Unmanned aerial vehicle-based intelligent pruning method for garden plants

Abstract

The application discloses a garden plant intelligent pruning method based on a UAV, and belongs to the technical field of intelligent maintenance of landscaping. The method comprises the following steps: collecting color images of a garden area by a UAV and obtaining a landscape planning digital model; identifying plant varieties and extracting morphological and physiological parameters of the plant varieties by using a deep learning model; calling a pruning knowledge base rule set corresponding to the plant varieties; fusing real-time parameters and design information, and generating an individualized scheme comprising pruning positions, intensities and processes through rule reasoning; and finally, superimposing the scheme and the images to form a visual guidance file, and sending the visual guidance file to a mobile terminal to guide on-site operation. The method solves the problems of low efficiency, insufficient accuracy and difficulty in implementing design intent caused by the dependence of traditional garden pruning on manual experience, and through data driving and intelligent decision-making, realizes the automation, precision and standardization of pruning operation, and significantly improves the quality and efficiency of garden maintenance.

Description

Technical Field

[0001] This invention relates to the field of intelligent maintenance technology for landscaping, and in particular to an intelligent pruning method for landscaping plants based on unmanned aerial vehicles (UAVs). Background Technology

[0002] Regular pruning of garden plants is an important maintenance measure to maintain the aesthetic effect of the landscape, ensure the healthy growth of plants, and enable them to perform specific functions such as shading, isolation, and guidance. This need is widespread in urban parks, residential green spaces, roadside green belts, and large scenic areas. In these settings, the plant species are diverse and numerous, and often have specific landscape design requirements, making pruning a continuous, labor-intensive, and technically complex task. Especially in the maintenance of some highly ornamental garden plants such as bougainvillea, due to their rapid growth, intricate branches, and frequent use in landscaping, the frequency, precision, and artistry of pruning are even higher, making traditional manual pruning methods inadequate.

[0003] Currently, mainstream garden plant pruning relies heavily on the personal experience and visual judgment of horticultural workers. Workers determine the pruning location, amount, and method for each plant based on their understanding of the plant's growth habits, seasonal characteristics, and desired shape. This approach has several inherent limitations. First, it is inefficient and inconsistent. For large-scale gardens, manual inspection and assessment are time-consuming and labor-intensive, and different workers may have varying understandings and implementations of the same standards, leading to inconsistent pruning results and making it difficult to achieve uniformity and harmony in the overall landscape. Second, it is highly experience-dependent and lacks scientific rigor. The quality of pruning decisions largely depends on the worker's skill level. Without objective, quantitative data support, judgments about the plant's internal physiological state (such as latent dead branches or abnormal leaf color) may be inaccurate, and the impact of the plant's long-term growth trend on current pruning decisions may be overlooked. Furthermore, landscape designers' design intentions are usually expressed in drawings or concepts; relying entirely on workers' interpretation and reproduction during on-site pruning can lead to information loss or bias, making it difficult to accurately reproduce or maintain the original design in the final pruning result.

[0004] With technological advancements, some research and applications have attempted to introduce modern equipment to assist in pruning. For example, drones are used for park inspections and photography, or ground-based image analysis technology is used to monitor the growth of individual plants. However, these technologies mostly focus on status monitoring or single-point analysis, and have not yet formed a closed-loop workflow from status perception and intelligent decision-making to precise execution. Specifically, existing technologies often lack deep integration with digital models of landscape design, resulting in a lack of direct, quantifiable correlation between monitoring data and maintenance objectives. In the decision-making stage, there is a lack of an intelligent reasoning system that can integrate botanical knowledge, growth status data, and landscape design rules. In the execution stage, there is still a general lack of effective means to translate abstract decision results into intuitive and unambiguous operational guidelines that on-site workers can understand. Therefore, the field of garden plant pruning still has a need for an intelligent overall solution that can systematically improve the accuracy, efficiency, and scientific nature of operations. Summary of the Invention

[0005] This invention overcomes the problems of low efficiency, insufficient precision, and difficulty in implementing design intentions caused by the reliance on manual experience in traditional garden pruning. Through data-driven and intelligent decision-making, it realizes the automation, precision, and standardization of pruning operations, significantly improving the quality and efficiency of garden maintenance.

[0006] To achieve the above objectives, the present invention adopts the following solution: A drone-based intelligent pruning method for garden plants includes the following steps: S1: Using a drone equipped with an RGB camera, fly along a preset route and collect color digital images of the target garden area. At the same time, obtain the landscape planning digital model of the area from the landscape management database. The landscape planning digital model includes the design location of plants, the design shape outline, and functional zoning information. S2: Input the acquired color digital images into the pre-trained first deep learning model to identify and output the variety identifier of each plant in the image. At the same time, input the color digital images into the pre-trained second deep learning model to perform image semantic segmentation and extract the morphological and structural parameters and physiological state parameters of each plant. The morphological and structural parameters include crown size, branch length and branch spatial distribution, and the physiological state parameters include leaf color distribution and dead branch area marking. S3: Retrieve the set of plant pruning knowledge base rules corresponding to the variety identifier from the landscape management database. The set of plant pruning knowledge base rules consists of plant physiological pruning rules and landscape shaping pruning rules. The plant physiological pruning rules are associated with the variety identifier and growth status parameters, while the landscape shaping pruning rules are associated with the design outline and functional zoning information. S4: For each identified plant, the corresponding extracted morphological and structural parameters, physiological state parameters, and the design outline and functional zoning information of the plant extracted from the landscape planning digital model are integrated. Logical matching and reasoning are performed according to the plant pruning knowledge base rule set to generate an individual plant pruning guidance plan that includes pruning location, pruning intensity, and pruning technique. S5: Overlay and register the individual plant pruning guidance plan with the corresponding color digital image to form a guidance document with visual pruning marks and send it to mobile terminal devices to guide staff in pruning operations.

[0007] Preferably, step S1 specifically includes: Based on the boundary vector data of the target garden area stored in the landscape management database, a parallel scanning route with full coverage and a set overlap rate is generated. The UAV is controlled to fly along the parallel scanning route during periods above a set light intensity threshold, with an RGB camera acquiring color digital images at set time intervals in a vertically downward orientation. During the UAV's flight, its onboard real-time differential global positioning system and inertial measurement unit record the acquisition position and attitude data of each frame of color digital image. After acquisition, based on the acquisition position and attitude data, a motion reconstruction structure algorithm is used to generate a 3D point cloud model of the park with geographic coordinates from the color digital images. The 3D point cloud model of the park is then matched with the landscape planning digital model using coordinate system one and spatial registration to establish a precise correspondence between visual features in the color digital images and design elements in the landscape planning digital model. The plant design outlines in the landscape planning digital model are defined by a 3D mesh model, and functional zoning information is stored as attribute data in the primitives of the 3D mesh model.

[0008] As a preferred option, in step S2, the pre-trained first deep learning model adopts a cascaded network structure. First, the region proposal network is used to locate the bounding box of each plant in the color digital image. Then, the image region within each bounding box is classified to output the variety identifier. The pre-trained second deep learning model is a fully convolutional network with an encoder-decoder structure, used for pixel-level semantic segmentation of color digital images. The segmentation categories include leaves, branches, flowers, fruits, and background. The crown size in the morphological structure parameters is obtained by calculating the diagonal length of the minimum bounding rectangle formed by the leaf and branch pixels in the semantic segmentation results. The spatial distribution of branches is obtained by calculating the angular variance of the skeletonized lines of the branch pixels in the image. Both the first and second deep learning models were trained using a multi-seasonal garden image dataset, which includes color digital images of the same plant in spring, summer, autumn, and winter, labeled with corresponding variety identifiers and semantic segmentation labels. During training, data augmentation techniques were employed, including random rotation, scaling, and color jitter.

[0009] As a preferred method, when extracting physiological state parameters, for leaf color distribution, the mean of the red channel, the mean of the green channel, the mean of the blue channel, and the corresponding standard deviation are extracted from the leaf pixel region in the semantic segmentation results, and the normalized difference vegetation index is calculated based on these extracted values; for dead branch region labeling, a support vector machine classifier is trained, and based on the gray-level co-occurrence matrix texture features and HSV color space features of the branch pixel region, the branches are distinguished into healthy branches and dead branches, and the pixels classified as dead branches are clustered into continuous dead branch regions through the region growing algorithm; Morphological and physiological parameters are stored as structured data records for each plant. The data records include variety identifier, bounding box coordinates, crown size, branch length, angular variance of branch spatial distribution, normalized difference vegetation index, and the coordinates of the bounding rectangle of the dead branch area. The data records are associated with the design outline of the corresponding plant extracted from the landscape planning digital model through a unique identifier.

[0010] Preferably, in step S3, the set of plant pruning knowledge base rules is stored in the knowledge base of the rule engine in the form of production rules, and each rule has an "IF-THEN" structure. The "IF" part of the plant physiological pruning rule consists of conditions connected by logical AND relations, including matching of variety identifiers, judgment of the numerical range of morphological structure parameters, and judgment of the threshold of physiological state parameters. The "IF" part of the landscape shaping pruning rule, in addition to containing the conditions of the plant physiological pruning rule, also includes matching of functional zoning information and judgment of deviation from the design outline. The "THEN" part of all rules defines one or more pruning operation suggestions. When performing logical matching and reasoning, the rule engine performs forward chain reasoning on the rules in the knowledge base based on the currently input variety identifier, morphological structure parameters, physiological state parameters, design outline, and functional zoning information, activating all rules whose conditions are met, and adjudicating potentially conflicting pruning operation suggestions according to preset rule priorities, outputting a set of conflict-free pruning operation suggestions.

