A method for controlling stellera chamaejasme

Abstract

The present application discloses a kind of prairie Stellera chamaejasme prevention method, comprising: obtaining hyperspectral image and laser radar point cloud of operation area;Get pixel level wolf poison classification result;Segment out plant object cluster;Wolf poison plant is identified and crown layer center coordinates and crown layer size are determined;According to crown layer center coordinates, the number of neighborhood wolf poison strains of each wolf poison plant is determined;The number of neighborhood wolf poison strains is compared with preset isolated plant threshold value, and physical uprooting control instruction is generated for isolated plant less than or equal to threshold value, and variable chemical spraying control instruction is generated for contiguous plant greater than threshold value, the present application relates to prairie poisonous grass prevention technical field, the present application is fused by spectral feature of hyperspectral image and spatial structure feature of laser radar point cloud, and wolf poison plant is determined doubly in spectrum and space, effectively solve the problem of high misjudgment rate of single sensor in dry grass, bare soil, similar green plant and other complex backgrounds, significantly improve the precision and robustness of wolf poison identification.

Description

Technical Field

[0001] This invention relates to the field of grassland poisonous weed control technology, specifically a method for controlling Daphne odora and Euphorbia fischeriana in grasslands. Background Technology

[0002] Stellera chamaejasme L., a perennial poisonous herb belonging to the Thymelaeaceae family, is the most widely distributed and extensive native invasive poisonous grass in my country's natural grasslands. Its population dynamics are considered an important indicator of grassland degradation. The entire plant is poisonous and can spread rapidly through seeds and root suckers. It also has allelopathic effects, inhibiting the growth of surrounding high-quality forage grasses, forming a monoculture community, leading to decreased grassland productivity and deterioration of grass quality, seriously threatening grassland ecological security and the sustainable development of animal husbandry.

[0003] Existing control technologies for *Euphorbia fischeriana* mainly include physical, biological, and chemical control. Physical control inhibits growth by severing roots and produces no chemical pollution, but it is inefficient and costly in large-scale, high-density outbreaks. Chemical control primarily involves spraying herbicides, which is fast and efficient, and is currently the main means of controlling large-scale outbreaks of *Euphorbia fischeriana*. However, long-term, excessive, and broadcast application of herbicides can have negative impacts on grassland ecosystems, posing environmental safety and non-target organism risks. To achieve precision control, existing research has utilized multispectral imaging and machine learning algorithms to identify *Euphorbia fischeriana* and has reduced herbicide usage based on variable spraying techniques. However, overall, existing methods lack a closed-loop approach that integrates multi-source sensing, intelligent decision-making, and targeted execution. They cannot autonomously decide on differentiated intervention strategies based on *Euphorbia fischeriana* distribution density, such as physical uprooting or chemical spraying, and are difficult to implement with centimeter-level precision in complex grassland environments with undulating terrain. Summary of the Invention

[0004] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a method for controlling Daphne odora and Elaeagnus pungens in grasslands, which solves the problems of insufficient accuracy in identifying Elaeagnus pungens, simplistic and crude control strategies, and the significant impact of terrain on the accuracy of implementation in existing technologies.

[0005] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: A method for controlling Daphne odora and Euphorbia fischeriana in grasslands, comprising the following steps: S1: Acquire hyperspectral images and lidar point clouds of the work area; S2: Perform feature band extraction and pixel classification on the hyperspectral image to obtain pixel-level wolfsbane classification results; S3: Filter and spatially cluster the lidar point cloud to segment plant object clusters and calculate the spatial position of each plant object cluster. S4: The pixel-level wolfsbane classification result is fused with the plant object cluster to identify the wolfsbane plant and determine the canopy center coordinates and canopy size of the identified wolfsbane plant. S5: Determine the number of neighboring *Euphorbia fischeriana* plants for each *Euphorbia fischeriana* plant based on the coordinates of the canopy center; S6: Compare the number of neighboring wolfsbane plants with a preset isolated plant threshold. If the number of neighboring wolfsbane plants of a wolfsbane plant is less than or equal to the preset isolated plant threshold, generate a physical uprooting control command for the wolfsbane plant. If it is greater than the preset isolated plant threshold, generate a variable chemical spraying control command for the wolfsbane plant. S7-1: For the wolfsbane plant that generates the physical uprooting control command, extract the local ground point cloud within a preset radius around the center of the canopy of the wolfsbane plant, fit the ground plane, and generate the end effector control command according to the normal vector of the ground plane to control the end effector to cut into the ground at a preset depth in a direction perpendicular to the ground plane to cut the root. S7-2: For the *Euphorbia fischeriana* plant that generates the variable chemical spraying control command, plant canopy outline information is generated based on the determined canopy size of the *Euphorbia fischeriana* plant, and a nozzle flow control signal is generated based on the canopy outline information to adjust the nozzle flow in real time for variable application.

