A variable pesticide spraying method and system based on crop canopy recognition

Abstract

This invention relates to the field of smart agriculture technology, and in particular to a variable pesticide spraying method and system based on crop canopy recognition. The method includes the following steps: acquiring canopy structure data; obtaining sparse anchor point data based on the canopy structure data; segmenting the canopy based on the sparse anchor point data to obtain canopy segmentation data; extracting canopy features from the canopy segmentation data to obtain canopy feature data; estimating penetration demand based on the canopy feature data to obtain penetration demand measurement data; acquiring nozzle parameter data; obtaining nozzle control data based on the nozzle parameter data and penetration demand measurement data; and monitoring operations based on the nozzle control data to obtain nozzle operation data, which is used for spray uniformity feedback to obtain spray uniformity data. This invention, by integrating canopy structure recognition and spray path modeling, achieves precise matching of pesticide spraying direction, spray volume, and penetration demand, significantly improving spray uniformity and pesticide utilization efficiency.

Description

Technical Field

[0001] This invention relates to the field of smart agriculture technology, and in particular to a variable pesticide spraying method and system based on crop canopy recognition. Background Technology

[0002] With the development of intelligent agricultural machinery and green pesticide application technologies, variable-rate pesticide spraying based on the actual crop condition has become an important means of reducing pesticide use and improving pesticide utilization. Traditional variable-rate spraying methods adjust the spray volume based on two-dimensional images, crop color, or NDVI index, which makes it difficult to fully perceive the canopy spatial structure and its impact on spray penetration. This can easily lead to spraying blind spots or repeated spraying, resulting in pesticide waste and uneven crop foliage coverage. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention proposes a variable pesticide spraying method and system based on crop canopy identification, thereby resolving at least one of the aforementioned technical issues.

[0004] This application provides a variable pesticide spraying method based on crop canopy identification, comprising the following steps: Step S1: Obtain canopy structure data; perform sparse anchor point calibration based on the canopy structure data to obtain sparse anchor point data; Step S2: Perform canopy segmentation based on sparse anchor point data to obtain canopy segmentation data; extract canopy features from the canopy segmentation data to obtain canopy feature data; estimate penetration demand based on the canopy feature data to obtain penetration demand measurement data. Step S3: Obtain nozzle parameter data; generate nozzle control data based on nozzle parameter data and penetration requirement measurement data; Step S4: Perform operation monitoring based on nozzle control data to obtain nozzle operation data; perform spray uniformity feedback based on nozzle operation data to obtain spray uniformity data.

[0005] This invention constructs a canopy structure representation guided by sparse anchor points, achieving lightweight modeling and precise segmentation of canopy spatial features. This avoids the excessive reliance on computing power and equipment accuracy inherent in traditional high-density mapping, and possesses good adaptability for field deployment. The method further integrates voxel orientation sensing, local connectivity estimation, and impedance path modeling to generate physically interpretable penetration demand indicators, providing quantitative driving basis for nozzle control and significantly improving spray point matching and pesticide utilization efficiency. By dynamically generating control commands based on nozzle parameters and introducing operation monitoring and spray uniformity feedback mechanisms, a closed-loop control system is constructed. This system enables robust and adaptive variable-rate application under various canopy structures, effectively reducing pesticide waste and environmental drift risks, and is suitable for precision plant protection applications in high-value crops.

[0006] Preferably, the sparse anchor point calibration specifically involves: Multi-scale geometric detection is performed based on canopy structure data to obtain canopy geometric data; Topological bifurcation points are extracted from the canopy geometric data to obtain canopy bifurcation point data; Voxel sparsity constraints are applied to the canopy bifurcation point data to obtain canopy sparse data; Sparse anchor point data is generated by using sparse canopy data.

[0007] This invention utilizes multi-scale geometric detection to effectively identify geometric feature regions in the canopy structure that signify morphological abrupt changes or structural transitions, providing a precise input basis for topological analysis. Topological bifurcation points are extracted as key points with strong structural stability and high spatial representativeness, avoiding redundant computational problems associated with traditional regular sampling or dense point clouds. A voxel sparsity constraint mechanism compresses the data representation scale, improving computational efficiency and response speed in subsequent segmentation and modeling stages.

[0008] Preferably, the canopy segmentation specifically includes: Anchor point structure transfer data is obtained by performing anchor point structure transfer processing based on sparse anchor point data. Similarity drift clustering is performed based on the anchor point structure transfer data to obtain anchor point structure clustering data; Spray projection data is obtained by performing spray direction view projection based on anchor point structure clustering data. Collaborative sensing and segmentation of the spray projection data yields canopy segmentation data.

[0009] This invention effectively preserves the structural continuity of sparse anchor points across different spatial locations and time frames through anchor point structure transfer processing. This ensures that the clustering process is based not only on geometric location but also on dynamic behavior consistency, improving the physical stability of the segmented regions. Employing similarity-drift clustering enhances adaptability to irregular canopy morphologies and avoids misclassification in complex structures by traditional Euclidean distance-based clustering methods. Combining spray direction-view projection operations makes the segmentation results closer to the actual spray incidence path, improving spray matching efficiency from a control execution perspective. Through a collaborative perception cutting strategy, spatial structural features and direction projection information are fused to generate canopy segmentation results with semantic integrity, directional accessibility, and spray adaptability.

[0010] Preferably, the canopy feature extraction specifically involves: Hierarchical structure voxel scanning was performed on the canopy segmentation data to obtain hierarchical voxel data; The voxel orientation density is calculated based on the hierarchical voxel data to obtain the voxel orientation data; Local connectivity porosity is calculated based on voxel orientation data to obtain local connectivity porosity data; The trunk tilt direction is estimated by performing a step-by-step estimation on the hierarchical voxel data to obtain the trunk tilt direction data. Canopy morphology symmetry data is obtained by processing the canopy morphology symmetry data based on the trunk tilt direction data; By integrating voxel orientation data, local connectivity porosity data, canopy morphology symmetry data, and trunk tilt direction data, canopy feature data are obtained.

[0011] This invention achieves multi-scale structural perception of the canopy region through hierarchical structure voxel scanning, enabling refined expression of spatial configuration features at different levels of "plant-row-area". Voxel directional density calculation characterizes the spatial distribution trend of leaf arrangement direction, helping to identify regions of consistent spray direction and improving the physical fit of penetration prediction; local connectivity porosity further integrates directional consistency and connectivity structure, constructing a flow-scale expression of canopy permeability, providing strong structural driving support for variable spray control. By estimating the trunk tilt direction and processing morphological symmetry, the overall tilt state and structural equilibrium characteristics of the canopy can be captured, making the nozzle control strategy more directionally adaptable.

[0012] Preferably, the calculation of locally connected porosity is specifically as follows: Voxel principal orientations are extracted based on voxel orientation data to obtain voxel principal orientation data; Flow penetration maps are constructed based on the principal direction data of voxels to obtain flow penetration map data; Flow connectivity calculations are performed on flow permeability map data to obtain flow connectivity data; Local directional porosity is calculated based on the flow connectivity data to obtain the first local connectivity porosity data. Voxel impedance direction sensing is performed based on voxel orientation data to obtain voxel impedance data; Minimum impedance path grid fitting is performed based on voxel impedance data to obtain minimum impedance path grid data; Voxel-level flowability is estimated based on the minimum impedance path grid data to obtain voxel-level flowability data. Based on the voxel-level flowability data, the penetration-enhancing porosity is calculated to obtain the second locally connected porosity data; Spatial fusion is performed based on the first and second local connected porosity data to obtain local connected porosity data.

[0013] This invention fully integrates the spatial directionality and structural accessibility characteristics of crop canopies. Through voxel principal direction extraction and flow permeability map construction, it effectively identifies regions with consistent spray direction. Connectivity calculations assess the canopy's accessibility along the spray axis, generating a first type of porosity index with structural continuity. Simultaneously, based on minimum impedance path fitting with directional impedance sensing, it captures microchannels and path bottlenecks in different canopies, improving the accuracy of flow characterization in difficult-to-penetrate areas, thus constructing a second type of enhanced porosity index with spray adaptability. Spatial fusion using flow accessibility multipliers and path entropy suppression factors results in locally connected porosity that possesses geometric connectivity, physical accessibility, and directional stability. This significantly improves the discriminative power of penetration demand estimation and the precision of variable spray control, making it suitable for generalized deployment and dynamic response regulation of various crop canopy structures.