[0011] As a preferred option, the rule priority is preset according to the pruning goal, wherein the priority of landscape pruning rules related to maintaining the design outline is higher than the priority of plant physiological pruning rules that are only related to the healthy growth of plants. In the logical matching and reasoning process of the rule engine, the deviation between the morphological structural parameters and the design outline is judged by comparing the crown size, branch length and branch spatial distribution in the morphological structural parameters with the corresponding design dimensions and design angle range defined in the three-dimensional mesh model of the design outline to calculate the quantitative deviation value. The rule engine has pre-defined threshold judgment sub-rules for different deviation ranges. When the quantified deviation value exceeds the preset first threshold, the rule with shape restoration as the primary pruning goal is activated. When the quantified deviation value is lower than the preset second threshold, the rule with routine health maintenance as the primary pruning goal is activated. The set of conflict-free pruning operation suggestions is combined with the predefined pruning operation parameterization template. The pruning operation suggestions are mapped to specific pruning position coordinates, pruning tool entry angle and branch cut length through the pruning operation parameterization template, thereby generating individual plant pruning guidance schemes.

[0012] Preferably, step S4, which integrates morphological parameters, physiological parameters, and the designed shape outline, specifically includes: A unified spatial coordinate system is established for each plant. The spatial distribution of branches in the morphological structure parameters, the dead branch area in the physiological state parameters, and the three-dimensional mesh model of the design outline are aligned and mapped in this spatial coordinate system. Then, the volume overlap between the current actual branch distribution and the branch design area in the design outline is calculated, and the actual branch parts that exceed the volume boundary of the design outline and the dead branch area that is within the volume boundary of the design outline are identified. Logical matching and reasoning are based on volume overlap, spatial coordinates of actual branches that exceed the design outline, and spatial coordinates of dead branch areas. The pruning position in the individual plant pruning guidance plan is determined based on the spatial coordinates of the actual branches that exceed the design outline and the spatial coordinates of dead branch areas. The pruning intensity is calculated based on the volume overlap and the preset pruning tolerance coefficient of the plant variety. The pruning process is selected from the predefined process library based on the level and diameter of the branch's spatial coordinates.

[0013] As a preferred approach, after generating individual plant pruning guidelines, a pre-evaluation is performed, specifically including: The morphological and structural parameters are loaded as the initial state, and the corresponding plant growth parameterization model is called according to the variety identifier. After simulating the pruning position, pruning intensity, and pruning process in the computer, the growth direction and length of new branches in the next growth cycle are predicted. The prediction results are used to calculate the predicted volume overlap between the simulated branch distribution and the designed shape outline. If the predicted volume overlap is lower than the current volume overlap, it is determined that the plant individual pruning guidance plan does not meet the long-term shape maintenance goal. At this time, the combination of pruning intensity and pruning position is adjusted, and the simulation and evaluation are repeated until the predicted volume overlap is not lower than the current volume overlap. Finally, the adjusted plant individual pruning guidance plan is output.

[0014] Preferably, in step S5, the guidance document is generated in the following manner: The system receives individual plant pruning guidelines, corresponding color digital images, and a 3D point cloud model of the park obtained from the landscape management database. It analyzes the pruning location coordinates in the individual plant pruning guidelines and finds the corresponding 3D spatial points in the park's 3D point cloud model. Based on the real-time geographic location, spatial attitude, and real-time scene images acquired by the mobile terminal device's built-in GPS receiver, inertial measurement unit, and camera, a feature-based image matching algorithm is used to match the real-time scene images with the color digital images, calculating the real-time relative position and attitude between the mobile terminal device's camera and the target plant. Based on the real-time relative position and attitude, the 3D spatial points corresponding to the pruning location coordinates are rendered as virtual pruning markers with 3D spatial attributes, which are then overlaid on the real-time scene image stream displayed on the mobile terminal device screen, forming a guidance document with visual pruning markers. After receiving the guidance document, the mobile terminal device displays a real-time augmented reality view on the screen with virtual trimming markers overlaid to assist the operator in positioning and trimming.

[0015] As a preferred approach, the landscape management database is constructed and updated in the following ways: Multi-angle images of the planted garden area are acquired using a multi-lens oblique camera mounted on a drone. The multi-angle images are then processed using a motion recovery structure algorithm to generate a high-precision real-scene 3D mesh model containing accurate color information. Based on this real-scene 3D mesh model, a pre-trained instance segmentation model identifies and labels the 3D outline of each plant and its spatial coordinates within the model, using these coordinates as initial data for the plant's design location. Horticultural designers then use 3D modeling software to refine and redesign the 3D outline of each identified plant, defining its long-term maintenance design outline 3D mesh model and assigning variety identification and functional zoning information to each plant. The initial design location data, the 3D mesh model of the design outline, the variety identification, and the functional zoning information are linked and stored as a plant record in the landscape planning digital model data table of the landscape management database. When plants are replanted or replaced in the garden area, the above acquisition, identification, design, and linking steps are repeated to generate a new plant record and update the landscape planning digital model data table.

[0016] The present invention has at least the following beneficial effects: (1) By rapidly collecting data through drones and combining deep learning models to automatically identify plant varieties, analyze morphology and physiological state, it replaces the traditional experience judgment that relies entirely on manual labor, significantly improving the efficiency and objectivity of state assessment. It generates individualized pruning plans through intelligent reasoning and transforms the plans into visual guidance documents, making on-site operations based on evidence and greatly reducing human error and operational differences, thereby improving the overall accuracy, speed and standardization of pruning operations; (2) By generating a high-precision real-world 3D model and registering it with the landscape planning digital model, a precise mapping relationship between the real world and digital design is established, ensuring that the generated pruning plan can directly serve the fundamental goal of adjusting and aligning the actual form of plants with the established design shape, effectively solving the problem of design and maintenance disconnect; (3) The robustness of state analysis is ensured by the deep learning model trained with multi-season data; and the plant physiological rules and landscape aesthetic rules are digitized through a structured knowledge base and rule engine. Encapsulation and integrated reasoning; By setting priority and conflict resolution mechanisms, the contradictions in multi-objective decision-making are properly handled, which can adapt to the needs of different seasons and different management priorities, and the scientificity and systematicness of decision-making are significantly enhanced; (4) By conducting detailed volume overlap analysis in three-dimensional space and simulating pre-evaluation and iterative optimization of pruning scheme based on plant growth model, it is possible to predict and avoid pruning decisions that may be detrimental to the long-term health or shape stability of plants. It not only focuses on solving the current morphological deviation problem, but also guides plants to grow healthily towards the ideal design shape, thereby enhancing the strategic nature and long-term benefits of maintenance work; (5) Using augmented reality technology, virtual pruning marks are accurately superimposed onto the real scene video stream captured by the mobile terminal, providing intuitive guidance for on-site workers, transforming complex spatial location instructions into clear visual prompts, greatly simplifying the positioning and operation process of workers, reducing the requirements for workers' personal experience and spatial imagination, and effectively avoiding mis-pruning and missed pruning, significantly improving the operability and first-time success rate of complex pruning tasks. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the principle of one method of the present invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings, so that those skilled in the art can implement it based on the description.

[0019] like Figure 1 As shown, the intelligent pruning method for garden plants based on unmanned aerial vehicles (UAVs) provided by this invention includes the following steps: S1: Using a drone equipped with an RGB camera, fly along a preset route and collect color digital images of the target garden area. At the same time, obtain the landscape planning digital model of the area from the landscape management database. The landscape planning digital model includes the design location of plants, the design shape outline, and functional zoning information.

[0020] During missions, drones fly along parallel scanning routes generated by the system for full coverage. Route design typically ensures an overlap of 20% to 30% between adjacent flight strips to guarantee the integrity of image stitching and 3D reconstruction. RGB cameras automatically capture color digital images at fixed time intervals or distances during periods of good lighting (e.g., midday when light intensity is high), positioned vertically downwards. The landscape management database stores digital models of the landscape planning, containing the intended design information for each plant within the area. This includes design outlines defined by 3D mesh models, design locations defined by coordinate points or bounding boxes, and functional zoning information associated with attributes (e.g., viewing areas, protected areas, rest areas, etc.). This step enables rapid, digital acquisition of the current state of the garden and establishes a link with the established design blueprint, providing a unified spatiotemporal and data foundation.

[0021] S2: Input the acquired color digital images into the pre-trained first deep learning model to identify and output the variety identifier of each plant in the image. At the same time, input the color digital images into the pre-trained second deep learning model to perform image semantic segmentation and extract the morphological and structural parameters and physiological state parameters of each plant. The morphological and structural parameters include crown size, branch length and branch spatial distribution, and the physiological state parameters include leaf color distribution and dead branch area marking.

[0022] The first model is a deep learning model for plant variety identification. It can automatically detect each plant in an image and output its variety identifier, such as "bougainvillea," "ginkgo," "osmanthus," and "privet." The second model is a deep learning model for image semantic segmentation. It performs pixel-level classification on the input image, distinguishing categories such as leaves, branches, flowers, fruits, and background. From the segmentation results, quantitative parameters for each plant are further extracted: morphological parameters include crown size calculated from pixels (e.g., the diagonal length of the envelope rectangle formed by the outermost branches and leaves), the estimated length of the main branches, and the spatial distribution dispersion (e.g., angular variance) obtained by analyzing the pixel skeleton of the branches; physiological state parameters include vegetation indices (e.g., NDVI) calculated by analyzing the color values ​​of leaf pixels (e.g., the mean and standard deviation of each RGB channel), and dead branch areas identified through texture and color feature classification. Both models are trained using a multi-season, multi-angle garden image dataset and employ data augmentation techniques to improve generalization ability.

[0023] S3: Retrieve the set of plant pruning knowledge base rules corresponding to the variety identifier from the landscape management database. The set of plant pruning knowledge base rules consists of plant physiological pruning rules and landscape shaping pruning rules. The plant physiological pruning rules are associated with the variety identifier and growth status parameters, while the landscape shaping pruning rules are associated with the design outline and functional zoning information.