[0006] In some embodiments, step S2, which involves feature band extraction and pixel classification of the hyperspectral image, includes the following steps: The hyperspectral image is subjected to black-and-white correction and noise removal; The characteristic bands that distinguish wolfsbane from the background are selected from the bands of the hyperspectral image by using a continuous projection algorithm or a competitive adaptive reweighted sampling algorithm. Using the aforementioned feature bands as input, a support vector machine classifier or a random forest classifier is used to perform binary classification of each pixel between wolfsbane and background, thereby obtaining the pixel-level wolfsbane classification result.

[0007] In some embodiments, step S3 involves filtering and spatial clustering the lidar point cloud, including the following steps: A statistical outlier removal filter is used to remove noise from the lidar point cloud. A cloth-simulation filtering algorithm is used to separate ground points from non-ground points; The non-ground points are clustered using an Euclidean distance clustering algorithm or a region growing algorithm to segment independent plant object clusters, and the three-dimensional centroid coordinates of each plant object cluster are calculated as the spatial location of the plant object cluster.

[0008] In some embodiments, step S4, which fuses the pixel-level wolfsbane classification result with the plant object cluster to identify wolfsbane plants, includes: The pixel-level wolfsbane classification results are projected onto the plant object cluster through coordinate mapping; If the number of pixels classified as wolfsbane within the projection area of ​​a plant cluster exceeds 60%, then the plant cluster is determined to be the wolfsbane plant.

[0009] In some embodiments, determining the number of neighboring *Euphorbia humifusa* plants in step S5 includes: Taking the center of the canopy of each of the aforementioned wolfsbane plants as the center, count the number of wolfsbane plants within a preset radius, which is taken as the number of neighboring wolfsbane plants of the aforementioned wolfsbane plant.

[0010] In some embodiments, the preset radius is 1 meter.

[0011] In some embodiments, step S7-1 specifically includes the following steps: Extract the ground point cloud within a 0.5-meter radius around the center of the canopy of the target *Euphorbia milii* plant; The random sampling consensus algorithm is used to perform plane fitting on the ground point cloud to obtain the normal vector and elevation of the local ground plane; Based on the normal vector and the elevation, instructions are generated to control the cutting angle and cutting depth of the end shovel, so that the cutting direction of the end shovel is parallel to the normal vector and cuts into the ground to a depth of 7-10 cm to sever the root.

[0012] In some embodiments, step S7-2 specifically includes the following steps: Pre-generate variable spray prescription maps; The basic spraying amount is determined according to the variable spraying prescription diagram, and the basic spraying amount is corrected by using the plant canopy outline information generated based on the canopy size of the wolfsbane plant. A pulse width modulation signal is generated to adjust the duty cycle of the solenoid valve, thereby controlling the instantaneous flow rate of the nozzle. When there are no *Euphorbia fischeriana* plants in the area to be sprayed, the instantaneous flow rate of the nozzle is controlled to be zero.

[0013] In some embodiments, the modulation frequency of the pulse width modulation signal is 10 Hz.

[0014] In some embodiments, the preset isolated plant threshold is 2 plants.

[0015] (III) Beneficial Effects The beneficial effects of this invention are: This invention integrates the spectral features of hyperspectral images with the spatial structure features of lidar point clouds to make dual spectral and spatial determinations of wolfsbane plants. This effectively solves the problem of high misjudgment rate of single sensors in complex backgrounds such as withered grass, bare soil, and similar green plants, and significantly improves the accuracy and robustness of wolfsbane identification.

[0016] This invention introduces a neighborhood density-based decision-making mechanism to automatically switch control strategies based on the density of plants within a certain range around the wolfsbane plant. It uses physical removal for sporadic wolfsbane and variable chemical spraying for patchy or high-density wolfsbane, achieving differentiated management with precise removal and reduced spraying. This avoids the overuse of chemical agents and ensures operational efficiency in high-density areas.

[0017] In the physical removal step, this invention extracts local ground point clouds around the target plant and fits them to the ground plane. Based on the ground plane normal vector, it controls the cutting direction of the end effector, ensuring that the digging tool always cuts into the ground at an angle perpendicular to the ground surface. This overcomes the interference of uneven grassland terrain on the root cutting depth and ensures the thoroughness and consistency of root cutting.

