Virtual reality (VR)-based landscape design dynamic display method and system

Abstract

The invention discloses a virtual reality (VR)-based landscape design dynamic display method and system, and the method comprises the steps: collecting real-time environment data and plant growth cycle information through a multi-source sensor, carrying out the processing through a particle filtering algorithm, and obtaining a preliminary time-space alignment dynamic data set; according to the dynamic data set subjected to the preliminary time-space alignment, a neural network model is adopted to train the mutual influence relation among the key feature parameters, and a coupling coefficient matrix is determined; obtaining a logic rule of environment interaction from the coupling coefficient matrix to obtain an interaction graph containing time-space consistency; on the basis of a node connection structure in the interaction map, updating a dynamic state value of each node through iterative simulation equipment to obtain an optimized scene evolution sequence; if a part with inconsistent data exists in the optimized scene evolution sequence, determining a collaborative driving model; and generating a rendering data stream of virtual display according to the collaborative driving model, and generating a continuous landscape dynamic result.

Description

Technical Field

[0001] This invention belongs to the field of virtual reality technology, and in particular relates to a method and system for dynamic display of landscape design based on virtual reality (VR). Background Technology

[0002] Landscape design, as a crucial element in building a harmonious living environment, aims to anticipate and shape the vitality and beauty of space over time. With the development of digital technology, design expression has evolved from traditional drawings to 3D visualization. However, existing virtual display methods still face significant limitations in conveying the "vitality" of a landscape. These methods are not merely lacking in dynamic effects; their deeper limitation lies in treating each dynamic element in the landscape as an isolated, independently programmable module, neglecting the complex coupling relationships and mutual influences between elements in the real world. This results in a lack of inherent logical unity in the simulated scene.

[0003] This fragmented approach raises a core technical challenge: how to ensure the spatiotemporal consistency of multi-source heterogeneous dynamic data. The realism of a landscape scene stems from the fact that all its internal changes follow unified temporal and physical laws. For example, the growth state of plants is not only related to their own biological cycles but is also affected by real-time environmental data such as light intensity, temperature, and humidity. If the seasonal changes of plants in the system cannot be precisely synchronized with the simulated data of the environmental climate, or if the wind data and the swaying amplitude of tree branches and leaves lack a precise physical logical connection, a false sense of "superficial harmony" will be generated. The root cause of this distortion in dynamic effects lies in the failure to establish an underlying framework that can effectively integrate and collaboratively drive dynamic information from different sources and time scales, making the various dynamic effects "dynamic" but not "alive."

[0004] Specifically, when a designer attempts to evaluate a waterfront design, they need to consider not only the fountain's spray pattern but also the dispersion range of the water flow at different wind speeds, the subtle impact of the water mist on the surrounding microclimate (such as local humidity), and the long-term changes in the texture of the waterside stone under sunlight and moisture erosion. If these dynamic changes are based on independent pre-set scripts rather than driven by a unified environmental physics engine, the evolution of the entire scene will lose credibility and fail to provide a reliable basis for design optimization.

[0005] Therefore, how to construct a system that can coordinate discrete, multi-source dynamic data (such as plant growth cycles, environmental and climate changes, and material physical feedback) in time and space, and transform it into a unified, logically self-consistent virtual landscape system, allowing designers and users to intuitively experience and verify the real evolution of landscape elements and the complex interaction with the environment at different time scales, has become a key issue in driving the landscape design industry into the era of high-fidelity dynamic simulation. Summary of the Invention

[0006] To address the aforementioned technical issues, this invention proposes a method and system for dynamic landscape design display based on virtual reality (VR). This method can construct a system that coordinates discrete, multi-source dynamic data (such as plant growth cycles, environmental climate changes, and material physical feedback) in time and space, and transforms them into a unified, logically self-consistent virtual landscape system. This allows designers and users to intuitively experience and verify the real evolution of landscape elements and the complex interaction with the environment at different time scales.

[0007] To achieve the above objectives, the present invention provides a method for dynamic display of landscape design based on virtual reality (VR), comprising:

[0008] Real-time environmental data and plant growth cycle information are collected by multi-source sensors and processed using a particle filter algorithm to obtain a preliminary spatiotemporally aligned dynamic dataset.

[0009] Based on the preliminary spatiotemporally aligned dynamic dataset, key feature parameters are extracted using a feature extraction device, and a neural network model is used to train the mutual influence relationship between the key feature parameters to determine the coupling coefficient matrix.

[0010] The logical rules of environmental interaction are obtained from the coupling coefficient matrix, and a node connection structure under a unified framework is constructed using a graph neural network to obtain an interaction graph containing spatiotemporal consistency.

[0011] Based on the node connection structure in the interaction graph, the dynamic state value of each node is updated by iteratively simulating the device, and it is determined whether the updated state value conforms to the constraints of physical laws to obtain an optimized scene evolution sequence.

[0012] If there are inconsistent data parts in the optimized scenario evolution sequence, supplementary information is obtained from multi-source heterogeneous data sources and fused into the sequence to determine the collaborative driving model;

[0013] The rendering data stream for virtual display is generated based on the collaborative driving model, and the rendering data stream is processed by a real-time rendering engine to produce continuous dynamic landscape results.

[0014] Optionally, real-time environmental data and plant growth cycle information are collected through multi-source sensors and processed using a particle filter algorithm to obtain a preliminary spatiotemporally aligned dynamic dataset, including:

[0015] Real-time environmental data and plant growth cycle information are collected through multi-source sensors;

[0016] A semantic segmentation network was used to extract the outline information of the plant canopy and stem from the plant growth cycle information, and combined with environmental data, an initial multidimensional data record containing temperature, humidity, light intensity and canopy area was constructed.

[0017] Based on the historical fluctuation range of each dimension in the initial multidimensional data record, calculate the information entropy of the current data point. If the information entropy exceeds the preset confidence interval, mark the corresponding data point as a potential outlier to be processed.

[0018] The particle filter algorithm is used to perform state estimation and weight update on the data marked as potential outliers based on the state transition equation of the plant growth model, so as to obtain a corrected smooth data sequence.

[0019] A unified spatiotemporal axis is established based on the global timing signal. The smoothed data sequences from different sources are resampled using the cubic spline interpolation method to obtain the preliminary spatiotemporally aligned dynamic dataset.

[0020] Optionally, based on the preliminary spatiotemporally aligned dynamic dataset, key feature parameters are extracted using a feature extraction device, and a neural network model is used to train the interaction relationships between the key feature parameters to determine the coupling coefficient matrix, including:

[0021] Key coupling coefficients are extracted from the initial spatiotemporally aligned dynamic dataset to generate a table of inter-feature correlation strength.

[0022] Based on the correlation strength table, a feature interaction network graph is constructed to determine the dynamic coupling mode between wind intensity and light intensity changes;

[0023] If the complexity of the dynamic coupling pattern exceeds a preset threshold, a simplified coupling pattern set is obtained by grouping the patterns using clustering analysis.

