Spatial suitability optimization method and system based on three-generation space re-identification

Abstract

The invention relates to the technical field of territorial space planning, in particular to a three-generation space re-identification-based space suitability optimization method and system, and the method comprises the steps: taking a three-generation space as a space carrier, constructing a three-generation space re-identification matrix through a space-time semantic graph network, and extracting a potential origin space unit; constructing a three-dimensional suitability tensor by taking spatial value evaluation as a research kernel, wherein each channel of the three-dimensional suitability tensor respectively represents an ecological space base line constraint value, a production space value matching feature and a living space demand matching feature; and on this basis, a multi-agent reinforcement learning origin space evolution model is constructed, real-time monitoring data is combined to form a dynamic punishment feedback mechanism, a corresponding space suitability layout result is output, and dynamic closed-loop updating is realized. According to the method, dynamic and automatic upgrading is realized on the basis of the evaluation result of the existing static index system, and quantitative decision support can be provided for territorial space planning and compilation, evaluation implementation and dynamic adjustment.

Description

Technical Field

[0001] This invention relates to the field of land and space planning technology, and more specifically, to a spatial adaptability optimization method and system based on the re-identification of three-dimensional space. Background Technology

[0002] With the rapid development of the economy and society and the in-depth advancement of ecological civilization construction, regional space, as the core carrier of production, living, and ecological activities, has seen its spatial pattern scientifically configured, becoming a key link in achieving sustainable regional development. In the macro-planning system of "three-life spaces" (production, living, and ecological spaces), production space is a core element for ensuring food security, promoting industrial agglomeration, and supporting the real economy. With the rise of modern agriculture, new energy industries, and diversified industries, the connotation and boundaries of production space are becoming increasingly rich, and traditional single-element evaluation has evolved towards multi-dimensional element integration. In recent years, with the widespread application of IoT, big data, and artificial intelligence technologies, spatial planning has gradually transformed from static macro-control to dynamic, refined governance. Introducing cutting-edge algorithms such as graph neural networks, multi-agent models, and cellular automata to mine the deep semantic relationships behind multi-source spatiotemporal data, simulate the coupling process of multiple elements in spatial evolution, and combine this with intelligent calculation of ecological carrying capacity has become an important development direction for optimizing the adaptability of production space and digital twin spatial governance.

[0003] However, in the process of compiling and implementing territorial spatial planning, the adaptability of land use is affected by factors such as multi-stakeholder games, factor flows, ecological risk disturbances, and real-time changes in monitoring data. The results of a single static evaluation have certain limitations in the dynamic management scenario during the planning implementation period. Therefore, it is necessary to further develop a dynamic and automated optimization methodology system oriented towards the implementation period, based on the existing static indicator system and evaluation results. This means expanding the static evaluation and result output into a technical path of static evaluation as a foundation—dynamic evolution and optimization—monitoring and feedback closed loop—continuous updating, in order to improve the feasibility and dynamic governance capabilities of territorial spatial planning. Summary of the Invention

[0004] The purpose of this invention is to provide a spatial adaptability optimization method and system based on three-dimensional spatial re-identification. By constructing a three-dimensional adaptability tensor and a multi-agent reinforcement learning evolution model, it not only achieves deep fusion of multi-dimensional spatial data, but also, relying on carbon sink constraints and a dynamic penalty closed-loop mechanism, completes the technical leap from static planning to dynamic adaptive game optimization of production space.

[0005] This invention is achieved through the following technical solution:

[0006] The spatial adaptability optimization method based on three-dimensional space re-identification includes the following steps: Acquire multi-source spatiotemporal data of the target area, construct a spatiotemporal semantic graph network, and aggregate the spatial features of each grid node through a graph convolutional neural network to output a three-life space re-identification matrix containing the fuzzy membership degree of each grid node belonging to production, living and ecological spaces. Based on the three-life space re-identification matrix, the carbon sink opportunity cost of the target area is calculated and an ecological rigid constraint mask is applied to extract the set of potential production space units in the target area. By integrating spatial value assessment data of potential production area spatial units, industrial evolution demand data, and natural background data, a three-dimensional suitability tensor is constructed with spatial unit coordinates as the plane and ecological constraints, value matching, and demand adaptation as characteristic channels. The characteristic values ​​of each channel of the three-dimensional suitability tensor are then normalized. The three-dimensional suitability tensor is input into the multi-agent game model as the environmental state. The action value output by the multi-agent game model is transformed into the state transition probability, which drives the cellular automaton to perform spatial evolution and output the corresponding spatial adaptability optimization layout.