[0014] Preferably, spatial fusion specifically refers to: Based on the first and second local connectivity porosity data, the flow accessibility multiplier and path entropy suppression factor are calculated to obtain the flow accessibility multiplier data and path entropy suppression factor data, respectively. The first local connectivity porosity data and the second local connectivity porosity data are fused and calculated based on the flow accessibility multiplier data and the path entropy suppression factor data to obtain the local connectivity porosity data.

[0015] This invention achieves multi-dimensional physical fusion of two types of locally connected porosity data through a flow accessibility multiplier and a path entropy suppression factor, effectively avoiding structural bias and overfitting risks under a single-index-driven approach. The flow accessibility multiplier enhances the weight of regions with high consistency between structural connectivity and directional accessibility, highlighting stable channels in the main spray direction; the path entropy suppression factor evaluates the stability and concentration of the minimum impedance path grid from the perspective of information entropy, reducing the interference of path divergence or unstable regions on the overall penetration judgment. This fusion no longer relies on linear weighting or simple averaging, but generates a composite porosity index based on a physical consistency control mechanism of direction-topology-information factors, significantly improving the robustness and local resolution of spray penetration feasibility assessment.

[0016] Preferably, the penetration demand metric estimation is specifically as follows: Spray direction simulation data is obtained by simulating the spray direction based on canopy feature data. Local void flow processing is performed on the spray direction simulation data to obtain local flow data; Directional occlusion data is obtained by performing directional occlusion calculations based on local flow data. Based on the directional masking data, the spray incident loss is calculated to obtain the penetration demand measurement data.

[0017] This invention effectively improves the spatial resolution and physical interpretability of penetration requirement assessment by simulating the incident path of the spray direction within the canopy structure and combining it with voxel-by-voxel analysis of local void flow. Compared to traditional simplified judgment methods based on canopy density or height, this method, based on directional shielding calculations, can accurately characterize the degree of structural shielding along the spray path, making it particularly suitable for handling complex scenarios with trunk skewing, canopy asymmetry, or highly dense areas. By calculating spray incident loss and combining it with the interaction between droplet propagation direction and structural impedance, a near-physical model of the drug decay process is formed. The output penetration requirement measurement data not only reflects "whether spraying is possible" but also provides a quantitative decision-making basis for "how much to spray."

[0018] Preferably, the nozzle control generation specifically includes: Based on nozzle parameter data and penetration demand measurement data, spray volume-penetration coupling mapping is performed to obtain coupling mapping data; Spray delay data is obtained by simulating the spray delay based on the coupling mapping data; Spray delay data is used to predict the spray width overlap distribution to obtain nozzle control data.

[0019] This invention constructs a spray volume-penetration coupling mapping model to achieve dynamic correlation and matching between the spray volume and the penetration requirements of the target area structure, overcoming the limitation of traditional fixed spray volume strategies that cannot adapt to spatially heterogeneous canopies. Through spray delay simulation, the system fully considers the time lag between the motion response of the sprayed liquid during high-altitude application and its actual arrival time, providing precise dynamic compensation for control commands. By predicting the overlap distribution of spray width, the system can anticipate the overlap area and boundary attenuation effect of multiple nozzles in spatial projection, thereby rationally adjusting the nozzle start / stop logic and output intensity to avoid redundant coverage and missed spray risks.

[0020] Preferably, step S4 specifically includes: Sprinkler control operations are performed and monitored based on sprinkler control data to obtain sprinkler operation data. The operation coverage is calculated based on the nozzle operation data and the preset nozzle operation space data to obtain the operation coverage data; Spray uniformity data are obtained by performing nozzle conductivity inversion based on the coverage data.

[0021] This invention deeply couples nozzle control data with the actual operation process, achieving not only dynamic driving of nozzle operation but also real-time monitoring to acquire the actual operating status of the nozzles, thus constructing a closed-loop operation data acquisition path. By spatially mapping and comparing nozzle operation data with a preset operation space model, the system can calculate the actual spray coverage area and identify existing spray blind spots or overlapping areas. Compared to traditional simple statistical or indirect estimation methods, the introduced nozzle electrical conductivity inversion mechanism utilizes the physical coupling relationship between nozzle energization, conduction, and liquid spraying to construct a mapping channel between electrical parameters and atomization intensity, thereby indirectly evaluating spray uniformity in a non-invasive and low-cost manner.

[0022] Preferably, this application also provides a variable pesticide spraying system based on crop canopy recognition for performing the variable pesticide spraying method based on crop canopy recognition as described above. The variable pesticide spraying system based on crop canopy recognition includes: The sparse anchor point calibration module is used to acquire canopy structure data; based on the canopy structure data, sparse anchor points are calibrated to obtain sparse anchor point data. The canopy segmentation and spray demand analysis module is used to segment the canopy based on sparse anchor point data to obtain canopy segmentation data; extract canopy features from the canopy segmentation data to obtain canopy feature data; and estimate penetration demand based on the canopy feature data to obtain penetration demand measurement data. The variable nozzle control generation module is used to acquire nozzle parameter data; and to generate nozzle control data based on the nozzle parameter data and penetration requirement measurement data. The spray operation monitoring and uniformity feedback module is used to monitor the operation based on the nozzle control data to obtain nozzle operation data; and to provide spray uniformity feedback based on the nozzle operation data to obtain spray uniformity data.

[0023] The beneficial effects of this invention are as follows: By sparsely anchoring the canopy structure data in step S1, the system can efficiently capture the spatial distribution skeleton of plants while ensuring lightweight modeling, providing a stable reference framework for segmentation and analysis. Step S2 further completes canopy segmentation through structural transfer and perceptual projection, and extracts multi-dimensional features including voxel directional density, connectivity porosity, and trunk tilt, constructing an integrated canopy information representation of structure, flow, and morphology, providing a physically deducible basis for penetration demand measurement. In step S3, through the coupled mapping of nozzle parameters and penetration demand, not only is responsive generation of nozzle control achieved, but also the spray pattern layout is optimized through spray delay simulation and overlap prediction. Step S4 constructs a coverage inversion channel based on operational monitoring, using nozzle conductivity parameters to achieve reverse perceptual feedback on spray uniformity, forming a complete closed loop. Attached Figure Description

[0024] Other features, objects, and advantages of this application will become more apparent from the following detailed description of the non-limiting embodiments, taken with reference to the accompanying drawings: Figure 1 A flowchart illustrating the steps of a variable pesticide spraying method based on crop canopy identification is shown in one embodiment. Figure 2 A flowchart illustrating the steps of a sparse anchor point calibration method according to one embodiment is shown. Figure 3 A flowchart illustrating the steps of a canopy segmentation method according to one embodiment is shown; Figure 4 A flowchart illustrating the steps of a nozzle control generation method according to one embodiment is shown. Figure 5 A flowchart illustrating the steps of a spray operation monitoring and uniformity feedback method according to one embodiment is shown. Detailed Implementation

[0025] The technical method of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0026] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. Functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0027] It should be understood that although the terms "first," "second," etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are used merely to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0028] Please see Figures 1 to 5 This application provides a variable pesticide spraying method based on crop canopy identification, comprising the following steps: Step S1: Obtain canopy structure data; perform sparse anchor point calibration based on the canopy structure data to obtain sparse anchor point data; Specifically, 3D imaging equipment such as LiDAR and structured light 3D scanners are used to perform a full-coverage scan of the crop canopy area, collecting 3D point cloud data of the canopy. The point cloud data should include the spatial coordinates, elevation values, local density, and neighborhood spatial distribution information of each sampling point. Multi-scale geometric detection is performed on the collected raw point cloud data, calculating the rate of change of local curvature radius at each point under multiple set spatial scales. The rate of change of local curvature radius represents the magnitude of curvature change at a point at different scales, thereby filtering out areas with significant local curvature changes. These areas correspond to transitional areas between the main trunk and branches or areas with complex structural changes. For the candidate areas with significant curvature changes (greater than a preset threshold), analysis is performed based on the topological bifurcation index. Topological bifurcation refers to the number of edges directly connected to a node in a 3D topological graph, and the directional distribution characteristics of these edges in space. Specifically, the set of nearest neighbors of the node is partitioned and statistically analyzed according to spatial direction (e.g., through azimuth and elevation angles in a spherical coordinate system) to obtain the number of connected edges and their directional distribution uniformity index. A higher topological bifurcation degree indicates that the node is located at the intersection of multiple structural branches. After the bifurcation points are extracted, the entire point cloud data is voxelized, mapping the irregular point cloud to a regular 3D voxel grid structure. The sparsity index of each voxel unit is calculated, such as the proportion of non-empty voxels, and sparse voxel regions are selected based on a preset sparsity threshold (e.g., the non-empty voxel ratio is below 0.2). A sparse anchor map is constructed within the retained sparse voxels. This sparse anchor map contains the 3D spatial position, local direction vector, and topological connectivity with surrounding anchors of each anchor, thus forming complete sparse anchor data.