[0024] By simplifying the set of coupling modes, the parameters of the multilayer perceptron neural network model are optimized to obtain the optimized model;

[0025] The optimized model is used to predict the interaction trend between wind intensity and light intensity changes in the new input dataset, and the prediction interaction results are obtained.

[0026] From the predicted interaction results, significant trend features are extracted to determine the coupling coefficient matrix.

[0027] Optionally, extracting significant trend features from the predicted interaction results to determine the coupling coefficient matrix includes:

[0028] If the value of an element in the coupling coefficient matrix exceeds a preset threshold, the set of elements exceeding the threshold is obtained by threshold judgment, and the list of dynamic elements to be adjusted is determined.

[0029] Based on the list of dynamic elements to be adjusted, the K-means clustering algorithm is used to group the dynamic elements to obtain the grouping results of the elements;

[0030] Based on the grouping results, each group of dynamic elements is sorted using a priority sorting algorithm to obtain the sorted element sequence;

[0031] Based on the sorted element sequence, adjust the positions of the corresponding elements in the coupling coefficient matrix to obtain the preliminary adjusted matrix;

[0032] If the matrix still contains elements whose values ​​exceed the preset threshold after initial adjustment, the matrix is ​​locally adjusted using the gradient descent algorithm to obtain the optimized matrix.

[0033] Based on the optimized matrix, verify whether all element values ​​meet the preset threshold conditions to obtain the final coupling coefficient matrix.

[0034] Optionally, logical rules for environmental interaction are obtained from the coupling coefficient matrix, and a node connection structure under a unified framework is constructed using a graph neural network to obtain an interaction graph containing spatiotemporal consistency, including:

[0035] Based on the coupling coefficient matrix, obtain a list of strongly coupled relationships;

[0036] Based on the list of strongly coupled relationships and the associated entity state data, the Apriori algorithm is used to analyze the co-occurrence patterns of entity pairs in the time dimension, and a preliminary set of interaction rules describing the influence of the association between entities is generated.

[0037] By parsing the logical relationships in the initial set of interaction rules, each independent entity is abstracted into a graph node, and the pointing relationships between entities in the rules are transformed into directed connections between nodes, thus obtaining an initial topology graph.

[0038] Based on each graph node in the initial topology graph, obtain the corresponding spatial location information and historical state time series data, use the spatial location information as the static attribute of the node, and use the time series data as the dynamic attribute of the node to obtain a spatiotemporal node set.

[0039] The topological graph containing the spatiotemporal node set is input into the graph attention network model to obtain new node embeddings containing dynamic evolution information;

[0040] Cosine similarity is used to calculate the similarity between new node embeddings, generating a weighted directed graph where nodes represent environmental entities and edge weights represent interaction strengths, thus obtaining an interaction graph with spatiotemporal consistency.

[0041] Optionally, based on the node connection structure in the interaction graph, the dynamic state value of each node is updated iteratively by simulating the device, and it is determined whether the updated state value conforms to the constraints of physical laws, to obtain an optimized scene evolution sequence, including:

[0042] Obtain the initial node set, connectivity, and structure information of the interaction graph, and assign an initial state value to each node to form an initial scene graph;

[0043] Based on the initial scene diagram, a set of constraints related to the physical quantities of nodes is extracted from the pre-established rule base to construct a discriminant for state evolution;

[0044] An iterative simulation method is used to calculate the potential rate of change of the state value of each node based on the connectivity between nodes, and a scenario graph containing temporary dynamic values ​​is generated.

[0045] The discriminant is used to perform compliance checks on the temporary dynamic values ​​of each node in the scenario graph to be tested. If the temporary dynamic value conforms to the constraint set, it is determined to be a valid state; if the temporary dynamic value does not conform to the constraint set, it is determined to be an invalid state.

[0046] Based on the determination results of valid and invalid states, update the evolution tree, take the scene graph to be verified containing valid states as the new branch node of the evolution tree, and discard the scene graph to be verified containing invalid states.

[0047] When the evolution tree reaches a preset simulation depth, multiple complete state evolution paths are extracted from the root node to all leaf nodes to form a candidate sequence set.

[0048] The Q-learning algorithm is used to evaluate the cumulative reward value of each evolution path in the candidate sequence set, and the evolution path with the highest cumulative reward value is determined as the final optimized scenario evolution sequence.

[0049] Optionally, if there are inconsistent data portions in the optimized scenario evolution sequence, supplementary information is retrieved from multi-source heterogeneous data sources and fused into the sequence to determine the collaborative driving model, including:

[0050] A time-series logic verification method is used to scan the continuous time nodes in the optimized scene evolution sequence, and to determine whether there are data breakpoints or logical conflicts according to preset logic rules, so as to obtain the location information of inconsistent segments.

[0051] Based on the location information of the inconsistent fragments, entities and relationships in a predefined knowledge graph are matched, and supplementary data streams related to the corresponding fragment timestamps and event types are obtained from multiple heterogeneous supplementary sources.

[0052] By performing structured parsing and semantic alignment on the supplementary data stream, and converting it into standardized information units with the same data specification as the original scene evolution sequence, a data set to be fused is obtained.

[0053] An attention-based fusion strategy is adopted to calculate the contribution weight of each standardized information unit in the dataset to be fused to the inconsistent segment, and an initial calibration sequence is generated by weighted summation.

[0054] If the initial calibration sequence passes the consistency check in the timing logic check, the corresponding sequence is determined as the final calibration sequence; if it fails, the weight parameters of the attention mechanism are adjusted and the fusion is re-executed until the final calibration sequence is obtained.

[0055] Based on the final calibration sequence, a long short-term memory network is used for model training. By learning the state transitions and cooperative relationships in the sequence, the cooperative driving model is obtained.

[0056] Optionally, generating a rendering data stream for the virtual display based on the collaborative driving model, and processing the rendering data stream using a real-time rendering engine to produce continuous dynamic landscape results includes:

[0057] The driving instruction set output by the collaborative driving model is obtained, and the driving instruction set is parsed one by one by a syntax parser to obtain a structured scene parameter set containing entity position, posture and environmental state;

[0058] Based on the entity identifiers in the structured scene parameter set, retrieve the corresponding 3D geometric data and surface texture map information from the digital asset library to obtain the original asset set required to construct the scene;

[0059] The original asset set is bound to predefined physical field parameters using a physics simulation module, and an enhanced scene description with added dynamic attributes is obtained through calculation.

[0060] The serialization processor encodes the geometric vertices, texture map indexes, and physical field vectors in the enhanced scene description in a unified manner, generating a continuous rendering data stream containing timestamps.

[0061] The continuous rendering data stream is received by a real-time rendering engine. If the data frame integrity check of the data stream passes, it is unpacked in the rendering pipeline to reconstruct the layered scene tree.

[0062] Based on the node information of the layered scene tree and the associated illumination field data, shading calculations and post-processing are performed to generate a single-frame image and append it to the output frame sequence to obtain the final high-fidelity simulated visual scene.

