A three-dimensional real scene-based garden landscape automatic design method and system

Abstract

This invention discloses an automated landscape design method and system based on 3D real-scene data, specifically relating to the fields of smart gardening and computer-aided design technology. The method includes: acquiring 3D real-scene data of the target site; constructing a semantic 3D real-scene model through semantic segmentation; automatically parsing and extracting multi-dimensional design constraints based on the model; generating multiple initial landscape design schemes using parametric rules and generative algorithms; collecting user interaction data; dynamically optimizing the scheme evaluation function through machine learning; and quantitatively evaluating and ranking the schemes; finally, outputting the optimized scheme sequence and a multi-dimensional quantitative evaluation report. The system includes data acquisition and semantic modeling, design constraint generation, parametric scheme generation, interactive scheme optimization, and scheme output and report generation. This invention achieves full-process automation and personalized optimization from data to scheme, improving design efficiency, scientific rigor, and user satisfaction.

Description

Technical Field

[0001] This invention relates to the fields of smart garden and computer-aided design technology, and more specifically, to an automated design method and system for garden landscape based on three-dimensional real-world scenes. Background Technology

[0002] In the current field of landscape design, traditional workflows heavily rely on designers' manual on-site surveys, experience-based judgments, and manual drawing and scheme refinement using CAD and other software. With technological advancements, methods are emerging that incorporate auxiliary tools such as 3D reality modeling, parametric design, and generative algorithms. For example, acquiring 3D site models through drone oblique photography or laser scanning, or generating partial layout schemes using pre-defined rule bases, aims to improve the initial efficiency and intuitiveness of the design process.

[0003] However, in practical use, it still has some shortcomings, such as the single data acquisition method or insufficient collaboration, resulting in a lack of semantic information in the 3D model, which cannot directly support automated analysis and design decisions; the degree of intelligence in the design process is limited, and key links from current situation analysis to solution generation still require a lot of manual intervention, making it difficult to achieve automated solution exploration and optimization under multiple constraints; solution generation and evaluation are often out of touch with the real-time preferences of end users, lacking a dynamic optimization mechanism based on real interactive feedback, which means that the practicality of the generated solutions and user satisfaction need to be improved. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, the present invention provides a method and system for automated design of garden landscape based on three-dimensional real scene, which solves the problems mentioned in the background art through the following solutions.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an automated landscape design method based on three-dimensional real-scene, comprising: S1: Acquire 3D real-scene data of the target site and perform semantic segmentation preprocessing to build a semantic 3D real-scene model that supports automated parsing; S2: Based on the semantic 3D real scene model, the terrain features, current element distribution and environmental analysis factors are automatically parsed and extracted through the pre-trained element recognition model to generate structured and computable multi-dimensional design constraints. S3: Based on parametric rules and generative algorithms, using the multi-dimensional design constraints as input, multiple initial landscape design schemes are automatically generated to form an initial design scheme set. S4: Based on user interaction data with the initial design scheme set, dynamically optimize the scheme generation strategy. The process includes: S401: Collect and analyze user interaction data on the initial design scheme set, the interaction data including scheme browsing trajectory, parameter adjustment sequence and scheme selection record; S402: Based on the interactive behavior data, dynamically adjust the weight coefficients in a scheme evaluation function through a machine learning model, wherein the scheme evaluation function is used to quantitatively evaluate multiple indicators in the technical performance, economic performance, and spatial perception performance of the design scheme. S403: Using the dynamically adjusted scheme evaluation function, calculate and evaluate each scheme in the initial design scheme set to obtain the comprehensive evaluation value of each scheme, and sort the schemes according to the comprehensive evaluation value to generate an optimized scheme sequence; S5: Output the optimized scheme sequence, in which each scheme is associated with a comprehensive evaluation value and a score of each performance index calculated based on the dynamically adjusted scheme evaluation function, and generate a complete quantitative evaluation report.

[0006] An automated landscape design system based on 3D reality, comprising: Data acquisition and semantic modeling module: acquire multi-source 3D real-scene data of the target site, and construct a semantic 3D real-scene model that supports automated parsing through hierarchical semantic segmentation and association with semantic ontology; Design constraint generation module: Based on the semantic 3D real scene model, through the pre-trained element recognition model and built-in parsing rules, automatically extracts and quantifies terrain features, current element distribution and environmental analysis factors, and generates structured multi-dimensional design constraints. Parametric scheme generation module: Based on the multi-dimensional design constraints and parametric rule library, it drives multiple generative algorithms to work together to automatically generate multiple differentiated initial landscape design schemes, forming an initial design scheme set. Interactive solution optimization module: Collects and analyzes user interaction data on the initial design solution set, dynamically adjusts the weights of the solution evaluation function through a machine learning model, and quantitatively evaluates and ranks the solutions accordingly, outputting an optimized solution sequence; Solution Output and Report Generation Module: Standardizes and hierarchically outputs the optimized solution sequence, and dynamically generates a complete report with multi-dimensional quantitative evaluation.

[0007] The technical effects and advantages of this invention are as follows: 1. It has achieved full automation and intelligence from data acquisition to solution output. Through collaborative acquisition of data from multiple heterogeneous devices and semantic modeling, combined with parametric rules and generative algorithms, it has improved the efficiency of landscape design and reduced the reliance on manual experience and repetitive labor. 2. The generated solutions are more scientific and practical. Based on semantic 3D real-world models, multi-dimensional and quantifiable design constraints are automatically extracted to ensure that the solutions are closely integrated with the actual site conditions. Compliance is verified in real time during the generation process, which improves the rationality of the design and the feasibility of the project. 3. The system has adaptive and continuous optimization capabilities. By collecting and learning user interaction behavior data, it dynamically adjusts the solution evaluation function and generation strategy, so that the output solution can accurately match user preferences and project requirements. Attached Figure Description

[0008] Figure 1 This is a schematic diagram of the overall structure of the present invention.