[0029] Step S2: Perform canopy segmentation based on sparse anchor point data to obtain canopy segmentation data; extract canopy features from the canopy segmentation data to obtain canopy feature data; estimate penetration demand based on the canopy feature data to obtain penetration demand measurement data. Specifically, the obtained sparse anchor point map undergoes structure transfer processing, which involves extending the local orientation vector information of each anchor point to its neighboring voxel units. By defining a certain spatial radius, the orientation information of the anchor points is propagated to neighboring voxels with weights, thus forming an anchor point structure tensor describing the local geometric orientation. Cluster analysis of the anchor point structure tensor is performed using a clustering method based on orientation similarity. For example, each sample itself is used as an initial window center. For the current center, neighboring samples within the radius are searched. Based on the distance between the neighboring samples and the center, weights are calculated using a kernel function (such as a Gaussian kernel) (smaller distance → larger weight). The weighted average of the neighboring samples is calculated using these weights to obtain the new center position. If the displacement between the old and new centers is less than a threshold, the center is considered converged; otherwise, the process returns to the step (calculating weights using a kernel function (such as a Gaussian kernel) based on the distance between the neighboring samples and the center) to continue iterating. Converged centers that are very close to each other are merged into a single cluster center. Each sample is assigned to its nearest (or highest weighted) cluster center, resulting in a cluster. Based on the angle difference between anchor point direction vectors or the density distribution of direction vectors on the sphere, anchor points with similar orientations are grouped into the same cluster, yielding the anchor point structure clustering result. Based on a pre-defined trajectory model for spraying operations (constructed based on empirical data or pre-defined operational parameters, such as nozzle motion control parameters (including travel speed, nozzle swing angle, spray coverage width, etc.) and operational path planning data), the above structure clustering result is projected onto a two-dimensional plane corresponding to the nozzle's field of view according to the spray incident direction. This projection process maps the three-dimensional canopy structure to the observation section in the spray direction. On the projection plane, combining the anchor point cluster density distribution with the voxel occlusion relationship, a collaborative sensing segmentation operation is performed, i.e., based on local density gradient changes and occlusion boundary detection results, the outer contour boundary of the canopy is extracted within a threshold range. The output canopy segmentation data includes the spatial extent of the canopy region and its cross-sectional distribution in the spray direction. After obtaining the canopy segmentation data, a layered voxel scan is performed on the canopy region. The canopy is divided into several height layers along the vertical direction with a preset interlayer spacing (e.g., 10 cm), and corresponding hierarchical voxel sets are generated. For each height layer, the principal direction vector distribution of all voxels within that layer is calculated. The principal direction vectors can be obtained through principal component analysis (PCA) of the voxel point cloud and encoded using a direction density histogram in a spherical coordinate system. Based on the above, the system constructs a locally connected graph guided by the principal directions of the voxels as edge weights, and calculates the local directional porosity index accordingly. The overall height distribution of the canopy is fitted to identify the central axis of the trunk, and the tilt direction of the trunk is estimated based on the fitted central axis point sequence. Using the central axis of the trunk as a reference axis of symmetry, the voxel distribution density and structural geometry on the left and right sides are compared, and the density deviation rate and morphological difference quantification value on the left and right sides are calculated to obtain the bilateral symmetry of the canopy morphology. By fusing directional density distribution data, local directional porosity data, trunk tilt angle data, and morphological symmetry data, complete canopy feature data are generated.

[0030] Based on the aforementioned canopy feature data, a spray path simulation model was constructed. This model uses the canopy trunk direction as the reference axis to simulate the propagation path of spray particles along each incident direction. For each simulated path, the connectivity of the voxels it traverses (i.e., the proportion of continuous voids traversed by the path), voxel density (the number of solid voxels per unit path length), and directional masking ratio (the masking rate corresponding to the cosine value of the angle between the voxel surface normal vector and the spray direction) were statistically analyzed. The above path information was integrated to construct a local flowability map. Based on the path masking ratio, an incident energy loss factor was calculated, which represents the degree of energy attenuation of the spray before reaching the target location. The loss factor was normalized to obtain a penetration demand metric, which was then used to generate a spray penetration demand map.

[0031] Step S3: Obtain nozzle parameter data; generate nozzle control data based on nozzle parameter data and penetration requirement measurement data; Specifically, the structural and performance parameters of the nozzle are read in real time from the spraying platform. These parameters include at least the nozzle model, atomization angle range (in °), maximum allowable flow rate (in L / min), adjustable injection pressure range (in MPa), and spray particle size distribution characteristics. The obtained penetration demand measurement data is jointly modeled with the nozzle parameters. Through a preset spray volume-penetration demand coupling mapping model, different penetration demand values ​​(dimensionless, ranging from 0 to 1) are mapped to the required spray flow rate range (in L / min). This mapping model can be obtained in two ways: first, by establishing a lookup table based on historical spraying test data, directly associating each penetration demand value with the optimal spray volume range; second, by using a nonlinear regression method (such as quadratic polynomial fitting or piecewise linear fitting) to train a continuous function relationship on the test data. Due to the mechanical response delay (on the order of milliseconds to hundreds of milliseconds) when the spraying platform performs spray volume and angle adjustments, the system constructs a nozzle delay compensation model. This model predicts the actual location area of ​​the spray particles within the delay time based on the response curve of the nozzle drive system, thereby correcting the target spray trajectory. Based on delay compensation, and combined with the movement trajectory data of the work platform and the geometric characteristics of the nozzles, the coverage distribution of the spray in the target area is estimated. Specifically, the projection range of the spray pattern on the work plane is calculated using the geometric model of the spray cone and the movement path, and the area of ​​the overlapping region between adjacent spray patterns is determined (unit: ...). The nozzle control command set is generated by combining the spray volume-penetration demand coupling results, delay compensation prediction results, and spray width overlap control results with a preset parameter table. The command set includes at least the spray start and stop time points (in milliseconds), spray volume change curves (target flow rate sequence over time), spray pressure control sequence (target pressure value changing over time), and spray width offset angle (target angle of nozzle rotation or oscillation).

[0032] Step S4: Perform operation monitoring based on nozzle control data to obtain nozzle operation data; perform spray uniformity feedback based on nozzle operation data to obtain spray uniformity data.

[0033] Specifically, during spraying operations, position and attitude sensors (such as GNSS positioning modules and IMU inertial measurement units) installed on the work platform and a nozzle execution status acquisition module are used to acquire parameters such as the nozzle's spatial position, orientation angle, spray output velocity (in L / min), and spray pressure (in MPa) in real time, recording the actual work path and the execution results of control commands. The acquired nozzle operation data is mapped to a three-dimensional spatial coordinate system corresponding to the work area and aligned with a pre-established nozzle operation spatial model. The spatial model can be divided using a regular grid (e.g., each unit area is...). The work area is discretized into multiple spatial units using a method that combines the frequency of spray coverage and the corresponding cumulative spray volume for each spatial unit. Overlap analysis is then used to generate a spray coverage distribution matrix, which is used to calculate the actual coverage frequency and spray volume distribution for each spatial unit, thus generating a spray coverage heatmap. During spraying, a conductivity-flow sensor installed in the nozzle pipeline collects the conductivity data of the sprayed liquid in real time. Based on a pre-established conductivity-liquid concentration calibration curve, the conductivity (unit: S / m) is converted into the spray concentration, and then combined with instantaneous flow rate measurements, an inversion calculation formula is used. ,in This refers to the actual sprayed dose during the operation period. This is the start time of the work period. This is the end time of the work period. For instantaneous flow, Instantaneous concentration, The time variable is used. The actual spray dose of each spatial unit is compared with the target spray dose at the corresponding position in the nozzle control data to calculate the spray uniformity error. ,in For the first The relative uniformity error of spatial units (units are dimensionless, value range 0-1). For the first The actual ejection dose of the space unit, For the first Target ejection dose of the space unit Index the spatial units. Assemble the uniformity error matrix from the error data of all spatial units, and use it as input parameters for feedback adjustment to optimize the spray volume-penetration demand mapping model and nozzle control commands. For example, if the relative uniformity error is too large, stop the operation and upload an alarm to the cloud platform.