[0018] In the variable chemical spraying step, this invention generates a canopy outline based on the canopy size information obtained in the previous identification stage, and adjusts the nozzle flow rate in real time accordingly. This ensures that the amount of pesticide applied is precisely matched with the canopy volume of the target plant, avoiding uneven application problems such as spraying large plants with small amounts of pesticide or small plants with large amounts of pesticide. This further reduces the amount of herbicide used and reduces environmental risks. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the method steps of the present invention. Detailed Implementation

[0020] To better explain and facilitate understanding of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0021] This method is achieved collaboratively by a hyperspectral imager and lidar scanning system mounted on an unmanned aerial vehicle (UAV) platform, a ground-based data processing system, and an autonomous walking platform. The autonomous walking platform carries a physical removal actuator and a variable spraying actuator. The physical removal actuator includes an end-effector and a clamping mechanism, while the variable spraying actuator includes a nozzle, a pesticide pipeline, and a solenoid valve. The data processing system receives hyperspectral images and lidar point clouds and executes steps S2 to S6 for processing and decision-making. The autonomous walking platform receives control commands from the data processing system and autonomously navigates to the target plant location based on the target coordinates in the commands to perform the corresponding operations. This invention achieves accurate identification and three-dimensional positioning of *Euphorbia fischeriana* plants by fusing hyperspectral images and lidar point clouds. It also assigns two operating modes—physical removal and variable chemical spraying—to isolated plants and clusters of plants based on the number of neighboring *Euphorbia fischeriana* plants, aiming to solve the problem that existing control methods cannot differentiate treatment based on the distribution pattern of *Euphorbia fischeriana*.

[0022] To better understand the above technical solutions, exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that the present invention can be understood more clearly and thoroughly, and that the scope of the present invention can be fully conveyed to those skilled in the art.

[0023] Example Please see Figure 1 The present invention provides a technical solution: a method for controlling Daphne odora and Euphorbia fischeriana in grasslands, comprising steps S1 to S7-2.

[0024] Step S1: Acquire hyperspectral images and lidar point clouds of the work area. Hyperspectral images and lidar point cloud data of the work area can be simultaneously acquired using a hyperspectral imager and lidar scanning system mounted on a drone platform. The hyperspectral image provides rich spectral information to distinguish *Euphorbia fischeriana* from background vegetation; the lidar point cloud provides accurate three-dimensional spatial information to obtain plant structure and terrain. Before acquisition, the two sensors need to be spatially calibrated and temporally synchronized to ensure subsequent fusion accuracy. Simultaneously, an autonomous walking work platform is equipped at the ground end of the work area, carrying a physical removal actuator and a variable chemical spraying actuator. The physical removal actuator includes an end-effector and a clamping mechanism; the variable chemical spraying actuator includes a nozzle, a pesticide pipeline, and a solenoid valve. The autonomous walking work platform receives physical removal control commands or variable chemical spraying control commands generated by the data processing system and autonomously navigates to the target plant location to perform the corresponding operation based on the target coordinates in the command.

[0025] Step S2 involves extracting feature bands and classifying pixels from the hyperspectral image to obtain pixel-level classification results for *Euphorbia fischeriana*. *Euphorbia fischeriana* exhibits spectral reflectance characteristics distinct from background vegetation such as pasture in specific bands. Selecting the most sensitive feature bands for distinguishing *Euphorbia fischeriana* from the numerous bands in the hyperspectral image effectively reduces data redundancy and improves classification efficiency. Based on the selected feature bands, a supervised classifier is used to perform binary classification on each pixel, determining whether it belongs to *Euphorbia fischeriana* or the background, ultimately obtaining pixel-level classification results consistent with the spatial resolution of the original image.

[0026] Step S3 involves filtering and spatial clustering the LiDAR point cloud to segment plant clusters and calculate the spatial location of each cluster. The LiDAR point cloud contains ground points, vegetation points, and noise points. Noise points are removed using a filtering algorithm, and the point cloud is separated into ground points and non-ground points, with non-ground points primarily corresponding to vegetation. Spatial clustering is then performed on the non-ground points, grouping spatially adjacent points belonging to the same plant or clump of plants into a single cluster. The spatial location of each cluster is calculated, providing spatial boundary and location information of individual plants for subsequent fusion with hyperspectral classification results.