[0063] The present invention also provides a landscape design dynamic display system based on virtual reality (VR), including: a data acquisition module, a feature extraction module, a map construction module, a scene evolution sequence optimization module, a model construction module, and a result output module;

[0064] The data acquisition module is used to collect real-time environmental data and plant growth cycle information through multi-source sensors, and process them using a particle filter algorithm to obtain a preliminary spatiotemporally aligned dynamic dataset.

[0065] The feature extraction module is used to extract key feature parameters based on the preliminary spatiotemporally aligned dynamic dataset using a feature extraction device, train the mutual influence relationship between the key feature parameters using a neural network model, and determine the coupling coefficient matrix.

[0066] The graph construction module is used to obtain the logical rules of environmental interaction from the coupling coefficient matrix, and to construct a node connection structure under a unified framework using a graph neural network to obtain an interaction graph containing spatiotemporal consistency.

[0067] The scene evolution sequence optimization module is used to update the dynamic state value of each node by iteratively simulating the device based on the node connection structure in the interaction graph, determine whether the updated state value conforms to the constraints of physical laws, and obtain an optimized scene evolution sequence.

[0068] The model building module is used to obtain supplementary information from multiple heterogeneous data sources and integrate it into the sequence if there are inconsistent parts in the optimized scene evolution sequence, thereby determining the collaborative driving model.

[0069] The result output module is used to generate a rendering data stream for virtual display based on the collaborative driving model, and to process the rendering data stream using a real-time rendering engine to produce continuous dynamic landscape results.

[0070] Compared with the prior art, the present invention has the following advantages and technical effects:

[0071] This invention processes multi-source data through particle filtering and utilizes neural networks to train the interaction between characteristic parameters such as wind force and illumination, quantifying it into a coupling coefficient matrix. Based on this matrix, a spatiotemporally consistent graph neural network interaction map is constructed. This invention optimizes the scene evolution sequence through iterative simulation and physical constraints, and corrects inconsistencies by fusing supplementary information. The resulting collaborative driving model provides a precise and physically consistent data stream for real-time rendering engines, thereby generating high-fidelity dynamic landscape effects. Attached Figure Description

[0072] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0073] Figure 1 This is a flowchart of a dynamic display method for landscape design based on virtual reality (VR) according to an embodiment of the present invention. Detailed Implementation

[0074] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0075] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0076] This embodiment proposes a dynamic display method for landscape design based on virtual reality (VR), such as... Figure 1 As shown, the specific steps include:

[0077] Real-time environmental data and plant growth cycle information are collected by multi-source sensors and processed using a particle filter algorithm to obtain a preliminary spatiotemporally aligned dynamic dataset.

[0078] Based on the preliminary spatiotemporally aligned dynamic dataset, key feature parameters are extracted using a feature extraction device, and a neural network model is used to train the mutual influence relationship between the key feature parameters to determine the coupling coefficient matrix.

[0079] Logical rules for environmental interaction are obtained from the coupling coefficient matrix, and a node connection structure under a unified framework is constructed using a graph neural network to obtain an interaction graph containing spatiotemporal consistency.

[0080] Based on the node connection structure in the interactive graph, the dynamic state value of each node is updated by iteratively simulating the device, and it is determined whether the updated state value conforms to the constraints of physical laws to obtain an optimized scene evolution sequence.

[0081] If there are inconsistent data parts in the optimized scenario evolution sequence, supplementary information is obtained from multi-source heterogeneous data sources and fused into the sequence to determine the collaborative driving model;

[0082] The rendering data stream for virtual display is generated based on the collaborative driving model, and a real-time rendering engine is used to process the rendering data stream to produce continuous dynamic landscape results.

[0083] Specifically, this invention processes multi-source data through particle filtering and utilizes neural networks to train the interaction between characteristic parameters such as wind force and illumination, quantifying it into a coupling coefficient matrix. Based on this matrix, a graph neural network interaction map with spatiotemporal consistency is constructed. This invention optimizes the scene evolution sequence through iterative simulation and physical constraints, and corrects inconsistencies by fusing supplementary information. The resulting collaborative driving model provides a precise and physically consistent data stream for the real-time rendering engine, thereby generating high-fidelity dynamic landscape effects.

[0084] Furthermore, real-time environmental data and plant growth cycle information were collected through multi-source sensors and processed using a particle filter algorithm to obtain a preliminary spatiotemporally aligned dynamic dataset, including:

[0085] Real-time environmental data and plant growth cycle information are collected through multi-source sensors;

[0086] Semantic segmentation network was used to extract the outline information of plant canopy and stem from plant growth cycle information, and combined with environmental data, an initial multidimensional data record containing temperature, humidity, light intensity and canopy area was constructed.

[0087] Based on the historical fluctuation range of each dimension in the initial multidimensional data record, calculate the information entropy of the current data point. If the information entropy exceeds the preset confidence interval, mark the corresponding data point as a potential outlier to be processed.

[0088] The particle filter algorithm is used to perform state estimation and weight update on the data marked as potential outliers based on the state transition equation of the plant growth model, so as to obtain a corrected smooth data sequence.

[0089] A unified spatiotemporal axis is established based on the global timing signal. Smooth data sequences from different sources are resampled using cubic spline interpolation to obtain a preliminary spatiotemporally aligned dynamic dataset.

[0090] Specifically, in one possible implementation, the device identifier and timestamp protocol ensures millisecond-level time synchronization accuracy by assigning a unique MAC address to each sensor and combining it with GPS timing signals. Specifically, the soil moisture sensor in the greenhouse is identified as A001, the light sensor as B002, and the camera as C003. All devices upload data every 10 seconds, with timestamps uniformly in UTC standard time, ensuring accurate correspondence between sensor data distributed across different areas. For example, the semantic segmentation network uses the DeepLab architecture to perform pixel-level classification of plant images, labeling the leaf area of ​​tomato plants as green pixels and the stem area as brown pixels. By calculating the total area of ​​green pixels, the canopy coverage area is determined to be 0.85 square meters. Combined with simultaneously collected data such as temperature (26.5 degrees Celsius), humidity (65%), and light intensity (450 lux), a four-dimensional data vector is constructed. In one embodiment, information entropy calculation is based on the historical distribution probability of data in each dimension. When the historical temperature fluctuation range is 20-30 degrees Celsius, a sudden outlier of 35 degrees Celsius occurs. This outlier has an extremely low probability of occurrence, and the calculated information entropy is 4.2 bits, exceeding the pre-set 3.5-bit threshold. Therefore, it is marked as a potential outlier. For example, a particle filter algorithm maintains 1000 particles to represent the possible distribution of plant growth states. For the outlier temperature of 35 degrees Celsius, the algorithm predicts, based on the plant growth model, that leaves will wilt at this temperature, and the canopy area should decrease to 0.75 square meters. By comparing the predicted value with the actual observed value, the weights of each particle are updated, and the final corrected temperature value is output as 28.5 degrees Celsius. Specifically, the cubic spline interpolation method addresses the problem of inconsistent sampling frequencies from different sensors. The temperature sensor samples every 5 seconds, and the image sensor samples every 15 seconds. Interpolation unifies all data onto a time axis with one sampling point every 10 seconds, ensuring complete alignment of data across all dimensions during subsequent analysis.