[0034] Preferably, the sparse anchor point calibration specifically involves: Step S11: Perform multi-scale geometric detection based on the canopy structure data to obtain canopy geometric data; Specifically, the collected canopy 3D structural data (which can be lidar point cloud data or 3D image reconstruction data) undergoes voxelization processing. This voxelization process divides the continuous space into regular cubic units (voxels), for example, setting the voxel side length to 5cm, thus obtaining a spatial voxel grid structure. Each voxel contains a certain number of original spatial points. Within each voxel window, local geometric feature extraction is performed, specifically including local curvature calculation: for the point set within each voxel, principal component analysis (PCA) is used to calculate its 3D covariance matrix. Let... This is the smallest eigenvalue of the matrix. Given the sum of the eigenvalues ​​of the three principal axes, the local curvature is defined as: Normal direction consistency calculation: For the current voxel, the principal direction vector obtained by PCA is used as the normal vector. The normal vector to adjacent voxels. Calculate the included angle: ,in The angle between the normal vectors (unit: radians) is calculated by statistically analyzing the angles between adjacent voxels. The mean or standard deviation of the values ​​is used to obtain the local normal direction consistency index for the voxel. The above curvature and normal direction consistency calculations are performed not only at a single scale but also in multiple scale windows, for example... voxel window, Voxel windows, etc. Geometric features under each scale window are calculated separately and made dimensionless through normalization (e.g., mapping eigenvalues ​​to the [0,1] interval).

[0035] Step S12: Extract topological bifurcation points from the canopy geometric data to obtain canopy bifurcation point data; Specifically, based on the obtained multi-scale curvature distribution map and normal uniformity distribution map, the feature response value of each voxel unit is calculated. Voxels with large curvature values ​​(e.g., more than 1.5 times the global curvature mean) or high normal uniformity change rates are marked as local candidate points. A set of bifurcation candidate regions is formed by searching for local extrema on the feature map. Within each candidate region, a local orientation map is constructed using the three-dimensional eight-neighbor path tracing method. The system connects voxels in eight spatially adjacent directions to the candidate voxel, extracting the principal direction vector (obtained by PCA) for each adjacent voxel. The system calculates the cosine similarity between all pairwise direction vectors within the region: ,in The cosine similarity is the similarity between the direction vectors. For the first The principal direction vector of an individual element unit. For the first The principal direction vector of each voxel unit. If a candidate region contains three or more principal direction groups (for multi-scale data, the existence of a similar situation at any scale is also considered acceptable), and the cosine similarity within each group is less than 0.7 (indicating a significant angular difference between these directions), then the region is determined to be a topological bifurcation point. A spatial density threshold is set (e.g., the number of principal directions within a unit cube volume is greater than 2). If the number of directions within a candidate region is insufficient, the region is removed. The spatial coordinates, local principal direction vectors, and feature response amplitudes (e.g., curvature values ​​or normal uniformity change values) of the voxels determined to be topological bifurcation points are recorded to form canopy bifurcation point data.

[0036] Step S13: Perform voxel sparsity constraints based on the canopy bifurcation point data to obtain canopy sparse data; Specifically, the overall 3D point cloud data of the canopy is voxelized according to rules to form a voxel raster model. The voxel raster is divided into several spatial partitions, each with a side length of 10cm. 10cm 10cm (or adjusted according to the average structural size of the crop canopy). Within each spatial partition, a representative bifurcation point is selected from the identified canopy bifurcation points as a sparse candidate point. For each bifurcation point within a partition, its characteristic response amplitude is compared, where the characteristic response amplitude is the curvature value, the orientation entropy value, or a weighted combination of the two. The bifurcation point with the largest characteristic response amplitude is selected as the priority retention object. If the difference in characteristic response amplitude of multiple bifurcation points within a partition is less than a preset difference threshold (e.g., 0.05), then among these points, the bifurcation point with the smallest angle between the principal direction vector and the vertical unit vector is selected as the retention object. After the above sparsification screening rules are applied, only one representative bifurcation point from each spatial partition is retained, forming a sparse point cloud set.

[0037] Step S14: Generate sparse anchor point map data based on the sparse canopy data to obtain sparse anchor point data.

[0038] Specifically, based on the 3D coordinates of each point in the sparse point cloud, the k-nearest neighbor method is used to establish the connection relationships between anchor points. The value of k can be set according to the canopy spatial density, taking either 3 or 5. For each pair of connected anchor points, their distance is calculated as the distance weight of the edge. Simultaneously, a direction angle index is calculated to represent the angle between the anchor point direction vectors. The above edge weights and direction indices together form the edge attributes of the weighted directed graph. For each anchor point in the graph, the following local structural feature indices are calculated: average edge length: the average Euclidean distance between the anchor point and all its connected edges; principal direction vector: the direction of the first principal component obtained through principal component analysis (PCA) based on the normal vector distribution of the anchor point's neighboring points; and neighborhood sparse point density: the number of sparse points per unit volume. The 3D coordinates, local principal direction vector, structural feature indices (including average edge length and neighborhood sparse point density), and its connecting edge information (including distance weights and direction angle indices) of each anchor point are uniformly stored to construct a complete sparse anchor point graph data structure.

[0039] Preferably, the canopy segmentation specifically includes: Step S21: Perform anchor point structure transfer processing based on sparse anchor point data to obtain anchor point structure transfer data; Specifically, the system takes a 3D sparse anchor graph as input and reconstructs the local topological relationships of the anchor points. A nearest neighbor set is defined for each anchor point. ,in anchor point The nearest neighbor set, i.e., the set of neighbors whose Euclidean distance to the anchor point is less than the structural transfer radius threshold. All anchor point sets, For the 3D sparse anchor point graph, the first The three-dimensional coordinate vector of each anchor point For the current computing center The three-dimensional coordinate vector of each anchor point This is the threshold for the structural transfer radius. For each Calculate the orientation tensor of its nearest neighbor. , anchor point The direction tensor, Nearest Neighbor Set The first in The three-dimensional coordinate vector of each anchor point anchor point The nearest neighbor set, For the current computing center The three-dimensional coordinate vectors of each anchor point are obtained, and the tensor is subjected to eigenvalue decomposition to obtain the principal direction axis. Through aggregation and transfer of the principal direction axis, the anchor points are associated with an approximate local structural skeleton (i.e., the orthogonal distance and the angle of orientation are determined to be within the threshold range), and the anchor point structure transfer data is output.

[0040] Step S22: Perform similarity drift clustering based on the anchor point structure transfer data to obtain anchor point structure clustering data; Specifically, the system calculates the structural similarity between any two anchor points: ,in anchor point and Structural similarity, It is a natural exponential function. anchor point The structural feature vector, anchor point The structural feature vector, This is the similarity scaling parameter. In the feature space, it is represented by the bandwidth parameter (which can be taken as...). (Similarity) is the kernel width, and the process is iterative: calculate the weighted mean of the feature vector of each anchor point within the kernel window (the weights are determined by similarity). (Given); translate the anchor point feature vector along the density gradient direction to the weighted mean position; repeat steps 1–2 until the translation amount of all anchor points is less than the convergence threshold (e.g. The cluster center is the drift convergence point, and all anchor points that converge to the same center are classified into the same cluster.

[0041] Step S23: Perform spray direction view projection based on anchor point structure clustering data to obtain spray projection data; Specifically, for each anchor point structure cluster, a corresponding pseudo-spray direction vector is defined. This vector can be set according to the spray axis of agricultural machinery (such as sprayers or drone nozzles) and uniformly normalized to a unit vector. Using the spray direction as the normal vector, a projection plane orthogonal to it is constructed in three-dimensional space. This plane serves as the imaging plane from the spray perspective, representing the outline of the canopy under the spray viewpoint. The clustered regions are projected onto the projection plane perpendicular to the spray direction using the perspective transformation formula: ,in Let the anchor point be a two-dimensional coordinate vector on the projection plane. This is the three-dimensional coordinate vector of the anchor point (unit: meters). The reference origin on the projection plane (which can be the position of the spray nozzle or the projection point on the operation path). This represents the unit spray direction vector. After performing the above projection calculation on all anchor points within the cluster, a set of two-dimensional coordinates on the projection plane is obtained. The system discretizes this coordinate set into pixels or voxels of fixed resolution (e.g., 1 cm). (1cm grid) to generate a dense two-dimensional view map.

[0042] Step S24: Perform collaborative sensing segmentation on the spray projection data to obtain canopy segmentation data.