[0027] Step S4 involves fusing the pixel-level *Euphorbia humifusa* classification results with the plant clusters to identify individual *Euphorbia humifusa* plants and determine the canopy center coordinates and canopy dimensions of the identified plants. This is achieved by projecting the spectral-dimensional pixel classification results onto the three-dimensional plant clusters. Using each plant cluster as the basic unit, the proportion of pixels classified as *Euphorbia humifusa* within its projection area is statistically analyzed. When the proportion meets the criteria, the cluster is identified as a single *Euphorbia humifusa* plant. For the identified *Euphorbia humifusa* plants, the coordinates of their canopy center and the canopy outline dimensions are extracted to provide a basis for subsequent operational decisions.

[0028] Step S5: Determine the number of neighboring *Euphorbia humifusa* plants for each *Euphorbia humifusa* plant based on the canopy center coordinates. Using the canopy center of each identified *Euphorbia humifusa* plant as a reference, count the number of other *Euphorbia humifusa* plants within a certain distance around it to quantify the local distribution density of each *Euphorbia humifusa* plant.

[0029] Step S6: Compare the number of neighboring *Euphorbia fischeriana* plants with a preset isolated plant threshold. The preset isolated plant threshold is the boundary for distinguishing between isolated growth and contiguous growth. If the number of neighboring *Euphorbia fischeriana* plants of a plant is less than or equal to the threshold, it indicates that the plant is in a sparse and isolated state, suitable for physical removal, and a physical removal control command is generated; if the number is greater than the threshold, it indicates that the plant is in a denser community, physical removal is inefficient and causes significant ecological disturbance, and a variable chemical spraying control command is generated.

[0030] Step S7-1: For the *Euphorbia humifusa* plant for which a physical uprooting control command has been generated, a physical uprooting operation is performed. The physical uprooting operation is completed by a physical uprooting actuator carried on an autonomous walking platform. A local ground point cloud within a certain radius around the center of the plant's canopy is extracted. The ground plane normal vector of this local area is obtained through a plane fitting algorithm. Based on this, a control command is generated to control the end-effector of the physical uprooting actuator to cut into the ground to a predetermined depth in a direction perpendicular to the ground plane, thereby severing the *Euphorbia humifusa*'s main root.

[0031] Step S7-2: For *Euphorbia milii* plants that have generated variable chemical spraying control commands, perform variable chemical spraying. The variable chemical spraying operation is completed by a variable spraying actuator carried on an autonomous walking platform. Based on the determined canopy size of the *Euphorbia milii* plants, information characterizing the canopy outline and coverage of individual plants is generated. Based on this, a signal is generated to control the nozzle flow rate of the variable spraying actuator, adjusting the spraying volume in real time so that plants with large canopies receive more pesticide, plants with small canopies receive less pesticide, and spraying is stopped in areas without *Euphorbia milii*, achieving targeted pesticide application on demand.

[0032] This invention proposes a method for controlling *Daphne odora* and *Euphorbia fischeriana* in grasslands. During operation, an unmanned aerial vehicle (UAV) system first acquires hyperspectral images and lidar point clouds. After feature band extraction, the hyperspectral images are used to generate pixel-level *Euphorbia fischeriana* identification maps using a classifier. The lidar point clouds are filtered and clustered to generate plant clusters. The two methods are then fused to identify *Euphorbia fischeriana* plants and obtain their precise canopy center, canopy size, and neighborhood distribution density. For isolated *Euphorbia fischeriana* plants, the system sends physical removal commands and target coordinates to an autonomous walking platform. This platform navigates to the target location, and the end effector of the physical removal actuator vertically penetrates the soil to sever the roots based on the local terrain. For contiguous *Euphorbia fischeriana* plants, the system sends variable chemical spraying commands and target coordinates to the autonomous walking platform. The variable spraying actuator adjusts the nozzle flow rate in real time according to the canopy size of each plant for targeted spraying. This method achieves a shift from indiscriminate, blanket-style application to intelligent, plant-specific management, balancing the eco-friendliness of physical removal with the high efficiency of chemical spraying.