[0009] Figure 2 This is a schematic diagram of the data acquisition and semantic modeling structure of the present invention.

[0010] Figure 3 This is a schematic diagram of the structure generated by the design constraints of the present invention.

[0011] Figure 4 This is a schematic diagram of the parameterization scheme generation and preliminary verification structure of the present invention. Detailed Implementation

[0012] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0013] refer to Figures 1-4 The illustrated method for automated landscape design based on 3D reality includes: S1: Acquire and process the 3D real-scene data of the target site, and construct a semantic 3D real-scene model. The implementation details are as follows: S101. Acquisition of 3D Reality Data of the Target Site: A dynamic temporal fusion acquisition method using multimodal heterogeneous devices is adopted to output a multi-dimensional 3D reality raw dataset. The specific implementation process is as follows: Data Acquisition Equipment Group Setup: Construct a three-tiered heterogeneous data acquisition equipment group consisting of air, space, and ground systems, including: The airborne equipment consists of a cluster of multi-rotor UAVs equipped with multispectral cameras, lidar (LiDAR), and inertial navigation units (IMU). The cluster contains 3-5 UAVs, which adopt a distributed formation flight mode, and each UAV is equipped with an independent timing synchronization module. The ground equipment consists of a mobile intelligent data acquisition vehicle equipped with a panoramic camera and a ground 3D laser scanner. At the same time, fixed millimeter-wave radar blind spot terminals are deployed at key nodes in the site, such as terrain inflection points, dense vegetation areas, and hard boundaries. The space-ground collaborative equipment is a GNSS (Global Navigation Satellite System) differential base station deployed at the edge of the site, providing centimeter-level positioning references for all mobile data acquisition devices, while establishing 5G+ edge computing communication links between devices to achieve real-time backhaul and synchronous calibration of acquired data.

[0014] Data Acquisition Path and Timing Planning: Based on the preliminary site survey map, a pre-trained path planning model is used to generate the equipment acquisition path. The path planning must meet the principles of "no repetition in overlapping areas, supplementary data acquisition in blind areas, and high precision in key areas." Drone swarm: A "layered and regional" time-series data acquisition strategy is adopted, which divides the data into three layers according to altitude from high to low: a macro-topography layer at 100-150m, a meso-vegetation and building structure layer at 30-50m, and a micro-surface detail layer at 5-10m. During the acquisition of each layer, the flight time difference between drones is kept to no more than 500ms to ensure that the data at different altitudes in the same area are aligned in the time dimension. Intelligent data collection vehicle: It cruises along site roads or passable areas while receiving real-time data feedback from the drone swarm. When the drones identify ground blind spots such as building shadow areas or dense forest obstructions, it automatically guides the data collection vehicle to the area for supplementary ground data collection. Fixed millimeter-wave radar terminal: continuously collects data at a frequency of 100ms / time, and performs time-series completion on the collection gaps of mobile devices to avoid data gaps caused by device movement.

[0015] Real-time preprocessing and fusion of multi-source data: Real-time preprocessing and fusion of raw data collected at edge computing nodes. LiDAR point cloud calibration: Real-time denoising and coordinate calibration are performed based on IMU attitude data using an attitude calibration formula, which is: , in, The original point cloud coordinates, Based on pitch angle Roll angle Heading angle The rotation matrix, It is a translation vector. To calibrate the point cloud coordinates; at the same time, combining the spectral information from the multispectral camera, attribute labels such as vegetation health and surface material reflectivity are assigned to the point cloud data. Ground data fusion: Point cloud-image registration is performed on ground laser point clouds and panoramic images to map texture information and remove invalid data of dynamic interference objects such as pedestrians and vehicles; All collected data are based on the timestamps of GNSS differential positioning data and time synchronization modules to construct a unified "space-time-attribute" three-dimensional index, forming a preliminary multi-source fused three-dimensional data volume.

[0016] Data Acquisition Quality Verification and Supplementary Acquisition: The fused data is verified using a data quality assessment model built into the edge computing nodes. Integrity verification: By comparing with a preset list of site elements, identify areas of elements that have not been collected and automatically trigger supplementary collection instructions for the corresponding devices; Accuracy verification: By checking the horizontal and vertical positioning errors, ensure that the horizontal accuracy is not less than 2cm and the vertical accuracy is not less than 3cm. If the accuracy does not meet the standards, recalibrate the equipment and take additional samples. The formula for planar positioning error is: ; The formula for elevation positioning error is: ; in To collect coordinate values, These are the coordinates of the GNSS reference point. Density verification: The data density is evaluated by point cloud density. The point cloud density is required to be no less than 500 points / ㎡ in key areas such as landscape nodes and areas with abrupt changes in terrain slope, and no less than 200 points / ㎡ in ordinary areas. If the density is insufficient, the drone is instructed to reduce its flight altitude or the data collection vehicle to slow down its cruising speed for encrypted data collection. Point cloud density formula: ,in Point cloud density, The total number of point clouds in the target area. The area of ​​the target region.

[0017] After completing the full-area data collection and supplementary data collection, the system outputs a multi-dimensional 3D real-scene raw dataset containing spatial coordinates, geometric shape, material properties, spectral features, and temporal information.

