Method for efficient removal of weeds in rape seed production breeding process

Abstract

The application provides a method for efficiently removing weeds in the breeding process of oilseed rape, and belongs to the technical field of agricultural seed production. The method comprises the following steps: obtaining the shape and area parameters of the oilseed rape field, combining the optimal plant spacing and the size of the intelligent mobile robot to divide the field into a grid shape, the grid being a sowing block, and sowing one oilseed rape seed; when the oilseed rape seed grows to a preset stage, moving the intelligent mobile robot carrying an image recognition and weed removal execution module to the front of a typical oilseed rape plant confirmed by a human, identifying an image and extracting features by the image recognition module to obtain target image features; driving the robot to move along a moving path, identifying all oilseed rape plants in the field according to the target image features by the image recognition module, and starting the weed removal execution module to remove weeds when the plants are identified as weeds. Through the steps of "precise planning, marker calibration and automatic identification and removal of weeds", the efficiency, accuracy and cost control of weed removal are considered, and technical support is provided for the intelligent and large-scale development of oilseed rape seed production.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of agricultural seed production, and in particular to a method for efficient removal of weeds in the process of oilseed rape seed production. BACKGROUND

[0002] In the process of oilseed rape seed production, the seedling stage is a key growth stage. Removing weeds (including grasses, plants with significant phenotypic differences from the planting population, etc.) in the field at this stage can optimize the breeding growth environment and effectively reduce the workload of subsequent patrol and weed removal during the bolting and flowering stages. Therefore, seedling weed removal is a key process in the process of hybrid oilseed rape seed production.

[0003] In the prior art, the removal of weeds in the seedling stage of oilseed rape seed production is mainly achieved by manually patrolling the field and visually comparing the weeds to identify and remove them. This method has the problem of low efficiency and is difficult to adapt to the needs of large-scale oilseed rape seed production.

[0004] The current seed industry has gradually developed towards intelligent seed production that integrates biotechnology, artificial intelligence, and big data information technology. Intelligent equipment developed based on machine vision, deep learning, and data fusion artificial intelligence technologies has been applied in field agriculture, providing technical feasibility for intelligent identification and removal of weeds in the seedling stage of oilseed rape seed production. However, the deployment cost of mature weed intelligent identification models is currently high, limiting the application of such artificial intelligence technologies in the seedling weed removal scenario of oilseed rape seed production and making it difficult to be widely applied. SUMMARY

[0005] The present application aims to provide a method for efficient removal of weeds in the process of oilseed rape seed production. Through the synergistic technical logic of "precise planning, marker calibration, and automated identification and removal of weeds", the efficiency, accuracy, and cost control of weed removal are considered, effectively solving the core technical problems of existing seedling weed removal in oilseed rape seed production and providing reliable technical support for the intelligent and large-scale development of oilseed rape seed production.

[0006] The technical solution adopted by the present application is as follows:

[0007] A method for efficient removal of weeds in the process of oilseed rape seed production, comprising the following steps:

[0008] Step S1: Obtain the shape and area parameters of the oilseed rape field, and based on the shape and area parameters, combine the preset optimal plant spacing and the size of the intelligent mobile robot to perform grid division. Each grid is a planting block, and the grid lines intersected vertically and horizontally are the movement path of the intelligent mobile robot. In each planting block, a predetermined number of oilseed rape seeds are sown;

[0009] Step S2, when the oilseed rape plants grow to a preset stage, the intelligent mobile robot carrying an image recognition module and a weed removal execution module is moved to a typical oilseed rape plant confirmed by manual operation to meet the cultivation requirements, and the image recognition module is used to perform image recognition and feature extraction to obtain target image features;

[0010] Step S3, the intelligent mobile robot is driven to move along the moving path, and the image recognition module is used to identify all the oilseed rape plants in the oilseed rape field based on the target image features; if the current oilseed rape plant is identified as a weed, the weed removal execution module is started to remove the weed;

[0011] In the step S2 and the step S3, when the image recognition module performs image recognition on the typical oilseed rape plant and the oilseed rape plant, the image recognition module adopts a fixed posture and uses a tilt angle to compensate in real time to keep the shooting angle constant, and performs scale normalization and perspective correction on the shooting image based on the grid physical size of the seeding block, and unifies the plant height measurement reference with the relative elevation of the field ground, so that the equivalent shooting distance and imaging scale between the lens and each plant are kept consistent.

[0012] Further, in the step S3, before the image recognition module identifies the oilseed rape plant, the actual growth position coordinates of the current oilseed rape plant are obtained and compared with the seeding position coordinates corresponding to the seeding block; if the actual growth position is not within a preset threshold range of the seeding position, the weed removal execution module is directly started to remove the weed; if the actual growth position is within the preset threshold range of the seeding position, the image recognition module is used to identify and determine based on the target image features.

[0013] Further, the preset threshold range of the seeding position is adjusted in real time according to the plant width size corresponding to the preset growth stage of the oilseed rape seedling, and the larger the plant width size of the oilseed rape seedling, the larger the preset threshold range of the seeding position.

[0014] Further, in the step S2, the image recognition module extracts multi-dimensional image features of the typical oilseed rape plant in terms of shape, texture and color, and removes abnormal pixel points in the extracted features to obtain refined target image features, and in the step S3, the weed identification is performed based on the refined target image features.

[0015] Further, in the step S3, when the image recognition module identifies the oilseed rape plant, a feature hierarchical comparison logic is used, the core features of the plant shape are compared first, and if the matching degree of the shape features is lower than a preset threshold, the plant is directly determined as a weed; and after the shape features are matched, the texture and color auxiliary features are compared.

[0016] Furthermore, in step S2, before performing image recognition and feature extraction, the image recognition module first performs adaptive correction of the rapeseed field environment on the collected typical rapeseed plant images. The adaptive correction of the rapeseed field environment includes light intensity compensation, soil background grayscale removal, and plant edge sharpening processing.

[0017] Furthermore, in step S3, when the intelligent mobile robot moves along the moving path, it dynamically adjusts its moving speed according to the actual emergence status of the rapeseed plants in each sowing block; when there are seedlings emerging in the sowing block and identification work needs to be completed, the moving speed is reduced; when there are no seedlings emerging in the sowing block, the moving speed is increased.

[0018] Furthermore, in step S3, the image recognition module identifies the current rapeseed plant as a hybrid plant based on the target image features. Before starting the hybrid removal execution module to remove the hybrid plant, the hybrid plant is first subjected to secondary image acquisition and feature verification. When the verification result is consistent with the first recognition result, the removal operation is then performed.