[0033] In this embodiment, step S2, which involves feature band extraction and pixel classification of the hyperspectral image, includes the following steps: First, the original hyperspectral image undergoes black-and-white correction and noise removal. Black-and-white correction is performed by acquiring dark current images of a standard white board and a lens cap-covered image, and normalizing the pixel brightness values ​​of each band to eliminate the influence of sensor dark current and sensitivity differences between bands. Noise removal employs a minimum noise separation transform, retaining the top few components with high signal-to-noise ratios, removing noise-dominant components, and then inversely transforming back to the image space. Then, a successive projection algorithm is used to select feature bands that distinguish *Euphorbia fischeriana* from the background from hundreds of bands in the hyperspectral image. The successive projection algorithm sequentially selects the bands with the highest degree of orthogonality to the selected band set from the original band set to minimize redundant information between bands. The final selected feature bands are concentrated in the red-near-infrared transition region (680nm to 750nm) and the short-wave near-infrared region (950nm to 1000nm), retaining 10 to 20 feature bands, significantly reducing data dimensionality. Finally, using the selected feature bands as input, a random forest classifier is used to perform binary classification of each pixel into wolfsbane and background. The random forest classifier is pre-trained using ground truth samples labeled with wolfsbane and background categories. The classifier outputs a label of wolfsbane or background for each pixel, resulting in a pixel-level wolfsbane classification result.

[0034] In this embodiment, step S3 involves filtering and spatial clustering the lidar point cloud, including the following steps: First, a statistical outlier removal filter is used to remove noise points from the lidar point cloud. For each point, the average distance to its 50 nearest neighbors is calculated. If the average distance of a point exceeds 1.0 times the standard deviation of the overall distribution, it is identified as an outlier and removed. Then, a cloth simulation filtering algorithm is used to separate ground points from non-ground points. This algorithm simulates a virtual cloth covering an inverted point cloud from top to bottom. By iteratively solving the displacement of the cloth nodes, the cloth is made to conform to the terrain surface. Points within a certain distance below the cloth nodes are classified as ground points, and points above them are classified as non-ground points. Ground points are used for subsequent ground plane fitting, while non-ground points contain the main vegetation information. Finally, the Euclidean distance clustering algorithm is used to cluster the non-ground points. Starting from any non-ground point, all neighboring points with a spatial Euclidean distance of less than 0.05 meters are searched, and the clusters are recursively expanded to form plant object clusters. Small clusters with fewer than 10 points are removed, and valid plant object clusters are retained. The three-dimensional centroid coordinates of each plant object cluster are calculated as the spatial location of the plant object cluster.

[0035] In this embodiment, step S4, which fuses the pixel-level *Euphorbia humifusa* classification results with plant clusters to identify *Euphorbia humifusa* plants, includes the following specific methods: First, the pixel-level *Euphorbia humifusa* classification results are projected onto the plant clusters via coordinate mapping: using the interior orientation elements of the hyperspectral imager and the sensor exterior orientation elements provided by RTK-GPS and IMU, the pixel coordinates of each pixel are calculated into object-space three-dimensional coordinates using collinearity equations. Then, using the extrinsic parameter calibration matrix between the hyperspectral imager and the lidar, the object-space three-dimensional coordinates are transformed into the lidar coordinate system. Next, each plant cluster is projected vertically onto a horizontal plane to form a two-dimensional projection area. The classification labels of all hyperspectral pixels falling within this projection area are counted. If the number of pixels classified as *Euphorbia humifusa* within the projection area of ​​a plant cluster exceeds 60% of the total number of valid pixels in that area, then the plant cluster is determined to be a *Euphorbia humifusa* plant. For plant clusters identified as *Euphorbia fischeriana*, the canopy center coordinates and canopy dimensions are extracted. The canopy center coordinates are the median of the two-dimensional coordinates of all points on the horizontal plane. The canopy dimensions include the equivalent canopy diameter, calculated using the formula... Calculate, where S is the convex hull area of ​​the horizontal projection of the cluster.

[0036] In this embodiment, step S5, determining the number of neighboring *Euphorbia humifusa* plants for each plant, specifically involves: using the two-dimensional coordinates of the canopy center of each *Euphorbia humifusa* plant as the center, counting the number of other *Euphorbia humifusa* plants within a preset radius, which is taken as the neighboring *Euphorbia humifusa* plant of that plant. If the horizontal distance between the canopy centers of two *Euphorbia humifusa* plants is less than or equal to the preset radius, the other plant is counted as a neighboring plant of that plant.

[0037] In this embodiment, the preset radius is 1 meter. The 1-meter radius covers the main spatial range of root competition and above-ground part crossing among wolfsbane plants. According to field measurements, the crown diameter of a mature wolfsbane plant is usually between 0.3 and 0.6 meters. The 1-meter radius contains the spatial scale of neighboring plants and generates significant interactions, which can effectively distinguish between solitary plants and clusters.