[0091] Furthermore, based on the preliminary spatiotemporally aligned dynamic dataset, key feature parameters are extracted using a feature extraction device. A neural network model is then used to train the interaction relationships between these key feature parameters, and the coupling coefficient matrix is ​​determined, including:

[0092] Key coupling coefficients are extracted from the initial spatiotemporally aligned dynamic dataset to generate a table of inter-feature correlation strength.

[0093] Based on the correlation strength table, a feature interaction network graph is constructed to determine the dynamic coupling mode between wind intensity and light intensity changes;

[0094] If the complexity of the dynamic coupling pattern exceeds a preset threshold, a simplified coupling pattern set is obtained by grouping the patterns using clustering analysis.

[0095] By simplifying the set of coupling modes, the parameters of the multilayer perceptron neural network model are optimized to obtain the optimized model;

[0096] The optimized model is used to predict the interaction trend between wind intensity and light intensity changes in the new input dataset, and the prediction interaction results are obtained.

[0097] From the predicted interaction results, significant trend features are extracted to determine the coupling coefficient matrix.

[0098] Specifically, in a plant growth environment monitoring system, cross-validation is used to evaluate the stability of the coupling coefficient matrix by dividing historical datasets into training and validation sets. Specifically, the system divides temperature, humidity, light, and wind data from the past 30 days into a 7:3 ratio and calculates the coupling coefficients between environmental factors using 5-fold cross-validation. When the validation error is less than 0.05, the matrix stability is confirmed. In one embodiment, key coefficients are extracted from the stable coupling coefficient matrix to construct a feature correlation strength table. For example, the correlation coefficient between wind intensity and light intensity is 0.73, and the correlation coefficient between temperature and humidity is -0.68; these values ​​reflect the degree of mutual influence between environmental factors. The system filters out significantly correlated factor pairs based on a correlation strength threshold of 0.6, forming a feature interaction network graph containing node and edge weights. For example, when the network graph contains more than 15 nodes and the connection density is greater than 0.8, the system determines that the complexity of the dynamic coupling pattern exceeds a preset threshold. In this case, K-means clustering analysis is used to classify similar environmental factor combinations, simplifying the originally complex multi-factor interaction patterns into 3-5 main coupling pattern groups. The first set of coupling patterns may include light-temperature coupling, the second set includes wind-humidity coupling, and the third set includes soil temperature-root growth coupling. In one possible implementation, the simplified set of coupling patterns is used to optimize the number of hidden layer nodes and the learning rate parameter of the multilayer perceptron neural network. The original network contained 128 hidden nodes, which were adjusted to 64 nodes after being guided by the coupling patterns, and the learning rate was optimized from 0.01 to 0.005, significantly improving the accuracy of the model's prediction of the plant growth environment. For example, the optimized model receives new 24-hour environmental monitoring data and predicts that the wind intensity will increase from the current 3.2 m / s to 4.8 m / s in the next 48 hours, while the light intensity will decrease from 35,000 lux to 28,000 lux. The system identifies this inverse coupling trend of increasing wind and decreasing light, providing a basis for timely adjustment of plant protection measures. It should be noted that the significant trend features extracted from the prediction interaction results include key indicators such as the magnitude of change, the rate of change, and the duration. When wind intensity changes by more than 1.5 meters per second for more than 6 hours, the system marks this as a significant trend feature. The resulting dynamic coupling coefficient matrix updates these feature weights in real time, ensuring that plant growth environment regulation strategies can adapt to constantly changing external conditions and providing reliable data support for precision agriculture management.

[0099] Furthermore, significant trend features are extracted from the predicted interaction results, and the coupling coefficient matrix is ​​determined as follows:

[0100] If the value of an element in the coupling coefficient matrix exceeds a preset threshold, the set of elements exceeding the threshold is obtained by threshold judgment, and the list of dynamic elements to be adjusted is determined.

[0101] Based on the list of dynamic elements to be adjusted, the K-means clustering algorithm is used to group the dynamic elements to obtain the grouping results of the elements;

[0102] Based on the grouping results, the dynamic elements in each group are sorted using a priority sorting algorithm to obtain the sorted element sequence;

[0103] Based on the sorted element sequence, adjust the positions of the corresponding elements in the coupling coefficient matrix to obtain the preliminary adjusted matrix;

[0104] If the matrix still contains elements whose values ​​exceed the preset threshold after initial adjustment, the matrix is ​​locally adjusted using the gradient descent algorithm to obtain the optimized matrix.

[0105] Based on the optimized matrix, verify whether all element values ​​meet the preset threshold conditions to obtain the final coupling coefficient matrix.

[0106] Specifically, in one embodiment, when an element in the coupling coefficient matrix between wind intensity and illumination change has a value of 0.85, exceeding a preset threshold of 0.8, the system automatically identifies and marks this element. The threshold judgment mechanism scans each element value in the matrix one by one, recording the positions of elements exceeding the threshold to form a dynamic list of elements to be adjusted. This mechanism can effectively identify abnormally strong coupling relationships in the interaction between wind and illumination, avoiding deviations in model predictions. Specifically, when processing dynamic element grouping, the K-means clustering algorithm clusters elements based on their numerical characteristics and location information. Assuming the list to be adjusted contains 12 elements exceeding the threshold, the algorithm may divide them into 3 groups, each containing elements with similar characteristics. The first group may contain highly coupled elements related to wind intensity, the second group contains elements related to illumination change, and the third group contains mixed interaction elements. This grouping method helps with subsequent targeted adjustments and improves the accuracy of matrix optimization. In one possible implementation, a priority ranking algorithm sorts elements according to their importance and influence. For example, elements that affect the accuracy of wind prediction are given higher priority and placed at the front of the sequence. The sorting criteria include the element's numerical value, its positional weight within the matrix, and its contribution to the overall model performance. This sorting mechanism allows the system to prioritize the most critical coupling relationships, ensuring the effectiveness of the adjustment process. It's important to note that the matrix position adjustment process involves rearranging elements and fine-tuning their values. When a high-priority element is moved from its original position to a new one, the system simultaneously updates the relevant coupling relationships. For example, the wind-light coupling coefficient, originally located in the 2nd row and 3rd column of the matrix, might be moved to the 1st row and 2nd column, with its value adjusted from 0.85 to 0.78 to meet the threshold requirement. For instance, the gradient descent algorithm plays a crucial role in the local adjustment phase. When the initially adjusted matrix still contains elements exceeding the threshold, the algorithm calculates the gradient direction for each element and gradually reduces its value. During the adjustment process, the algorithm maintains the overall structural stability of the matrix, ensuring that the fundamental coupling pattern between wind and light is not disrupted. Through multiple iterations, the elements exceeding the threshold gradually converge to a reasonable range. In one embodiment, the final matrix verification process ensures that all element values ​​meet preset conditions. The verification mechanism rescans the entire matrix to ensure no element exceeds the threshold of 0.8. Through this rigorous verification process, the system generates a stable and reliable coupling coefficient matrix, providing an accurate parameter basis for subsequent wind and solar illumination predictions. Specifically, the priority allocation directly impacts the execution order of the business processes. High-priority wind monitoring tasks are executed first, followed by solar illumination change analysis, and finally, interaction effect assessment. This orderly execution mechanism improves overall prediction efficiency and ensures that key parameters are processed promptly and accurately.