[0007] Optionally, the output includes a three-dimensional spatial re-identification matrix containing the fuzzy membership degrees of each grid node to production, living, and ecological spaces, specifically: Define the spatiotemporal semantic graph network as ,in, For a set of grid nodes, Let be the set of edges. This is the edge weight matrix based on the intensity of material flow and information flow; The graph convolutional neural network aggregates node features and outputs arbitrary nodes. Fuzzy membership vectors belonging to production, living, and ecological spaces The calculation formula is as follows:

[0008] in, Let D be the set of neighboring nodes of node i, and let D be the degree matrix. Let be the input feature vector of node j. The weight parameter matrix, For activation functions; Extracting satisfaction Node i serves as the fuzzy transition zone in the three-dimensional space, where... This is the preset confidence threshold.

[0009] Optionally, the opportunity cost of carbon sequestration in the target area is calculated and an ecologically rigid constraint mask is applied, specifically as follows: Define the carbon sink opportunity cost for any unit k in the set of potential production spatial units. The calculation formula is as follows:

[0010] in, The amount of carbon sequestration per unit area that maintains the natural background state for unit k. This represents the estimated carbon sink after unit k is transformed into its production location. For the unit price equivalent in the carbon trading market, This is the amplification factor for ecological sensitivity. Construct an ecologically rigid constraint mask matrix M, if the region where cell k is located... If the cost exceeds a preset threshold or is within the legally mandated ecological red line, then the mask matrix elements... ,otherwise The potential set of spatial units is extracted by performing a Hadamard product operation on the spatial grid and the mask matrix M.

[0011] Optionally, a three-dimensional suitability tensor is constructed with spatial unit coordinates as the plane and ecological constraints, value matching, and demand adaptation as characteristic channels, specifically as follows: Constructing tensors ,in, These are the latitude and longitude grid coordinates of the spatial unit; The first channel of the tensor Fill as cells The ecological constraint index is obtained by weighted summation and normalization of soil erosion sensitivity and biodiversity maintenance index; The second channel of tensor Fill as cells The value matching index is obtained by subtracting the opportunity cost of carbon sink from the economic value of basic production. The third channel of tensor Fill as cells The demand matching index is obtained by combining the potential index of industrial scale contiguous area and the resource footprint matching degree.

[0012] Optional, a third channel of the tensor Demand Fit Index The calculation formula, based on the composite industry compatibility assessment model, is as follows:

[0013] in, and Units Soil physical and chemical suitability and agricultural water resource carrying capacity; and These are the solar thermal radiation resource index and the capacity margin of the grid access node, respectively. and These are the agricultural demand weight and the new energy access weight, respectively. As a factor for adjusting the spatial compatibility of composite industries, and .

[0014] Optionally, the three-dimensional suitability tensor can be used as the environmental state input into the multi-agent game model, specifically as follows: The developers and protectors of the production area are instantiated as intelligent agents in a game of mutual competition, and the state space of each agent is a three-dimensional suitability tensor T. Define the local joint reward function for an agent developing intelligence at any origin at time t. The calculation formula is as follows:

[0015] in, The set of potential production space units occupied by the agent at time t; , , Units The three corresponding channel eigenvalues ​​in the tensor; As a positive incentive weight, This is the penalty coefficient for crossing ecological boundaries. This is an ecological sensitivity parameter.