[0019] Further, after extracting the target image features in step S2, the soil fertility and light duration microenvironment parameters of the sowing block where the typical rapeseed plant is located are simultaneously acquired and stored; in step S3, when identifying each rapeseed plant, the real-time microenvironment parameters of the sowing block where the rapeseed plant is located are first acquired, and the difference between the real-time microenvironment parameters of the sowing block where the typical rapeseed plant is located is calculated. The comparison threshold of the target image features is dynamically corrected according to the difference ratio. The larger the difference of the microenvironment parameters, the wider the fault tolerance range of the comparison threshold.

[0020] Furthermore, in step S3, after the intelligent mobile robot completes a round of whole-field identification and weed removal, it also performs secondary feature extraction and comparison on the rapeseed plants that have not been removed. For the rapeseed plants whose feature matching is blurry in the first identification, the secondary extracted image features are finely matched with the target image features. Based on the matching result, it is determined whether the plant is a weed and the corresponding removal operation is performed.

[0021] The beneficial effects of this invention are:

[0022] 1. By dividing the fields into grids and setting the movement path of intelligent mobile robots, the robot can automatically inspect the entire field and identify weeds, replacing the traditional manual field inspection and visual comparison weed removal methods. This significantly reduces the amount of manual operation and solves the problems of low efficiency and high labor intensity of manual weed removal. It can be adapted to rapeseed seed production fields of different sizes and meet the operational needs of large-scale seed production.

[0023] 2. Without relying on large-scale sample training or high-cost general intelligent recognition models, a typical rapeseed plant is manually identified as the standard plant, and the target image features are extracted by the image recognition module. This simplifies the feature acquisition process, eliminates the need for complex model training procedures, effectively reduces the application cost of intelligent impurity removal technology, and facilitates its promotion and application in rapeseed seed production scenarios.

[0024] 3. Using manually confirmed typical rapeseed plants as standard plants ensures that the target image features match the cultivation requirements, avoids the problem of poor compatibility between general standard plants and actual cultivated varieties, provides an accurate reference standard for subsequent whole-field identification of hybrid plants, reduces identification deviations caused by unclear standard plants, and lays the foundation for accurate removal of hybrid plants.

[0025] 4. By identifying and removing weeds and plants that do not meet the cultivation requirements in advance, a good growth competition environment is created for the target rapeseed plants, reducing the competition between weeds and target plants for nutrients and light. At the same time, it can effectively reduce the workload of patrolling and removing weeds during the bolting and flowering periods, indirectly reducing the labor cost of the entire breeding process and ensuring the purity and breeding quality of rapeseed seeds.

[0026] 5. The core process revolves around field planning, standard plant identification, and automated identification and weed removal. Each step is logically coherent and technologically mature. The movement path of the intelligent mobile robot is designed based on field parameters and plant spacing, adapting to rapeseed seedling growth scenarios with different plant widths and densities. It does not require significant modifications to the existing seed production process and is highly practical. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 This is a flowchart illustrating an efficient method for removing hybrid plants during rapeseed seed production and cultivation, as provided in Example 1. Detailed Implementation

[0029] The embodiments of the invention will now be described in detail with reference to the accompanying drawings.

[0030] Example 1

[0031] Please see Figure 1 This embodiment provides a method for efficiently removing hybrid plants during rapeseed seed production and cultivation. The method includes the following steps:

[0032] Step S1: Obtain the shape and area parameters of the rapeseed field, and based on the shape and area parameters, combined with the preset optimal plant spacing and the size of the intelligent mobile robot, divide the field into grids. Each grid is a sowing block, and the crisscrossing grid lines are the movement path of the intelligent mobile robot. Sow a preset number of rapeseed seeds in each sowing block and record the sowing position coordinates simultaneously. In this embodiment, one rapeseed seed is placed in the center of each sowing block as an example.

[0033] Among them, "shape parameters" can be the boundary contour information of the rapeseed field, such as the coordinate sequence of polygon vertices or the aspect ratio of the minimum bounding rectangle, used to characterize the geometric shape of the field. "Area parameters" can be the total area of ​​the field, in square meters, used to support the calculation of grid density and grid size. "Preset optimal plant spacing" is an empirical parameter set based on the minimum nutrient space distance required for the target rapeseed variety during the seedling stage, which can reflect the lateral spacing required between plants to avoid competition. "Intelligent mobile robot size" refers to the maximum projected size of its physical shape, including the overall length and width, wheel width, etc., used to avoid structural interference in path planning. Specifically, the structure of the intelligent mobile robot used in this embodiment is existing technology, such as an adaptive addition based on the functions required in this embodiment of "An Automatic Rapeseed Transplanter and Its Automatic Transplanting Method" (CN118104444B), which does not require creative effort from those skilled in the art. "Grid-like division" involves dividing the field into several equidistant or approximately equidistant rectangular units. The centers of each unit constitute a theoretical set of planting points. The intervals between these divisions in both the longitudinal and transverse directions are constrained by the optimal plant spacing and the robot's dimensions, ensuring sufficient clearance between adjacent planting blocks. The planting position of each block is defined by the geometric center coordinates of that grid unit, serving as the theoretical landing point for a single seed. This step, through spatial structuring and pre-setting, enables spatial predictability of subsequent plant distribution, providing prior positional constraints for image recognition and a deterministic topological basis for robot path planning.

[0034] In one alternative implementation, the grid-like partitioning method can be as follows: first, estimate the total number of theoretically sown blocks based on the field area parameters and the optimal plant spacing; then, combine the shape parameters to fit the boundary constraints; and finally, use an adaptive grid generation algorithm to divide the field area, outputting the center coordinates and coverage of each sown block.

[0035] Step S2: Once the rapeseed seed has grown and developed to the preset stage, the intelligent mobile robot equipped with an image recognition module and a weed removal module is moved to a typical rapeseed plant that has been manually confirmed to meet the cultivation requirements. The image recognition module then performs image recognition and feature extraction to obtain the target image features.

