source sensor data acquired by the UAV during flight operations includes: The image data, spectral data, and three-dimensional point cloud data on the target crop area are simultaneously acquired through the multispectral imaging unit, the hyperspectral imaging unit, and the millimeter-wave radar unit, respectively, to form the real-time multi-source sensing data. Feature extraction and feature fusion are performed on the real-time multi-source sensor data to obtain a multi-dimensional feature vector characterizing crop canopy nutrients, water stress and canopy geometry. The feature fusion includes spatial alignment and numerical normalization of the first spectral feature, the second spectral feature and the three-dimensional geometric structure feature obtained by feature extraction from the real-time multi-source sensor data. Based on the real-time flight condition parameters and real-time environmental parameters of the UAV, the multi-dimensional feature vector is weighted and information compensated to obtain an enhanced feature vector. The information compensation includes numerical correction of the first spectral feature, the second spectral feature and the three-dimensional geometric structure feature by calling a preset compensation model. The preset compensation model refers to a series of mathematical models used to correct specific physical interferences, and quantitative compensation is performed on the interfered features through calculation. The enhanced feature vector is fused with the historical feature vector of the previous time step in the current work cycle through a sliding window in the time domain to generate a fused feature vector; The fused feature vector is interpreted to obtain the real-time crop status data.

3. The artificial intelligence-based drone-based precision variable fertilization method according to claim 2, characterized in that, The step of interpreting the fused feature vector to obtain the real-time crop status data includes: By using a pre-defined crop physiological model, the fused feature vectors are subjected to dot product and nonlinear transformation calculations to obtain an initial set of crop physiological parameters. The initial crop physiological parameter set is temporally correlated with the historical crop physiological parameter set of the previous time in the current operation cycle to obtain the target crop physiological parameter set. Based on a preset calibration mapping relationship, the target crop physiological parameter set is converted into real-time crop status data.

4. The artificial intelligence-based drone-based precision variable fertilization method according to claim 1, characterized in that, The step of simulating the predicted growth status of the target crop area under at least two different fertilization schemes in the future time period using the preset digital twin model based on the input real-time crop status data includes: Based on the real-time crop status data, initialize the crop growth status and soil nutrient status of the target crop area in the preset digital twin model; Based on the spatiotemporal distribution of fertilizer application for each fertilization scheme, and combined with meteorological forecast data for the future period, the predicted growth state is obtained by coupling the crop growth kinetics equation and the soil nutrient transport equation in the preset digital twin model through time-step simulation. The crop growth kinetics equation is a mathematical model describing the mechanism of changes in physiological processes such as crop dry matter accumulation, organ formation, and yield formation with time and environmental conditions, and the soil nutrient transport equation is a mathematical model describing the mechanism of changes in the physicochemical processes of nutrients in the soil with time and space.

5. An artificial intelligence-based drone precision variable fertilization system, characterized in that, include: The data acquisition module is used to obtain real-time crop status data of the target crop area based on real-time multi-source sensor data acquired by the UAV during flight operations. The real-time multi-source sensor data includes image data, spectral data and three-dimensional point cloud data of the target crop area. The data calculation module is used to input the real-time crop status data into a preset digital twin model for calculation to obtain dynamic variable fertilization control instructions for crop growth in future time periods. The operation control module is used to control the UAV to perform fertilization operations on the target crop area according to the dynamic variable fertilization control command; The step of inputting the real-time crop status data into a preset digital twin model for calculation to obtain dynamic variable fertilization control instructions for crop growth in future time periods includes: Using the preset digital twin model, based on the input real-time crop status data, the predicted growth status of the target crop area under at least two different fertilization schemes is simulated in the future time period; Calculate the degree of deviation between each predicted growth state and the preset predicted growth state, and take the fertilization scheme corresponding to the degree of deviation being less than the preset degree of deviation as the target fertilization scheme; Based on the UAV's operational parameters and fertilizer tank status, the fertilizer application rate and flight path of the target fertilization plan are adjusted to generate the dynamic variable fertilization control command. The operational parameters include flight speed, operational width, and current hovering position. The total fertilizer application constraint is calculated based on the initial fertilizer requirement of each zone within the target crop area in the target fertilization plan and the total remaining fertilizer in the fertilizer tank. Based on this constraint, the initial fertilizer requirement of each zone is adjusted to obtain an optimized fertilizer requirement distribution. The flight path is then adjusted according to the optimized fertilizer requirement distribution, the flight speed, operational width, and current hovering position to obtain an optimized flight path. The dynamic variable fertilization control command includes the optimized fertilizer requirement distribution and the optimized flight path. The step of controlling the drone to perform fertilization operations on the target crop area according to the dynamic variable fertilization control command includes: The optimized flight path is calculated into a real-time control signal for the UAV in terms of horizontal position and flight speed. The UAV is then driven to perform the current flight operation along the optimized flight path using the real-time control signal. Based on the optimized fertilizer demand distribution, and the real-time flight position and real-time flight speed of the current flight operation, the real-time spraying flow rate of the fertilizer nozzle unit on the UAV is calculated. Based on the real-time spray flow rate, a spray control signal is generated for the valve opening and liquid pump speed on the fertilizer nozzle unit, so as to drive the fertilizer nozzle unit to perform spraying operation based on the spray control signal; The measured spray flow rate of the spraying operation is obtained, and the measured spray flow rate is compared with the real-time spray flow rate to obtain the real-time flow error. The spraying control signal is then adjusted in real time based on the real-time flow error to optimize the spraying operation.

Images