[0016] Optionally, the action value output by the multi-agent game model can be converted into state transition probabilities, and the state transition probabilities of its cellular automaton can be used as the basis for further analysis. The calculation formula is:

[0017] in, The action value Q function output by the agent, which transforms unit i into place type j. Represents the set of neighboring nodes of cell i. The proportion of cells whose internal state is j. As an indicator function, when the state of cell k... hour ,otherwise , This is the global spatial suitability control function, where element i corresponds to the mask matrix element. hour, Otherwise, it is 1.

[0018] Optionally, output the corresponding space-adaptive optimized layout, which is as follows: Extracting the eigenvectors of potential production space units that have evolved to a stable state in the three-dimensional suitability tensor ; Define the positive ideal solution vector for space adaptation. Calculate the weighted Euclidean distance ,in, These are the weighting coefficients for each dimension; Based on weighted Euclidean distance The spatial units are arranged in ascending order, and based on the allocation ratio of the total control indicators for the target area, the first suitable scale development zone, the second suitable characteristic compatibility zone, and the third suitable restricted retreat zone are delineated in sequence.

[0019] Optionally, it also includes: acquiring real-time ecological monitoring data of the Internet of Things, generating a dynamic penalty factor through a nonlinear decay model, and feeding the dynamic penalty factor back to the value matching feature channel of the three-dimensional suitability tensor to update the feature value, so as to trigger the re-evolution of the spatial layout and complete the dynamic closed-loop management of spatial adaptability. Specifically, a dynamic penalty factor is generated through a nonlinear decay model, and this dynamic penalty factor is fed back into the value matching feature channel of the three-dimensional suitability tensor to update the feature values. Computing unit Ecological stress index at time t ; If unit The ecological stress index in continuous If the preset ecological tolerance threshold is exceeded within a monitoring cycle, a dynamic penalty factor for that unit will be triggered and calculated. The calculation formula is as follows:

[0020] in, For unit The dynamic penalty factor at the current time t. As a penalty decay rate constant, For the summation index variable of historical monitoring times, The set number of continuous monitoring cycles, For unit At historical monitoring moments Ecological stress index. This is the ecological tolerance threshold; Apply a dynamic penalty factor to the second channel of the 3D fitness tensor, updating the local tensor value. This is to suppress the game expansion probability of the unit in the spatial evolution model.

[0021] A spatial adaptability optimization system based on three-dimensional spatial re-identification includes: The graph network spatial recognition module is used to acquire multi-source spatiotemporal data of the target area and construct a spatiotemporal semantic graph network. It outputs a three-life space re-identification matrix containing the fuzzy membership degree of each grid node belonging to production, living and ecological spaces through a graph convolutional neural network. The mask constraint and set extraction module is used to calculate the opportunity cost of carbon sinks and apply an ecologically rigid constraint mask to extract a set of potential production site spatial units; The 3D suitability tensor construction module is used to fuse multidimensional data, construct a 3D suitability tensor with ecological constraints, value matching and demand adaptation as feature channels, and normalize the feature values. The game evolution optimization module is used to input the three-dimensional suitability tensor as the environmental state into the multi-agent game model, and convert the action value into the state transition probability to drive the spatial evolution of the cellular automaton to output the adaptive optimized layout. The digital twin closed-loop feedback module is used to generate dynamic penalty factors based on real-time ecological monitoring data and feed back to update the value matching feature channels of the three-dimensional suitability tensor to trigger dynamic correction of spatial layout.

[0022] The technical solution of the present invention has at least the following advantages and beneficial effects: On the one hand, this invention constructs a three-dimensional suitability tensor that includes ecological constraints, value matching, and demand adaptation, and integrates the fuzzy membership degree extracted from graph networks with a carbon sink opportunity cost mask. This not only achieves dimensionality reduction mapping of complex multi-source spatiotemporal data, but also accurately delineates the ecological security boundary of habitat evolution. On the other hand, this invention innovatively couples deep reinforcement learning with cellular automata, using the action value output by multi-agent game to dynamically drive spatial evolution. This not only effectively simulates the Nash equilibrium between ecological protection and habitat development, but also constructs a digital twin closed-loop feedback mechanism by introducing a dynamic penalty factor based on real-time ecological monitoring data. This enables adaptive triggering of spatial layout re-evolution and precise correction when facing ecological stress. Attached Figure Description