[0036] The "preset stage" refers to a stable growth stage in rapeseed seedlings, such as when the number of true leaves reaches 3-5 and the plant height is within the range of 8-15 cm. At this time, the plant morphological characteristics tend to be stable and background interference is relatively controllable. A "typical rapeseed plant" is a representative healthy individual manually selected by breeding technicians based on variety standards, field growth, and the absence of pests and diseases. It serves as a visual reference for whole-field identification and can be one or more plants. This embodiment uses one typical rapeseed plant as an example. The "image recognition module" is a common industrial-grade vision component integrating a visible light CMOS sensor, a multi-degree-of-freedom robotic arm, a fixed-focus lens, and an embedded image processing unit. It has automatic white balance, exposure compensation, and distortion correction capabilities. Simultaneously, to reduce interference from factors such as the distance between the lens and the plant, ground flatness, and shooting angle, various sensors mounted on an intelligent mobile robot are used for real-time detection and dynamic adjustment of the "image recognition module." During the shooting process, a unified measurement benchmark is used to eliminate interference from changes in shooting angle, differences in ground flatness, and differences in shooting distance between the lens and each plant on the measurement of the plant's basic visual characteristics and subsequent statistical calculation results. "Target image features" is a set of low-dimensional vectors extracted from typical plant images by this module, which can characterize the variety-specific morphology, texture and color distribution patterns, and is used for subsequent comparison and recognition.

[0037] In this embodiment, this step avoids the deployment threshold of general deep learning models by using manual calibration and local feature extraction, forming a lightweight, low-cost, and highly adaptable recognition benchmark construction mechanism.

[0038] In one alternative implementation, the image recognition and feature extraction method can be: controlling the image recognition module to acquire five images (left view, right view, front view, rear view, and top view) of a typical plant at a fixed distance and orthophoto angle, and sequentially performing grayscale normalization, background segmentation, contour extraction, and Hu moment calculation to generate a 7-dimensional invariant moment feature vector as the target image feature.

[0039] Step S3: Drive the intelligent mobile robot to move along the movement path, and the image recognition module identifies all rapeseed plants in the rapeseed field based on the target image features; if the current rapeseed plant is identified as a weed, the weed removal module is activated to remove it; wherein, when the typical rapeseed plant and the rapeseed plant are image recognized in steps S2 and S3 respectively, the image recognition module adopts a fixed posture and maintains a constant shooting angle through real-time tilt compensation, and performs scale normalization and perspective correction on the captured image based on the physical size of the grid of the sowing block, and combines the relative elevation of the ground of the plot to unify the plant height measurement benchmark, so that the equivalent shooting distance and imaging scale between the lens and each plant are consistent.

[0040] The "movement path" is a set of deterministic trajectories formed by the grid lines in step S1, possessing topological continuity and full spatial coverage, supporting the robot's sequential traversal of all sowing blocks. "Identification based on target image features" refers to the image recognition module calculating the similarity between the real-time acquired images of the rapeseed plants to be identified and the target image features obtained in step S2, determining whether they belong to the same phenotypic category based on a preset threshold. "Hybrids" include weeds that are not the target rapeseed variety, as well as variant individuals or hybrid varieties that, although belonging to the same family, significantly deviate from typical rapeseed plants in morphology, texture, or color. The "weed removal execution module" is a miniature gripper, rotary cutting blade, or pneumatic suction device mounted at the end of a multi-degree-of-freedom robotic arm. Its actions are controlled by the image recognition module's feedback recognition results to execute the removal action, and by real-time pose information for coordinated linkage, ensuring that it only acts on the target plant without damaging adjacent normal plants.

[0041] In this embodiment, this step achieves an integrated "identification-decision-action" workflow through path-driven + feature comparison + closed-loop execution, ensuring that the entire field is identified without omission and that the rejection response is timely.

[0042] In one alternative implementation, the method for identification based on target image features can be: extracting feature vectors of the same dimension from each frame of the acquired image, calculating the Euclidean distance between the feature vector and the target image features, and determining that the feature vector is a hybrid when the distance is greater than a preset threshold.

[0043] In this embodiment, a spatial mapping relationship between sowing blocks and movement paths is established through grid-like division, enabling the location of rapeseed plant distribution. Lightweight target image features are constructed using manually identified typical rapeseed plants, avoiding reliance on complex models and reducing layout costs. Intelligent mobile robots perform whole-field identification and removal along predetermined paths, forming a technical closed loop of synergistic effect between structured field management and localized visual recognition. Based on this, only a few calibration operations are needed to support automated weed removal across the entire field, significantly reducing deployment costs and operational barriers, while ensuring identification accuracy and operational coverage. This effectively solves the core contradictions of existing technologies, such as low manual efficiency, difficulty in implementing intelligent systems, and weak scalability.

[0044] For situations where two or more rapeseed seeds (e.g., three seeds) are simultaneously sown in a single planting hole within each sowing block to propagate into multiple rapeseed plants, or where a single rapeseed seed is simultaneously sown in a planting hole at different locations within each sowing block to propagate into multiple rapeseed plants (e.g., three planting holes, one seed per hole), or where two or more rapeseed seeds are simultaneously sown in a planting hole at different locations within each sowing block to propagate into multiple rapeseed plants (e.g., three planting holes, two seeds per hole), the process for each rapeseed plant should be performed according to the aforementioned procedure. After the image recognition module acquires the image of the rapeseed plants within the corresponding sowing block, it is necessary to first segment the image containing multiple rapeseed plants to obtain the image of each individual rapeseed plant before performing subsequent processing. It should be noted that the relative positions of each planting hole within each sowing block should remain consistent.

[0045] Example 2

[0046] This embodiment provides a method for efficiently removing hybrid plants during rapeseed seed production. Compared with Embodiment 1, it further provides that before the image recognition module in step S3 identifies the rapeseed plant, it first obtains the actual growth position coordinates of the current rapeseed plant and compares them with the sowing position coordinates of the corresponding sowing block. If the actual growth position is not within the preset threshold range of the sowing position, the hybrid removal execution module is directly activated to remove it. If it is within the preset threshold range of the sowing position, the image recognition module then identifies and determines it based on the target image features.

[0047] The "actual growth position coordinates" refer to the geospatial coordinates obtained by mapping the pixel coordinates output by the positioning module (e.g., RTK-GNSS module or visual-inertial joint positioning unit) and image recognition module on the intelligent mobile robot. These coordinates represent the actual landing point of the rapeseed plant in the Cartesian coordinate system of the field. The "sowing position coordinates of the sowing block" are preset coordinates corresponding to the sowing positions within each sowing block determined by the grid division in step S1, and are known fixed values. The comparison between these two coordinates is used to determine whether the current rapeseed plant falls within the spatial distribution tolerance range of its sowing block. For example, in this embodiment, the method for obtaining and comparing these coordinates can be as follows: During the movement of the intelligent mobile robot along the path, the position information output by its own high-precision positioning system is called in real time. Combined with the calibration relationship between the current camera's field of view center and the robot's body coordinate system, the center of the plant pixels identified in the image is back-projected onto the ground coordinate system to obtain the actual growth position coordinates. Then, the Euclidean distance between these coordinates and the pre-stored sowing position coordinates of the sowing block is calculated to determine whether it is less than or equal to a preset threshold.