[0107] Furthermore, the logical rules of environmental interaction are obtained from the coupling coefficient matrix, and a node connection structure under a unified framework is constructed using a graph neural network to obtain an interaction graph containing spatiotemporal consistency, including:

[0108] Based on the coupling coefficient matrix, obtain a list of strongly coupled relationships;

[0109] Based on the list of strongly coupled relationships and the state data of the associated entities, the Apriori algorithm is used to analyze the co-occurrence patterns of entity pairs in the time dimension, and a preliminary set of interaction rules describing the influence of the association between entities is generated.

[0110] By parsing the logical relationships in the initial set of interaction rules, each independent entity is abstracted into a graph node, and the pointing relationships between entities in the rules are transformed into directed connections between nodes, thus obtaining an initial topology graph.

[0111] Based on each graph node in the initial topology graph, obtain the corresponding spatial location information and historical state time series data. Use the spatial location information as the static attribute of the node and the time series data as the dynamic attribute of the node to obtain a spatiotemporal node set.

[0112] The topological graph containing the spatiotemporal node set is input into the graph attention network model to obtain new node embeddings containing dynamic evolution information;

[0113] Cosine similarity is used to calculate the similarity between new node embeddings, generating a weighted directed graph where nodes represent environmental entities and edge weights represent interaction strengths, thus obtaining an interaction graph with spatiotemporal consistency.

[0114] Specifically, in one embodiment, the identification of strongly coupled pairs is based on a threshold judgment of the values ​​in the coupling coefficient matrix. Specifically, when the value of a matrix element reaches 0.8 or higher, it indicates a significant interdependence between the corresponding two business entities. For example, in a supply chain management system, if the coupling coefficient between supplier A and manufacturer B is 0.85, this entity pair is marked as strongly coupled and included in the strongly coupled pair list for subsequent analysis. It should be noted that the Apriori algorithm in this scenario is used to mine the co-occurrence patterns of entity pairs over time. For example, by analyzing the correlation data between supplier delivery times and manufacturer production plans over the past 30 days, the algorithm can identify time-dimensional interaction patterns such as "when the supplier delays delivery by more than 2 days, the manufacturer's capacity utilization rate decreases by 15%". After confidence screening, these patterns form a preliminary set of interaction rules describing the causal influence between entities. In one possible implementation, the initial topology graph construction abstracts each business entity as a graph node, transforming the influence relationships in the rule set into directed edges. Specifically, if the rule indicates that "a change in the state of entity A causes a response from entity B," then a directed connection from node A to node B is established in the graph. Through this transformation, the originally scattered interaction rules are integrated into a relational network with a clear topological structure. For example, the construction of a spatiotemporal node set requires assigning dual attribute features to each node. Static attributes include fixed information such as the entity's geographical coordinates and capacity, while dynamic attributes cover the entity's state change sequence over a period of time, such as time-series data like inventory level fluctuations and processing capacity changes. This dual-attribute design allows nodes to maintain the stability of their spatial location while reflecting the dynamic characteristics of temporal evolution. In one embodiment, the graph attention network model quantifies the intensity of dynamic influence by calculating attention weights between nodes. The model assigns differentiated attention scores to different neighboring nodes based on the similarity of their historical states and their spatial proximity. Preferably, the model employs a multi-head attention mechanism to capture complex interaction patterns between nodes from multiple perspectives, generating richer node embedding representations. For example, cosine similarity calculation is used to evaluate the consistency of the updated node embeddings. When the similarity of the embedding vectors of two connected nodes exceeds 0.75, the system increases the weight of the corresponding connecting edge by 20%, and vice versa. This dynamic adjustment mechanism ensures that the connection structure accurately reflects the true interaction strength between entities, avoiding the limitations of static topology graphs that cannot adapt to environmental changes. Understandably, the final interaction graph integrates spatiotemporal consistency constraints, enabling the entire system to accurately capture the dynamic evolution patterns between entities while maintaining a reasonable spatial layout, providing a reliable basis for subsequent business decisions.

[0115] Furthermore, based on the node connection structure in the interaction graph, the dynamic state value of each node is updated iteratively by simulating the device. It is then determined whether the updated state value conforms to physical constraints, resulting in an optimized scene evolution sequence including:

[0116] Obtain the initial node set, connectivity, and structure information of the interaction graph, and assign an initial state value to each node to form an initial scene graph;

[0117] Based on the initial scene graph, extract the constraint set related to the physical quantities of the nodes from the pre-established rule base, and construct the discriminant for state evolution;

[0118] An iterative simulation method is used to calculate the potential rate of change of the state value of each node based on the connectivity between nodes, and a scenario graph containing temporary dynamic values ​​is generated.

[0119] The discriminant is used to perform compliance checks on the temporary dynamic values ​​of each node in the scenario graph to be tested. If the temporary dynamic value conforms to the constraint set, it is determined to be a valid state; if the temporary dynamic value does not conform to the constraint set, it is determined to be an invalid state.

[0120] Based on the determination results of valid and invalid states, update the evolution tree, take the scene graph to be verified containing valid states as the new branch node of the evolution tree, and discard the scene graph to be verified containing invalid states.

[0121] When the evolution tree reaches the preset simulation depth, multiple complete state evolution paths are extracted from the root node to all leaf nodes to form a candidate sequence set.

[0122] The Q-learning algorithm is used to evaluate the cumulative reward value of each evolution path in the candidate sequence set, and the evolution path with the highest cumulative reward value is determined as the final optimized scenario evolution sequence.

