Landscape design data processing method based on artificial intelligence

Abstract

The invention relates to the technical field of landscape design, and provides a landscape design data processing method based on artificial intelligence, which comprises the steps of multi-source heterogeneous data acquisition and fusion, knowledge graph driven design constraint extraction, dynamic demand perception and weight distribution, generative adversarial design of a coupling physical model, multi-objective optimization and scheme decision. And virtual-real fusion scheme verification can be carried out. The method comprises the following steps of: constructing a multi-modal data set by acquiring geographic information, real-time environment monitoring and other data, identifying constraint generation boundary conditions by using a knowledge graph, analyzing requirements and adjusting weights by using natural language processing and a reinforcement learning model, inputting constraint conditions and weight parameters into a generative adversarial network to generate a candidate scheme set, and constructing a multi-modal data set; and adopting a Pareto frontier algorithm to optimize and output an optimal solution set, and finally importing the optimal solution set into an augmented reality terminal for verification and adjustment. According to the method, landscape design data can be efficiently processed, a scheme meeting multi-aspect requirements is generated, and the design quality and efficiency are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of landscape design, and in particular to a landscape design data processing method based on artificial intelligence. Background Art

[0002] Landscape design data processing refers to the entire process of collecting, organizing, analyzing, visualizing, and applying various data related to the site, environment, and user needs in the field of landscape design. By processing these data, designers can more accurately grasp the site characteristics and user needs. With the help of tools such as data analysis models, geographic information systems, and big data technologies, complex data can be transformed into the basis for design decisions, thereby optimizing spatial layout, ecological functions, and user experience, making landscape design more scientific and sustainable while meeting aesthetic and functional requirements.

[0003] When faced with multi-source heterogeneous data, existing landscape design data processing methods have difficulty in effectively constructing multimodal datasets with temporal and spatial correlations. The extraction of design constraints such as terrain features and ecologically sensitive areas is not comprehensive and structured enough. At the same time, when processing user design needs, there is a lack of a mechanism for dynamically adjusting the weights of various indicators. The generated design schemes may have problems with insufficient physical feasibility, and the multi-objective optimization and scheme verification processes are not scientific and reasonable enough, resulting in the need to improve design efficiency and quality. Summary of the Invention

[0004] In response to the shortcomings of the existing technology, the present invention provides a landscape design data processing method based on artificial intelligence, which solves the problems of insufficient design efficiency and quality of existing landscape design data processing.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solutions: A landscape design data processing method based on artificial intelligence includes the following steps: S1. Multi-source heterogeneous data acquisition and fusion: Acquire geographic information data of the target site, real-time environmental monitoring data, user behavior trajectory data, and historical design case libraries to construct a multimodal dataset with temporal and spatial correlation; S2. Knowledge-graph-driven design constraint extraction: Utilizing a pre-trained landscape design knowledge graph, identify terrain features, ecologically sensitive areas, functional zoning hard constraints, and cultural attribute soft constraints in the multimodal dataset to generate structured design boundary conditions. S3. Dynamic Demand Perception and Weighting: This approach uses natural language processing to parse user-entered design requirements and, combined with a reinforcement learning model, dynamically adjusts the priority weights of ecological indicators, cost coefficients, and aesthetic evaluation factors. S4. Generative Adversarial Design with Coupled Physical Models: The constraints from step S2 and the weight parameters from step S3 are fed into a conditional generative adversarial network. The generator integrates a fluid dynamics simulation module and a vegetation growth prediction model to iteratively generate a set of candidate design solutions that meet physical feasibility. S5. Multi-objective optimization and solution decision-making: The Pareto frontier algorithm is used to perform multi-objective optimization on the carbon sequestration efficiency, construction cost, and spatial connectivity of the candidate solution set, and the Pareto optimal solution set is output for user selection.

[0006] Preferably, the geographic information data in step S1 includes laser point cloud terrain data and oblique photography three-dimensional model, the user behavior trajectory data is obtained by integrating Wi-Fi probe and GPS positioning to obtain a thermal map, and the real-time environmental monitoring data includes PM2.5, temperature and humidity, light intensity and noise decibel value.

[0007] Preferably, the method for constructing the knowledge graph in step S2 includes: S2.1. Extract entity-relationship triples from the landscape design specification text and build a regulatory constraint subgraph. S2.2. Identify spatial layout patterns in historical design case images using a convolutional neural network and construct visual feature subgraphs. S2.3. The regulatory constraint subgraph and the visual feature subgraph are integrated to form a cross-modal knowledge graph.

[0008] Preferably, the reinforcement learning model in step S3 uses the number of design modifications as a reward signal, and automatically reduces the weight of the corresponding factor when the user continuously rejects similar modification plans.