[0024] Based on the identified plant variety identifiers, the system retrieves the corresponding rule set from the plant pruning knowledge base associated with the landscape management database. The knowledge base rule set consists of two parts: first, physiological pruning rules focusing on healthy plant growth, such as thinning principles for specific varieties in specific seasons and pruning suggestions for signs of pests and diseases; second, aesthetic and functional pruning rules focusing on landscape aesthetics and maintenance, such as pruning guidelines for maintaining specific geometric shapes (spherical, columnar, etc.) or meeting the requirements of specific functional areas (e.g., controlling crown width next to pedestrian walkways). Rules are typically stored as production rules in the form of "IF-THEN," where the IF part specifies the triggering conditions (e.g., variety A, crown width exceeding value B, NDVI below value C, located in functional area D), and the THEN part specifies the recommended pruning operations (e.g., shortening main branches, thinning inner branches). This step digitizes and formalizes botanical knowledge, horticultural experience, and landscape design intentions.

[0025] S4: For each identified plant, the corresponding extracted morphological and structural parameters, physiological state parameters, and the design outline and functional zoning information of the plant extracted from the landscape planning digital model are integrated. Logical matching and reasoning are performed according to the plant pruning knowledge base rule set to generate an individual plant pruning guidance plan that includes pruning location, pruning intensity, and pruning technique.

[0026] For each identified plant, the system integrates and analyzes the plant's current morphological and structural parameters and physiological state parameters obtained in the preceding steps, along with the target design outline and functional zoning information extracted from the landscape planning digital model. This integration process is conducted within a unified spatial reference system, such as aligning and comparing the 3D point cloud of the current branches with the 3D mesh model of the design outline. Logical matching and reasoning are performed based on the plant pruning knowledge base rule set, such as determining whether the current crown width exceeds the design range, whether there are dead branches that need to be removed, and whether the current form deviates from the target shape. The reasoning engine (e.g., a forward-chain-based rule engine) activates all rules that meet the conditions and resolves potential rule conflicts according to preset priorities (usually, shape maintenance has a higher priority than routine health maintenance). Finally, a specific pruning guideline is generated for the plant, clearly indicating the specific locations requiring pruning (e.g., 3D coordinates or descriptions relative to the trunk), the pruning intensity (e.g., the percentage of the original branch length to be cut off), and the recommended pruning techniques (e.g., flat pruning, oblique pruning, thinning, etc.).

[0027] S5: Overlay and register the individual plant pruning guidance plan with the corresponding color digital image to form a guidance document with visual pruning marks and send it to mobile terminal devices to guide staff in pruning operations.

[0028] The system overlays and registers the individual pruning guidelines for each plant with the initially acquired color digital image of the area. The registration process utilizes pose data recorded during image acquisition or is achieved through feature point matching to ensure that the pruning locations in the guidelines are accurately marked on the image. After overlay, a guidance document is generated containing visual pruning markers (such as highlighted circles for branches to be pruned and arrows indicating pruning points). This document can be wirelessly transmitted to the mobile devices of gardening staff (such as tablets, mobile phones, or augmented reality glasses with the application installed). Staff can then view the overlaid pruning guidance image or real-time augmented reality view on-site via their mobile devices, enabling them to perform pruning tasks intuitively and accurately, avoiding misjudgments or omissions that might result from relying on personal experience.

[0029] Compared to traditional pruning methods that rely on the experience and visual assessment of gardeners, this solution automates, digitizes, and intelligently manages the pruning decision-making process. Through drone data collection and deep learning analysis, the status of each plant within a large-scale garden can be quickly, objectively, and comprehensively assessed, overcoming the problems of low efficiency, strong subjectivity, and difficulty in quantification associated with manual inspections. By fusing design models, botanical rules, and real-time status data, the generated pruning plan considers both the general laws of healthy plant growth and the specific aesthetic and functional requirements of landscape design, transforming pruning operations from experience-driven to science- and data-driven. Finally, on-site guidance through visual instruction documents significantly improves the accuracy, consistency, and overall efficiency of pruning operations, reduces the professional skill threshold and labor costs, and facilitates the refined maintenance and management of large-scale gardens.

[0030] In another technical solution, step S1 specifically includes: Based on the boundary vector data of the target garden area stored in the landscape management database, a parallel scanning route with full coverage and a set overlap rate is generated. The UAV is controlled to fly along the parallel scanning route during periods above a set light intensity threshold, with an RGB camera acquiring color digital images at set time intervals in a vertically downward orientation. During the UAV's flight, its onboard real-time differential global positioning system and inertial measurement unit record the acquisition position and attitude data of each frame of color digital image. After acquisition, based on the acquisition position and attitude data, a motion reconstruction structure algorithm is used to generate a 3D point cloud model of the park with geographic coordinates from the color digital images. The 3D point cloud model of the park is then matched with the landscape planning digital model using coordinate system one and spatial registration to establish a precise correspondence between visual features in the color digital images and design elements in the landscape planning digital model. The plant design outlines in the landscape planning digital model are defined by a 3D mesh model, and functional zoning information is stored as attribute data in the primitives of the 3D mesh model.

[0031] First, the system automatically plans the drone's flight path based on the boundary vector data of the target garden area stored in the landscape management database. This boundary data defines the precise geographical range of the area to be worked on. The system generates a set of parallel scanning routes based on this range, ensuring that the drone's flight path can completely cover the entire area. For the accuracy of subsequent image stitching and 3D reconstruction, the images between adjacent flight strips need to maintain a certain overlap rate, for example, set between 20% and 40%. This effectively ensures sufficient matching feature points during the motion reconstruction algorithm processing. The system controls the drone equipped with an RGB camera to perform tasks during periods of good ambient lighting conditions, for example, setting the light intensity threshold above 30,000 lux, to avoid image blurring or color distortion due to insufficient light. Of course, adjustments can be made appropriately if weather conditions do not meet the requirements for an extended period. In principle, image acquisition is carried out during the best lighting conditions of the day. The drone flies autonomously along the planned parallel flight path, with its onboard RGB camera set to a vertically downward attitude and automatically triggering shooting at set time intervals (such as every 0.5 seconds, or at fixed distance intervals based on flight speed), thereby acquiring a series of continuous and ordered color digital image sequences. During the drone's flight, its onboard positioning and attitude determination systems work synchronously, recording precise spatiotemporal labels for each frame of the acquired image. Specifically, by integrating a high-precision real-time differential global positioning system (RTK-GNSS) and an inertial measurement unit (IMU), the system can acquire the drone's three-dimensional spatial coordinates (longitude, latitude, altitude) and the aircraft's attitude angles (pitch, roll, yaw) in real time for each image captured. This position and attitude data is stored synchronously as metadata for the image. After all images have been acquired, the system uses the Structure for Motion Reconstruction (SfM) algorithm to process all the color digital images and their corresponding pose data. This algorithm uses computer vision technology to match feature points from multi-view images with overlapping areas, calculates the camera's motion trajectory, and reconstructs a three-dimensional sparse point cloud of the scene. Then, through a multi-view stereo vision algorithm, it generates a dense three-dimensional point cloud model of the park. This model not only has realistic visual texture, but each of its three-dimensional points also carries real-world geographic coordinates, thus forming a digital three-dimensional replica of the garden scene.

[0032] After generating the real-world 3D point cloud model, it is spatially aligned and fused with the landscape planning digital model pre-stored in the landscape management database. The landscape planning digital model is a digital twin model containing design intent, in which the design outline of plants is precisely defined through a 3D mesh model (e.g., a surface model composed of multiple triangular facets), describing the 3D spatial morphology of plants under ideal maintenance conditions. Simultaneously, information on different functional zones of the garden (such as viewing areas, road isolation zones, and private spaces) is stored as attribute data in the corresponding 3D mesh primitives of the areas or plants. Using coordinate system one and spatial registration techniques (such as the ICP iterative nearest point algorithm based on feature points), the real-world 3D point cloud model and the landscape planning digital model are precisely overlaid. This process establishes a precise correspondence between each visual feature in the acquired real-world image (e.g., a leaf, a branch) and the corresponding design element in the planning model (e.g., the design outline facet of a plant), providing a spatial benchmark for subsequent comparisons between the actual growth state and the design objectives.

[0033] This solution ensures high-quality and consistent source data through automated and standardized planning of UAV flight paths and strict lighting and attitude control during image acquisition. By integrating high-precision positioning, attitude determination, and structure-of-motion (SOR) algorithms, it can automatically generate centimeter-level accurate 3D models with georeferenced data from 2D image sequences, significantly improving the efficiency and accuracy of spatial information acquisition. Precise spatial registration between the real-world model and the design planning model achieves seamless integration between the physical world and the digital design blueprint. This allows subsequent intelligent analysis to quantitatively compare the actual plant growth status with the expected design goals within a unified and precise coordinate system, providing a crucial spatial framework for precise pruning decisions and overcoming the problems of design-reality disconnect and inaccurate positioning inherent in traditional methods.

[0034] As a more specific example, taking bougainvillea, a common plant in gardens, as the pruning target, we provide an example of an important operational process: 1) Data Acquisition and Recognition: Images acquired by the drone were identified by a first deep learning model, confirming the plant species as "bougainvillea". A second deep learning model performed semantic segmentation on the images, accurately extracting the clustered flowers, slender branches, and oval leaf regions. The current crown size was calculated to be 2.3 meters, and the large variance in the spatial distribution angle of the branches indicated that the branches were growing outwards in a scattered manner. At the same time, the normalized difference vegetation index calculated based on the leaf pixel color was slightly lower than the health threshold, and two areas of dead branches were identified.