[0123] Specifically, in one embodiment, the construction of the initial scene graph requires extracting key topological information from the interaction graph. Specifically, each environmental entity node is assigned initial values ​​for physical states such as temperature, humidity, and pressure, while connectivity reflects the number of direct associations between the node and its neighboring entities. For example, in an intelligent environmental monitoring system, sensor node A has an initial temperature set of 25 degrees Celsius and a connectivity of 3, indicating a direct physical influence relationship with its three neighboring nodes. For instance, the constraint set in the rule base contains reasonable ranges of variation and mutual constraints for various physical quantities. Temperature change constraints might stipulate that the temperature change in a single iteration cannot exceed 2 degrees Celsius, and there might be a negative correlation between humidity and temperature, meaning that humidity should decrease accordingly when temperature rises. These constraint sets are transformed into mathematical discriminant expressions for subsequent state verification processes. In one possible implementation, iterative simulation predicts state changes by calculating the propagation of influence between nodes. Node A is affected by the temperature of its three neighboring nodes, and its potential rate of change is calculated using a weighted average. Assuming adjacent node temperatures are 23, 27, and 26 degrees Celsius, with corresponding connection weights of 0.4, 0.3, and 0.3, the expected temperature change for node A is an adjustment towards 25.1 degrees Celsius, forming a temporary dynamic value. For example, during compliance checks, the discriminant verifies whether each temporary dynamic value meets preset constraints. If node B's temporary humidity value is 85%, but based on its corresponding temperature increase trend, the humidity constraint requires its value to be below 80%, then this state is deemed invalid. Conversely, state changes that conform to physical laws and constraints are retained as valid states. In one embodiment, the evolution tree is constructed using a branching expansion mechanism, where each valid scenario becomes a new node in the tree structure. After five rounds of iterative simulation, the evolution tree may contain 32 leaf nodes, each representing a possible final state configuration. The paths from the root node to each leaf node constitute different evolution sequences, reflecting multiple possible development trajectories of the system state. For example, the Q-learning algorithm evaluates the merits of each evolution path by setting a reward function. System stability, energy efficiency, and response speed are quantified as reward values. Path A may receive a higher cumulative reward value of 8.7 due to achieving rapid temperature stabilization, while path B only receives a reward value of 6.2 due to oscillations. By comparing the cumulative rewards of all candidate paths, the system automatically selects the optimal evolution sequence as the reference scheme for environmental regulation.

[0124] Furthermore, if there are inconsistent data portions in the optimized scenario evolution sequence, supplementary information is retrieved from multi-source heterogeneous data sources and integrated into the sequence to determine the collaborative driving model, which includes:

[0125] The temporal logic verification method is used to scan the continuous time nodes in the optimized scene evolution sequence, and to determine whether there are data breakpoints or logical conflicts according to the preset logic rules, so as to obtain the location information of inconsistent segments.

[0126] Based on the location information of inconsistent fragments, entities and relationships in a predefined knowledge graph are matched, and supplementary data streams related to the timestamps and event types of the corresponding fragments are obtained from multiple heterogeneous supplementary sources.

[0127] By performing structured parsing and semantic alignment on the supplementary data stream, and converting it into standardized information units with the same data specifications as the original scene evolution sequence, a set of data to be fused is obtained.

[0128] An attention-based fusion strategy is adopted to calculate the contribution weight of each standardized information unit in the dataset to be fused to the inconsistent segments, and an initial calibration sequence is generated by weighted summation.

[0129] If the initial calibration sequence passes the consistency check in the timing logic check, the corresponding sequence is determined as the final calibration sequence; if it fails, the weight parameters of the attention mechanism are adjusted and the fusion is re-executed until the final calibration sequence is obtained.

[0130] Based on the final calibration sequence, a long short-term memory network is used for model training. By learning the state transitions and cooperative relationships in the sequence, a cooperative-driven model is obtained.

[0131] Specifically, the temporal logic verification method checks the continuity of the scene evolution sequence based on preset logical rules. In physical system simulation, when node A's energy state is detected to be 100 joules at time t1, but suddenly drops to 10 joules at time t2 without any energy transfer events in between, the system marks this time period as an inconsistent segment. Specifically, the verification algorithm scans the rate of state change within each time window, triggering an anomaly marker when the rate of change exceeds the physical constraints. In one embodiment, the knowledge graph matching process locates supplementary data sources through entity recognition and relational reasoning. After the system identifies an abnormal energy segment, it searches for entity nodes related to "energy transfer" and "state decay" in a predefined graph, and then obtains supplementary information with corresponding timestamps from heterogeneous sources such as sensor databases and historical record databases. This targeted acquisition mechanism can significantly improve the accuracy and efficiency of data completion. Structured parsing and semantic alignment ensure that information from different data sources can be processed uniformly. For example, sensor data might be stored in the format "Temperature: 25 degrees Celsius," while historical records might be recorded as "Thermodynamic state value: 298 Kelvin." Semantic alignment converts these into unified, standardized information units, facilitating subsequent fusion processing. The attention-based fusion strategy generates a calibration sequence by calculating the contribution weight of each information unit. Specifically, the system assigns different weights based on the reliability, temporal relevance, and content matching of the data source. For example, direct sensor data has a weight of 0.7, and historical inferred data has a weight of 0.3; the initial calibration value is generated through weighted summation. It should be noted that consistency verification employs an iterative optimization strategy. When logical conflicts still exist in the initial calibration sequence, the system adjusts the attention weight parameters, for example, increasing the sensor data weight to 0.8, and re-executes the fusion calculation until the verification passes. Long Short-Term Memory (LSTM) network training performs state transition learning based on the final calibration sequence. In one possible implementation, the network learns the collaborative relationship patterns between nodes, enabling it to predict the future state evolution trend of the system, thereby constructing a predictive collaborative driving model and providing a reliable state prediction foundation for subsequent scenario simulations.

[0132] Furthermore, the rendering data stream for the virtual display is generated based on the collaborative driving model, and a real-time rendering engine is used to process the rendering data stream to produce continuous dynamic landscape results, including:

[0133] The driving instruction set output by the collaborative driving model is obtained, and the driving instruction set is parsed one by one by the syntax parser to obtain a structured scene parameter set containing entity position, pose and environmental state;

[0134] Based on the entity identifiers in the structured scene parameter set, retrieve the corresponding 3D geometric data and surface texture map information from the digital asset library to obtain the original asset set required to construct the scene;

[0135] The physical simulation module binds the original asset set with predefined physical field parameters and calculates an enhanced scene description with added dynamic attributes.

[0136] The serialization processor encodes the geometry vertices, texture map indices, and physics vectors in the enhanced scene description in a unified manner, generating a continuous rendering data stream containing timestamps.

[0137] The real-time rendering engine receives continuous rendering data streams. If the data frame integrity check of the data stream passes, it is unpacked within the rendering pipeline to reconstruct a layered scene tree.

[0138] Based on the node information of the layered scene tree and the associated lighting field data, shading calculations and post-processing are performed to generate a single-frame image and append it to the output frame sequence to obtain the final high-fidelity simulated visual scene.