[0009] Preferably, the generator in step S4 includes a cascade structure: the first-level generator outputs a site topology relationship diagram, encoding the spatial connection relationship between roads, water systems and green spaces; the second-level generator generates a three-dimensional geometric scheme based on the topology relationship diagram, and calls a physical model to verify drainage efficiency and vegetation shade coverage.

[0010] Preferably, the physical model verification adopts a real-time feedback mechanism: if the drainage efficiency is lower than a threshold, the gradient penalty term in the adversarial training is activated to force the generator to relearn the fluid dynamics characteristics.

[0011] Preferably, the Pareto front algorithm in step S5 adopts NSGA-III optimizer, wherein the spatial connectivity index calculates node accessibility through a graph theory algorithm, and the carbon sink efficiency index is associated with carbon sequestration parameters through a tree species database.

[0012] Preferably, it also includes: S6. Virtual-reality fusion solution verification: importing the selected solution into the augmented reality terminal, superimposing the real site image to perform spatial scale verification, and triggering the automatic adjustment module based on constraint propagation when it is detected that the solution elements collide with the on-site objects.

[0013] Preferably, the automatic adjustment module adopts a graph neural network to model the dependencies between design elements and maintains minimal changes to the original design intent through a node message passing mechanism.

[0014] The present invention provides a landscape design data processing method based on artificial intelligence. It has the following beneficial effects: 1. This invention collects and integrates multi-source heterogeneous data to construct a spatiotemporal multimodal dataset, and extracts design constraints in combination with knowledge graphs. It can accurately identify hard constraints such as terrain and ecology, as well as soft constraints such as cultural attributes, providing comprehensive and structured boundary conditions for landscape design and avoiding design defects caused by incomplete consideration of constraints.

[0015] 2. The dynamic demand perception and weight allocation of the present invention, combined with the reinforcement learning model, can dynamically adjust the weights of factors such as ecology and cost based on user feedback. At the same time, the generative adversarial design coupled with the physical model can ensure the physical feasibility of the solution. Multi-objective optimization and virtual-reality fusion verification further enhance the scientificity and rationality of the solution, and improve design efficiency and quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is the process flow of the present invention. DETAILED DESCRIPTION

[0017] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0018] Example 1: Please see the attached Figure 1 The embodiment of the present invention provides a landscape design data processing method based on artificial intelligence. Taking the design of a 50-hectare urban comprehensive park as an example, the specific implementation steps include: S1. Multi-source heterogeneous data collection and fusion: A Trimble UX5 drone equipped with a Riegl VUX-1 lidar was used to acquire topographic data with a point cloud density of 50 points / m². TerraSolid processing was used to generate a 0.5m resolution DEM. A DJI P4RTK was used to acquire oblique photographic imagery, and ContextCapture was used to generate a 3D model with a 0.03m accuracy, including topographic features at a scale of 1:500. Thirty Cisco AIR-CT2504 Wi-Fi probes were deployed around the park, integrating GPS positioning data from the Baidu Maps API. Seven days of data collection generated a heat map, identifying hotspots such as the main entrance and children's activity areas. Fifteen EnviroPro monitoring stations were deployed to collect real-time PM2.5, temperature and humidity, light intensity, and noise data, which was uploaded to the cloud via a LoRa network. 300 park design cases were selected from the ASLA award-winning case library, and CAD drawings, energy consumption simulation reports, and user satisfaction survey data were structured and stored. The data uses the UTMZone50N coordinate system and is associated through spatiotemporal indexing to form a 350GB multimodal dataset.

[0019] S2. Knowledge graph-driven design constraint extraction: S2.1. Use the spaCyNLP tool to parse 12 design specifications, extract entity-relationship triples, and construct a regulatory constraint subgraph containing 4,200 triples. S2.2. Vectorize the historical case CAD drawings and use the FasterR-CNN model to identify elements such as roads, water bodies, and green spaces, constructing a visual feature subgraph containing 6,500 nodes. S2.3. Use the TransR entity alignment algorithm to fuse the regulations and visual subgraph to generate a cross-modal knowledge graph containing 12,000 nodes. The constraint extraction results show that the 12-hectare wetland in the northeast of the site is an ecological protection red line, and the building setback must be ≥50m. The south side is adjacent to the historical district, and the paving material must be bluestone slabs and match local intangible cultural heritage patterns.

[0020] S3. Dynamic Demand Perception and Weight Assignment: A user inputs "Build a sponge city demonstration park with low maintenance costs and high ecological benefits, with a budget of ≤ 30 million yuan." The RoBERTa model is used to extract ecological indicators, cost coefficients, and aesthetic factors, with initial weights set to [0.4, 0.35, 0.25]. The reinforcement learning model uses the DQN algorithm, with the number of user modifications serving as a reward signal. If the rainwater garden proposal is rejected twice in a row as being too expensive, the cost weight is increased to 0.45 and the ecological weight is reduced to 0.35 using the weight adjustment formula: ,in: is the weight vector at time t+1, and the learning rate , is the reward signal at time t, is the gradient of the loss function with respect to the weights. The loss function is defined as: ,in: Forecast the cost for the current solution, For budgetary goals, Cost deviation coefficient is the weight variance coefficient, which is used to maintain the stability of weight distribution.