[0023] Figure 1 A flowchart illustrating the spatial adaptability optimization method based on three-dimensional space re-identification provided by the present invention; Figure 2 This is a schematic diagram illustrating the principle of the spatial adaptability optimization system based on three-dimensional space re-identification provided by the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0025] In this embodiment of the invention, a spatial adaptability optimization method based on the re-identification of three-dimensional spaces is provided. These three-dimensional spaces refer to production space, living space, and ecological space. Existing static evaluation and planning models based on indicator systems play a crucial role in current status diagnosis, pattern identification, and scheme demonstration, and can serve as a basic framework for spatial adaptability assessment. Building upon this, to adapt to the needs of dynamic monitoring, rolling evaluation, and timely optimization during the implementation of national land spatial planning, this embodiment further introduces spatiotemporal semantic graph networks, three-dimensional suitability tensors, multi-agent deep reinforcement learning, and a digital twin closed-loop feedback mechanism to achieve an automated upgrade from data perception and state modeling to dynamic evolution optimization. Therefore, the method provided in this embodiment achieves a complete closed loop from data perception, state modeling, dynamic game theory, to adaptive correction.

[0026] Specifically, the method includes the following steps performed in sequence: Step 1: Spatiotemporal Data Acquisition and Output of the Three-Life Space Re-identification Matrix. This step is used to establish the correspondence between basic spatial units and functional dimensions within the framework of the three-life space carrier, providing input for subsequent spatial value assessment and dynamic optimization.

[0027] Acquire multi-source spatiotemporal data for the target area. For example, multi-source spatiotemporal data includes, but is not limited to, high-resolution remote sensing imagery, urban point-of-interest distribution data, mobile phone signaling trajectory data, meteorological observation data, and basic land use status vector data. To extract deep features characterizing spatial functional attributes from massive and unstructured multi-source spatiotemporal data, this embodiment constructs a spatiotemporal semantic graph network and aggregates the spatial features of each grid node using a graph convolutional neural network.

[0028] The spatiotemporal semantic graph network is defined as a topological structure consisting of a set of grid nodes, a set of edges, and an edge weight matrix based on the intensity of material and information flows. In implementation, the target region is divided into uniformly scaled grids, with each grid serving as a node. Edges between nodes are constructed through spatial adjacency and physical connectivity based on the transportation network. The elements of the edge weight matrix are not simply the reciprocal of distance, but rather incorporate the frequency of population flow, the intensity of logistics transportation, and the connectivity of ecological corridors between adjacent grids, thus objectively and realistically reflecting the flow of spatial elements.

[0029] Node features are deeply aggregated using a multi-layer graph convolutional neural network. The node feature vector input to the graph convolutional neural network is composed of remote sensing spectral features, vegetation index, population density, and road network density. In each layer of the network, the features of the target node and the features of its neighboring nodes are weighted and summed according to the edge weight matrix, multiplied by a learnable weight parameter matrix, and finally mapped through a non-linear activation function to achieve high-dimensional feature purification.

[0030] After network propagation, the final output layer is processed using a normalized exponential function, outputting a three-dimensional space re-identification matrix containing the fuzzy membership degrees of each grid node belonging to the production, living, and ecological spaces. Specifically, the fuzzy membership degree vector of any node belonging to the production, living, and ecological spaces consists of three probability values ​​between zero and one that sum to one. The calculation formula is obtained by multiplying the feature and weight parameter matrices of the target node and its neighboring nodes, performing symmetric normalization weighting based on the degree matrix, processing with an activation function, and then performing normalized exponential mapping.

[0031] Furthermore, after obtaining the fuzzy membership vectors of all nodes, the maximum value in each probability dimension is extracted. Nodes whose maximum probability value is less than a preset confidence threshold are selected as fuzzy transition zones in the three-dimensional space. This confidence threshold is typically set between 0.6 and 0.7. These fuzzy transition zone nodes are usually located in urban fringe areas or agroforestry ecotones, possessing extremely high potential for spatial function transformation and are active areas of focus for subsequent place-of-origin spatial evolution models.