[0048] The "preset threshold range for sowing location" is a spatial tolerance area with the coordinates of the sowing location in the sowing block as the center and a preset radius. This length can be dynamically set according to the developmental stage of the rapeseed seedlings. For example, it can be ±2 cm at the cotyledon stage, ±4 cm at the early true leaf stage, and ±6 cm at the 5-leaf stage. This threshold range does not depend on the image content; it only reflects the reasonable displacement boundary of the rapeseed seedlings under normal growth conditions due to soil compaction, micro-topographical undulations, or sowing machinery errors. When the actual growth location coordinates exceed this range, it indicates that the plant is highly likely not to belong to the typical rapeseed plant species identified manually, or is a non-target type such as a weed or a self-grown seedling. Therefore, it does not need to enter the image recognition stage; the weed removal module is directly triggered to perform physical removal, significantly reducing the risk of misjudgment and saving recognition resources.

[0049] The phrase "within the range of..." indicates that the spatial deviation between the actual growth location coordinates and the center coordinates of the sowing block meets the preset distance constraint, constituting a necessary prerequisite for the image recognition module to initiate the recognition task. This condition ensures that all plants participating in the image comparison possess basic spatial rationality, allowing the image recognition process to focus on candidate objects that may truly exhibit phenotypic differences, thereby improving the targeting and reliability of feature comparison results. Based on this, the image recognition module can then call upon the target image features extracted in step S2 to perform similarity analysis on the current plant image in dimensions such as morphology, texture, and color, completing the final determination of hybrid plants.

[0050] In this embodiment, a spatial consistency screening layer is constructed by introducing a comparison mechanism between the sowing location coordinates and the actual growth location coordinates of the sowing block, using geographical distribution patterns as the first identification threshold. Based on this, the image recognition process is initiated only for spatially compliant plants, realizing a hierarchical decision-making logic of "coarse screening followed by fine judgment." This not only reduces the computational load and response latency of the image recognition module but also effectively avoids the risk of misclassifying off-center seedlings as qualified plants due to environmental factors such as wind disturbance, soil compaction, and seedling lodging. This enhances the robustness and engineering practicality of the overall impurity removal strategy, while ensuring the stable implementation of the lightweight identification paradigm established in this application, which uses typical single-plant samples as a reference.

[0051] Meanwhile, this embodiment also provides a technical solution that adjusts the preset threshold range of sowing position in real time according to the plant width size corresponding to the preset growth stage of rapeseed seedlings. The larger the plant width size of rapeseed seedlings, the larger the preset threshold range of sowing position.

[0052] The "preset threshold range for sowing location" refers to a circular tolerance area centered on the sowing location of each sowing block and with a preset distance as the radius. This area is used to determine whether the current rapeseed plant's actual growth position is within the reasonable emergence space of the sowing block; the radius of this tolerance area is the preset threshold range. "Rapeseed seedling width" refers to the maximum horizontal projection width of the rapeseed plant canopy at a specific preset growth stage, representing the spatial scale occupied by the plant during that stage of natural growth. "Real-time adjustment" refers to dynamically retrieving the corresponding plant width reference value based on the current actual growth stage of the rapeseed seedling and updating the preset threshold range accordingly, without relying on manual settings or fixed values.

[0053] In this embodiment, the preset threshold range for sowing location is positively correlated with the size of the rapeseed seedling. Its function is to match the positional tolerance to the biological expansion pattern of the plant—in the early seedling stage, the plant width is small, and the plant is sensitive to center positioning, so the tolerance range is correspondingly reduced to avoid false rejection due to slight deviations; in the middle and late seedling stage, the plant width increases significantly, and the natural extension of the plant easily causes center point deviation, so the tolerance range expands synchronously to prevent misjudgment of non-hybrid plants caused by plant width expansion. This adjustment mechanism does not change the physical division structure of the sowing block, but only dynamically corrects the position determination boundary to ensure that the recognition logic always adapts to the actual growth state.

[0054] This application establishes a positive mapping relationship between a preset threshold range for sowing location and the size of rapeseed seedlings, introducing a plant growth time dimension into the initial location screening stage. This ensures that rapeseed plants at different developmental stages within the same field can obtain location tolerance matching their biological characteristics. Based on this, a two-stage judgment process of location comparison and image recognition is combined, ensuring the rigor of early seedling identification while improving the robustness of mid-to-late-stage large seedling identification. This synergistically enhances the accuracy and rationality of weed identification and removal, avoiding systematic misjudgment bias caused by static threshold settings.

[0055] Example 3

[0056] This embodiment provides a method for efficient removal of hybrid plants during rapeseed seed production. Compared with Embodiment 1, it also provides that the image recognition module in step S2 extracts multi-dimensional image features of the typical rapeseed plant in terms of morphology, texture, and color, and removes abnormal pixels from the extracted features to obtain refined target image features. In step S3, hybrid plant identification is carried out based on the refined target image features.

[0057] "Morphological features" refer to visual attributes that reflect the overall outline structure and spatial distribution patterns of rapeseed plants, including but not limited to leaf extension directionality, leaf margin curvature changes, plant height-to-canopy ratio, leaf arrangement, leaf extension angle distribution, canopy outline compactness, and main stem-lateral branch topology. In this embodiment, these morphological features are used to characterize the standard plant structure of typical rapeseed plants in the seedling stage, providing a geometric benchmark for subsequent identification. "Texture features" refer to visual attributes that reflect the alternating light and dark areas, density distribution, and local grayscale variations on the surface of rapeseed leaves. They can characterize the density of leaf surface tissue, waxy layer coverage, and micro-deformation features. In this embodiment, these texture features are used to enhance the ability to distinguish similar background interference in the field (such as bare soil areas and dead leaf debris). "Color features" refer to visual attributes that reflect the spectral reflectance response characteristics of rapeseed seedling leaves in the visible light band. These features may include the RGB three-channel mean, HSV hue component concentration, and the difference between the a* and b* components in the Lab space. In this embodiment, these color features are used to capture the stable color performance of typical plants under current light and soil conditions, improving the sensitivity to non-target plants (such as yellowing weed seedlings or closely related hybrids of the cruciferous family). The above three types of features together constitute a multi-dimensional joint feature vector. The dimensional combinations can be concatenated, weighted fusion, or principal component mapping to form a more comprehensive and robust digital representation of typical rapeseed plants.