[0018] S102. Semantic Segmentation Preprocessing of 3D Reality Data: A hierarchical semantic segmentation preprocessing method with multi-feature linkage is adopted to transform the multi-dimensional 3D reality dataset into a standardized semantic segmentation dataset containing feature semantic labels, geometric parameters, attribute features, and topological relationships. The specific implementation process is as follows: Hierarchical decoupling and feature extraction of multi-dimensional 3D reality raw dataset: The multi-dimensional 3D reality raw dataset is subjected to targeted hierarchical decoupling, and is divided into four feature layers according to data dimension and feature type. Among them, the spectral attribute layer completes feature quantization through professional formulas. The core features of each layer are extracted as follows: Geometric feature layer: Extracts geometric parameters such as 3D coordinates, normal vectors, curvature, neighborhood point density, and elevation difference from LiDAR point cloud data; extracts geometric features such as buildings and terrain contours from the 3D reconstruction results of panoramic images; Spectral attribute layer: Extracts reflectance of each band from data acquired by multispectral cameras, and calculates core indicators using spectral index formulas. The formula for the normalized vegetation index is: , The formula for the normalized water index is: , in, These are the reflectances of the red, green, and near-infrared bands, respectively, and are also correlated with millimeter-wave radar echo intensity data to form a spectral-radar attribute feature vector set. Temporal dynamic layer: Based on the timestamp of the built-in temporal synchronization module of the original dataset, the dynamic change features of the data in different collection periods are extracted, such as the positional shift of temporary obstructions and the slight swaying of vegetation branches and leaves, to form a temporal dynamic feature vector set. Material reflection layer: Texture features such as texture roughness, texture direction, and repeating units are extracted from high-resolution images from panoramic cameras. Combined with the differences in echo reflection intensity from LiDAR, different materials are distinguished to form a set of material reflection feature vectors.

[0019] Multi-model collaborative hierarchical semantic segmentation: Based on four feature layers decoupled from the multi-dimensional 3D real-scene raw dataset, a three-level semantic segmentation model system is constructed, as follows: First-level coarse segmentation (feature category segmentation): Using the improved PointNet++ point cloud segmentation model, with geometric feature layer and spectral attribute layer data as input, the original dataset is divided into six major feature categories: terrain, vegetation, water bodies, buildings and structures, hard paving, and dynamic disturbances, ensuring that the category segmentation accuracy is not less than 95%. Model training is optimized using the following loss function: , in, For cross-entropy loss, For regularization loss, For loss balance weighting coefficients; Secondary subdivision (feature subclass segmentation): For each major feature category after primary coarse segmentation, different subdivision models are used for subclass labeling: Terrain subclasses: Based on the elevation difference and curvature data of the geometric feature layer, the terrain is divided into subclasses such as flat terrain, gentle slope terrain, steep slope terrain, depression terrain, and ridge terrain using the U-Net3D model; Vegetation subclasses: The NDVI index of the spectral attribute layer and the leaf texture data of the material reflection layer are integrated and classified into subclasses such as trees, shrubs, herbs, vines, and dead trees through a two-branch CNN model, while the vegetation health level is labeled. Building subclasses: Based on the outline shape of the geometric feature layer and the exterior texture of the material reflection layer, the YOLOv8-3D model is divided into subclasses such as landscape pavilions, walls, pergolas, and management rooms; Hard paving subclass: Combining the texture of the material's reflective layer and the intensity of the LiDAR echo, it is divided into subclasses such as stone paving, concrete paving, permeable asphalt, and wooden boardwalk using a random forest classification model; Level 3 Refinement (Cross-Layer Feature Cross-Validation): The segmentation results are validated and corrected using time-series dynamic layer data. Segmentation accuracy is evaluated through Intersection over Union (IoU) and overall classification accuracy, eliminating invalid interference information and ensuring the reliability of subclass segmentation. The formula for overall classification accuracy is: , The formula for overall classification accuracy is: ,in True positive True negative False positive, To detect false negatives, invalid interference information in the original data is removed to ensure the accuracy of subclass segmentation.

[0020] Topology construction and data optimization: Establish a topology map of elements, simplify redundant data, and review and correct conflicting data.

[0021] The final output is a standardized semantic segmentation dataset containing feature semantic labels, geometric parameters, attribute features, and topological relationships.

[0022] S103. Constructing a semantic 3D reality model supporting automated parsing: An integrated construction scheme of "feature semantic ontology + computable association rules + dynamic parsing interface" is adopted to transform the semantic segmentation dataset into a semantic 3D reality model supporting automated parsing. The specific implementation process is as follows: Construction of a Semantic Ontology Library for Landscape Elements: Based on industry standards for landscape design and the needs of automated design, a dedicated semantic ontology library for landscape elements is constructed as the underlying logical support for semantic association in the model. The ontology library includes: Element classification ontology: The segmented elements are defined into an ontology according to a three-level system of "major category - subcategory - detailed category", and each category is bound to a corresponding industry standard attribute; Attribute-related ontology: Define multi-dimensional related attributes for each feature ontology, which are divided into basic geometric attributes, ecological attributes, engineering attributes, and spatial relationship attributes, and establish constraint rules between attributes; The parsing rules ontology: embeds the core normative rules of landscape design, providing a rule basis for subsequent automated parsing.

[0023] Hierarchical Construction of Semantic 3D Model: Based on a semantic ontology library, hierarchical modeling is performed on the semantic segmentation dataset to achieve deep integration of geometric shape and semantic information. Specifically, it consists of a three-layer structure: Basic geometric layer: Using a parametric modeling engine, point cloud data and image data are transformed into editable 3D geometric models. Terrain features are constructed using triangular irregular networks (TINs) combined with parametric surface algorithms. Vegetation features are modeled using a hybrid approach of "geometric model + attribute labels". Buildings and hard paving are constructed using parametric models based on BIM lightweight technology. Semantic association layer: Through ontology mapping algorithm, the association rules and constraint logic of semantic ontology library are embedded into the basic geometry layer to establish semantic association graph of elements and add semantic query interface to each element model; Dynamic parsing layer: A dynamic parsing engine is built at the top level of the model, consisting of two core components: Rule matching component: Scans the feature association chains in the model in real time, matches them with the parsed rule ontology of the semantic ontology library, and marks associations that do not conform to the specification; Parameter calculation component: Includes a built-in multi-dimensional calculation function library, capable of automatically quantifying key engineering and ecological parameters of calculation elements. For example, it automatically calculates the catchment capacity of topographic units using the catchment area formula. ,in The total catchment area, For the first Area of ​​each terrain unit The system automatically calculates key parameters for elements such as the catchment area of ​​the terrain, the ecological benefit value of vegetation, and the amount of hard paving work, and synchronizes the calculation results to the model attribute panel.