[0038] In this embodiment, step S7-1 specifically includes the following steps. First, all ground points within a 0.5-meter radius around the center of the canopy of the target *Euphorbia fischeriana* plant are extracted from the ground point cloud as a local ground point cloud. This range is sufficient to cover the distribution area of ​​the main root of *Euphorbia fischeriana* and its surrounding micro-topography. Then, a random sampling consensus algorithm is used to perform planar fitting on the local ground point cloud. Candidate planar models are calculated by randomly sampling three points multiple times. The number of local points that match the model is counted, and the model with the most local points is iteratively selected as the final fitting result, obtaining the local ground plane normal vector n = (nx, ny, nz) and elevation value. Next, instructions for controlling the cutting angle and cutting depth of the end-effector are generated based on the normal vector and elevation. The cutting angle is set to the angle α = arccos(nz) between the normal vector and the vertical direction. By controlling the end-rotor joint of the robotic arm, the blade of the shovel is made parallel to the normal vector, meaning the cutting direction is perpendicular to the local ground plane. The cutting depth is set to 7 to 10 cm, covering the root collar and the upper 10 cm of the main root of the wolfsbane. Cutting this area effectively inhibits its asexual reproduction and regeneration. During the cutting process, the torque on the shovel is monitored in real time. When the resistance torque exceeds 50 N·m, the cutting is stopped immediately and the plant is withdrawn. The current plant is marked as a failed removal and reported. If the ground plane cannot be reliably fitted due to insufficient ground point cloud data (less than 20 points), the vertical direction is used directly as the cutting direction, and the preset default value of 8 cm is used as the cutting depth. An alarm is generated simultaneously. After cutting to the target depth, the robotic arm uses the shovel as a fulcrum to perform a small prying motion, using the lever principle to loosen the soil around the roots, separating the main root of the wolfsbane from the soil. Then, the clamping mechanism clamps the root collar of the wolfsbane and pulls up the entire plant.

[0039] In this embodiment, step S7-2 specifically includes the following steps. First, a variable spraying prescription map is generated in advance. The prescription map covers the entire operation area using a 0.5m × 0.5m grid as the basic unit. The basic spraying amount for each grid is determined based on the density and canopy size of the *Euphorbia humifusa* plants within that grid: the number of *Euphorbia humifusa* plants falling into that grid is counted. If there are no *Euphorbia humifusa* plants, the basic spraying amount is zero; if there are *Euphorbia humifusa* plants, the basic spraying amount is the average value d of the equivalent canopy diameter of all *Euphorbia humifusa* plants within that grid. avg and according to formula Q base = 0.3×d avg The calculation is based on 0.2 ml / m², where 0.3 ml / m² is the recommended spraying amount for a nominal plant with a canopy equivalent diameter of 0.2 m. During actual spraying, the base spraying amount is adjusted in real-time using the canopy equivalent diameter d of each *Euphorbia fischeriana* plant, with the adjustment coefficient k calculated using the formula... The calculation shows that the value of k is limited to between 0.5 and 2.0. The corrected instantaneous spraying rate Q = k × Q baseThen, a pulse width modulation signal is generated based on the corrected instantaneous spray volume to adjust the duty cycle of the solenoid valve, thereby controlling the instantaneous flow rate of the nozzle. The flow rate when the solenoid valve is fully open is F. max The vehicle's speed is v, the nozzle's spray width is W, and the relationship between the instantaneous spray rate Q and the duty cycle D is Q = (F max From this, we can calculate the duty cycle: D = Q × v × W / F. max When the nozzle travels to a grid in the prescription map where the base spray volume is zero, the duty cycle is set to zero, the solenoid valve is fully closed, and the instantaneous flow rate of the nozzle is zero.

[0040] In this embodiment, the modulation frequency of the pulse width modulation signal is 10 Hz. The response time of the solenoid valve is less than 50 milliseconds, and each modulation cycle is 100 milliseconds at the 10 Hz modulation frequency, which can stably achieve the switching between opening and closing actions. At the 10 Hz modulation frequency and a travel speed of 1 m / s, the vehicle advances 0.1 meters in each modulation cycle, which is less than the diameter of a typical wolfsbane canopy, thus achieving spray positioning accuracy at the canopy scale.