[0035] 2) Rule Invocation and Fusion Analysis: The system invokes the pruning knowledge base rule set associated with the "Bougainvillea" variety identifier. This rule set contains its unique physiological and shaping rules: for example, physiological rules include "IF the variety is Bougainvillea AND there are dead branches THEN recommend removing dead branches" and "IF NDVI is below the threshold THEN recommend light pruning to promote new shoot growth"; shaping rules include "IF it is located in a shaped hedge area AND the crown width exceeds the design value by 15% THEN recommend reducing the excess portion to within the design outline." The system fuses and spatially compares the current parameters of the Bougainvillea plant (crown width 2.3 meters, dead branches present, low NDVI) with the design outline of its "spherical shaping area" (target crown width of a 1.8-meter diameter sphere) in the landscape planning digital model.

[0036] 3) Intelligent Decision-Making and Solution Generation: The rule engine performs reasoning. Due to the detection of dead branches, physiological rules are activated; simultaneously, the current crown width (2.3 meters) significantly exceeds the design target (1.8 meters), with deviation exceeding the threshold, so the shape restoration rule is activated with higher priority. The system calculates the specific spatial locations requiring pruning: first, the three-dimensional coordinates of the two dead branches; second, the point cloud of the ends of all branches exceeding the boundary of the spherical design grid. Combining the bougainvillea's tolerance to pruning (high tolerance coefficient), the system generates an individualized pruning guideline: it recommends completely removing dead branches using a "thinning" technique; for branches exceeding the design shape, a "heavy pruning" technique is used, shortening them to approximately 20 centimeters within the design curve to promote new growth, which is beneficial for later shaping and flowering.

[0037] 4) Pre-assessment and Visual Guidance: The system uses a parametric model of bougainvillea growth for simulation. The pre-assessment shows that after the above pruning, new shoots will more densely fill the design sphere's outline, and the predicted volume overlap will be significantly improved, meeting the long-term shaping goals. Finally, the scheme is overlaid with on-site images to generate an augmented reality guidance document. On-site workers can clearly see the virtual highlighted marks superimposed on dead and excessive branches through mobile terminals, thus achieving rapid positioning and precise pruning.

[0038] In another technical solution, in step S2, the pre-trained first deep learning model adopts a cascaded network structure. First, the region proposal network is used to locate the bounding box of each plant in the color digital image. Then, the image region within each bounding box is classified to output the variety identifier. The pre-trained second deep learning model is a fully convolutional network with an encoder-decoder structure, used for pixel-level semantic segmentation of color digital images. The segmentation categories include leaves, branches, flowers, fruits, and background. The crown size in the morphological structure parameters is obtained by calculating the diagonal length of the minimum bounding rectangle formed by the leaf and branch pixels in the semantic segmentation results. The spatial distribution of branches is obtained by calculating the angular variance of the skeletonized lines of the branch pixels in the image. Both the first and second deep learning models were trained using a multi-seasonal garden image dataset, which includes color digital images of the same plant in spring, summer, autumn, and winter, labeled with corresponding variety identifiers and semantic segmentation labels. During training, data augmentation techniques were employed, including random rotation, scaling, and color jitter.

[0039] In the two deep learning models, the model for plant variety recognition employs a cascaded network structure, divided into two stages to balance detection accuracy and speed. First, the region proposal network in the model scans the entire input color digital image, quickly generating a series of candidate rectangular regions, or bounding boxes, that may contain plants. These candidate boxes exclude a large number of background regions, focusing on potential targets. Then, for each candidate bounding box, the model extracts the image region within the box and feeds it into a more refined classification network for identification. This classification network ultimately outputs the specific variety identifier of the plant within the region, such as "bougainvillea," "malus," "Japanese late-blooming cherry," or "boxwood." This two-stage structure enables the model to efficiently locate and identify each individual plant in complex garden backgrounds. The deep learning model for image semantic segmentation uses a fully convolutional network with an encoder-decoder structure. The encoder part progressively extracts high-level semantic features of the image through multiple convolutional layers and downsampling operations; the decoder part progressively restores the feature map to the original image size through upsampling and skip connections, assigning a class label to each pixel. This model performs pixel-level fine-grained classification on the input color digital image. In the output, each pixel in the image is labeled as belonging to one of the following categories: leaf, branch, flower, fruit, or background. From this fine-grained segmentation result, key plant morphological and structural parameters can be further calculated. For example, by extracting all pixel sets belonging to "leaf" and "branch" and calculating the diagonal length of their minimum bounding rectangle, the crown size of the plant can be estimated. Simultaneously, by refining the segmented "branch" pixels to obtain their skeletal lines, and then statistically analyzing the directional distribution of these skeletal lines in the image plane, calculating their angular variance, this quantitatively describes the spatial dispersion of the branches, whether it is relatively concentrated or spread out in all directions.

[0040] To ensure high accuracy and robustness of these two deep learning models in complex and ever-changing garden environments, they were both trained using a specially constructed multi-seasonal garden image dataset. This dataset contains color digital images of the same garden area or similar plants taken in spring, summer, autumn, and winter, with detailed annotations for each image: the bounding box of each plant and its corresponding variety identifier, as well as semantic segmentation labels for each pixel. During model training, data augmentation techniques were employed to increase data diversity and prevent overfitting. These techniques included, but were not limited to, random rotation of training images (within ±30 degrees), scaling at different ratios (e.g., 0.8x to 1.2x), and slight random perturbations to the color channels (i.e., color jitter) to simulate imaging effects under different lighting and weather conditions. Models trained in this way are better able to adapt to the challenges posed by plant appearance changes at different growth stages and seasons, as well as by varying shooting conditions.

[0041] This solution employs a cascaded network structure for variety identification, enabling rapid and accurate localization and classification of multiple plants in complex scenarios, providing accurate identification criteria for subsequent personalized pruning rule invocation. By using an encoder-decoder structure for pixel-level semantic segmentation, it achieves fine-grained analysis of plant organs, making it possible to automatically extract key morphological parameters such as crown width and branch distribution from images, replacing the inefficient traditional method relying on manual measurement. More importantly, by utilizing a professional dataset covering multi-seasonal variations and combining various data augmentation techniques for model training, the generalization ability and environmental adaptability of the deep learning model in practical applications are greatly improved. This allows it to stably cope with the annual morphological changes of plants and the influence of different light conditions, ensuring the year-round validity and reliability of the state analysis results.

[0042] When extracting physiological state parameters, for leaf color distribution, the mean of the red channel, the mean of the green channel, the mean of the blue channel, and the corresponding standard deviation are extracted from the leaf pixel region in the semantic segmentation results, and the normalized difference vegetation index is calculated based on these extracted values; for dead branch region labeling, a support vector machine classifier is trained, and based on the gray-level co-occurrence matrix texture features and HSV color space features of the branch pixel region, the branches are distinguished into healthy branches and dead branches, and the pixels classified as dead branches are clustered into continuous dead branch regions through the region growing algorithm; Morphological and physiological parameters are stored as structured data records for each plant. The data records include variety identifier, bounding box coordinates, crown size, branch length, angular variance of branch spatial distribution, normalized difference vegetation index, and the coordinates of the bounding rectangle of the dead branch area. The data records are associated with the design outline of the corresponding plant extracted from the landscape planning digital model through a unique identifier.

[0043] When extracting the physiological parameter of leaf color distribution, all pixel regions labeled "leaf" are first precisely separated from the semantic segmentation results. For these pixels, the average value and standard deviation of these values ​​in the red (R), green (G), and blue (B) color channels are calculated. These statistical values ​​reflect the overall hue, saturation, and color consistency of the leaves. To more scientifically assess the greenness and photosynthetic activity of the leaves, the Normalized Difference Vegetation Index (NDVI) is further calculated based on these color values. This index is calculated using near-infrared and red light reflectance, but when using only an RGB camera, it can be estimated using a specific formula (e.g., an approximate calculation based on the G and R channels). The value of the NDVI can indirectly reflect the chlorophyll content and growth vigor of the leaves, providing a quantitative indicator for judging plant health.

[0044] For the identification and labeling of dead branch regions, a scheme combining traditional image features and machine learning classifiers was adopted. First, texture and color features were extracted from the "branch" pixel regions obtained from semantic segmentation. Texture features can be obtained by calculating the gray-level co-occurrence matrix, such as contrast, correlation, energy, and homogeneity; color features can be obtained by converting the image to the HSV color space and extracting statistical values ​​of hue, saturation, and brightness. Then, a pre-trained support vector machine (SVM) classifier was used to classify each branch pixel region based on these extracted texture and color feature combinations, determining whether it is a "healthy branch" or a "dead branch." The SVM classifier distinguishes between the two classes of samples by finding the optimal hyperplane in the feature space. To aggregate the potentially discrete dead branch pixels into meaningful continuous regions, a region growing algorithm was used. Starting from the seed pixels classified as dead branches, the region growing algorithm aggregated pixels based on their proximity and similarity, ultimately forming several connected dead branch region blocks. The coordinates of the bounding rectangle of each dead branch region were then labeled.