[0058] Here, "abnormal pixels" refer to outlier data units in the image feature space that significantly deviate from their neighborhood statistical distribution. These may originate from lens smudges, strong reflective spots, shadow occlusion causing mis-sampling, or sensor noise during image acquisition. In this embodiment, abnormal pixel identification is achieved by combining sliding window local variance analysis with global Z-score threshold determination. Specifically, the local standard deviation of each feature channel is first calculated using a 3×3 pixel neighborhood as a unit. Then, the absolute Z-score of all pixels in each channel is compared with a preset threshold. If the Z-score of any channel exceeds the threshold and continuously appears in multiple consecutive neighborhoods, it is determined to be an abnormal pixel. This abnormal pixel removal operation can be implemented using interpolation repair, channel masking, or feature dimension clipping, aiming to reduce the impact of noise on the stability of subsequent feature comparisons and ensure the data purity and representativeness of the target image features.

[0059] In one alternative implementation, the multi-dimensional image feature extraction method may be: extracting the morphological contour parameters of the plant region based on the image segmentation results, using the contrast and entropy values ​​derived from the gray-level co-occurrence matrix (GLCM) as texture features, and using the peak range of the H component histogram in the HSV space as color features.

[0060] This embodiment constructs a composite feature representation with high discriminative power and environmental adaptability by simultaneously extracting three complementary visual features: morphology, texture, and color. An abnormal pixel identification and removal mechanism controls data quality at the feature generation source, preventing noise propagation to subsequent comparison stages. Based on this, the obtained refined target image features not only accurately depict the inherent phenotypic characteristics of typical rapeseed plants in the seedling stage but also effectively suppress interference from complex field lighting, background clutter, and image acquisition disturbances. This supports the high-confidence identification and judgment of rapeseed plants across the field in step S3, improving the accuracy and robustness of weed removal.

[0061] Meanwhile, this application also provides a feature-layer comparison logic used by the image recognition module in step S3 when recognizing rapeseed plants. First, the core features of plant morphology are compared. If the morphological feature matching degree is lower than the preset threshold, it is directly determined to be a hybrid plant. After the morphological feature matching meets the standard, the texture and color auxiliary features are compared.

[0062] The "feature hierarchical comparison logic" refers to dividing the image recognition process into multiple comparison levels with priority order. Each level has a logical dependency, and the execution of the next level depends on the current level's judgment result meeting preset conditions. This logic does not change the composition of the target image features themselves, but only adjusts their calling order and judgment path in the recognition process. "Core features of plant morphology" are key image features characterizing the overall geometric structure and spatial distribution of rapeseed plants, including but not limited to leaf arrangement, leaf extension angle distribution, canopy outline compactness, and main stem-lateral branch topological relationship. In the field of plant phenotypic recognition, morphological features typically possess strong species differentiation ability and low environmental sensitivity, stably reflecting the inherent genetic phenotype of a variety. In this embodiment, this feature is used to construct the first line of defense for rapid initial screening—when the similarity between the core morphological features of the plant to be identified and the target image features is lower than a preset threshold, the system can terminate the recognition process and output a "hybrid" judgment result without continuing subsequent comparisons. "Texture and color auxiliary features" are image features that supplement the description of the microstructure and spectral reflectance of the plant surface. Among them, texture features can reflect the roughness of the leaf surface, the consistency of the vein direction and the changes in tissue density, while color features can reflect the chlorophyll content level, the distribution of lesions or yellowing areas and other state information. In this embodiment, these features are activated and called only after the morphological core features are matched and meet the criteria. As a secondary verification method, they are used to finely distinguish individuals with similar morphology but abnormal physiological state or mixed varieties, thereby avoiding omissions or false rejections caused by misjudgment due to a single dimension.

[0063] In one optional implementation, the feature-layer comparison method can be as follows: The image recognition module first constructs a reference morphological template based on the morphological core features extracted from the target image features, and after scale normalization and pose correction of the image to be recognized, calculates the structural similarity index (SSIM) between it and the reference template; if the SSIM value is lower than a preset threshold, it is directly determined to be a hybrid; if the SSIM value reaches or exceeds the threshold, the local binary pattern (LBP) texture features and the hue mean and saturation variance in the HSV color space of the image to be recognized are further extracted, and the cosine similarity is compared with the corresponding texture and color auxiliary features in the target image features to comprehensively determine whether it is a hybrid.

[0064] This embodiment constructs a two-level recognition architecture of "morphology first, auxiliary verification later," which significantly reduces the computing resources and response time required for single-plant recognition while ensuring recognition accuracy. It leverages the high robustness of core morphological features to complete rapid initial screening, greatly reducing the number of plants that need to enter the fine comparison stage. On this basis, texture and color auxiliary features are used to perform differentiated verification on plants that pass the initial screening, which avoids the computing power redundancy caused by full-dimensional synchronous comparison and maintains the sensitivity to recognize subtle phenotypic differences. Finally, it realizes the efficient, stable, and low-power operation of the seedling hybrid plant recognition task on the field edge computing device, adapting to the computing power constraints and real-time requirements of the intelligent mobile robot platform.

[0065] Example 4

[0066] This embodiment provides a method for efficient removal of hybrid plants during rapeseed seed production and cultivation. Compared with embodiment 1, it also provides that in step S2, before performing image recognition and feature extraction, the image recognition module first performs adaptive correction of the rapeseed field environment on the collected typical rapeseed plant images. The adaptive correction of the rapeseed field environment includes light intensity compensation, soil background grayscale removal, and plant edge sharpening.