[0043] Specifically, the projection results from clusters of different anchor point structures are represented as a set of binary masks. ,in This represents the mask value of pixel (x, y) of the k-th cluster on the projection plane (a value of 1 indicates that the pixel is covered by this cluster, and a value of 0 indicates that it is not covered). The overlap intensity map R(x, y) is constructed, and its calculation formula is as follows: ,in How many clusters simultaneously cover the pixel (x, y)? For cluster indexing, As an indicator function, when (x,y) belongs to The value is 1 if the condition is met, and 0 otherwise. The system sets an overlap threshold (e.g., 2, meaning that only areas covered by at least two clusters are considered highly overlapping regions). When a pixel meets the overlap threshold, it is marked as a highly overlapping region, and boundary refinement processing is performed on that region, such as tracing the boundary curve pixel by pixel along the edge of the highly overlapping region; or, performing dilation-erosion operations on the mask of the highly overlapping region to fill boundary cracks and remove isolated artifact pixels. After completing the cutting on the two-dimensional projection plane, the system inversely projects the pixel coordinates of each segmented region back to the three-dimensional voxel space: based on the geometric relationship between the projection plane and the spray direction, the two-dimensional coordinates are restored to the corresponding three-dimensional spatial positions; the reconstructed three-dimensional points are mapped onto the voxel grid to form a three-dimensional canopy segmentation voxel set.

[0044] Preferably, the canopy feature extraction specifically involves: Hierarchical structure voxel scanning was performed on the canopy segmentation data to obtain hierarchical voxel data; Specifically, the obtained three-dimensional canopy segmentation region is voxelized according to a regular three-dimensional mesh. Each voxel unit is a cube with a side length ranging from 2cm to 5cm, the specific value being set according to the spatial resolution of the spraying equipment, canopy geometric details, and computational resource constraints. If a voxel unit contains at least one canopy segmentation voxel point, the unit is marked as "occupied". Using the Z-axis as the height reference, each horizontal layer of the voxel mesh (i.e., the z=k plane layer, where z is the Z-axis value, which can also be considered a horizontal layer, and k is the height data) is scanned sequentially from bottom to top. During the scanning process, adjacent occupied voxels in the same horizontal layer are identified and grouped into continuous voxel blocks; the spatial position, occupied area, and density distribution of each voxel block in the layer are recorded; the average height of the voxel blocks in the layer (the height value relative to the bottom of the entire canopy) is calculated. The system constructs a hierarchical index tree, which is implemented using an octree, voxel tree, or other hierarchical spatial index structure. Each node of the tree corresponds to a voxel unit or voxel block and records its level ID (indicating its height level), a list of adjacent voxels (including adjacent voxels in the same level and connected voxels in the upper and lower levels), and a density estimate of its level (e.g., the ratio of the number of occupied voxels to the total number of primes in that level).

[0045] The voxel orientation density is calculated based on the hierarchical voxel data to obtain the voxel orientation data; Specifically, for each voxel neighborhood (e.g., 3...) 3 3) Perform PCA (Principal Component Analysis) to extract local principal direction vectors, i.e., the eigenvectors corresponding to the largest eigenvalues. Project the direction vectors onto a spherical coordinate system, count their frequency of occurrence in different direction intervals, and construct a direction histogram. Use a direction clustering algorithm (such as spherical K-means or mean-shift clustering) to group the principal direction vectors, obtaining the typical direction set of that layer. For each voxel, the system calculates the consistency ratio between the principal direction vectors of all voxels in its neighborhood and the direction of its cluster: the consistency criterion is that the angle between the directions is less than a set threshold (e.g., 15°); the direction density value is defined as the ratio of the number of neighborhood voxels that meet the consistency condition to the total number of neighborhood voxels; use this direction density value as the direction density label of that voxel.

[0046] Local connectivity porosity is calculated based on voxel orientation data to obtain local connectivity porosity data; Specifically, based on the obtained voxel orientation data, the system selects a set of neighboring voxels (e.g., 3) for each voxel as a node. 3 (Neighborhood range of 3). For any adjacent voxels, calculate the directional consistency between their principal direction vectors: ,in voxels and The directional consistency coefficient of the principal direction vector. For the first in the local analysis window Individual factors, To be within the local analysis window and Adjacent voxels, Let cosine be the value of the two direction vectors. voxels The principal direction unit vector, voxels The principal direction unit vector. Using this consistency value as the edge weight, construct a direction weight graph. In each local analysis window (e.g., 5... 5 Within 5 voxels, perform connected component extraction based on the orientation weight graph: set an orientation consistency threshold (e.g., 0.85), retain only edges that meet the consistency requirements; count the total number of connected components formed by orientation-consistent paths within this window; count the total number of primes within this window; the first connectivity porosity is calculated as follows: ,in The first connectivity porosity represents the proportion of voxels formed by directionally consistent paths to the total number of voxels. The total number of voxels contained in a connected path that satisfies the directional consistency threshold. To analyze the total number of voxels within a window, the system calculates the local density gradient (density change rate) of each voxel within the window; it also calculates the directional offset (principal direction angle) between adjacent voxels. Based on these two types of data, the system calculates the flow impedance between voxels using a weighted average; a smaller impedance indicates easier flow. The system then identifies the path with the minimum impedance in this impedance diagram and calculates the proportion of the window it occupies, defining it as the second connectivity porosity. The two types of porosity data are then analyzed. and To merge: ,in The second connectivity porosity represents the proportion of the minimum impedance path within the analysis window. As an accessibility adjustment factor, the specific generation process involves calculating the average impedance of all adjacent voxel pairs within a local analysis window, thus obtaining the average impedance. The average directional consistency is obtained by taking the average directional consistency coefficient of all adjacent voxel pairs. Then, according to the formula Calculate the local accessibility index T, which ranges from 0 to 1; a higher value indicates better liquidity in the local area. Then, convert the local accessibility index T into an accessibility adjustment factor using a linear mapping. The mapping range is limited to [0.9, 1.1] (or set to [0.8, 1.2] as needed). A locally connected porosity map is output, where each voxel unit corresponds to a fused porosity value (range 0-1).

[0047] The trunk tilt direction is estimated by performing a step-by-step estimation on the hierarchical voxel data to obtain the trunk tilt direction data. Specifically, the canopy segmentation region is divided into several height layers according to elevation coordinates (Z-axis direction). Within each height layer, the geometric centroid position of all voxels in that layer (i.e., the average of the three-dimensional coordinates of all voxels in that layer) is calculated. The centroids of each layer are connected sequentially according to elevation to form an initial central axis representing the spatial orientation of the main trunk. Least-squares line fitting is performed on the centroid sequence of the initial central axis to obtain a straight line that best represents the overall orientation: the direction vector of this line is defined as the trunk tilt direction vector. The angle between this direction vector and the ground normal vector (taking the positive Z-axis direction) is calculated; this angle is the trunk tilt angle. The trunk tilt direction vector and the trunk tilt angle are integrated to obtain the trunk tilt direction data.

[0048] Canopy morphology symmetry data is obtained by processing the canopy morphology symmetry data based on the trunk tilt direction data; Specifically, the obtained trunk tilt direction vector is used as the central axis; a plane passing through the midpoint of the trunk is arbitrarily chosen on this axis as a symmetry reference plane, which is perpendicular to the trunk tilt direction; using this reference plane as a boundary, the three-dimensional voxel structure of the canopy is divided into two parts (both sides of the reference plane). The voxel set of the canopy on one side of the reference plane is mirrored, and the mapping rule is that each voxel point is translated in the opposite direction along the normal vector direction with an equal vertical distance to the plane, so that it corresponds to the other side, resulting in a mirrored voxel set. The Hausdorff distance is calculated for the original voxel set and its mirror image. The Hausdorff distance is defined as the larger of the maximum shortest distance from any point in set A to set B, and the maximum shortest distance from set B to set A. Here, set A is the original voxel set, and set B is the mirror image voxel set. The Hausdorff distance is normalized by setting a maximum reference scale (e.g., maximum canopy width or diagonal length), and the morphological symmetry index is calculated: Morphological symmetry = 1 − (Hausdorff distance / maximum reference scale). This index ranges from 0 to 1, with values ​​closer to 1 indicating greater symmetry. The canopy is divided into several horizontal layers along its height, and the morphological symmetry index is calculated for each layer to form a layered symmetry curve. When the upper layers have higher symmetry and the lower layers have lower symmetry, the canopy morphology tends towards an "umbrella" shape; when the difference in symmetry between the upper and lower layers is small and the overall value is high, the canopy morphology tends towards a "tower" shape. The output canopy morphology symmetry data includes global morphological symmetry values; layered symmetry curves and corresponding morphological classification labels; and parameter information on the symmetry reference plane and the trunk tilt direction.