[0045] All extracted morphological and physiological parameters need to be organized and stored in a structured manner to facilitate subsequent retrieval, matching, and fusion. The system creates an independent structured data record for each identified plant. This record contains the plant's core information: variety identifier, its bounding box coordinates in the original image, calculated crown size, estimated branch length, branch spatial distribution angular variance describing branch dispersion, normalized difference vegetation index reflecting leaf vitality, and the set of coordinates of the bounding rectangles of all identified dead branch areas. The system generates a unique identifier for each such plant data record. This unique identifier allows for precise association and binding between this data record, reflecting the plant's current state obtained from image analysis, and information representing the plant's design goals (its 3D design outline mesh) extracted from the landscape planning digital model. This facilitates a convenient one-to-one comparison and analysis of the "current state" and the "goal" in subsequent steps. By combining color statistics, vegetation index estimation, and machine learning-based texture classification, an automated and objective diagnosis of plant leaf health and branch dieback was achieved. This transformed the experience-based judgment, previously reliant on horticulturists' visual observation, into quantifiable data indicators, significantly improving the scientific rigor and consistency of the condition assessment. Clustering dead branch pixels using a region growing algorithm accurately delineates the actual extent and shape of dead branches, providing clear guidance for precisely locating pruning areas. Finally, by designing a structured data recording format and using unique identifiers for association, a seamless data pipeline from image recognition to design models was established. This ensures that the multidimensional information (identity, current status, and target) of each plant can be efficiently and accurately integrated, effectively avoiding information confusion or matching errors.

[0046] In another technical solution, in step S3, the set of plant pruning knowledge base rules is stored in the knowledge base of the rule engine in the form of production rules. Each rule has an "IF-THEN" structure. The "IF" part of the plant physiological pruning rule consists of conditions connected by logic and relations. The conditions include matching the variety identifier, judging the numerical range of morphological structure parameters, and judging the threshold of physiological state parameters. The "IF" part of the landscape shaping pruning rule includes not only the conditions of the plant physiological pruning rule, but also matching the functional zoning information and judging the deviation of the design outline. The "THEN" part of all rules defines one or more pruning operation suggestions. When performing logical matching and reasoning, the rule engine performs forward chain reasoning on the rules in the knowledge base based on the currently input variety identifier, morphological structure parameters, physiological state parameters, design outline, and functional zoning information. It activates all rules whose conditions are met and, according to the preset rule priority, adjudicates any conflicting pruning operation suggestions, outputting a set of conflict-free pruning operation suggestions.

[0047] The rules in the plant pruning knowledge base are stored in the form of production rules, a knowledge representation method widely used in expert systems. Each rule has a clear "IF (condition) - THEN (conclusion)" structure, clearly expressing what action should be recommended under what circumstances. The rule set of the plant pruning knowledge base consists of two types of rules: plant physiological pruning rules and landscape shaping pruning rules. Plant physiological pruning rules mainly focus on the healthy growth patterns of the plant itself. Its "IF" part consists of a series of conditions connected by logical "AND" relationships. These conditions may include precise matching of the identified variety identifier (e.g., whether the variety is "crape myrtle"), numerical range judgment of extracted morphological parameters (e.g., whether the current crown size exceeds the normal growth upper limit range of the variety in spring, such as 1.5 meters to 2 meters), and threshold judgment of physiological state parameters (e.g., whether the calculated normalized difference vegetation index is lower than the lower limit threshold indicating health status, such as 0.3). When all these conditions are met simultaneously, the rule is activated. Its "THEN" section defines one or more pruning suggestions aimed at restoring or maintaining plant health, such as "removing overly dense branches in the inner canopy" or "shortening one-third of vigorous shoots." Landscape shaping pruning rules, based on physiological rules, further incorporate landscape aesthetic design requirements. Its "IF" section, in addition to the conditions related to variety and plant status mentioned above, also includes a matching judgment of the functional zoning information of the plant (e.g., whether the plant is located "from the main entrance to the scenic area"), and a judgment of the deviation between the plant's current actual form and the designed shape outline. Deviation is a quantitative indicator used to measure the size of the difference between the actual growth form and the target shape. This type of rule is only triggered when the plant's variety, growth status, location, and morphological deviation all meet the rule's preset conditions. The pruning suggestions defined in its "THEN" section focus more on shaping and restoring the form, such as "retracting branches exceeding the outer edge of the spherical outline into the design surface" or "to maintain the flatness of the hedge, uniformly pruning protruding points to the reference plane."

[0048] All rules are integrated into a knowledge base of a rule engine. When the system reasons about a particular plant, the rule engine receives all input information about that plant, including its variety identifier, specific values ​​of various morphological and physiological parameters, its design outline, and functional zoning information. The rule engine employs a forward chain reasoning strategy, traversing all rules in the knowledge base and checking whether the conditions in the "IF" part of each rule are satisfied by the currently input facts. Rules whose conditions are fully satisfied are activated. Since different rules may be activated simultaneously, and their "THEN" parts may offer conflicting operational suggestions (for example, a health rule suggests retaining a branch to promote growth, while a shaping rule suggests pruning the branch to maintain the outline), the rule engine has a pre-defined conflict resolution mechanism. This mechanism typically decides based on pre-defined rule priorities; for example, "the priority of landscape shaping pruning rules related to maintaining the design outline can be set higher than the priority of plant physiological pruning rules related only to plant health and growth." The engine will filter or merge conflicting suggestions based on priority, and finally output a set of pruning operation suggestions that are internally consistent and logically free of conflict, as the initial decision result for the plant.

[0049] This solution encapsulates horticultural knowledge and design specifications using structured production rules, transforming implicit experience previously held by horticulturists into explicit digital knowledge that can be stored, managed, and reused, thus systematizing and standardizing pruning knowledge. Through a rule engine, automated logical matching and forward chain reasoning efficiently and accurately compare complex plant status data with a vast knowledge base, automatically generating preliminary pruning suggestions, significantly reducing the workload and subjective arbitrariness of manual analysis and decision-making. Furthermore, by pre-setting rule priorities and introducing conflict resolution mechanisms, the system ensures that when faced with potential conflicts between multiple objectives such as health maintenance and shape preservation, it can make consistent decisions based on established management principles (such as prioritizing shape). This results in pruning plans that are both scientifically sound and effectively implement the core intent of the landscape design, enhancing the systematic nature and authority of the decisions.

[0050] The priority of the rules is preset according to the pruning objectives. Among them, the priority of landscape pruning rules related to maintaining the design outline is higher than the priority of plant physiological pruning rules that are only related to the healthy growth of plants. In the logical matching and reasoning process of the rule engine, the deviation between the morphological structural parameters and the design outline is judged by comparing the crown size, branch length and branch spatial distribution in the morphological structural parameters with the corresponding design dimensions and design angle range defined in the three-dimensional mesh model of the design outline to calculate the quantitative deviation value. The rule engine has pre-defined threshold judgment sub-rules for different deviation ranges. When the quantified deviation value exceeds the preset first threshold, the rule with shape restoration as the primary pruning goal is activated. When the quantified deviation value is lower than the preset second threshold, the rule with routine health maintenance as the primary pruning goal is activated. The set of conflict-free pruning operation suggestions is combined with the predefined pruning operation parameterization template. The pruning operation suggestions are mapped to specific pruning position coordinates, pruning tool entry angle and branch cut length through the pruning operation parameterization template, thereby generating individual plant pruning guidance schemes.

[0051] To ensure that pruning decisions effectively serve the overall aesthetic goals of the landscape, a clear priority order is preset in the rule engine. Landscape pruning rules directly related to maintaining the design outline are given higher priority, while plant physiological pruning rules primarily focused on the healthy internal growth of plants have relatively lower priority. This priority setting reflects that in landscape maintenance, when an irreconcilable conflict arises between the healthy growth of plants and the intended landscape form, the system's decision-making tends to prioritize restoring and maintaining the design shape. This aligns with the practical management needs of many landscape maintenance scenarios (such as formal gardens and themed green sculptures). Of course, the priority strategy can be configured and adjusted according to different garden types and management objectives.

[0052] When executing landscape pruning rules, the deviation between morphological parameters and the designed shape outline is quantitatively calculated. This process involves precisely comparing the plant's current measured morphological data with its digital design model. For example, the system extracts the currently identified crown size value and compares it with the designed crown size of the plant defined in the 3D mesh model of the designed shape outline (e.g., a sphere with a diameter of 1.2 meters), calculating the percentage exceeding or falling short as the quantitative deviation value. Similarly, for branch length, it is compared with the design length of the corresponding branch in the design model; for the angular variance of the branch spatial distribution, it is compared with the expected range of branch angle distribution in the design (e.g., for an umbrella-shaped crown, the angle between the main branch and the trunk may be between 40 and 60 degrees), calculating the degree of deviation from the design range. Each comparison generates one or more specific, calculable quantitative deviation values. To provide more refined guidance for pruning, the rule engine's knowledge base also includes pre-set threshold judgment sub-rules specifically for different deviation ranges, which constitute the refined branches of the decision-making process. For example, a first threshold can be set at 15% deviation from the design dimensions, and a second threshold at 5%. When the system calculates that the quantitative deviation of a plant's crown exceeds the first threshold (i.e., it is too large or too small by more than 15%), it indicates that its shape has been severely distorted. At this time, a strong intervention rule with shape restoration as the primary pruning objective will be activated, and its suggested pruning intensity will be relatively large. When the quantitative deviation is below the second threshold (e.g., less than 5%), it indicates that the plant's morphology basically meets the design requirements. At this time, a weak intervention rule with routine health maintenance as the primary objective will be activated, mainly for tasks such as clearing dead branches and making minor adjustments. Finally, the system combines the set of conflict-free pruning operation suggestions obtained after conflict resolution with a predefined parameterized template for pruning operations. This template acts as a translator, mapping abstract pruning suggestions (such as "remove inner branches") into specific, executable operation parameters, such as the specific pruning location coordinates in three-dimensional space (based on the branch point cloud), the suggested cutting angle of the pruning tool (such as pruning shears) (relative to the branch axis), and the specific length of the branch to be cut (e.g., 15 cm). In this way, highly operable individual plant pruning guidelines are ultimately generated. By explicitly prioritizing shaping rules over health rules, the intelligent decision-making system ensures that it firmly grasps the core aesthetic goals of landscape maintenance, ensuring that the results of automated decisions do not deviate from the original intention of the landscape design and are applicable to garden scenarios with strict shaping requirements. Through multi-dimensional (size, length, angle) precise quantitative calculations of deviation and setting multi-level thresholds to trigger different sub-rules, the decision logic can flexibly adopt pruning strategies of varying intensity, from fine-tuning maintenance to intensive shaping, based on the severity of the deviation, greatly improving the precision and adaptability of the decision-making.Ultimately, by using parametric templates to transform logical suggestions into specific work instructions in three-dimensional space, the key connection from intelligent analysis to on-site construction was completed. This made the generated guidance plan no longer a principle description, but a precise work order that workers could directly refer to and execute, significantly improving the feasibility of the plan and the standardization of the operation.