[0021] S4. Generative Adversarial Design for Coupled Physical Models: A two-level cascade generator structure is used: the first-level generator generates a site topology map through the GAT graph attention network, defining the main garden path width as 6m and the water area as 15%-25%, satisfying the graph theory connectivity index λ(G) ≥ 2; the second-level generator integrates the OpenFOAM fluid module, simulates runoff based on the rainstorm intensity formula, and activates the gradient penalty when the drainage efficiency is less than 60mm / h.

[0022] S5. Multi-objective optimization and solution decision-making: Using the NSGA-III optimizer, the objective functions include: Carbon sequestration efficiency: , For the The number of tree species planted, For the The annual carbon sequestration of trees is 0.22 t / tree for ginkgo and 0.35 t / tree for camphor. The target value is ≥800 t / year. Construction cost: , For the The engineering quantity of each sub-item of the project, For the The comprehensive unit price of each sub-item of the project, is the total number of sub-items. This example includes earthwork, landscaping, and drainage projects, with a constraint of ≤ RMB 30 million. Spatial connectivity: , Functional Node With function node The reachability matrix element between nodes is 1. and Reachable, a value of 0 indicates unreachable; is the total number of functional nodes. In this embodiment, m=12, including nodes such as the main entrance, children's activity area, and rest square; The three objective functions were iteratively optimized using the NSGA-III optimizer, ultimately outputting seven Pareto optimal solutions. The optimal solution achieved an annual runoff control rate of 88%, a construction cost of RMB 28.5 million, and a spatial connectivity of 0.93, while meeting all the objectives and constraints. S6. Virtual-real fusion scheme verification: Through MagicLeap2AR device verification, when it is detected that the rainwater garden overlaps with the existing tree root system range, a design element dependency graph is constructed, the nodes include the rainwater garden, the tree, and the footpath, the edge defines the distance constraint >= 3m, and the GNN message passing algorithm is used to translate the rainwater garden by 3.5m and adjust the footpath curvature, while keeping the sponge facility proportion at 18%.

[0023] Embodiment two: Please refer to the attached Figure 1 The embodiment of the present application provides a landscape design data processing method based on artificial intelligence. For a residential area with a plot ratio of 2.5 and a land area of 8 hectares, the specific implementation steps include: S1. Multi-source heterogeneous data acquisition and fusion: 0.02m precision ground point cloud is obtained through a vehicle-mounted mobile measurement system to generate DEM data required for building setback and sunshine analysis; a three-dimensional model including building contours and hard site ranges is generated through oblique photography modeling with an accuracy of 0.01m. The peak time of children's activities and the hot path of the elderly's walking are identified through time and space clustering using community access card data and mobile phone signaling data. Eight micro monitoring stations are arranged between residential buildings to collect wind speed and sunshine duration data to support wind environment and sunshine analysis. S2. Knowledge graph driven design constraint extraction: The "residential sunshine standard >= 2 hours on the Da Han day" regulatory constraint is extracted from the "Urban Residential District Planning and Design Standard" GB50180-2018, combined with the minimum width of 4m of the inter-residential green space and the visual features of the safety distance of children's facilities and roads >= 2m identified from historical cases, a residential district exclusive knowledge graph is constructed, including the cross-modal fusion of the regulatory constraint subgraph and the visual feature subgraph.

[0024] S3. Dynamic demand perception and weight distribution: After analyzing the user demand "create a safe and comfortable, parent-child friendly residential landscape, and control the hard ground ratio <= 40%", the safety index, function index, and economy index are determined, and the initial weight is [0.3, 0.4, 0.3]. The reinforcement learning model takes the "number of safety hazard modifications" as the reward signal, and automatically increases the safety index weight to 0.45 when the distance between the children's slide and the road is adjusted for three consecutive times. Priority adjustment is achieved through a dynamic weight algorithm. S4. Coupling physical model of generative adversarial design: The first generator outputs the topological relationship of inter-residential green space, activity site, and fire access, defines the fire access width >= 4m and the turning radius >= 9m; the second generator integrates the Osimertinib wind environment simulation module, sets the wind speed in the children's activity area in winter <= 5m / s, and triggers gradient penalty optimization layout when it is not satisfied to ensure wind environment comfort. S5. Multi-objective optimization and solution decision-making: The optimization goal is to increase the sunshine compliance rate, requiring ≥90% of residents to have sunshine for ≥2 hours on the Great Cold Day. The NSGA-III algorithm is used to calculate the sunshine blocking relationship between buildings through graph theory. Combined with the hard ground ratio constraint and the unit cost ≤1,200 yuan / m2, a set of candidate solutions that meet the "Building Sunlight Calculation Standard" are generated. S6. Virtual-Real Integration Solution Verification: When the AR verification detects that the distance between fitness equipment and the residential building entrance steps is less than 1.5m, the automatic adjustment module is activated. A design element dependency graph is constructed, with nodes including fitness equipment, steps, and paving, and edges defining an accessibility distance constraint of ≥ 1.5m. Using a graph neural network message passing mechanism, the equipment is translated 2m and the paving dividing line is updated, maintaining the same equipment type and minimizing design changes.