[0032] Step 2: For the set of potential production space units obtained in Step 1, a three-dimensional suitability tensor is constructed using the three-dimensional space as the spatial carrier and spatial value assessment as the core. Among them, the tensor channels respectively represent the bottom-line constraint value of ecological space, the value matching characteristics of production space, and the demand adaptation characteristics of living space, so as to realize the computable expression of the value of production space in the three-dimensional space.

[0033] In traditional space development, the hidden costs of ecological degradation are often overlooked. This embodiment, based on the aforementioned three-dimensional space re-identification matrix, further calculates the carbon sink opportunity cost of the target area and applies an ecologically rigid constraint mask.

[0034] A carbon sequestration opportunity cost calculation model is defined for any potential carbon-producing spatial unit within the target area. The core logic of this model lies in assessing the lost carbon sequestration and absorption value when a unit transforms from its natural background state to a carbon-producing space. The specific calculation formula is: the difference between the carbon sequestration per unit area of ​​the unit maintaining its natural background state and the estimated carbon sequestration after the unit transforms into a carbon-producing space (if this difference is less than zero, it is rounded to zero), then multiplied by the carbon trading market unit price equivalent and an ecological sensitivity amplification factor. The ecological sensitivity amplification factor is normalized and dynamically assigned based on natural conditions such as the unit's slope and water system buffer distance, ensuring that the model can be quantitatively calculated.

[0035] After determining the carbon sequestration opportunity cost for each spatial unit, an ecologically rigid constraint mask matrix is ​​constructed. Iterating through all spatial units, if the carbon sequestration opportunity cost of a unit's region exceeds a preset cost threshold, or if the unit's geographical coordinates fall directly within the legally defined ecological protection red line area designated by the national or local government, the region is deemed to possess inviolable ecological rigidity, and the corresponding mask matrix element is assigned a value of zero. Conversely, if the above constraints are not triggered, the value is assigned a value of one. Finally, by performing a Hadamard product operation on the global spatial grid matrix of the target region and the constructed mask matrix (i.e., multiplying corresponding elements one by one), regions with mask matrix elements of zero are eliminated, accurately extracting the set of potential production spatial units in the target region. This process eliminates redundant computational load and establishes a safety baseline for subsequent tensor construction.

[0036] Step 3: Based on the three-dimensional suitability tensor, construct a multi-agent reinforcement learning-based origin space evolution model, enabling the origin space to dynamically adapt and seek optimization in the three-life space carrier around the spatial value goal, and output candidate layout schemes.

[0037] After extracting the set of potential production area spatial units, spatial value assessment data, industrial evolution demand data and natural background data are further integrated to construct a three-dimensional suitability tensor with the latitude and longitude grid coordinates of the spatial units as the plane and ecological constraint channels, value matching channels and demand adaptation channels as the feature dimensions.

[0038] Construct the dimensional structure of the tensor. The first and second dimensions correspond to the horizontal and vertical coordinates of the spatial unit, and the third dimension is the channel dimension with a depth of three.

[0039] The first channel of the tensor is filled with the ecological constraint index of each spatial unit. This index is obtained by weighted summation and normalization of the soil erosion sensitivity assessment results and the biodiversity maintenance index. The value of the first channel reflects the vulnerability of the grid in the ecological dimension; the larger the value, the stronger the ecological constraint.

[0040] The second channel of the tensor is filled with a value matching index for each spatial unit. This index is obtained by extracting the basic economic value of the target grid and subtracting the carbon sink opportunity cost mapping calculated in the preceding steps. This channel intuitively reflects the true economic development potential of the spatial unit after eliminating the costs of ecological damage.