[0049] By integrating voxel orientation data, local connectivity porosity data, canopy morphology symmetry data, and trunk tilt direction data, canopy feature data are obtained.

[0050] Specifically, a feature fusion tensor is constructed by combining voxel orientation data, local connectivity porosity data, canopy morphology symmetry data, and trunk tilt direction data to obtain canopy feature data.

[0051] Preferably, the calculation of locally connected porosity is specifically as follows: Voxel principal orientations are extracted based on voxel orientation data to obtain voxel principal orientation data; Specifically, for each voxel neighborhood (e.g., 5) 5 PCA analysis was performed on 5 voxel blocks to extract the principal direction vector corresponding to the largest eigenvalue. If the eigenvalue variance ratio (largest / second largest) is lower than the threshold (e.g., 1.2), it is considered a region with ambiguous orientation and is removed or marked as insufficient anisotropy. The principal direction data of each effective voxel is output in the form of a vector field and normalized for storage.

[0052] Flow penetration maps are constructed based on the principal direction data of voxels to obtain flow penetration map data; Specifically, each voxel unit in the entire three-dimensional canopy space is considered as a node in a graph structure, denoted as... The set of nodes is denoted as For any two spatially adjacent voxels (sharing a face, edge, or vertex), if their principal direction vectors are nearly parallel, then establish a connecting edge between them. The edge set is denoted as The resulting preliminary graph structure can be represented as G=(V,E). For each pair of connected nodes... and Calculate from point to The unit vector, denoted as and from point to The unit vector, denoted as ; Calculate the nodes respectively Main direction vector With unit vector The direction cosine between them (i.e., the cosine of the included angle); similarly, the nodes are calculated. Main direction vector With unit vector The direction cosine value between the two is used to obtain the weight value of the connecting edge. A weight threshold is set (e.g., 0.5); only edges with weights greater than the threshold are retained, and the remaining low-weight edges are deleted; the resulting sparse graph structure is the flow-penetrable graph, which represents the distribution of channels inside the canopy that are directionally continuous and suitable for fluid or spray penetration.

[0053] Flow connectivity calculations are performed on flow permeability map data to obtain flow connectivity data; Specifically, each voxel node in the flow-penetrable graph is taken as the starting node. Starting from this node, the process traverses along the connecting edges whose weights satisfy a threshold condition. The traversal method can be depth-first search, exploring adjacent nodes layer by layer until no further progress is possible; or an improved shortest path search (similar to Dijkstra's algorithm) can be used, introducing a direction preference during the search process, i.e., prioritizing paths with higher consistency with the current node's main direction. During the traversal, only voxel nodes whose direction consistency with the starting node is greater than a set threshold are counted, forming the reachable region of that starting node. For each starting voxel node, the number of voxel nodes contained in its reachable region is counted, denoted as the number of direction-consistent reachable voxels. Direction consistency means that the cosine of the angle between the voxel and the starting node in the path propagation direction is not less than a preset threshold (e.g., 0.5), and the weights of all edges in the path satisfy the retention conditions after sparsification. The maximum number of reachable voxels is found among all voxel nodes, denoted as... Consistent orientation of each node can achieve a certain number of voxels. and Divide them to get the local accessibility of the node: , For nodes The local accessibility is calculated, ranging from 0 to 1, where 1 indicates that the reachable region is the same as the global optimum, and 0 indicates that it is unreachable. The output flow connectivity data includes the local accessibility value of each voxel node, a list of reachable voxel nodes, and the corresponding path information.

[0054] Local directional porosity is calculated based on the flow connectivity data to obtain the first local connectivity porosity data. Specifically, for each target voxel node v, a local computation window is established centered on it; the three-dimensional size of the window is set according to the analysis requirements, for example, a side length of 7. 7 7 voxels, 9 9 9 voxels, etc.; Within this local window, count the number of voxels that are aligned with the target voxel and meet the connectivity condition; the alignment is determined by the cosine of the angle between the path directions of the voxel and the target voxel in the flow penetration map not being lower than a set threshold (e.g., 0.5); and the path is confirmed as a reachable path in the connectivity calculation. This number is recorded as the number of aligned connected voxels. Calculate the total number of voxels within the window, and use the ratio of the number of aligned connected voxels to the total number of voxels in the window as the first type of local connectivity porosity: ,in For target voxels The first type of locally connected porosity, This represents the number of oriented connected voxels within the window. This represents the total number of elements in the window.

[0055] Voxel impedance direction sensing is performed based on voxel orientation data to obtain voxel impedance data; Specifically, let the principal direction unit vector of each voxel be... Select a mainstream reference direction. A global reference direction can be taken (e.g., the direction of trunk inclination). Alternatively, within each height level / job partition, the local reference direction can be obtained using the mode direction of the PCA / orientation histogram. Define the angle between the voxel direction and the reference direction. , voxels The directional angle, that is, the angle between the principal direction of the voxel and the reference mainstream direction. This is the inverse cosine function operator; the absolute value is taken to eliminate the effect of positive and negative signs. Calculate the directional deviation impedance: ,in The direction of the deviation from the impedance coefficient is such that the closer it is to 0, the smaller the impedance and the easier it is to pass through.

[0056] Minimum impedance path grid fitting is performed based on voxel impedance data to obtain minimum impedance path grid data; Specifically, the 3D voxel data is treated as a graph structure G=(V,E,W), where the node set V corresponds to a node for each voxel position, and the edge set E connects adjacent voxels (6-neighborhood, 18-neighborhood, or 26-neighborhood can be used, depending on the resolution and application). The weight of each edge is determined by the impedance values ​​of adjacent voxels, i.e., the sum of the impedance values ​​of adjacent voxels divided by 2. A penetration direction (the overall penetration direction (e.g., positive and negative directions along the main trunk)) or a target point set (given a set of starting points and ending points (e.g., top and bottom of the canopy, nozzle position and target canopy region), with their coordinate indices marked in the node set) is specified, and Dijkstra's or A* path algorithm is used to search for the set of minimum impedance paths.

[0057] Voxel-level flowability is estimated based on the minimum impedance path grid data to obtain voxel-level flowability data. Specifically, in the minimum impedance path grid data, all the searched path sets are traversed. The number of paths covering each voxel is counted, denoted as the path density. The path density of the voxels is then combined. Its impedance value (Range from 0 to 1, with larger values ​​indicating higher resistance), calculate the flowability index of voxels. : The system normalizes all flowability indices to obtain a voxel-level flowability map, and stores the flowability value of each voxel in the form of a three-dimensional scalar field.

[0058] Based on the voxel-level flowability data, the penetration-enhancing porosity is calculated to obtain the second locally connected porosity data; Specifically, the average voxel-level flowability data of all voxels within the window area is calculated to obtain the second local connectivity porosity data.

[0059] Spatial fusion is performed based on the first and second local connected porosity data to obtain local connected porosity data.

[0060] Specifically, spatial fusion computing: ,in for, To integrate the weighting coefficients, This is the porosity data for the first locally connected area. This is the porosity data for the second locally connected region. This is a path entropy suppression factor, and its specific generation process is shown in the local analysis window of the voxel (e.g., 7). 7 Within a window of 7 voxels, the distribution of all penetration paths is statistically analyzed. For each path, its relative frequency across voxels is calculated, which is the ratio of the number of times the path passes through that voxel to the sum of the number of times all paths pass through that voxel. Based on this frequency distribution, the path distribution entropy is calculated and mapped to the range [0,1] to obtain the path entropy suppression factor. The above formula is applied to all voxel positions to generate a local connectivity porosity map in the form of a three-dimensional scalar field.

[0061] Preferably, spatial fusion specifically refers to: Based on the first and second local connectivity porosity data, the flow accessibility multiplier and path entropy suppression factor are calculated to obtain the flow accessibility multiplier data and path entropy suppression factor data, respectively. Specifically, at the voxel location, the formula for calculating the flow accessibility multiplier is: ,in For the accessibility multiplier data, It is a minimum value function. This is the porosity data for the second locally connected region. This is the porosity data for the first locally connected area. To prevent small constants with zero denominators (e.g.) ); In each local analysis window (e.g., 7) 7 Within a set of 7 voxels, the distribution of all penetration paths is statistically analyzed. For the k-th path in the path set, its probability distribution at voxel v is defined as: Calculate the path distribution entropy H(v) at the voxel location: , The path number. To prevent small constants with zero logarithm, a higher entropy value indicates a more dispersed path distribution (no clear main path); a lower entropy value indicates a more concentrated path distribution (the existence of a stable main path). The path entropy suppression factor is calculated as follows: ,in This is path entropy suppression factor data. For the natural index term, This is an adjustment factor (e.g., 0.5).