[0067] "Light intensity compensation" refers to the dynamic equalization of the overall brightness distribution of an image to eliminate local overexposure or underexposure caused by changes in the sun's altitude angle, cloud cover, or shadow projection. This technical feature is a basic illumination normalization operation in the field of image processing, and its inherent function is to improve the consistency of the image's grayscale dynamic range. In this embodiment, its role is to ensure that key areas such as leaf color and stem nodes of a typical rapeseed plant can present stable and comparable brightness responses under different natural lighting conditions, providing a light-robust input image for subsequent feature extraction. "Soil background grayscale removal" refers to the process of separating the main plant area using an adaptive threshold segmentation method based on the significant difference in grayscale distribution between the target plant and the soil background in the image spatial domain, while suppressing or filtering out soil background pixels dominated by low grayscale values. This technique is often used in agricultural image analysis for foreground-background decoupling, and its inherent function is to reduce the interference of non-target areas on the feature extraction process. In this embodiment, its role is to eliminate the erroneous interference caused by background components such as bare soil, gravel, and dead leaves on the extraction of morphological and texture features, ensuring that the extracted target image features only reflect the phenotypic information of the typical rapeseed plant itself. Among them, "plant edge sharpening" refers to enhancing the gray-level gradient difference between the pixels of the plant outline edge and its neighborhood in the image to highlight geometric structural details such as leaf extension direction, leaf edge serrations, and stem branching points. This technical feature belongs to the classic spatial domain enhancement operation in computer vision, and its inherent function is to enhance the spatial structural expression ability of the target object. In this embodiment, its role is to improve the visualization clarity of key morphological features of typical rapeseed plants (such as rosette leaf arrangement, basal leaf arrangement density, leaf tip directionality, etc.), thereby supporting the high-confidence comparison and judgment based on morphological features in subsequent steps.

[0068] This application forms a progressively adaptive environmental correction chain by sequentially performing light intensity compensation, soil background grayscale removal, and plant edge sharpening: light intensity compensation provides a base image with consistent brightness for subsequent processing; on this basis, soil background grayscale removal further focuses on the main area of ​​the plant, reducing feature deviations introduced by background noise; finally, plant edge sharpening enhances the expression of structural details on the purified image, enabling the target image features to more fully carry the variety-specific morphological semantics of rapeseed plants; the synergistic effect of these three processes significantly improves the image quality stability and feature expression reliability of the image recognition module under varying field lighting and complex soil backgrounds, thereby ensuring the accuracy and generalization ability of heterophyte identification in step S3.

[0069] Example 5

[0070] This embodiment provides a method for efficiently removing hybrid plants during rapeseed seed production and cultivation. Compared with Embodiment 1, it also provides that when the intelligent mobile robot moves along the moving path in step S3, the moving speed is dynamically adjusted according to the actual emergence status of the rapeseed plants in each sowing block; when there are seedlings in the sowing block and identification work needs to be completed, the moving speed is reduced; when there are no seedlings in the sowing block, the moving speed is increased.

[0071] The "actual seedling emergence status" refers to the status information of whether visible rapeseed seedlings exist, determined by a rapid visual scan of the central area of ​​each sowing block using an image recognition module. This status information does not rely on the recognition of complete plant features; it can be obtained solely based on pixel brightness distribution, green vegetation index threshold response, and the existence of the target outline in the central area. It is a coarse-grained perception result with low computational overhead. In this embodiment, it serves as a pre-trigger condition for movement speed control, reserving sufficient image acquisition time for subsequent high-precision recognition and avoiding redundant dwelling and image acquisition operations on empty blocks, thereby shortening the overall inspection cycle.

[0072] In one alternative implementation, the method for dynamically adjusting the moving speed is as follows: the image recognition module, in conjunction with the output of the infrared thermal imaging sensor, simultaneously acquires visible light images and surface micro-temperature difference images in the central area of ​​the sowing block. By fusing and analyzing the reflectance of green vegetation and the characteristics of local temperature gradient changes, it determines whether there are signs of live seedling growth. If both support the conclusion of seedling emergence, the deceleration logic is triggered.

[0073] This embodiment uses the actual seedling emergence status of the sowing area as the basis for adjusting the movement speed. In areas with seedlings requiring identification, the intelligent mobile robot proactively reduces its speed to ensure image acquisition frame rate and clarity, while in areas without seedlings, it accelerates its movement to reduce wasted travel time. Based on this, by utilizing a coarse-grained rapid seedling emergence status discrimination mechanism and a multi-source perception / prediction collaborative strategy, it avoids mechanical wear and increased energy consumption caused by frequent start-stop operations, while ensuring the timeliness and completeness of the identification work. Ultimately, without increasing hardware costs, it significantly improves the overall operational efficiency and system energy efficiency ratio of weed identification and removal across the entire field.

[0074] Example 6

[0075] This embodiment provides a method for efficiently removing hybrid plants during rapeseed seed production and cultivation. Compared with Embodiment 1, it also provides that in step S3, the image recognition module identifies the current rapeseed plant as a hybrid plant based on the target image features. Before starting the hybrid removal execution module to remove the hybrid plant, the hybrid plant is first subjected to secondary image acquisition and feature verification. When the verification result is consistent with the first recognition result, the removal operation is then performed.

[0076] The initial identification result includes the identification confidence level, matching feature dimension, and judgment criterion category. The imposters to be verified are suspected imposter candidates output by the image recognition module; their identification results have not yet triggered the impurity removal action and are only used as input for secondary verification. The initial identification result drives the subsequent image resampling and feature comparison process and is the trigger condition for initiating the verification mechanism.

[0077] The "second image acquisition" refers to the independent image data re-acquired for the same physical plant within a similar spatial location and time window after the initial identification and before the removal of impurities. The image acquisition process can adjust the focal length, exposure parameters, or shooting angle to enhance the robustness of feature representation, but does not change the semantic content of the image. The verification image is used to provide redundant visual information and alleviate the feature distortion problem caused by instantaneous occlusion, sudden changes in lighting, or motion blur in a single acquisition.

[0078] The "secondary image acquisition and feature verification of hybrid plants" refers to a standardized feature extraction path consisting of the same processing logic, the same feature dimensions (morphology, texture, color), and the same abnormal pixel removal strategy. The verified image features are structured feature vectors output by this path, and their dimensions and semantic levels correspond to the target image features. The verification recognition result is an independent judgment conclusion based on this comparison, and its judgment logic and threshold setting are completely equivalent to the initial recognition. Only when both recognition results are "hybrid plants" will the output be "consistent," and all other cases (including "normal plant / hybrid plant," "hybrid plant / normal plant," and "normal plant / normal plant") are considered "inconsistent." The removal instruction is a control signal with a clear execution priority, including target coordinates, execution sequence, and action type. The inconsistency event record includes the initial recognition confidence level, the verification recognition confidence level, the time difference between the acquisition of the two frames of images, the change in ambient light intensity, and the feature matching difference dimension, which is used to support system reliability analysis and misjudgment attribution.