[0041] In this embodiment, the preset threshold for isolated plants is 2 plants. This threshold is determined based on a combination of the ecological characteristics of *Euphorbia fischeriana* and the efficiency of control operations: when the number of *Euphorbia fischeriana* plants within a 1-meter radius is less than or equal to 2, the average minimum distance between plants is usually more than 40 centimeters, indicating that they are independent growing individuals with a low probability of root cross-linking, making them suitable for physical removal one by one; when the number of plants is greater than 2, the root systems are mostly intertwined, and physical removal is likely to cause large-area turf damage and significantly reduce operational efficiency. In this case, targeted chemical spraying is more efficient.

[0042] Those skilled in the art will recognize that the specific technical means involved in this case, such as UAV remote sensing, spectral analysis, point cloud processing, classification and clustering algorithms, robotic arm control, and variable spraying, are well-known technologies in the field. The following mainly introduces the working principles and processes.

[0043] The workflow of this method is as follows: Before the operation, the joint extrinsic parameter calibration and hardware time synchronization of the hyperspectral imager and lidar are completed. During the peak flowering or vegetative growth period of *Euphorbia humifusa*, a drone flies over the work area to simultaneously acquire hyperspectral images and lidar point clouds. The hyperspectral images are sequentially processed through black-and-white correction, minimum noise separation and denoising, continuous projection algorithm to select feature bands, and random forest classifier for pixel-by-pixel binary classification to obtain a pixel-level *Euphorbia humifusa* classification map. The lidar point clouds are sequentially processed through statistical outlier removal filter for denoising, cloth simulation filter to separate ground points, and Euclidean distance clustering to segment plant clusters, and the three-dimensional centroid of each cluster is calculated. The pixel classification results are mapped to plant clusters through collinearity equations and extrinsic parameter calibration matrices. *Euphorbia humifusa* plants are identified as having a *Euphorbia humifusa* pixel proportion exceeding 60% within the projection area, and the canopy center coordinates and canopy equivalent diameter are extracted. Using the canopy center of each *Euphorbia humifusa* plant as the center, the number of other *Euphorbia humifusa* plants within a 1-meter radius is counted as the neighboring *Euphorbia humifusa* plant count. The system compares the number of neighboring *Euphorbia fischeriana* plants with the isolated plant threshold of 2. If there are 2 or fewer plants, the system sends a physical removal control command and target coordinates to the autonomous walking platform. After navigating to the target location, the platform controls the end effector of the physical removal actuator to cut vertically into the soil to sever the roots and remove the plant, using the normal vector of the ground point cloud fitted to the ground plane within 0.5 meters around the center of the plant's canopy as the cutting direction and a cutting depth of 7 to 10 centimeters. Safety protection logic is executed when encountering high resistance or insufficient point cloud. If there are more than 2 plants, the system sends a variable chemical spraying control command and target coordinates to the autonomous walking platform. After navigating to the target location, the platform determines the instantaneous spraying volume based on the variable spraying prescription map and the correction coefficient calculated from the equivalent diameter of the canopy. A 10 Hz pulse width modulation signal is generated to adjust the duty cycle of the solenoid valve of the variable spraying actuator to control the nozzle flow rate. The nozzles are automatically shut off in areas without *Euphorbia fischeriana*. After the entire operation is completed, an operation report is generated. This method organically integrates hyperspectral fine identification, lidar spatial positioning, intelligent distribution pattern discrimination, and differentiated precision operation, realizing intelligent and precise prevention and control based on individual plant conditions.

[0044] In the description of this specification, the terms "one embodiment," "some embodiments," "embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0045] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make modifications, alterations, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for controlling Daphne odora and Euphorbia fischeriana in grasslands, characterized in that, Includes the following steps: S1: Acquire hyperspectral images and lidar point clouds of the work area; S2: Perform feature band extraction and pixel classification on the hyperspectral image to obtain pixel-level wolfsbane classification results; S3: Filter and spatially cluster the lidar point cloud to segment plant object clusters and calculate the spatial position of each plant object cluster. S4: The pixel-level wolfsbane classification result is fused with the plant object cluster to identify the wolfsbane plant and determine the canopy center coordinates and canopy size of the identified wolfsbane plant. S5: Determine the number of neighboring *Euphorbia fischeriana* plants for each *Euphorbia fischeriana* plant based on the coordinates of the canopy center; S6: Compare the number of neighboring wolfsbane plants with a preset isolated plant threshold. If the number of neighboring wolfsbane plants of a wolfsbane plant is less than or equal to the preset isolated plant threshold, generate a physical uprooting control command for the wolfsbane plant. If it is greater than the preset isolated plant threshold, generate a variable chemical spraying control command for the wolfsbane plant. S7-1: For the wolfsbane plant that generates the physical uprooting control command, extract the local ground point cloud within a preset radius around the center of the canopy of the wolfsbane plant, fit the ground plane, and generate the end effector control command according to the normal vector of the ground plane to control the end effector to cut into the ground at a preset depth in a direction perpendicular to the ground plane to cut the root. S7-2: For the *Euphorbia fischeriana* plant that generates the variable chemical spraying control command, plant canopy outline information is generated based on the determined canopy size of the *Euphorbia fischeriana* plant, and a nozzle flow control signal is generated based on the canopy outline information to adjust the nozzle flow in real time for variable application.