[0139] Specifically, in actual autonomous driving test scenarios, the node updates of the hierarchical scene tree can be further optimized to support real-time interaction. Suppose a traffic light D needs to be temporarily added to the scene, with nodes containing the red light status and position coordinates (150, 250, 10). The rendering engine quickly locates the parent node "Environment" of traffic light D through the scene tree, updates only the relevant lighting and geometry data, and generates a new frame showing the red light on and vehicle A decelerating. This local update method significantly reduces computation and ensures real-time performance. In one possible implementation, the rendering pipeline can dynamically allocate resources based on scene complexity. For example, for dense urban scenes, high-precision texture and shading calculations are prioritized for near-field objects such as vehicle A and pedestrian C, while low-resolution textures are used for far-field objects such as background buildings. This resource allocation strategy ensures a balance between rendering efficiency and visual quality. It should be noted that the generation of continuous rendering data streams also needs to consider precise timestamp synchronization. For example, at time t2, the data stream records that vehicle A's speed drops to 5 meters per second, and traffic light D switches to green. The rendering engine unpacks frames sequentially according to timestamps, ensuring the frame sequence reflects correct temporal changes, such as vehicle A accelerating through an intersection at time t2. This synchronization mechanism avoids visual temporal distortions and enhances the realism of the simulated scene. For example, the output frame sequence can be used to verify the decision-making of the simulation system. Suppose an autonomous driving algorithm needs to adjust vehicle A's driving strategy based on traffic light status; the rendered scene intuitively presents vehicle A's braking posture at a red light and its acceleration trajectory at a green light. This intuitive feedback facilitates developers' analysis of the algorithm's response accuracy. Understandably, the generation of high-fidelity simulated scenes also supports multi-view output. For example, the rendering engine can simultaneously generate the driver's view, the overhead view, and the bystander's view of vehicle A, each view achieved by adjusting the camera node parameters of the scene tree. This multi-view capability facilitates comprehensive verification of the autonomous driving system's performance in different scenarios. Ultimately, the generated simulated scene not only provides a reliable basis for algorithm testing but can also serve as a visualization tool for user interface display, enhancing the human-computer interaction experience.

[0140] This embodiment also provides a landscape design dynamic display system based on virtual reality (VR), including: a data acquisition module, a feature extraction module, a map construction module, a scene evolution sequence optimization module, a model construction module, and a result output module;

[0141] The data acquisition module is used to collect real-time environmental data and plant growth cycle information through multi-source sensors, and process them using a particle filter algorithm to obtain a preliminary spatiotemporally aligned dynamic dataset.

[0142] The feature extraction module is used to extract key feature parameters from the initial spatiotemporally aligned dynamic dataset using feature extraction equipment, train the interaction relationship between key feature parameters using a neural network model, and determine the coupling coefficient matrix.

[0143] The graph construction module is used to obtain the logical rules of environmental interaction from the coupling coefficient matrix, and to construct the node connection structure under a unified framework using graph neural network to obtain an interaction graph with spatiotemporal consistency.

[0144] The scene evolution sequence optimization module is used to obtain an optimized scene evolution sequence by iteratively simulating the device to update the dynamic state value of each node based on the node connection structure in the interaction graph, judging whether the updated state value conforms to the physical law constraints.

[0145] The model building module is used to obtain supplementary information from multiple heterogeneous data sources and integrate it into the sequence if there are inconsistent parts in the optimized scenario evolution sequence, in order to determine the collaborative driving model.

[0146] The results output module is used to generate a rendering data stream for virtual display based on the collaborative driving model. The rendering data stream is processed by a real-time rendering engine to produce continuous dynamic landscape results.

[0147] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A dynamic display method for landscape design based on virtual reality (VR), characterized in that, include: Real-time environmental data and plant growth cycle information are collected by multi-source sensors and processed using a particle filter algorithm to obtain a preliminary spatiotemporally aligned dynamic dataset. Based on the preliminary spatiotemporally aligned dynamic dataset, key feature parameters are extracted using a feature extraction device, and a neural network model is used to train the mutual influence relationship between the key feature parameters to determine the coupling coefficient matrix. The logical rules of environmental interaction are obtained from the coupling coefficient matrix, and a node connection structure under a unified framework is constructed using a graph neural network to obtain an interaction graph containing spatiotemporal consistency. Based on the node connection structure in the interaction graph, the dynamic state value of each node is updated by iteratively simulating the device, and it is determined whether the updated state value conforms to the constraints of physical laws to obtain an optimized scene evolution sequence. If there are inconsistent data parts in the optimized scenario evolution sequence, supplementary information is obtained from multi-source heterogeneous data sources and fused into the sequence to determine the collaborative driving model; The rendering data stream for virtual display is generated based on the collaborative driving model, and the rendering data stream is processed by a real-time rendering engine to produce continuous dynamic landscape results.

2. The method for dynamic display of landscape design based on virtual reality (VR) according to claim 1, characterized in that, Real-time environmental data and plant growth cycle information were collected by multiple sensors and processed using a particle filter algorithm to obtain a preliminary spatiotemporally aligned dynamic dataset, including: Real-time environmental data and plant growth cycle information are collected through multi-source sensors; A semantic segmentation network was used to extract the outline information of the plant canopy and stem from the plant growth cycle information, and combined with environmental data, an initial multidimensional data record containing temperature, humidity, light intensity and canopy area was constructed. Based on the historical fluctuation range of each dimension in the initial multidimensional data record, calculate the information entropy of the current data point. If the information entropy exceeds the preset confidence interval, mark the corresponding data point as a potential outlier to be processed. The particle filter algorithm is used to perform state estimation and weight update on the data marked as potential outliers based on the state transition equation of the plant growth model, so as to obtain a corrected smooth data sequence. A unified spatiotemporal axis is established based on the global timing signal. The smoothed data sequences from different sources are resampled using the cubic spline interpolation method to obtain the preliminary spatiotemporally aligned dynamic dataset.

3. The method for dynamic display of landscape design based on virtual reality (VR) according to claim 1, characterized in that, Based on the preliminary spatiotemporally aligned dynamic dataset, key feature parameters are extracted using a feature extraction device. A neural network model is then used to train the interrelationships among these key feature parameters, and the coupling coefficient matrix is ​​determined, including: Key coupling coefficients are extracted from the initial spatiotemporally aligned dynamic dataset to generate a table of inter-feature correlation strength. Based on the correlation strength table, a feature interaction network graph is constructed to determine the dynamic coupling mode between wind intensity and light intensity changes; If the complexity of the dynamic coupling pattern exceeds a preset threshold, a simplified coupling pattern set is obtained by grouping the patterns using clustering analysis. By simplifying the set of coupling modes, the parameters of the multilayer perceptron neural network model are optimized to obtain the optimized model; The optimized model is used to predict the interaction trend between wind intensity and light intensity changes in the new input dataset, and the prediction interaction results are obtained. From the predicted interaction results, significant trend features are extracted to determine the coupling coefficient matrix.

4. The method for dynamic display of landscape design based on virtual reality (VR) according to claim 3, characterized in that, From the predicted interaction results, significant trend features are extracted to determine the coupling coefficient matrix, including: If the value of an element in the coupling coefficient matrix exceeds a preset threshold, the set of elements exceeding the threshold is obtained by threshold judgment, and the list of dynamic elements to be adjusted is determined. Based on the list of dynamic elements to be adjusted, the K-means clustering algorithm is used to group the dynamic elements to obtain the grouping results of the elements; Based on the grouping results, each group of dynamic elements is sorted using a priority sorting algorithm to obtain the sorted element sequence; Based on the sorted element sequence, adjust the positions of the corresponding elements in the coupling coefficient matrix to obtain the preliminary adjusted matrix; If the matrix still contains elements whose values ​​exceed the preset threshold after initial adjustment, the matrix is ​​locally adjusted using the gradient descent algorithm to obtain the optimized matrix. Based on the optimized matrix, verify whether all element values ​​meet the preset threshold conditions to obtain the final coupling coefficient matrix.