[0079] This embodiment introduces a dual verification mechanism by introducing two independent image acquisition-recognition-comparison closed loops, significantly improving decision robustness without changing the original recognition paradigm. The initial recognition result triggers a review process, avoiding computational redundancy caused by repeated recognition of all plants. The correlation between the review image and the initial image in the spatiotemporal neighborhood ensures that the two feature extractions have a comparable basis. Finally, a logical interlock is formed at the recognition result level, initiating physical intervention only when both independent recognitions point to the same conclusion. This effectively suppresses the risk of false rejection caused by transient interference, protects the integrity of the target rapeseed population, and improves field operation safety and user trust.

[0080] Example 7

[0081] This embodiment provides a method for efficiently removing hybrid plants during rapeseed seed production. Compared with Embodiment 1, it further provides that after extracting the target image features in step S2, the soil fertility and light duration microenvironment parameters of the sowing block where the typical rapeseed plant is located are simultaneously acquired and stored; in step S3, when identifying each rapeseed plant, the real-time microenvironment parameters of the sowing block where the rapeseed plant is located are first acquired, and the difference is calculated with the microenvironment parameters of the sowing block where the typical rapeseed plant is located. The comparison threshold of the target image features is dynamically adjusted according to the difference ratio. The larger the difference of the microenvironment parameters, the wider the fault tolerance range of the comparison threshold.

[0082] "Soil fertility" refers to a comprehensive indicator reflecting the effective content of major nutrients such as nitrogen, phosphorus, and potassium, as well as the level of organic matter in the soil. In agricultural monitoring, it is usually indirectly characterized by physicochemical parameters such as soil electrical conductivity, pH value, and concentration of available nutrients. "Light duration" refers to the cumulative effective photosynthetic radiation received by the sowing block within a unit growth cycle, usually measured in average daily hours or the length of continuous light periods. Together, they constitute key microenvironmental variables affecting the phenotypic expression of rapeseed seedlings. In this embodiment, soil fertility parameters and light duration parameters are used to establish an environmental benchmark reference system for typical plants, serving as the basis for dynamic adjustment of image feature comparison thresholds during subsequent identification. The parameters are collected in real time by a multi-source sensor array deployed at the bottom of the intelligent mobile robot, and stored in a local database after being bound to the spatial coordinates of the corresponding sowing block for use in step S3.

[0083] Among them, "real-time microenvironment parameters" refers to the soil fertility and light duration data collected by the same sensor array for the planting block to which the rapeseed plant to be identified belongs at the moment of performing step S3 identification operation; "difference calculation" refers to performing absolute difference calculation on parameters of the same type, i.e. ΔF = |F current - F reference |、ΔL = |L current - L reference| where F represents the soil fertility parameter, L represents the light duration parameter, and the subscripts current and reference represent the current plant's block and the typical plant's block, respectively. "Difference ratio" is a normalized comprehensive deviation measure, obtained by weighted summation after interval mapping of ΔF and ΔL, used to quantify the degree of deviation of the current plant's microenvironment from the standard plant's environment. "Comparison threshold" is the critical value for the image recognition module to determine whether the target image feature match is valid; its initial setting is based on the stability of the feature distribution extracted from the typical plant in a standard microenvironment. "Expanded tolerance range" refers to positively shifting the original comparison threshold according to a preset nonlinear function relationship, allowing for moderate dispersion of target image features in morphology, texture, and color dimensions under conditions of significant microenvironmental differences, still resulting in a successful match. This avoids false rejection caused by normal plant phenotypic variations due to environmental stress.

[0084] In one optional implementation, the dynamic correction method for the comparison threshold is as follows: the correction coefficient is obtained by looking up a table based on the difference ratio, and then the original comparison threshold is multiplied by the coefficient to obtain the updated comparison threshold; the table is constructed based on several pre-calibrated typical environmental deviation samples and their corresponding manual confirmation and identification results to ensure that the correction logic conforms to agronomic experience judgment.

[0085] This embodiment upgrades the original static image feature comparison mechanism into an environment-aware dynamic adaptation mechanism by introducing microenvironmental parameters as contextual constraint variables in the image recognition process. By calculating the difference between two key microenvironmental parameters, soil fertility and light duration, an environmental similarity metric between the current plant and typical plants is established. Based on this, the target image feature comparison threshold is corrected in a hierarchical, continuous, or lookup-table manner according to the difference ratio, enabling the recognition logic to actively respond to phenotypic disturbances caused by the heterogeneity of the microenvironment within the field. Ultimately, while ensuring the accuracy of identifying hybrid plants, the risk of misjudging normal plants induced by insufficient nutrient supply or limited light is significantly reduced, improving the robustness and adaptability of the system in real and complex farmland scenarios.

[0086] Example 8

[0087] This embodiment provides a method for efficient removal of hybrid plants during rapeseed seed production. Compared with Embodiment 1, it further includes the following steps: after the intelligent mobile robot completes a round of whole-field identification and removal in step S3, it performs secondary feature extraction and comparison on the rapeseed plants that have not been removed. For rapeseed plants with ambiguous feature matching in the first identification, the image features extracted in the second step are finely matched with the target image features. Based on the matching results, it is determined whether the plant is a hybrid plant and the corresponding removal operation is performed.

[0088] Among them, "rapeseed plants with ambiguous feature matching in the first identification" refers to plants in the first round of identification in step S3 where the matching degree between the current rapeseed plant image features and the target image features calculated by the image recognition module is within the preset fuzzy judgment interval. The fuzzy judgment interval is between the threshold for miscellaneous plants and the threshold for qualified plants, and its numerical range can be dynamically set according to the actual identification confidence distribution. This state condition feature is used to screen the pending objects that need to enter the closed-loop review process, avoid repeated processing of high-confidence identification results, and improve the system operating efficiency. "Feature secondary extraction and comparison" refers to the process where, without changing the shooting angle, lighting conditions, and imaging parameters, the image recognition module re-captures an image of the same plant and re-executes feature extraction and comparison operations. This technique has a clear functional purpose in the field of image recognition technology: to enhance recognition robustness through repeated sampling and feature reconstruction, and to reduce the risk of misjudgment caused by single-imaging noise, posture deviation, or local occlusion. In this embodiment, this action is limited to specific plants whose first-round recognition results are in the fuzzy range. The input is the location coordinates of the plant marked as "to be verified" in the first-round recognition and the corresponding sowing block number. The output is the updated recognition confidence and binary judgment result, thereby supporting subsequent differentiated treatment decisions. "Refined matching" refers to the image feature comparison process performed in the secondary comparison stage by using higher-dimensional feature space mapping, finer-grained similarity measurement strategies, and stricter logical judgment rules. In the field of computer vision, this technical feature is usually manifested as expanding the feature vector dimension, introducing local structural constraints, increasing the comparison level, or extending the matching path. In this embodiment, its specific function is to improve the ability to distinguish between hybrid types that are similar in shape but sensitive to subtle differences in texture / color, so that the secondary judgment result is more effective in distinguishing compared to the first round of recognition, thereby improving the accuracy and stability of the final hybrid removal judgment.