2. The method for controlling Daphne odora and Euphorbia fischeriana in grasslands according to claim 1, characterized in that, Step S2 involves feature band extraction and pixel classification of the hyperspectral image, including the following steps: The hyperspectral image is subjected to black-and-white correction and noise removal; The characteristic bands that distinguish wolfsbane from the background are selected from the bands of the hyperspectral image by using a continuous projection algorithm or a competitive adaptive reweighted sampling algorithm. Using the aforementioned feature bands as input, a support vector machine classifier or a random forest classifier is used to perform binary classification of each pixel between wolfsbane and background, thereby obtaining the pixel-level wolfsbane classification result.

3. The method for controlling Daphne odora and Euphorbia fischeriana in grasslands according to claim 1, characterized in that, Step S3 involves filtering and spatial clustering the lidar point cloud, including the following steps: A statistical outlier removal filter is used to remove noise from the lidar point cloud. A cloth-simulation filtering algorithm is used to separate ground points from non-ground points; The non-ground points are clustered using an Euclidean distance clustering algorithm or a region growing algorithm to segment independent plant object clusters, and the three-dimensional centroid coordinates of each plant object cluster are calculated as the spatial location of the plant object cluster.

4. The method for controlling Daphne odora and Euphorbia fischeriana in grasslands according to claim 1, characterized in that, Step S4 involves fusing the pixel-level wolfsbane classification result with the plant object cluster to identify the wolfsbane plant, including: The pixel-level wolfsbane classification results are projected onto the plant object cluster through coordinate mapping; If the number of pixels classified as wolfsbane within the projection area of ​​a plant cluster exceeds 60%, then the plant cluster is determined to be the wolfsbane plant.

5. The method for controlling Daphne odora and Euphorbia fischeriana in grasslands according to claim 1, characterized in that, Step S5 involves determining the number of neighboring *Euphorbia humifusa* plants for each of the aforementioned *Euphorbia humifusa* plants, including: Taking the center of the canopy of each of the aforementioned wolfsbane plants as the center, count the number of wolfsbane plants within a preset radius, which is taken as the number of neighboring wolfsbane plants of the aforementioned wolfsbane plant.

6. The method for controlling Daphne odora and Euphorbia fischeriana in grasslands according to claim 5, characterized in that, The preset radius is 1 meter.

7. The method for controlling Daphne odora and Euphorbia fischeriana in grasslands according to claim 1, characterized in that, Step S7-1 specifically includes the following steps: Extract the ground point cloud within a 0.5-meter radius around the center of the canopy of the target *Euphorbia milii* plant; The random sampling consensus algorithm is used to perform plane fitting on the ground point cloud to obtain the normal vector and elevation of the local ground plane; Based on the normal vector and the elevation, instructions are generated to control the cutting angle and cutting depth of the end shovel, so that the cutting direction of the end shovel is parallel to the normal vector and cuts into the ground to a depth of 7-10 cm to sever the root.

8. The method for controlling Daphne odora and Euphorbia fischeriana in grasslands according to claim 1, characterized in that, Step S7-2 specifically includes the following steps: Pre-generate variable spray prescription maps; The basic spraying amount is determined according to the variable spraying prescription diagram, and the basic spraying amount is corrected by using the plant canopy outline information generated based on the canopy size of the wolfsbane plant. A pulse width modulation signal is generated to adjust the duty cycle of the solenoid valve, thereby controlling the instantaneous flow rate of the nozzle. When there are no *Euphorbia fischeriana* plants in the area to be sprayed, the instantaneous flow rate of the nozzle is controlled to be zero.

9. The method for controlling Daphne odora and Euphorbia fischeriana in grasslands according to claim 8, characterized in that, The modulation frequency of the pulse width modulation signal is 10 Hz.

10. The method for controlling Daphne odora and Euphorbia fischeriana in grasslands according to claim 1, characterized in that, The preset threshold for isolated plants is 2 plants.

Images