[0025] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A landscape design data processing method based on artificial intelligence, characterized in that: The following steps are involved: S1. Multi-source heterogeneous data acquisition and fusion: Acquire geographic information data of the target site, real-time environmental monitoring data, user behavior trajectory data, and historical design case libraries to construct a multimodal dataset with temporal and spatial correlation; S2. Knowledge-graph-driven design constraint extraction: Utilizing a pre-trained landscape design knowledge graph, identify terrain features, ecologically sensitive areas, functional zoning hard constraints, and cultural attribute soft constraints in the multimodal dataset to generate structured design boundary conditions. S3. Dynamic Demand Perception and Weighting: This approach uses natural language processing to parse user-entered design requirements and, combined with a reinforcement learning model, dynamically adjusts the priority weights of ecological indicators, cost coefficients, and aesthetic evaluation factors. S4. Generative Adversarial Design with Coupled Physical Models: The constraints from step S2 and the weight parameters from step S3 are fed into a conditional generative adversarial network. The generator integrates a fluid dynamics simulation module and a vegetation growth prediction model to iteratively generate a set of candidate design solutions that meet physical feasibility. S5. Multi-objective optimization and solution decision-making: The Pareto frontier algorithm is used to perform multi-objective optimization on the carbon sequestration efficiency, construction cost, and spatial connectivity of the candidate solution set, and the Pareto optimal solution set is output for user selection.

2. The artificial intelligence-based landscape design data processing method according to claim 1, characterized in that: The geographic information data in step S1 includes laser point cloud terrain data and oblique photography three-dimensional model, the user behavior trajectory data is obtained by integrating Wi-Fi probes with GPS positioning to obtain a thermal map, and the real-time environmental monitoring data includes PM2.5, temperature and humidity, light intensity and noise decibel value.

3. The artificial intelligence-based landscape design data processing method according to claim 1, characterized in that: The method for constructing the knowledge graph in step S2 includes: S2.

1. Extract entity-relationship triples from the landscape design specification text and build a regulatory constraint subgraph. S2.

2. Identify spatial layout patterns in historical design case images using a convolutional neural network and construct visual feature subgraphs. S2.

3. The regulatory constraint subgraph and the visual feature subgraph are integrated to form a cross-modal knowledge graph.

4. The artificial intelligence-based landscape design data processing method according to claim 1, characterized in that: The reinforcement learning model described in step S3 uses the number of design modifications as a reward signal, and automatically reduces the weight of the corresponding factor when the user continuously rejects similar modification plans.

5. The artificial intelligence-based landscape design data processing method according to claim 1, characterized in that: The generator in step S4 includes a cascade structure: the first-level generator outputs a site topology diagram, encoding the spatial connection relationship between roads, water systems, and green spaces; the second-level generator generates a three-dimensional geometric scheme based on the topology diagram, and calls a physical model to verify drainage efficiency and vegetation shade coverage.

6. The artificial intelligence-based landscape design data processing method according to claim 5, characterized in that: The physical model verification adopts a real-time feedback mechanism: if the drainage efficiency is lower than a threshold, the gradient penalty term in the adversarial training is activated to force the generator to relearn the fluid dynamics characteristics.

7. The artificial intelligence-based landscape design data processing method according to claim 1, characterized in that: The Pareto front algorithm described in step S5 adopts the NSGA-III optimizer, wherein the spatial connectivity index calculates the node accessibility through a graph theory algorithm, and the carbon sink efficiency index is associated with the carbon sequestration parameter through the tree species database.

8. The artificial intelligence-based landscape design data processing method according to claim 1, characterized in that: Also includes: S6. Virtual-reality fusion solution verification: The selected solution is imported into the augmented reality terminal and the real site image is superimposed for spatial scale verification. When a collision between the solution elements and the on-site objects is detected, the automatic adjustment module based on constraint propagation is triggered.

9. The artificial intelligence-based landscape design data processing method according to claim 8, characterized in that: The automatic adjustment module uses a graph neural network to model the dependencies between design elements and maintains minimal changes to the original design intent through a node message passing mechanism.

Images