[0024] The model is empowered by automated parsing capabilities: It integrates multiple automated parsing interfaces into the semantic 3D reality model, including rapid element retrieval, rule conflict detection, multi-dimensional calculation output, and dynamic model update interface, enabling it to proactively respond to design requirements and output structured parsing results. It can achieve accurate retrieval of site elements, investigation of design rule conflicts, automatic extraction and calculation of key parameters, and incremental model updates, and can directly connect to the subsequent design constraint generation module.

[0025] Model Validation and Optimization: After model construction is completed, multi-dimensional validation is conducted to ensure that its automated parsing capabilities meet the standards. Semantic consistency verification: Through semantic consistency verification, we ensure that the semantic labels of model elements are more than 98% consistent with the semantic ontology library definition; Semantic consistency formula: ,in For semantic consistency, The number of element labels that were successfully matched. This represents the total number of element labels; Accuracy verification: The result must have at least 95% overlap with the manually selected results; Computational efficiency verification: The response time for parsing association rules and parameters of 100 elements in a single batch should not exceed 30 seconds; Analytical efficiency formula: ,in For parsing efficiency, The number of elements parsed in a single session. The time taken for parsing.

[0026] The final output is a semantic 3D reality model that supports automated parsing.

[0027] S2: Generate multi-dimensional design constraints based on a semantic 3D reality model. The specific implementation process is as follows: S201. Bidirectional Adaptation and Connection between Semantic 3D Reality Model and Feature Recognition Model: This addresses the incompatibility and semantic information loss issues between the semantic 3D reality model and the pre-trained feature recognition model. The specific process is as follows: Semantic label mapping calibration: First, the three-level label system of the semantic ontology of garden landscape elements in the semantic 3D reality model is extracted and matched with the feature classification labels of the pre-trained element recognition model. At the same time, a quantitative evaluation logic for label mapping matching degree is introduced. The core formula is as follows: , in, For the overall mapping matching degree of the tag, For the first The matching status of each tag (a match is marked as 1, and a non-match is marked as 0). The industry weights of the tags are determined by the priority of landscape design, with the weights set as follows: terrain tags 0.3, vegetation tags 0.25, buildings and structures tags 0.2, water bodies tags 0.15, and hard paving tags 0.1. In this embodiment, the overall matching degree is required to be no less than 97%. Unmatched tags are supplemented by manual annotation and then re-entered into the middleware to complete the iterative optimization of the mapping relationship.

[0028] Lightweight encapsulation of input data: The basic geometric layer and semantic association layer data of the semantic 3D reality model are sliced ​​and processed to extract only the core semantic attributes and key geometric parameters required by the feature recognition model, generating a dedicated input data package. At the same time, a parameter calculation interface for the dynamic parsing layer is reserved to realize the real-time calling of the model's built-in parameters during the extraction process.

[0029] S202. Hierarchical Analysis and Extraction of Multi-Dimensional Features and Factors: Based on the adapted model, the analysis and extraction of terrain features, current feature distribution, and environmental analysis factors are completed in modules, as follows: Accurate analysis and quantitative extraction of terrain features: A multi-parameter fusion-based hierarchical extraction method for terrain features is adopted, and the process is as follows: The parameter calculation component of the dynamic parsing layer of the semantic 3D reality model is invoked to first extract the basic geometric parameters of the terrain, such as elevation, slope, aspect, and curvature. Then, in conjunction with landscape engineering standards, core derived features such as terrain undulation and water catchment potential are quantified. The terrain relief is determined by the difference between the maximum and minimum elevations within the target terrain unit, and is used to divide the terrain into flat areas, gentle slope areas, and steep slope areas. The water catchment potential is combined with the total catchment area calculated by the model, and the soil permeability coefficient and surface cover of the site are simultaneously correlated for comprehensive judgment, providing a constraint basis for the subsequent layout of rain gardens, grassed swales and other facilities.

[0030] Based on the extracted terrain features, the terrain functional zoning was completed in accordance with the garden landscape design specifications, forming a structured terrain feature dataset that includes zoning range, core parameters, and engineering suggestions.

[0031] Semantic parsing and density quantification of current element distribution: Combining the topological relationship map of the pre-trained element recognition model and the semantic 3D reality model, the accurate positioning and distribution quantification of current elements are achieved. The process is as follows: The feature recognition model accurately identifies existing features such as vegetation, buildings, hard paving, and water bodies within the model, and outputs the three-dimensional coordinates, semantic category, and spatial coverage of each feature. Quantify the distribution status of elements: on the one hand, calculate the distribution density of a single element within the target unit; on the other hand, combine the element synergy of the garden landscape to calculate the coupling distribution density of the target element and related elements, such as the spatial synergy density of trees and shrubs, and mark the element dense area, blank area and spatial conflict area to form the basic constraint data of the current element distribution.

[0032] Coupled extraction and threshold quantification of environmental analysis factors: Combining the spectral attribute layer and time-series dynamic layer data of the semantic model, the extraction and threshold division of environmental analysis factors are completed. The process is as follows: Basic environmental data such as vegetation NDVI index and water body NDWI index are obtained from the spectral attribute layer, and dynamic data such as site illumination duration and wind direction frequency are extracted from the temporal dynamic layer. Based on the requirements of landscape ecological design, coupled factors such as vegetation ecological suitability and site microclimate comfort are quantitatively determined: vegetation ecological suitability is comprehensively determined by combining NDVI index, site light suitability, and water accessibility; site microclimate comfort is graded according to the suitability range of daily average temperature, average wind speed, and daily average humidity. At the same time, threshold ranges for each factor are set according to industry standards. For example, vegetation ecological suitability ≥0.7 is a high suitability zone, and <0.4 is a low suitability zone, thus completing the quantitative labeling of environmental factors.