[0053] In another technical solution, step S4, which involves integrating morphological parameters, physiological parameters, and the designed shape outline, specifically includes: A unified spatial coordinate system is established for each plant. The spatial distribution of branches in the morphological structure parameters, the dead branch area in the physiological state parameters, and the three-dimensional mesh model of the design outline are aligned and mapped in this spatial coordinate system. Then, the volume overlap between the current actual branch distribution and the branch design area in the design outline is calculated, and the actual branch parts that exceed the volume boundary of the design outline and the dead branch area that is within the volume boundary of the design outline are identified. Logical matching and reasoning are based on volume overlap, spatial coordinates of actual branches that exceed the design outline, and spatial coordinates of dead branch areas. The pruning position in the individual plant pruning guidance plan is determined based on the spatial coordinates of the actual branches that exceed the design outline and the spatial coordinates of dead branch areas. The pruning intensity is calculated based on the volume overlap and the preset pruning tolerance coefficient of the plant variety. The pruning process is selected from the predefined process library based on the level and diameter of the branch's spatial coordinates.

[0054] When integrating and analyzing the current state parameters of plants with design goals to generate specific pruning instructions, a unified spatial coordinate system is established for each individual plant. This coordinate system uses the base of the plant's trunk or a recognized reference point as its origin, dividing the three-dimensional space. Within this coordinate system, the system precisely aligns and maps the branch spatial distribution point cloud data obtained from image analysis, the three-dimensional coordinate set of marked dead branch areas, and the three-dimensional mesh model of the design outline describing the ideal form of the plant, extracted from the landscape planning digital model. The alignment operation ensures that all data are at the same scale, orientation, and spatial reference, allowing a real branch to be directly compared spatially with a corresponding curved surface or volume region in the design model. After alignment, spatial analysis calculations are performed, specifically calculating the overlap between the spatial volume occupied by the actual branch point cloud and the target volume defined by the design outline three-dimensional mesh model. This can be achieved by voxelizing the three-dimensional space and statistically analyzing the proportion of voxels simultaneously occupied by the actual point cloud and the design model. Simultaneously, the system performs two key identifications: first, it identifies the portions of the actual branch point cloud that lie outside the design outline volume boundary. These exceeding branches are the direct cause of the plant's shape deviating from the design, appearing cluttered or excessively large; second, it identifies areas that, although within the design outline volume boundary, have been classified as dead branches. These dead branches do not affect the external outline but impact the plant's internal health and aesthetics. These two types of areas are the priority targets for pruning. Based on the results of the spatial analysis above, logical matching and reasoning are performed to determine the pruning location most directly, namely, the spatial coordinates of the identified "actual branch portions exceeding the design outline" and "spatial coordinates of dead branch areas." These three-dimensional coordinate points or areas can be directly converted into specific operational points in the pruning guidance plan. The calculation of pruning intensity requires a more comprehensive judgment. The system combines the calculated volume overlap (reflecting the overall deviation) with the "pruning tolerance coefficient" pre-existing in the knowledge base and associated with the plant variety identifier to jointly determine the pruning intensity. The tolerance coefficient is an empirical value. For example, some plants with strong sprouting ability (such as privet) can be assigned a higher coefficient, allowing them to tolerate heavy pruning; while some slow-growing tree species (such as conifers) are assigned a lower coefficient, allowing only light pruning. The system uses a fusion algorithm to calculate the recommended pruning amount based on the deviation and tolerance level, which may be expressed as the proportion of branch length cut off or the proportion of branches to be removed. Finally, the pruning technique is selected from a predefined technique library based on the spatial coordinates of the branch being pruned (such as the trunk, primary branch, secondary twig) and its estimated diameter. For example, for trunk-level pruning with a large diameter (such as more than 3 cm), "thinning (removing from the base)" can be selected in conjunction with wound application; for terminal twigs with a small diameter (such as less than 1 cm), "light heading back (removing the tip)" can be selected.This series of steps generates a plant individual pruning guide that includes precise location, appropriate intensity, and suitable techniques.

[0055] This solution establishes a unified coordinate system for each plant and performs precise spatial alignment, achieving deep fusion and comparison of real-world growth data and virtual design models in three-dimensional space. This allows subsequent analysis to be conducted within a rigorous spatial framework, significantly improving the accuracy and reliability of the comparison. By calculating volume overlap and intelligently identifying parts exceeding the outline and internal dead branches, the system can accurately locate all specific spatial targets requiring pruning, avoiding omissions or misjudgments that may occur in traditional experience-based pruning, resulting in highly targeted pruning plans. By dynamically determining pruning intensity and techniques by comprehensively considering volume overlap, plant variety characteristics, and branch physical properties, the generated pruning guidance plans not only aim to restore the shape but also fully consider the plant's own growth characteristics and pruning safety. The scientific rigor, personalization, and operability of the plans are significantly enhanced, enabling workers to perform effective and safe precision pruning operations.

[0056] After generating individual plant pruning guidelines, a pre-assessment is performed, which includes: The morphological and structural parameters are loaded as the initial state, and the corresponding plant growth parameterization model is called according to the variety identifier. After simulating the pruning position, pruning intensity, and pruning process in the computer, the growth direction and length of new branches in the next growth cycle are predicted. The prediction results are used to calculate the predicted volume overlap between the simulated branch distribution and the designed shape outline. If the predicted volume overlap is lower than the current volume overlap, it is determined that the plant individual pruning guidance plan does not meet the long-term shape maintenance goal. At this time, the combination of pruning intensity and pruning position is adjusted, and the simulation and evaluation are repeated until the predicted volume overlap is not lower than the current volume overlap. Finally, the adjusted plant individual pruning guidance plan is output.

[0057] After generating preliminary individual plant pruning guidelines, a crucial pre-assessment and optimization phase is conducted to ensure that the pruning plan not only addresses the current problem but also promotes healthy plant growth towards the designed goals in the future. The system loads the current plant's morphological and structural parameters (such as branch length and distribution angle) as the initial state for the plant's growth simulation. Simultaneously, based on the plant's variety identifier, the system retrieves the corresponding parametric growth model from a plant growth model library. This model, built upon known growth habit data for the variety, simulates the growth rate, budding location, and extension direction of its branches under specific conditions in different seasons. In the virtual computer environment, the system performs simulated pruning operations on the loaded initial state model according to the preliminary pruning guidelines. This involves removing or shortening corresponding virtual branches based on the pruning location, intensity, and technique specified in the guidelines. Then, the system drives the parametric growth model to predict the likely budding location, dominant growth direction, and potential length of new shoots on the remaining branches after this simulated pruning within a complete future growth cycle (e.g., the next six months or a growing season). These predictions reflect the plant's biological response to this pruning. The system then merges these predicted new shoots with the existing branch structure to form a 3D model simulating the branch distribution after growth. This predicted future branch distribution model is then compared with the 3D mesh model of the designed shape of the same plant to calculate the predicted volume overlap. If the calculated predicted volume overlap is lower than the current volume overlap before pruning, it means that after implementing the current pruning plan, the new growth in the next growth cycle may cause the overall shape to deviate further from the design goal. This is a temporary fix that doesn't address the root cause and may even be counterproductive. In this case, the system will determine that the preliminary plan does not meet the long-term shape maintenance goal. Once the determination is incorrect, the system will not directly output the plan but will initiate an optimization loop to adjust the key parameters in the pruning plan. For example, while keeping the main pruning positions unchanged, the pruning intensity may be moderately increased or decreased, or the positions of some pruning points may be fine-tuned. Then, based on the adjusted new plan, the entire process from simulated pruning to growth prediction and then to calculating the predicted volume overlap is repeated. This cycle iterates, continuously adjusting the plan and evaluating its long-term effects until a pruning plan is found that ensures the predicted volume overlap is no less than (and usually aims to be higher than) the current volume overlap. Ultimately, the system outputs this optimized and verified pruning guide for individual plants that is beneficial for maintaining their long-term shape.

[0058] This solution, by introducing simulated pruning and growth prediction based on a parametric growth model, expands the perspective of pruning decisions from solving current static problems to considering the future impact of dynamic plant growth. This represents a leap from correcting the current state to guiding trends, significantly enhancing the predictability and strategic value of pruning plans. By comparing the predicted future form with the design goals and calculating the degree of overlap, an objective and quantitative indicator is provided for evaluating the long-term effectiveness of pruning plans, giving decisions a long-term scientific basis. Finally, by establishing an automated evaluation and adjustment iterative optimization cycle, the system can proactively identify and output optimal or optimized pruning plans that not only provide good immediate results but also contribute to long-term shape stability. This effectively avoids repeated pruning or greater future shape restoration costs caused by improper one-time decisions, realizing scientific management of garden maintenance from short-term treatment to long-term planning, and significantly improving the foresight and economic efficiency of maintenance work.