[0062] The first local connectivity porosity data and the second local connectivity porosity data are fused and calculated based on the flow accessibility multiplier data and the path entropy suppression factor data to obtain the local connectivity porosity data.

[0063] Specifically, fusion computing: ,in This is data on locally connected porosity. This is the structure / path balance factor, with an empirical value between 0.4 and 0.6. This is the porosity data for the first locally connected area. This is the porosity data for the second locally connected region. For the accessibility multiplier data, This is path entropy suppression factor data.

[0064] Preferably, the penetration demand metric estimation is specifically as follows: Spray direction simulation data is obtained by simulating the spray direction based on canopy feature data. Specifically, based on the canopy geometry and trunk attitude, the incident direction distribution of the spray in three-dimensional space is simulated. A vertically downward direction vector is used as the default spray incident direction; for canopy areas with significantly tilted trunks, a deflection correction term is set, adjusting the spray direction according to the following formula: ,in voxels The incident direction vector of the spray. To normalize the vector so that its direction length is 1, This is the default spray direction. This is the deflection adjustment coefficient, with a value ranging from 0.2 to 0.4. The canopy tilt direction vector is calculated based on the tilt posture of the trunk or branches in the local canopy to which the voxel belongs. After calculation, the system outputs a set of spray incidence direction vectors for each voxel in the 3D scene, i.e., the spray direction simulation data.

[0065] Local void flow processing is performed on the spray direction simulation data to obtain local flow data; Specifically, for each voxel In its direction Upsampling step size is Forward search length L (e.g., 10cm): If there are consecutive empty voxels on the path ( , Porosity (This is a porosity threshold, such as 0.5) and the spacing does not exceed [a certain value]. , The maximum allowable spacing is recorded as the effective flow segment; the cumulative length of all effective flow segments is calculated to obtain the effective flow length. The liquidity is calculated as follows: , voxels The corresponding local circulation data.

[0066] Directional occlusion data is obtained by performing directional occlusion calculations based on local flow data. Specifically, the shielding angle is the angle at which the first dense element (defined as its porosity) is encountered on the spray path. , The local blocking angle is formed by the density threshold (e.g., 0.2). For each voxel, the angle at which the first occlusion occurs is recorded as it moves along the direction (spray direction) from its starting point. This angle is calculated using the following occlusion function: ,in For directional occlusion data, It is a natural exponential function. This is the angular sensitivity coefficient (e.g., k=2.0). To block the angle, This is a task voxel unit.

[0067] Based on the directional masking data, the spray incident loss is calculated to obtain the penetration demand measurement data.

[0068] Specifically, the system calculates the combined loss coefficient of spray droplets as they pass through the crop canopy due to insufficient local flow and directional shading. ,in For loss coefficient data, For localized circulation data, This is localized liquidity weight data, with a value of 1.0. For directional occlusion data, The directional shading weight data is set to 1.2. The penetration demand metric is the spray compensation intensity required for this voxel. The penetration demand metric represents the spray intensity enhancement factor required to compensate for spray loss in the canopy path; it is obtained by multiplying by the standard demand. Spray compensation intensity calculation: ,in For spray compensation intensity data, For the first One canopy voxel unit, For loss coefficient data, The porosity influence factor (e.g., 0.8). This is data on porosity of locally connected structures.

[0069] Preferably, the nozzle control generation specifically includes: Step S31: Perform spray volume-penetration coupling mapping based on nozzle parameter data and penetration demand measurement data to obtain coupling mapping data; Specifically, based on the penetration requirements of different areas, Converted into injection volume control commands, the specific calculation is as follows: ,in To allocate to voxels The target spray volume, The minimum spray volume threshold, The maximum spray volume threshold, To penetrate demand metrics, As a work voxel unit, The mapping curvature parameter (value range [0.8, 1.5]) is used. To avoid excessive differences in spray volume between adjacent voxels, which could lead to frequent nozzle adjustments and mechanical vibration, smoothing filtering or sliding window averaging is performed after spray volume mapping to ensure a continuous spatial transition in spray volume distribution. The system is equipped with a pressure feedback module, which uses an empirical model to determine the target spray volume. Converted to corresponding spray pressure : ,in Determined by the nozzle's hydraulic characteristic curve or experimental calibration data. The median diameter of the droplets is the diameter that constitutes 50% of the cumulative volume in the spray droplet size distribution. The target spray volume per voxel. The values ​​are used to form a spray volume distribution map covering the operation area; if pressure feedback is available, the target spray pressure for each voxel is output. This value is used by the nozzle actuator.

[0070] Step S32: Perform spray delay simulation based on the coupling mapping data to obtain spray delay data; Specifically, the system simulates the time delay from the spray nozzle exit to the target crop canopy voxel. The system inputs or sets the spray velocity, the spatial distance between the crop and the nozzle, the wind direction angle, and the drag disturbance factor; calculations are then performed. ,in The delay time corresponding to each spray control point The spatial distance between the crop and the sprinkler head. For spray speed, It is a cosine function. The angle between wind directions, This is the drag disturbance factor (empirically set as 0.05–0.2).

[0071] Step S33: Predict the spray width overlap distribution from the spray delay data to obtain nozzle control data.

[0072] Specifically, for each nozzle, the following calculations are performed within one control cycle: combining the nozzle's current spray direction, spray volume control value, and delay time, the spatial coverage area of ​​the spray at the expected time (current time and delay time) is predicted, and the corresponding spray volume distribution is calculated. At each grid cell, the spray volume contributions from all operating nozzles are summed to obtain the total spray volume: ,in For in grid Total spray volume For nozzle sequence items, For the first The spray volume control value of each nozzle at the current moment. Let j be the spray distribution function of the j-th nozzle. The horizontal coordinates of the work area. The vertical coordinates of the work area. For the current calculation time, This represents the time delay for the spray from the j-th nozzle to reach the target position. If the total spray volume of a certain grid cell... If the spray volume exceeds a preset threshold, the area is identified as a spray overlap zone. The system uses a preset strategy library to trigger coordinated control strategies for this area, such as stopping the nozzles, reducing the spray volume, or adjusting the spray pressure, to avoid over-spraying.

[0073] Preferably, step S4 specifically includes: Step S41: Perform nozzle control operations based on nozzle control data and monitor the operations to obtain nozzle operation data; Specifically, based on the nozzle control data generated in the aforementioned steps, actual spraying operations are executed. During the operation, the nozzle operating status and spray parameters are monitored and recorded in real time, forming nozzle operation data. The nozzle control data includes parameters such as nozzle number, start / stop time, target spray volume, and spray pressure. The control commands are sent to the nozzle actuator via a programmable logic controller (PLC) or microcontroller unit (MCU), driving components such as solenoid valves and pumps to operate according to set modes, achieving spray start / stop, spray volume adjustment, and pressure control. During the spraying operation, the system collects actual spray volume, actual spray pressure, nozzle operating status, and other operating status codes in real time through various sensors. The system records these monitoring results in chronological order, forming a nozzle operation dataset.

[0074] Step S42: Perform operation coverage calculation based on the nozzle operation data and the preset nozzle operation space data to obtain operation coverage data; Specifically, based on nozzle operation data and preset nozzle operation space data, the actual spray range and spatial coverage are calculated and analyzed to form operation coverage data. This involves acquiring the spatial coordinates of each nozzle in the operation area, spray range, spray width angle, spray shape model (e.g., fan shape, elliptical cone shape), nozzle on / off status at different time periods, actual spray flow rate or controlled spray volume, and dynamic attitude information such as spray direction (if applicable). The system sets the instantaneous spray coverage area for each nozzle at any time t. This area is determined by spray width model parameters (spray width angle, range, etc.) and changes with nozzle attitude or control status. The farmland operation area is divided into two-dimensional or three-dimensional grid cells, each representing a fixed spatial location. For each grid cell, all nozzles are traversed, and it is determined whether the cell is within the instantaneous spray coverage range of a certain nozzle: if within the coverage range, the current spray flow rate of that nozzle is added to the cumulative spray volume value of that cell; if not within the coverage range, no accumulation is performed. After accumulation, the spray coverage matrix of the entire operation area is obtained. For each grid cell, a target spray volume is set, which can be preset based on crop type, growth stage, and application requirements. The cumulative spray volume is compared with the target spray volume to calculate the coverage deviation for each grid cell: if the cumulative spray volume is higher than the target value and exceeds the allowable error range, it is marked as an oversprayed area; if the cumulative spray volume is lower than the target value and exceeds the allowable error range, it is marked as an undersprayed area.