[0033] It should be further explained that the pre-trained feature recognition model can be a three-dimensional point cloud semantic segmentation model (such as PointNet++ or RandLA-Net) trained on a large-scale garden scene dataset, which can identify and classify typical landscape features such as terrain, vegetation, water bodies, buildings, and roads.

[0034] S203. Structured Quantification and Cross-Dimensional Correlation Verification of Feature Factors: Standardization and cross-dimensional correlation verification are performed on multi-dimensional feature factors. The process is as follows: Single-factor data standardization: Standardize the terrain features, current element distribution, and environmental analysis factors to unify the data dimensions of various factors, eliminate the differences in dimensions between different factors, provide a unified data foundation for subsequent correlation verification, and ensure that factors of different dimensions can be compared and verified horizontally.

[0035] Cross-dimensional factor correlation verification: Construct a factor correlation verification matrix, calculate the correlation between factors of different dimensions, and compare the verification results with the threshold range. If the threshold is exceeded, trigger the data re-extraction process to ensure the rationality of the data logic.

[0036] Structured dataset encapsulation: All validated factors are categorized into "terrain type, current feature type, and environment type" to generate a structured dataset with unique identifiers, quantified values, and threshold ranges.

[0037] S204. Hierarchical Construction and Computable Encapsulation of Multi-Dimensional Design Constraints: Based on a structured factor dataset, a hierarchical and computable design constraint system is constructed. The process is as follows: Constraint hierarchy: Design constraints are divided into three levels, each level being bound to a corresponding quantitative judgment criterion. Mandatory constraints: Rigid constraints based on site safety and industry standards, such as prohibiting the construction of large structures in steep slope protection areas and requiring vegetation coverage of ≥80% in water source protection areas; Priority constraints: Flexible priority constraints based on ecological and functional needs, such as prioritizing greening planning in areas with suitable high vegetation and prioritizing the design of rainwater facilities in areas with high water catchment potential; Reference constraints: Suggestive constraints based on spatial perception and aesthetic needs, such as prioritizing the maintenance of landscape continuity in areas with high element coupling density.

[0038] Dynamic weighting of constraint priority: To enable differentiated application of constraints, a comprehensive priority weight calculation logic is introduced. The core formula is as follows: , in, The overall priority weight of the constraints. These are the weighting coefficients for the security, ecosystem, and functionality dimensions, respectively (default values ​​are 0.4, 0.35, and 0.25, which can be adjusted adaptively according to the project type). These are the basic weights of the constraint in the corresponding dimension.

[0039] Computable interface encapsulation: Convert all constraints into a parameterized computable format, and establish a constraint conflict warning interface. When a constraint conflict occurs, the conflict factor and priority-based solution will be automatically output.

[0040] Finally, the output includes a list of mandatory constraints, a list of preferred constraints, a list of reference constraints, and corresponding computable parameters, forming a multi-dimensional design constraint condition.

[0041] S3: Generate an initial landscape design scheme based on parametric rules and generative algorithms. The specific implementation process is as follows: S301. Adaptation and Reconstruction of the Parametric Design Rule Library for Landscape Architecture: The pre-built parametric design rule library for landscape architecture is adapted and reconstructed in a targeted manner. The specific process is as follows: Mapping and binding of rule base and design constraints: Extract the list of mandatory constraints, priority constraints and reference constraints, and map and bind them to the rule entries in the pre-set parametric design rule base.

[0042] Dynamic calibration of rule parameters: A quantitative evaluation logic for rule fit is introduced to calibrate the mapped rules. The core formula is as follows: , in, To ensure the fit and suitability of rules and constraints, For the first Priority weights of design constraints For the first The matching status of the constraints and corresponding rules is divided into complete matching (1), partial matching (0.6), and non-matching (0). The overall fit is required to be no less than 92%. Rules that do not meet the standard are corrected by manual intervention through parameter iteration.

[0043] Layered reconstruction of the rule base: The calibrated rules are reconstructed in layers.

[0044] S302. Selection and Collaborative Deployment of Multimodal Generative Algorithms: A collaborative approach using multimodal generative algorithms is adopted, as detailed below: Targeted selection of algorithm modules: Terrain Modification Module: Employs a parametric topology optimization algorithm to generate micro-terrain modification schemes based on terrain feature constraints; Vegetation configuration module: A vegetation community generation algorithm based on generative adversarial network (GAN) is adopted, which integrates vegetation ecological suitability constraints and plant seasonal matching rules to generate diverse vegetation community combination schemes. Structure layout module: Using a reinforcement learning-driven layout algorithm, combined with the functional requirements of buildings and site space constraints, the optimal location and form of facilities such as pavilions and pergolas are generated. Hard paving module: Using fractal algorithms combined with material adaptation rules, it generates paving styles and layout schemes that match the texture of the site.

[0045] Build an algorithm collaborative scheduling middleware to realize the time-series linkage and parameter exchange of algorithms in various modules.

[0046] S303. Step-by-step generation of multi-scale schemes: The initial scheme is constructed step by step using multi-scale step-by-step generation logic. The specific process is as follows: Macro-layout scheme generation: Taking the overall functional zoning of the site as the target, the terrain functional zoning and the current element distribution constraints are input, and the algorithm collaboration middleware is called to first complete the macro-spatial layout division, while ensuring the connectivity and rationality of each functional area.

[0047] Mesoscopic morphology scheme generation: Based on the macroscopic layout, mesoscopic morphology schemes are generated for each functional zone.

[0048] Micro-level detail scheme generation: Supplementing micro-level details for core nodes in the meso-level scheme.