[0059] In another technical solution, the method for forming the guidance document in step S5 is as follows: The system receives individual plant pruning guidelines, corresponding color digital images, and a 3D point cloud model of the park obtained from the landscape management database. It analyzes the pruning location coordinates in the individual plant pruning guidelines and finds the corresponding 3D spatial points in the park's 3D point cloud model. Based on the real-time geographic location, spatial attitude, and real-time scene images acquired by the mobile terminal device's built-in GPS receiver, inertial measurement unit, and camera, a feature-based image matching algorithm is used to match the real-time scene images with the color digital images, calculating the real-time relative position and attitude between the mobile terminal device's camera and the target plant. Based on the real-time relative position and attitude, the 3D spatial points corresponding to the pruning location coordinates are rendered as virtual pruning markers with 3D spatial attributes, which are then overlaid on the real-time scene image stream displayed on the mobile terminal device screen, forming a guidance document with visual pruning markers. After receiving the guidance document, the mobile terminal device displays a real-time augmented reality view on the screen with virtual trimming markers overlaid to assist the operator in positioning and trimming.

[0060] The system's backend server aggregates the final individual pruning guidelines for each plant, color digital images of the corresponding area initially collected by drones, and a high-precision 3D point cloud model of the park generated using a motion reconstruction algorithm. The system first parses the coordinates of each pruning location contained in the pruning guidelines file. These coordinates are defined in a unified world coordinate system shared with the park's 3D point cloud model. Then, within this dense 3D point cloud model, the system finds the corresponding or nearest 3D spatial point for each pruning coordinate. These 3D points, carrying real spatial location information, are key anchor points connecting virtual pruning instructions with the real physical world. When workers enter the garden site using handheld or worn mobile terminal devices (such as smartphones, tablets, or augmented reality glasses), the devices begin collecting multi-source data in real time to perceive their own and their environment's state. The device's built-in GPS receiver provides its approximate latitude and longitude location, with an accuracy potentially in the meter range. More precise positioning and attitude determination rely on an inertial measurement unit (IMU), which continuously measures the device's acceleration and angular velocity, and calculates the terminal's real-time attitude (pitch, roll, yaw angles) and short-distance displacement changes through integral calculations. Meanwhile, the terminal device's camera continuously captures real-time images of the scene ahead. To accurately overlay the virtual pruning marks onto the real plant seen by the user, the system needs to calculate the precise relative relationship between the camera lens and the target plant. This is achieved through a feature-based image matching algorithm: the system extracts feature points (such as leaf edges and branch intersections) from the real-time scene image and quickly matches them with feature points in the corresponding high-definition color digital image of the area sent by the server. Through successfully matched feature point pairs, combined with the known acquisition pose of the color digital image, the system can calculate the three-dimensional spatial position and lens orientation of the mobile terminal camera relative to the target plant at the current moment, i.e., the real-time relative position and orientation.

[0061] Once the precise relative pose is calculated, augmented reality rendering can begin. The previously identified 3D spatial points representing the pruning location are projected onto the 2D coordinate system of the mobile terminal device screen based on the calculated real-time relative pose. The system renders these projected points as virtual graphic markers with 3D spatial attributes, such as a flashing red circle, an arrow pointing to a specific branch, or a highlighted branch outline. These virtual markers are bound to 3D spatial coordinates. When the operator moves the terminal or changes their viewpoint, the system continuously updates the real-time relative pose and recalculates the projection position of the virtual markers on the screen accordingly, making the virtual markers appear as stable as if they were nailed to a branch in the real world. Finally, the real-time video stream is merged with these stably superimposed virtual pruning markers to form a dynamic augmented reality view on the terminal screen, which is the final visual guidance document sent to the operator. Workers no longer need to refer to drawings or imagine; they can clearly and intuitively see which branch needs pruning and at which specific location simply by looking at the screen, greatly simplifying the positioning process.

[0062] This solution integrates high-precision 3D point cloud, real-time computer vision, and augmented reality rendering technologies to transform abstract coordinate commands into intuitive, stable, and precisely aligned visual guidance that mirrors real-world scenes. This completely transforms traditional pruning operations that rely on paper drawings, verbal descriptions, or manual marking. On-site staff no longer need to interpret complex drawings or perform time-consuming on-site measurements, enabling more intuitive operations and significantly reducing the technical barrier to entry and reliance on personnel experience. Simultaneously, this real-time, visual guidance method significantly reduces the occurrence of incorrect or missed pruning, greatly improving the accuracy and success rate of complex pruning or high-density plant pruning, providing a highly efficient and reliable technical means for the refined maintenance and management of large-scale gardens.

[0063] In another technical solution, the landscape management database is built and updated in the following way: Multi-angle images of the planted garden area are acquired using a multi-lens oblique camera mounted on a drone. The multi-angle images are then processed using a motion recovery structure algorithm to generate a high-precision real-scene 3D mesh model containing accurate color information. Based on this real-scene 3D mesh model, a pre-trained instance segmentation model identifies and labels the 3D outline of each plant and its spatial coordinates within the model, using these coordinates as initial data for the plant's design location. Horticultural designers then use 3D modeling software to refine and redesign the 3D outline of each identified plant, defining its long-term maintenance design outline 3D mesh model and assigning variety identification and functional zoning information to each plant. The initial design location data, the 3D mesh model of the design outline, the variety identification, and the functional zoning information are linked and stored as a plant record in the landscape planning digital model data table of the landscape management database. When plants are replanted or replaced in the garden area, the above acquisition, identification, design, and linking steps are repeated to generate a new plant record and update the landscape planning digital model data table.

[0064] When constructing the landscape management database, a comprehensive initial digital survey is conducted on the planted garden areas. Using drones equipped with oblique cameras (e.g., five lenses), aerial photography is performed from vertical and multiple oblique angles, acquiring a multi-angle image sequence covering the entire area and rich in lateral texture information. These images are processed using a motion reconstruction algorithm to generate a high-precision 3D mesh model that not only includes top information but also accurately reflects the lateral morphology of plants and the details of building facades. This model records the garden's true 3D state at a specific moment and includes realistic color and texture. Based on this initial 3D mesh model, a pre-trained deep learning instance segmentation model automatically identifies and segments the 3D point cloud or mesh cluster of each individual plant in the model, and labels each segmented plant instance with its 3D outline and spatial coordinates in the model's global coordinate system. These automatically identified position coordinates serve as the initial data for the plant's design location in the digital world. However, the real-world model reflects the initial state of planting or the current state at a specific moment, and may not represent the ideal long-term shape. Therefore, the involvement of a horticultural designer is necessary for blueprint design. The designers used professional 3D modeling software to load the real-world 3D mesh model as a base reference. For each plant automatically identified in the model, the designers refined, adjusted, and redesigned its current 3D outline. For example, they created a 3D mesh for the ideal umbrella-shaped crown of a tree five years later, or defined the smooth, curved surface of a row of shrubs that should be maintained as a hedge. This process defined the 3D mesh model of each plant's design outline under long-term maintenance goals. Simultaneously, the designers selected or confirmed the species identifier for each plant and assigned functional zoning information (such as "central lawn focal tree" or "roadside tree") according to the landscape plan.

[0065] Finally, for four key pieces of information—initial design location data automatically generated from the instance segmentation model, a 3D mesh model of the design outline created by the designer, variety identification, and functional zoning information—the system structurally correlates these four key pieces of information and packages them into a complete plant record. This record is stored in the landscape planning digital model data table of the landscape management database. Each record is identified by a unique ID for a specific plant, thus establishing its archive in the digital world. Gardens are not static; when plants are replanted, die, or the planning and design are adjusted, the database needs to be updated synchronously. At this time, for the areas where changes have occurred, the entire process of oblique photogrammetry acquisition, real-scene 3D modeling, instance segmentation and identification, manual design modification, and data correlation is repeated to generate new or update existing plant records and modify the landscape planning digital model data table, ensuring that the digital blueprint always remains synchronized with the latest design intent of the garden and significant changes in the physical world. By automatically extracting data from the real-scene 3D model and combining it with professional human design to construct the database, the system efficiently transforms the physical garden into an authoritative, accurate, and design-intended digital twin, providing an indispensable benchmark of trust for all intelligent analyses. This method leverages the efficiency of automatic identification technology to quickly acquire initial spatial data, while also incorporating the professional judgment and aesthetic creativity of landscape designers to ensure that the digital blueprint conforms to the principles of landscape art and long-term management goals, achieving a balance between efficiency and quality. By establishing a conveniently updatable process, the landscape management database can adapt to the dynamic changes in the garden, maintaining consistency between the digital and physical worlds, thereby ensuring the continued effectiveness of all intelligent pruning decisions based on this database.

[0066] It should be noted that although the steps are described in a specific order above, this does not mean that they must be performed in that order. In fact, some of these steps can be executed concurrently, or even in a different order, as long as the required functionality is achieved. The number of devices and processing scale described herein are for simplification of the invention; applications, modifications, and variations of this invention will be readily apparent to those skilled in the art.

[0067] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and illustrations shown and described herein.

Claims

1. A method for intelligent pruning of garden plants based on unmanned aerial vehicles (UAVs), characterized in that, Includes the following steps: S1: Using a drone equipped with an RGB camera, fly along a preset route and collect color digital images of the target garden area. At the same time, obtain the landscape planning digital model of the area from the landscape management database. The landscape planning digital model includes the design location of plants, the design shape outline, and functional zoning information. S2: Input the acquired color digital images into the pre-trained first deep learning model to identify and output the variety identifier of each plant in the image. At the same time, input the color digital images into the pre-trained second deep learning model to perform image semantic segmentation and extract the morphological and structural parameters and physiological state parameters of each plant. The morphological and structural parameters include crown size, branch length and branch spatial distribution, and the physiological state parameters include leaf color distribution and dead branch area marking. S3: Retrieve the set of plant pruning knowledge base rules corresponding to the variety identifier from the landscape management database. The set of plant pruning knowledge base rules consists of plant physiological pruning rules and landscape shaping pruning rules. The plant physiological pruning rules are associated with the variety identifier and growth status parameters, while the landscape shaping pruning rules are associated with the design outline and functional zoning information. S4: For each identified plant, the corresponding extracted morphological and structural parameters, physiological state parameters, and the design outline and functional zoning information of the plant extracted from the landscape planning digital model are integrated. Logical matching and reasoning are performed according to the plant pruning knowledge base rule set to generate an individual plant pruning guidance plan that includes pruning location, pruning intensity, and pruning technique. S5: Overlay and register the individual plant pruning guidance plan with the corresponding color digital image to form a guidance document with visual pruning marks and send it to mobile terminal devices to guide staff in pruning operations.