[0089] This embodiment establishes a closed-loop verification mechanism for secondary feature extraction and comparison of rapeseed plants with ambiguous feature matching in the first round of identification, and strengthens the discrimination ability of boundary samples with the help of a refined matching strategy. On this basis, it combines multiple optional implementation methods to adapt to different operational constraints, which not only ensures the rigor of identification and judgment, but also takes into account the flexibility of system deployment. Finally, without relying on large-scale labeled data and complex model training, it significantly improves the thoroughness of weed removal and the purity and reliability of remaining plants, effectively supporting the intelligent and standardized upgrade of the entire rapeseed seed production process.

Claims

1. A method for efficiently removing hybrid plants during rapeseed seed production and cultivation, characterized in that, Includes the following steps: Step S1: Obtain the shape and area parameters of the rapeseed field, and based on the shape and area parameters, combined with the preset optimal plant spacing and the size of the intelligent mobile robot, divide the field into grids. Each grid is a sowing block, and the crisscrossing grid lines are the movement path of the intelligent mobile robot. Sow a preset number of rapeseed seeds in each sowing block. Step S2: When the rapeseed seed has grown and developed into a rapeseed plant at a preset stage, the intelligent mobile robot equipped with an image recognition module and a weed removal execution module is moved to a typical rapeseed plant that has been manually confirmed to meet the cultivation requirements, and the image recognition module performs image recognition and feature extraction to obtain the target image features. Step S3: Drive the intelligent mobile robot to move along the movement path, and the image recognition module identifies all the rapeseed plants in the rapeseed field based on the target image features; if the current rapeseed plant is identified as a weed, the weed removal module is activated to remove it. In step S2 and step S3, when performing image recognition on the typical rapeseed plant and the rapeseed plant, the image recognition module adopts a fixed posture and maintains a constant shooting angle through real-time tilt compensation. At the same time, the physical size of the grid of the sowing block is used as a reference to perform scale normalization and perspective correction on the captured image, and the plant height measurement reference is unified by combining the relative elevation of the ground of the plot, so that the equivalent shooting distance and imaging scale between the lens and each plant are consistent. After extracting the target image features in step S2, the soil fertility and light duration microenvironment parameters of the sowing block where the typical rapeseed plant is located are simultaneously acquired and stored. In step S3, when identifying each rapeseed plant, the real-time microenvironment parameters of the sowing block where the rapeseed plant is located are first acquired, and the difference between the real-time microenvironment parameters of the sowing block where the typical rapeseed plant is located is calculated. The comparison threshold of the target image features is dynamically adjusted according to the difference ratio. The larger the difference of the microenvironment parameters, the wider the fault tolerance range of the comparison threshold.

2. The method for efficient removal of hybrid plants during rapeseed seed production and cultivation according to claim 1, characterized in that, In step S3, before the image recognition module identifies the rapeseed plant, it first obtains the actual growth position coordinates of the current rapeseed plant and compares them with the sowing position coordinates of the corresponding sowing block; if the actual growth position is not within the preset threshold range of the sowing position, the impurity removal execution module is directly activated to remove it. If the target image is within a preset threshold range at the planting location, the image recognition module will then make a determination based on the target image features.

3. The method for efficient removal of hybrid plants during rapeseed seed production and cultivation according to claim 2, characterized in that, The preset threshold range for sowing location is adjusted in real time according to the plant width size corresponding to the preset growth stage of the rapeseed seedling. The larger the plant width size of the rapeseed seedling, the larger the preset threshold range for sowing location.

4. The method for efficient removal of hybrid plants during rapeseed seed production and cultivation according to claim 1, characterized in that, In step S2, the image recognition module extracts multi-dimensional image features of the typical rapeseed plant in terms of morphology, texture, and color, and removes abnormal pixels from the extracted features to obtain refined target image features. In step S3, hybrid plant identification is carried out based on the refined target image features.

5. The method for efficient removal of hybrid plants during rapeseed seed production and cultivation according to claim 4, characterized in that, In step S3, when the image recognition module identifies rapeseed plants, it uses a feature-layered comparison logic to first compare the core features of plant morphology. If the morphological feature matching degree is lower than a preset threshold, it is directly determined to be a hybrid plant. After the morphological features meet the matching criteria, the auxiliary features of texture and color are compared.

6. The method for efficient removal of hybrid plants during rapeseed seed production and cultivation according to claim 1, characterized in that, In step S2, before performing image recognition and feature extraction, the image recognition module first performs adaptive correction of the rapeseed field environment on the collected typical rapeseed plant images. The adaptive correction of the rapeseed field environment includes light intensity compensation, soil background grayscale removal, and plant edge sharpening.

7. The method for efficient removal of hybrid plants during rapeseed seed production and cultivation according to claim 1, characterized in that, In step S3, when the intelligent mobile robot moves along the moving path, it dynamically adjusts its moving speed according to the actual emergence status of the rapeseed plants in each sowing block; when there are seedlings emerging in the sowing block and identification work needs to be completed, the moving speed is reduced; when there are no seedlings emerging in the sowing block, the moving speed is increased.

8. The method for efficient removal of hybrid plants during rapeseed seed production and cultivation according to claim 1, characterized in that, In step S3, the image recognition module identifies the current rapeseed plant as a hybrid plant based on the target image features. Before starting the hybrid removal execution module to remove the hybrid plant, the hybrid plant is first subjected to secondary image acquisition and feature verification. When the verification result is consistent with the first recognition result, the removal operation is then performed.

9. The method for efficient removal of hybrid plants during rapeseed seed production and cultivation according to any one of claims 1 to 8, characterized in that, In step S3, after the intelligent mobile robot completes a round of whole-field identification and weed removal, it also performs secondary feature extraction and comparison on the rapeseed plants that have not been removed. For the rapeseed plants whose feature matching is blurry in the first identification, the secondary extracted image features are finely matched with the target image features. Based on the matching result, it is determined whether the plant is a weed and the corresponding removal operation is performed.

Images