[0049] S304. Diversity Guarantee and Preliminary Verification of Initial Scheme: Establish diversity guarantee mechanism and preliminary verification process, as follows: Scheme diversity assurance: By adjusting the random seed and floating threshold of the rule parameters in the generative algorithm, such as setting the floating range of the rainwater facility area threshold to ±10%, multiple sets of differentiated schemes are generated; at the same time, a quantitative evaluation of the scheme differentiation index is introduced to ensure that the difference in the core parameters of any two sets of schemes is ≥20%, forming an initial scheme pool with sufficient diversity.

[0050] Preliminary compliance verification of the scheme: The constraint conflict warning interface is called to perform preliminary verification of the schemes in the scheme pool, remove schemes that violate mandatory constraints, and mark schemes with defects in the adaptation of priority constraints. Finally, a preset number (8-12 groups) of schemes are selected from the scheme pool to form an initial landscape design scheme set. Each scheme is accompanied by core design parameters, constraint adaptation instructions and generation logic notes.

[0051] Ultimately, the output is a set of initial design schemes containing multiple sets of differentiated and compliant initial solutions.

[0052] S4: Optimization of landscape design schemes based on user interaction behavior, the specific implementation process is as follows: S401: Collection and Analysis of User Interaction Behavior Data: Through multi-source data collection and hierarchical analysis, accurately extract user preference features for the initial solution. The specific process is as follows: Multi-dimensional interactive data collection: A full-link interactive data collection system is established, deploying data collection modules on the solution demonstration terminal. The collected interactive behavior data covers three core types: Browsing trajectory data for each plan: including browsing duration, browsing order, dwell time in key areas, zoom / rotation and other operation records; Parameter adjustment sequence data: including user adjustments to core parameters of the plan, such as vegetation planting density; Option selection record data: including user-marked "preferred options", "alternative options", "excluded options", and selection tendencies during the option comparison process.

[0053] Interactive data preprocessing: The collected raw interactive data is cleaned and standardized, outliers are removed, and unstructured data is transformed into structured features. For browsing trajectory data, extract quantitative features such as "percentage of time spent on a particular solution" and "frequency of interaction in key areas", where the percentage of time spent on a particular solution = browsing time of a single solution / total browsing time; For parameter adjustment sequence data, extract features such as "adjustment parameter type preference" and "core parameter adjustment direction", such as statistically analyzing the parameter types that users adjust frequently; The selected options are recorded and assigned quantitative weights to different selection results, such as 1 for preferred options, 0.6 for alternative options, and 0 for excluded options, thus forming a selection preference vector.

[0054] User preference feature extraction: Construct an interaction feature extraction model to perform in-depth analysis on the preprocessed structured data and output user preference feature vectors, specifically including: functional preference features, morphological preference features, and performance preference features.

[0055] S402: Dynamic Adjustment of Solution Evaluation Function Weights: Based on extracted user preference features, the weight coefficients of the solution evaluation function are dynamically optimized through a machine learning model to achieve accurate matching between the evaluation logic and user needs. The specific process is as follows: Basic scheme evaluation function construction: The core framework of the preset scheme evaluation function is used to quantitatively evaluate the technical performance, economic performance, and spatial perception performance of the design scheme in three dimensions. The basic expression is: , in, This is the comprehensive evaluation value of the plan. These are the initial weighting coefficients for technical performance, economic performance, and spatial perception performance (default values ​​are 0.35, 0.3, and 0.35, respectively). These are the quantitative scores for the corresponding dimensions (score range 0-1).

[0056] Machine learning model training and weight adjustment: A gradient boosting tree model is used as the core model for weight adjustment to achieve dynamic adaptation of the evaluation weights. Model input and output definition: The extracted user preference feature vector is used as the model input, and the weight adjustment increment of the evaluation function is used as the model output; Training data construction: Using historical user interaction data and corresponding solution evaluation feedback results as training samples, the mapping relationship between "user preference features - optimal weight combination" is labeled, and the training model learns the correlation between user preferences and evaluation weights; Dynamic weight generation: Input the current user's preference feature vector into the trained model, output the weight adjustment increment, and update the weight coefficients of the basic evaluation function. The update formula is as follows: , in, The weighting coefficients are dynamically adjusted and satisfy the following conditions: .

[0057] Validation of weight adjustment effectiveness: Introduce the quantitative evaluation logic of "weight-preference fit" to verify the degree of matching between the adjusted weight and user preferences. If the fit is less than 85%, the current user's interaction data will be added to the training sample set, the model will be retrained and weights will be generated until the fit reaches the standard.

[0058] S403: Scheme Evaluation, Ranking, and Optimization Sequence Generation: Based on the dynamically adjusted scheme evaluation function, the initial scheme set is quantitatively evaluated and ranked to generate an optimized scheme sequence. The specific process is as follows: Multi-dimensional quantitative evaluation of the schemes: Substitute each scheme in the initial set of landscape design schemes into the adjusted evaluation function to calculate the comprehensive evaluation value of each scheme. The specific process is as follows: Calculate the technical performance score of each scheme separately. Economic performance score Spatial perception performance score ; Substitute the adjusted weighting coefficients to calculate The comprehensive evaluation value of each scheme is obtained.

[0059] Scheme ranking and optimization sequence generation: based on the comprehensive evaluation value of each scheme. Sort the results in descending order to form a preliminary sequence of optimized solutions; simultaneously, verify the reasonableness of the sorting results. Excluding comprehensive evaluation value A poor solution; Ensure that at least three different schemes are retained in the sorted sequence to avoid sequence homogenization.

[0060] Optimization Solution Sequence Output: The final output of the optimization solution sequence includes three core pieces of information: the comprehensive evaluation value of each solution, the core advantage indicators, and the user preference matching explanation, providing precise guidance for users' subsequent decision-making.

[0061] Finally, the output is a sequence of landscape design schemes optimized based on the user's personalized preferences.