[0075] Step S43: Perform nozzle conductivity inversion based on the operation coverage data to obtain spray uniformity data.

[0076] Specifically, a trace electrolyte (such as potassium chloride solution) or a pesticide solution with weak conductivity is added to the spray liquid. An electrode array is deployed in the target operating area (soil surface or crop leaves), with electrodes evenly distributed in a grid pattern to monitor the change in conductivity before and after spraying in real time. For each monitoring point, the conductivity value before spraying is recorded; after spraying, the conductivity value after spraying is recorded again. The change in conductivity is calculated. An empirical mapping model is used to convert the conductivity increment into the amount of sprayed material deposited per unit area at that monitoring point. The mapping model can be obtained through pre-calibration experiments to ensure conversion accuracy under different liquid formulations and environmental conditions. The inverse distance weighted method (IDW) or Kriging interpolation method is used to convert the spray deposition amount at discrete monitoring points into a continuous spray deposition distribution map of the entire operating area. The obtained spray distribution map is compared with the preset target spray distribution to analyze the spatial location and area range of local overspraying or underspraying. Calculate the average spray deposition and the standard deviation of the spray distribution; the output spray uniformity data includes a spray deposition distribution map (2D or 3D); the spray coefficient of variation (CV) and the standard deviation of each region; and spatial markers for overspray and underspray areas. If the CV exceeds a threshold, one of the following control strategies may be triggered: automatically adjust the nozzle pressure, spray volume, or spray angle; generate a prompt to rework; or update the spray model parameters to optimize subsequent operations.

[0077] Preferably, this application also provides a variable pesticide spraying system based on crop canopy recognition for performing the variable pesticide spraying method based on crop canopy recognition as described above. The variable pesticide spraying system based on crop canopy recognition includes: The sparse anchor point calibration module is used to acquire canopy structure data; based on the canopy structure data, sparse anchor points are calibrated to obtain sparse anchor point data. The canopy segmentation and spray demand analysis module is used to segment the canopy based on sparse anchor point data to obtain canopy segmentation data; extract canopy features from the canopy segmentation data to obtain canopy feature data; and estimate penetration demand based on the canopy feature data to obtain penetration demand measurement data. The variable nozzle control generation module is used to acquire nozzle parameter data; and to generate nozzle control data based on the nozzle parameter data and penetration requirement measurement data. The spray operation monitoring and uniformity feedback module is used to monitor the operation based on the nozzle control data to obtain nozzle operation data; and to provide spray uniformity feedback based on the nozzle operation data to obtain spray uniformity data.

[0078] Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended application documents rather than the foregoing description. Thus, it is intended that all variations falling within the meaning and scope of the equivalents of the application documents be incorporated into the invention.

[0079] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. A variable pesticide spraying method based on crop canopy recognition, characterized in that, Includes the following steps: Step S1: Obtain canopy structure data; perform sparse anchor point calibration based on the canopy structure data to obtain sparse anchor point data; Step S2: Perform canopy segmentation based on sparse anchor point data to obtain canopy segmentation data; perform hierarchical structure voxel scanning on the canopy segmentation data to obtain hierarchical voxel data; calculate voxel orientation density based on the hierarchical voxel data to obtain voxel orientation data; calculate local connectivity porosity based on the voxel orientation data to obtain local connectivity porosity data; estimate the trunk tilt direction based on the hierarchical voxel data to obtain trunk tilt direction data; process canopy morphology symmetry based on the trunk tilt direction data to obtain canopy morphology symmetry data; integrate the voxel orientation data, local connectivity porosity data, canopy morphology symmetry data, and trunk tilt direction data to obtain canopy feature data; estimate penetration demand based on the canopy feature data to obtain penetration demand measurement data. Step S3: Obtain nozzle parameter data; generate nozzle control data based on nozzle parameter data and penetration requirement measurement data; Step S4: Perform operation monitoring based on nozzle control data to obtain nozzle operation data; perform spray uniformity feedback based on nozzle operation data to obtain spray uniformity data; The calculation of locally connected porosity is as follows: Voxel principal directions are extracted from voxel orientation data to obtain voxel principal direction data. A flow penetration map is constructed based on the voxel principal direction data to obtain flow penetration map data. Flow connectivity is calculated on the flow penetration map data to obtain flow connectivity data. Local directional porosity is calculated based on the flow connectivity data to obtain first locally connected porosity data. Voxel impedance direction sensing is performed based on the voxel orientation data to obtain voxel impedance data. Minimum impedance path mesh fitting is performed based on the voxel impedance data to obtain minimum impedance path mesh data. Voxel-level flowability is estimated based on the minimum impedance path mesh data to obtain voxel-level flowability data. Penetration-enhancing porosity is calculated based on the voxel-level flowability data to obtain second locally connected porosity data. Spatial fusion is performed on the first and second locally connected porosity data to obtain locally connected porosity data. In fact, spatial integration specifically refers to: Flow accessibility multipliers and path entropy suppression factors are calculated based on the first and second local connectivity porosity data to obtain flow accessibility multiplier data and path entropy suppression factor data, respectively. The first and second local connectivity porosity data are then fused based on the flow accessibility multiplier data and path entropy suppression factor data to obtain local connectivity porosity data.

2. The method according to claim 1, characterized in that, The sparse anchor point calibration is specifically as follows: Multi-scale geometric detection is performed based on canopy structure data to obtain canopy geometric data; Topological bifurcation points are extracted from the canopy geometric data to obtain canopy bifurcation point data; Voxel sparsity constraints are applied to the canopy bifurcation point data to obtain canopy sparse data; Sparse anchor point data is generated by using sparse canopy data.

3. The method according to claim 1, characterized in that, The canopy segmentation specifically includes: Anchor point structure transfer data is obtained by performing anchor point structure transfer processing based on sparse anchor point data. Similarity drift clustering is performed based on the anchor point structure transfer data to obtain anchor point structure clustering data; Spray projection data is obtained by performing spray direction view projection based on anchor point structure clustering data. Collaborative sensing and segmentation of the spray projection data yields canopy segmentation data.

4. The method according to claim 1, characterized in that, The specific estimation of penetration demand measurement is as follows: Spray direction simulation data is obtained by simulating the spray direction based on canopy feature data. Local void flow processing is performed on the spray direction simulation data to obtain local flow data; Directional occlusion data is obtained by performing directional occlusion calculations based on local flow data. Based on the directional masking data, the spray incident loss is calculated to obtain the penetration demand measurement data.

5. The method according to claim 1, characterized in that, The nozzle control generation specifically involves: Based on nozzle parameter data and penetration demand measurement data, spray volume-penetration coupling mapping is performed to obtain coupling mapping data; Spray delay data is obtained by simulating the spray delay based on the coupling mapping data; Spray delay data is used to predict the spray width overlap distribution to obtain nozzle control data.

6. The method according to claim 1, characterized in that, Step S4 is as follows: Sprinkler control operations are performed and monitored based on sprinkler control data to obtain sprinkler operation data. The operation coverage is calculated based on the nozzle operation data and the preset nozzle operation space data to obtain the operation coverage data; Spray uniformity data are obtained by performing nozzle conductivity inversion based on the coverage data.

7. A variable pesticide spraying system based on crop canopy recognition, characterized in that, For performing the variable pesticide spraying method based on crop canopy identification as described in claim 1, the variable pesticide spraying system based on crop canopy identification comprises: The sparse anchor point calibration module is used to acquire canopy structure data; based on the canopy structure data, sparse anchor points are calibrated to obtain sparse anchor point data. The canopy segmentation and spray demand analysis module is used to segment the canopy based on sparse anchor point data to obtain canopy segmentation data; extract canopy features from the canopy segmentation data to obtain canopy feature data; and estimate penetration demand based on the canopy feature data to obtain penetration demand measurement data. The variable nozzle control generation module is used to acquire nozzle parameter data; and to generate nozzle control data based on the nozzle parameter data and penetration requirement measurement data. The spray operation monitoring and uniformity feedback module is used to monitor the operation based on the nozzle control data to obtain nozzle operation data; and to provide spray uniformity feedback based on the nozzle operation data to obtain spray uniformity data.

Images