5. A method for dynamic display of landscape design based on virtual reality (VR) according to claim 1, characterized in that, The logical rules of environmental interaction are obtained from the coupling coefficient matrix, and a node connection structure under a unified framework is constructed using a graph neural network to obtain an interaction graph containing spatiotemporal consistency, including: Based on the coupling coefficient matrix, obtain a list of strongly coupled relationships; Based on the list of strongly coupled relationships and the associated entity state data, the Apriori algorithm is used to analyze the co-occurrence patterns of entity pairs in the time dimension, and a preliminary set of interaction rules describing the influence of the association between entities is generated. By parsing the logical relationships in the initial set of interaction rules, each independent entity is abstracted into a graph node, and the pointing relationships between entities in the rules are transformed into directed connections between nodes, thus obtaining an initial topology graph. Based on each graph node in the initial topology graph, obtain the corresponding spatial location information and historical state time series data, use the spatial location information as the static attribute of the node, and use the time series data as the dynamic attribute of the node to obtain a spatiotemporal node set. The topological graph containing the spatiotemporal node set is input into the graph attention network model to obtain new node embeddings containing dynamic evolution information; Cosine similarity is used to calculate the similarity between new node embeddings, generating a weighted directed graph where nodes represent environmental entities and edge weights represent interaction strengths, thus obtaining an interaction graph with spatiotemporal consistency.

6. The method for dynamic display of landscape design based on virtual reality (VR) according to claim 1, characterized in that, Based on the node connection structure in the interaction graph, the dynamic state value of each node is updated iteratively by simulating the device. The updated state value is then judged to determine whether it conforms to physical constraints, resulting in an optimized scene evolution sequence including: Obtain the initial node set, connectivity, and structure information of the interaction graph, and assign an initial state value to each node to form an initial scene graph; Based on the initial scene diagram, a set of constraints related to the physical quantities of nodes is extracted from the pre-established rule base to construct a discriminant for state evolution; An iterative simulation method is used to calculate the potential rate of change of the state value of each node based on the connectivity between nodes, and a scenario graph containing temporary dynamic values ​​is generated. The discriminant is used to perform compliance checks on the temporary dynamic values ​​of each node in the scenario graph to be tested. If the temporary dynamic value conforms to the constraint set, it is determined to be a valid state; if the temporary dynamic value does not conform to the constraint set, it is determined to be an invalid state. Based on the determination results of valid and invalid states, update the evolution tree, take the scene graph to be verified containing valid states as the new branch node of the evolution tree, and discard the scene graph to be verified containing invalid states. When the evolution tree reaches a preset simulation depth, multiple complete state evolution paths are extracted from the root node to all leaf nodes to form a candidate sequence set. The Q-learning algorithm is used to evaluate the cumulative reward value of each evolution path in the candidate sequence set, and the evolution path with the highest cumulative reward value is determined as the final optimized scenario evolution sequence.

7. A method for dynamic display of landscape design based on virtual reality (VR) according to claim 1, characterized in that, If there are inconsistent data parts in the optimized scenario evolution sequence, supplementary information is retrieved from multi-source heterogeneous data sources and fused into the sequence to determine the collaborative driving model, which includes: A time-series logic verification method is used to scan the continuous time nodes in the optimized scene evolution sequence, and to determine whether there are data breakpoints or logical conflicts according to preset logic rules, so as to obtain the location information of inconsistent segments. Based on the location information of the inconsistent fragments, entities and relationships in a predefined knowledge graph are matched, and supplementary data streams related to the corresponding fragment timestamps and event types are obtained from multiple heterogeneous supplementary sources. By performing structured parsing and semantic alignment on the supplementary data stream, and converting it into standardized information units with the same data specification as the original scene evolution sequence, a data set to be fused is obtained. An attention-based fusion strategy is adopted to calculate the contribution weight of each standardized information unit in the dataset to be fused to the inconsistent segment, and an initial calibration sequence is generated by weighted summation. If the initial calibration sequence passes the consistency check in the timing logic check, the corresponding sequence is determined as the final calibration sequence; if it fails, the weight parameters of the attention mechanism are adjusted and the fusion is re-executed until the final calibration sequence is obtained. Based on the final calibration sequence, a long short-term memory network is used for model training. By learning the state transitions and cooperative relationships in the sequence, the cooperative driving model is obtained.

8. A method for dynamic display of landscape design based on virtual reality (VR) according to claim 1, characterized in that, The rendering data stream for virtual display is generated based on the collaborative driving model, and the rendering data stream is processed by a real-time rendering engine to produce continuous dynamic landscape results, including: The driving instruction set output by the collaborative driving model is obtained, and the driving instruction set is parsed one by one by a syntax parser to obtain a structured scene parameter set containing entity position, posture and environmental state; Based on the entity identifiers in the structured scene parameter set, retrieve the corresponding 3D geometric data and surface texture map information from the digital asset library to obtain the original asset set required to construct the scene; The original asset set is bound to predefined physical field parameters using a physics simulation module, and an enhanced scene description with added dynamic attributes is obtained through calculation. The serialization processor encodes the geometric vertices, texture map indexes, and physical field vectors in the enhanced scene description in a unified manner, generating a continuous rendering data stream containing timestamps. The continuous rendering data stream is received by a real-time rendering engine. If the data frame integrity check of the data stream passes, it is unpacked in the rendering pipeline to reconstruct the layered scene tree. Based on the node information of the layered scene tree and the associated illumination field data, shading calculations and post-processing are performed to generate a single-frame image and append it to the output frame sequence to obtain the final high-fidelity simulated visual scene.

9. A landscape design dynamic display system based on virtual reality (VR) implemented according to the method of any one of claims 1-8, characterized in that, include: The system includes a data acquisition module, a feature extraction module, a map construction module, a scene evolution sequence optimization module, a model construction module, and a result output module. The data acquisition module is used to collect real-time environmental data and plant growth cycle information through multi-source sensors, and process them using a particle filter algorithm to obtain a preliminary spatiotemporally aligned dynamic dataset. The feature extraction module is used to extract key feature parameters based on the preliminary spatiotemporally aligned dynamic dataset using a feature extraction device, train the mutual influence relationship between the key feature parameters using a neural network model, and determine the coupling coefficient matrix. The graph construction module is used to obtain the logical rules of environmental interaction from the coupling coefficient matrix, and to construct a node connection structure under a unified framework using a graph neural network to obtain an interaction graph containing spatiotemporal consistency. The scene evolution sequence optimization module is used to update the dynamic state value of each node by iteratively simulating the device based on the node connection structure in the interaction graph, determine whether the updated state value conforms to the constraints of physical laws, and obtain an optimized scene evolution sequence. The model building module is used to obtain supplementary information from multi-source heterogeneous data sources and integrate it into the sequence if there are inconsistent parts in the optimized scene evolution sequence, thereby determining the collaborative driving model. The result output module is used to generate a rendering data stream for virtual display based on the collaborative driving model, and to process the rendering data stream using a real-time rendering engine to produce continuous dynamic landscape results.

Images