[0062] S5: The detailed refinement, multi-dimensional verification, and final output of the landscape design scheme, the specific implementation process is as follows: S501. Standardization and Hierarchical Output of Optimization Solution Sequence: The generated optimization solution sequence is standardized and standardized, hierarchically divided according to the priority of the comprehensive evaluation value, and the delivery positioning of each level of solution is clarified. The specific process is as follows: Standardization and regularization of the solution sequence: The optimized solution sequence is encapsulated in a unified format, and each solution contains three core modules: Core design parameter set: covering terrain modification parameters, vegetation configuration parameters, structure layout parameters, hard paving parameters, and rainwater facility parameters. All parameters are presented in structured table format with the source of the parameters indicated. Visualized solution deliverables include a parametric 3D model, a plan layout diagram, and renderings of core nodes. The 3D model supports lightweight preview and interactive query of core parameters. Constraint adaptation description: For each design constraint, explain how the solution satisfies the mandatory constraints, the adaptation strategy for the preferred constraints, and the response method for the reference constraints.

[0063] Hierarchical division and output of the solution sequence: The solution sequence is divided into three levels according to the comprehensive evaluation value from high to low, and the delivery positioning of each level is clearly defined: Level 1 Solution (Top 1 in Comprehensive Evaluation): The core recommended solution, marked "Priority Implementation Recommendation", with a complete parametric model and detailed design specifications; Secondary solutions (Top 2-3 in comprehensive evaluation): Alternative optimization solutions, marked "Alternative Adaptation Solutions", retaining core design parameters and simplified visualization results; Level 3 Scheme (Top 4-6 in Comprehensive Evaluation): Differentiated reference scheme, marked "Style Differentiated Reference", outputting a summary of core parameters and effect diagrams of key nodes; The final output is in the form of a "scheme sequence data package". The data package has a built-in search function, which supports quick filtering and viewing by parameter type and constraint adaptation type.

[0064] S502. Dynamic Generation of Multi-Dimensional Assessment Reports Based on the evaluation logic and the actual performance of the solution, a multi-dimensional evaluation report covering four dimensions—technology, economics, ecosystem, and user preferences—is generated to provide quantitative basis for solution decision-making. The specific process is as follows: Dynamic adaptation of the evaluation indicator system: The evaluation indicator system is dynamically adjusted based on the solution type and user interaction preference characteristics. The core essential indicators are: technical compatibility, economic rationality, ecological compatibility, and user preference matching, which are universal evaluation indicators for all solutions; Personalized optional indicators: Supplement unique indicators based on project positioning.

[0065] Accurate calculation and summarization of assessment data: The results are calculated by calling the dynamic evaluation function, and the technical performance score, economic performance score, and spatial perception performance score of each scheme are extracted. Supplement the quantitative data on ecological benefits and match it with details of user preferences; Introducing quantitative evaluation logic, the core formula is as follows: , in, To assess the completeness of the report, For the first The completion status of each evaluation indicator is recorded as follows: 1 for complete completion, 0.6 for partial completion, and 0 for incomplete completion. For the first The decision importance weight of each indicator is determined, such as the weight of core mandatory indicators ≥ 0.8 and the weight of personalized indicators ≤ 0.3, and the completeness is required to be no less than 98%.

[0066] Structured output of the evaluation report: The evaluation report specifically includes: Report Summary: Summarizes the overall performance of the solution series, the core recommended solutions, and key evaluation conclusions; Detailed evaluation by dimension: The scores, quantitative data, and advantages and disadvantages of each solution are presented one by one according to the dimensions of technology, economy, ecology, and user preference; Solution Comparison and Decision Recommendations: A horizontal comparison of the core indicators of each solution is conducted, and targeted decision recommendations are provided based on actual needs such as project budget and implementation cycle. Attachment: Includes explanations of evaluation indicators, data calculation basis, and a summary of user interaction preference analysis.

[0067] S503. Lightweight Output of Supporting Documentation: Output lightweight supporting documentation to facilitate the implementation of the solution. The core documentation includes a benchmark manual of core design parameters, key risk warnings for implementation, and instructions for connecting the parametric model with mainstream design and construction platforms. This provides concise and accurate core guidance for subsequent implementation, avoiding redundant construction details.

[0068] Finally, the output includes an evaluation report containing "optimization scheme sequence data package, multi-dimensional evaluation report, and lightweight implementation support materials".

[0069] This invention also provides an automated landscape design system based on 3D reality, which includes the following modules: Data acquisition and semantic modeling module: acquire multi-source 3D real-scene data of the target site, and construct a semantic 3D real-scene model that supports automated parsing through hierarchical semantic segmentation and association with semantic ontology; Design constraint generation module: Based on the semantic 3D real scene model, through the pre-trained element recognition model and built-in parsing rules, automatically extracts and quantifies terrain features, current element distribution and environmental analysis factors, and generates structured multi-dimensional design constraints. Parametric scheme generation module: Based on the multi-dimensional design constraints and parametric rule library, it drives multiple generative algorithms to work together to automatically generate multiple differentiated initial landscape design schemes, forming an initial design scheme set. Interactive solution optimization module: Collects and analyzes user interaction data on the initial design solution set, dynamically adjusts the weights of the solution evaluation function through a machine learning model, and quantitatively evaluates and ranks the solutions accordingly, outputting an optimized solution sequence; Solution Output and Report Generation Module: Standardizes and hierarchically outputs the optimized solution sequence, and dynamically generates a complete report with multi-dimensional quantitative evaluation.