2. The intelligent pruning method for garden plants based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, Step S1 specifically includes: Based on the boundary vector data of the target garden area stored in the landscape management database, a parallel scanning route with full coverage and a set overlap rate is generated. The UAV is controlled to fly along the parallel scanning route during periods above a set light intensity threshold, with an RGB camera acquiring color digital images at set time intervals in a vertically downward orientation. During the UAV's flight, its onboard real-time differential global positioning system and inertial measurement unit record the acquisition position and attitude data of each frame of color digital image. After acquisition, based on the acquisition position and attitude data, a motion reconstruction structure algorithm is used to generate a 3D point cloud model of the park with geographic coordinates from the color digital images. The 3D point cloud model of the park is then matched with the landscape planning digital model using coordinate system one and spatial registration to establish a precise correspondence between visual features in the color digital images and design elements in the landscape planning digital model. The plant design outlines in the landscape planning digital model are defined by a 3D mesh model, and functional zoning information is stored as attribute data in the primitives of the 3D mesh model.

3. The intelligent pruning method for garden plants based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, In step S2, the pre-trained first deep learning model adopts a cascaded network structure. First, the region proposal network is used to locate the bounding box of each plant in the color digital image. Then, the image region within each bounding box is classified to output the variety identifier. The pre-trained second deep learning model is a fully convolutional network with an encoder-decoder structure, used for pixel-level semantic segmentation of color digital images. The segmentation categories include leaves, branches, flowers, fruits, and background. The crown size in the morphological structure parameters is obtained by calculating the diagonal length of the minimum bounding rectangle formed by the leaf and branch pixels in the semantic segmentation results. The spatial distribution of branches is obtained by calculating the angular variance of the skeletonized lines of the branch pixels in the image. Both the first and second deep learning models were trained using a multi-seasonal garden image dataset, which includes color digital images of the same plant in spring, summer, autumn, and winter, labeled with corresponding variety identifiers and semantic segmentation labels. During training, data augmentation techniques were employed, including random rotation, scaling, and color jitter.

4. The intelligent pruning method for garden plants based on unmanned aerial vehicles (UAVs) according to claim 3, characterized in that, When extracting physiological state parameters, for leaf color distribution, the mean of the red channel, the mean of the green channel, the mean of the blue channel, and the corresponding standard deviation are extracted from the leaf pixel region in the semantic segmentation results, and the normalized difference vegetation index is calculated based on these extracted values. For the labeling of dead branch areas, a support vector machine classifier is trained to distinguish between healthy branches and dead branches based on the gray-level co-occurrence matrix texture features and HSV color space features of the branch pixel areas. Pixels classified as dead branches are then clustered into continuous dead branch areas using a region growing algorithm. Morphological and physiological parameters are stored as structured data records for each plant. The data records include variety identifier, bounding box coordinates, crown size, branch length, angular variance of branch spatial distribution, normalized difference vegetation index, and the coordinates of the bounding rectangle of the dead branch area. The data records are associated with the design outline of the corresponding plant extracted from the landscape planning digital model through a unique identifier.

5. The intelligent pruning method for garden plants based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, In step S3, the set of plant pruning knowledge base rules is stored in the knowledge base of the rule engine in the form of production rules. Each rule has an "IF-THEN" structure. The "IF" part of the plant physiological pruning rule consists of conditions connected by logic and relations. The conditions include matching the variety identifier, judging the numerical range of morphological structure parameters, and judging the threshold of physiological state parameters. The "IF" part of the landscape shaping pruning rule includes not only the conditions of the plant physiological pruning rule, but also matching the functional zoning information and judging the deviation of the design outline. The "THEN" part of all rules defines one or more pruning operation suggestions. When performing logical matching and reasoning, the rule engine performs forward chain reasoning on the rules in the knowledge base based on the currently input variety identifier, morphological structure parameters, physiological state parameters, design outline, and functional zoning information. It activates all rules whose conditions are met and, according to the preset rule priority, adjudicates any conflicting pruning operation suggestions, outputting a set of conflict-free pruning operation suggestions.

6. The intelligent pruning method for garden plants based on unmanned aerial vehicles (UAVs) according to claim 5, characterized in that, The priority of the rules is preset according to the pruning objectives. Among them, the priority of landscape pruning rules related to maintaining the design outline is higher than the priority of plant physiological pruning rules that are only related to the healthy growth of plants. In the logical matching and reasoning process of the rule engine, the deviation between the morphological structural parameters and the design outline is judged by comparing the crown size, branch length and branch spatial distribution in the morphological structural parameters with the corresponding design dimensions and design angle range defined in the three-dimensional mesh model of the design outline to calculate the quantitative deviation value. The rule engine has pre-defined threshold judgment sub-rules for different deviation ranges. When the quantified deviation value exceeds the preset first threshold, the rule with shape restoration as the primary pruning goal is activated. When the quantified deviation value is lower than the preset second threshold, the rule with routine health maintenance as the primary pruning goal is activated. The set of conflict-free pruning operation suggestions is combined with the predefined pruning operation parameterization template. The pruning operation suggestions are mapped to specific pruning position coordinates, pruning tool entry angle and branch cut length through the pruning operation parameterization template, thereby generating individual plant pruning guidance schemes.

7. The intelligent pruning method for garden plants based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, Step S4, the process of integrating morphological parameters, physiological parameters, and the designed shape outline, specifically includes: A unified spatial coordinate system is established for each plant. The spatial distribution of branches in the morphological structure parameters, the dead branch area in the physiological state parameters, and the three-dimensional mesh model of the design outline are aligned and mapped in this spatial coordinate system. Then, the volume overlap between the current actual branch distribution and the branch design area in the design outline is calculated, and the actual branch parts that exceed the volume boundary of the design outline and the dead branch area that is within the volume boundary of the design outline are identified. Logical matching and reasoning are based on volume overlap, spatial coordinates of actual branches that exceed the design outline, and spatial coordinates of dead branch areas. The pruning position in the individual plant pruning guidance plan is determined based on the spatial coordinates of the actual branches that exceed the design outline and the spatial coordinates of dead branch areas. The pruning intensity is calculated based on the volume overlap and the preset pruning tolerance coefficient of the plant variety. The pruning process is selected from the predefined process library based on the level and diameter of the branch's spatial coordinates.

8. The intelligent pruning method for garden plants based on unmanned aerial vehicles (UAVs) according to claim 7, characterized in that, After generating individual plant pruning guidelines, a pre-assessment is performed, which includes: The morphological and structural parameters are loaded as the initial state, and the corresponding plant growth parameterization model is called according to the variety identifier. After simulating the pruning position, pruning intensity, and pruning process in the computer, the growth direction and length of new branches in the next growth cycle are predicted. The prediction results are used to calculate the predicted volume overlap between the simulated branch distribution and the designed shape outline. If the predicted volume overlap is lower than the current volume overlap, it is determined that the plant individual pruning guidance plan does not meet the long-term shape maintenance goal. At this time, the combination of pruning intensity and pruning position is adjusted, and the simulation and evaluation are repeated until the predicted volume overlap is not lower than the current volume overlap. Finally, the adjusted plant individual pruning guidance plan is output.

9. The intelligent pruning method for garden plants based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, In step S5, the guidance document is generated in the following way: The system receives individual plant pruning guidance plans, corresponding color digital images, and a 3D point cloud model of the park obtained from the landscape management database; it analyzes the pruning position coordinates in the individual plant pruning guidance plans and finds the corresponding 3D spatial points in the park's 3D point cloud model; based on the terminal's geographical location, spatial attitude, and real-time scene images acquired in real time by the mobile terminal device's built-in GPS receiver, inertial measurement unit, and camera, it uses a feature-based image matching algorithm to match the real-time scene images with the color digital images, and calculates the real-time relative position and attitude between the mobile terminal device's camera and the target plant; Based on the real-time relative position and attitude, the three-dimensional spatial points corresponding to the trimming position coordinates are rendered as virtual trimming marks with three-dimensional spatial attributes, and superimposed on the real-time scene image stream displayed on the mobile terminal device screen to form a guidance document with visual trimming marks. After receiving the guidance document, the mobile terminal device displays a real-time augmented reality view on the screen with virtual trimming markers overlaid to assist the operator in positioning and trimming.

10. The intelligent pruning method for garden plants based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, The landscape management database is built and updated in the following ways: Multi-angle images of the planted garden area are captured by a multi-lens oblique camera mounted on a drone. The multi-angle images are processed using a motion recovery structure algorithm to generate a high-precision real-scene 3D mesh model containing accurate color information. Based on this real-scene 3D mesh model, a pre-trained instance segmentation model identifies and labels the 3D outline of each plant and its spatial coordinates within the model, using these coordinates as initial data for the plant's design location. Horticultural designers then use 3D modeling software to refine and redesign the 3D outline of each identified plant, defining its long-term maintenance target design shape outline 3D mesh model. Each plant is also assigned a variety identifier and functional zoning information. The initial design location data, the 3D mesh model design shape outline, the variety identifier, and the functional zoning information are linked and stored as a plant record in the landscape planning digital model data table of the landscape management database. When plants are replanted or replaced in the garden area, repeat the above steps of collection, identification, design and association to generate new plant records and update the landscape planning digital model data table.

Images