[0041] The third channel of the tensor is filled with the demand adaptation index for each spatial unit. To address the layout needs of modern integrated production areas, such as emerging industrial models like agro-solar integration and fishery-solar integration, the demand adaptation index is calculated using a pre-defined integrated industry adaptation assessment model. The specific calculation formula is as follows: The soil physicochemical suitability of the unit is extracted and multiplied by the agricultural water resource carrying capacity, then multiplied by the agricultural demand weight; simultaneously, the unit's light and temperature radiation resource index is extracted and multiplied by the grid access node capacity margin, then multiplied by the new energy access weight; the sum of these two weighted results is then summed, the square root is taken, and finally multiplied by the integrated industry spatial compatibility adjustment factor. The compatibility adjustment factor is a parameter greater than zero and less than or equal to one, used to characterize the probability of simultaneously developing agriculture and new energy construction on the site under policy or physical constraints.

[0042] After the data filling of the three channels is completed, in order to eliminate the dimensional differences between different evaluation indicators and ensure the convergence stability of the deep reinforcement learning model, the maximum-minimum mapping or standard deviation standardization method is used to strictly normalize the feature values ​​of each channel of the three-dimensional suitability tensor to the numerical range of zero to one, forming a standardized environment state input matrix.

[0043] Step 4: Construct a reward function and state update mechanism to uniformly map the three types of value signals—ecological space constraints, production space value, and living space demand—into decision feedback, and iteratively obtain stable spatial allocation results under Nash equilibrium constraints.

[0044] Traditional cellular automata rely on human experience to set transition rules. This embodiment innovatively uses the standardized three-dimensional suitability tensor as the environmental state input into the deep reinforcement learning multi-agent game model, transforming the action value output by the multi-agent into the state transition probability, thereby driving the cellular automata to perform spatial evolution.

[0045] First, we define the roles of the intelligent agents. The stakeholders within the system are instantiated as intelligent agents engaging in a game of interaction. These primarily include resource development agents aiming to maximize economic gains, and ecological protection agents aiming to minimize negative environmental impacts. The joint perception state space of each agent at any evolutionary time step is the three-dimensional suitability tensor updated at the current moment.

[0046] A local joint reward function is designed for the agent. The reward function of an agent developing in any production area at a specific time includes a positive incentive term and a negative penalty term. The positive incentive term is the weighted sum of the value matching feature value of the second channel and the demand adaptation feature value of the third channel within the set of potential production area units currently occupied by the agent, with the weights controlled by a preset positive incentive weight. The negative penalty term is obtained by extracting the ecological constraint feature value of the first channel, multiplying it by the ecological sensitivity parameter, performing a natural exponential operation, and then multiplying it by the ecological constraint over-limit penalty coefficient. This penalty mechanism, which includes exponential growth properties, ensures that when the agent attempts to expand into a region with high ecological constraints, its total reward drops sharply, thereby forcing the model to approximate a multi-objective Nash equilibrium state at the underlying level.

[0047] During deep reinforcement learning training, the agent outputs the action value evaluation of each unit's state transition through a deep Q-network. To distribute the macroscopic game results of deep reinforcement learning to the microscopic spatial grid, this embodiment proposes a core computational rule that transforms action value into state transition probability. The final probability of any unit in the cellular automaton transitioning to the target origin type is composed of the product of three parts.

[0048] The first part is the action value-based strategy distribution. This is achieved by calculating the exponent of the action value Q function value of the unit when it is transformed into the target type, and dividing this exponent by the sum of the action value indices of all possible types. This smoothly maps action value to probability. The second part is the neighborhood interaction effect. It calculates the proportion of cells within the unit's preset neighborhood whose state is already of the target type. An indicator function is introduced here: when the state of a neighboring cell matches the target type, the indicator function is one; otherwise, it is zero. All indicator function values ​​are summed and divided by the total number of neighboring cells. The third part is the global spatial suitability control function. This function is directly linked to the mask matrix in step two. When the mask matrix element corresponding to the unit is zero, the control function is forced to zero, thus vetoing the development and transformation of the unit; otherwise, the control function is one.