[0070] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other. In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A three-dimensional real scene-based automatic design method for landscape, characterized in that, include: S1: Acquire 3D real-scene data of the target site and perform semantic segmentation preprocessing to build a semantic 3D real-scene model that supports automated parsing; S2: Based on the semantic 3D real scene model, the terrain features, current element distribution and environmental analysis factors are automatically parsed and extracted through the pre-trained element recognition model to generate structured and computable multi-dimensional design constraints. S3: Based on parametric rules and generative algorithms, using the multi-dimensional design constraints as input, multiple initial landscape design schemes are automatically generated to form an initial design scheme set. S4: Based on user interaction data with the initial design scheme set, dynamically optimize the scheme generation strategy. The process includes: S401: Collect and analyze user interaction data on the initial design scheme set, the interaction data including scheme browsing trajectory, parameter adjustment sequence and scheme selection record; S402: Based on the interactive behavior data, dynamically adjust the weight coefficients in a scheme evaluation function through a machine learning model, wherein the scheme evaluation function is used to quantitatively evaluate multiple indicators in the technical performance, economic performance, and spatial perception performance of the design scheme. S403: Using the dynamically adjusted scheme evaluation function, calculate and evaluate each scheme in the initial design scheme set to obtain the comprehensive evaluation value of each scheme, and sort the schemes according to the comprehensive evaluation value to generate an optimized scheme sequence; S5: Output the optimized scheme sequence, in which each scheme is associated with a comprehensive evaluation value and a score of each performance index calculated based on the dynamically adjusted scheme evaluation function, and generate a complete quantitative evaluation report.

2. The method according to claim 1, wherein, The three-dimensional real-scene data includes: The system employs a three-tiered heterogeneous data acquisition equipment group consisting of air, ground, and space. The airborne equipment comprises a cluster of drones equipped with multispectral cameras and lidar, the ground-based equipment comprises mobile intelligent data acquisition vehicles and fixed blind spot filling terminals, and the space-ground collaborative equipment comprises GNSS differential base stations and 5G communication links. The model generates acquisition paths based on a pre-trained path planning model, and performs real-time fusion and quality verification on the acquired multi-source data, outputting a multi-dimensional 3D real-scene raw dataset containing spatial coordinates, geometric shape, material properties, spectral features, and time series information.

3. The method of claim 1, wherein the method is characterized by: The semantic 3D reality model includes: performing hierarchical semantic segmentation preprocessing on the multi-dimensional 3D reality original dataset, extracting geometric, spectral, temporal and material features, and outputting a standardized semantic segmentation dataset containing element semantic labels, geometric parameters, attribute features and topological relationships through multi-model collaborative segmentation. Based on the semantic ontology library of garden landscape elements, the semantic segmentation dataset is constructed into a semantic 3D real-world model containing a basic geometric layer, a semantic association layer, and a dynamic parsing layer. The dynamic parsing layer has built-in rule matching and parameter calculation components, which endow the model with automatic parsing capabilities.

4. The method of claim 1, wherein the method is characterized by, The multi-dimensional design constraints include: semantically mapping the semantic 3D real-world model to the pre-trained element recognition model and adapting the data interface; based on the adapted model, hierarchically analyzing and quantifying the extraction of terrain features, current element distribution, and environmental analysis factors; after standardizing and cross-dimensional correlation verification of the extracted factors, constructing structured design constraints at three levels: mandatory, priority, and reference, and assigning comprehensive priority weights based on safety, ecology, and functionality dimensions to each constraint.

5. The method of claim 1, wherein the method is characterized by: The initial design scheme set includes: mapping and calibrating the multi-dimensional design constraints with the parametric design rule base; driving multiple generative algorithms to work collaboratively, the algorithms including at least a parametric topology optimization algorithm for terrain modification, a generative adversarial network for vegetation configuration, a reinforcement learning algorithm for structure layout, and a fractal algorithm for hard paving; using a multi-scale hierarchical generation logic, first generating macroscopic layout schemes, then generating mesoscopic morphological schemes and microscopic detail schemes, and ensuring scheme diversity by adjusting algorithm parameters, ultimately forming the initial design scheme set.

6. The method of claim 1, wherein the method is characterized by: The dynamic optimization scheme generation strategy includes: Collect user browsing history, parameter adjustment sequence, and selection records for the plan, and extract user preference feature vectors; Using the user preference feature vector as input, the weight coefficients of technical performance, economic performance and spatial perception performance in the scheme evaluation function are dynamically adjusted through a machine learning model; The adjusted evaluation function is used to calculate the comprehensive evaluation value of each scheme, and the schemes are sorted accordingly to generate an optimized scheme sequence.

7. The method of claim 1, wherein the method is based on three-dimensional real scene. The pre-trained feature recognition model is a 3D point cloud semantic segmentation model trained on a large-scale garden scene dataset. 8.The three-dimensional real scene based garden landscape automatic design method according to claim 1, wherein, The complete quantitative evaluation report includes: standardizing and hierarchically outputting the optimized scheme sequence to form a scheme sequence data package containing a set of core design parameters, visualization results, and constraint adaptation instructions; generating a multi-dimensional quantitative evaluation report covering technical, economic, ecological, and user preference dimensions based on the dynamic evaluation results; and outputting lightweight support materials for scheme implementation.

9. A three-dimensional real scene-based garden landscape automatic design system, characterized in that, include: Data acquisition and semantic modeling module: acquire multi-source 3D real-scene data of the target site, and construct a semantic 3D real-scene model that supports automated parsing through hierarchical semantic segmentation and association with semantic ontology; Design constraint generation module: Based on the semantic 3D real scene model, through the pre-trained element recognition model and built-in parsing rules, automatically extracts and quantifies terrain features, current element distribution and environmental analysis factors, and generates structured multi-dimensional design constraints. Parametric scheme generation module: Based on the multi-dimensional design constraints and parametric rule library, it drives multiple generative algorithms to work together to automatically generate multiple differentiated initial landscape design schemes, forming an initial design scheme set. Interactive solution optimization module: Collects and analyzes user interaction data on the initial design solution set, dynamically adjusts the weights of the solution evaluation function through a machine learning model, and quantitatively evaluates and ranks the solutions accordingly, outputting an optimized solution sequence; Solution Output and Report Generation Module: Standardizes and hierarchically outputs the optimized solution sequence, and dynamically generates a complete report with multi-dimensional quantitative evaluation.

Images