[0049] The cellular automaton is driven to undergo multiple rounds of iterative evolution. After sufficient iterations until the state transition probabilities of each unit tend to stabilize and large-scale spatial function flips no longer occur, the feature vectors of the potential origin spatial units that have evolved to a stable state in the three-dimensional suitability tensor are extracted, and the initial spatial adaptability optimization layout is output.

[0050] To transform abstract feature vectors into concrete planning diagrams, the ideal solution vector for spatial adaptation is defined as a column vector containing three elements, each of which is one. For each evolutionarily stable spatial unit, the weighted Euclidean distance between its feature vector and the ideal solution vector is calculated. This distance is equal to the square of the difference between each channel's feature value and one, multiplied by the weight coefficient of each dimension, summed, and the square root taken. The smaller the distance value, the closer the spatial adaptability of the unit is to a perfect state. Subsequently, based on the calculated weighted Euclidean distance, all spatial units are arranged in ascending order, and according to the allocation ratio of the total control indicators of the target area, high-suitability large-scale development zones, medium-suitability characteristic compatible zones, and low-suitability restricted retreat zones are successively delineated, completing the physical spatial placement of the initial layout.

[0051] Step 5: Connect to IoT / monitoring data and construct dynamic penalty factors. Feed back the ecological risks and disturbances during the planning and implementation period as dynamic constraints within the three-life space carrier to the evolution model, triggering layout re-evolution and update, and realizing dynamic closed-loop management of spatial adaptability.

[0052] The implementation of planning schemes often faces dynamic changes in the natural environment. To overcome the risk of failure of static planning methods, this embodiment constructs a digital twin closed-loop feedback control mechanism based on the Internet of Things.

[0053] During the initial deployment and production or virtual operation phase, real-time ecological monitoring data, such as soil heavy metal ion concentration, groundwater level subsidence, and surface runoff pollution index, is continuously acquired from IoT sensor nodes deployed in the target area. The ecological stress index of each grid unit at the current moment is extracted through time-series data analysis.

[0054] Establish a warning triggering logic based on historical time windows. Monitor the ecological stress index of each unit in real time. If the ecological stress index of a unit exceeds the preset ecological tolerance threshold for multiple consecutive monitoring periods, the system determines that there is a risk of ecological overload in the area, and then triggers a nonlinear decay model to generate a dynamic penalty factor.

[0055] The dynamic penalty factor is calculated using a modified logistic function variant. First, the difference between the moving average of the ecological stress index and the ecological tolerance threshold over a continuous monitoring period is calculated. This difference is multiplied by the penalty decay rate constant and negatively taken. Then, a natural exponentiation is performed and incremented by one. The reciprocal is then subtracted from this value to ensure the penalty factor is confined within a certain range. This nonlinear decay model guarantees a moderate increase in penalty intensity when ecological overload is slight, while once the overload severely exceeds the critical point, the penalty factor will rapidly approach zero.

[0056] After generating the dynamic penalty factor, the system immediately initiates a state reset procedure, using the penalty factor as a feedback control signal to directly apply to the second channel of the three-dimensional suitability tensor constructed and stored in memory in step three, namely the value matching feature channel. By multiplying the original local tensor eigenvalues ​​of this unit by the dynamic penalty factor, the updated eigenvalues ​​will be significantly reduced.

[0057] As the local tensor eigenvalues ​​abruptly change, the environmental state alters. The system re-inputs the updated 3D suitability tensor into the deep reinforcement learning multi-agent game model. Due to the sharp drop in the value matching eigenvalue of the overloaded region, the agent's reward function is significantly reduced, leading to a sharp decrease in the action value Q-function value corresponding to that region. In the next round of re-evolution of the cellular automaton, according to the state transition probability formula, the probability of this unit continuing to maintain the production site development state is greatly suppressed, and the game probability of the surrounding neighborhood expanding into this region is also effectively cut off, prompting the system to spontaneously revert the state of this region or convert it into an ecological restoration space.

[0058] Finally, the system outputs a dynamically corrected, space-adaptive optimized layout. This step upgrades the traditional offline, one-off planning to a life-like solution that evolves synchronously with the real physical world, completely completing the dynamic closed-loop management of space adaptability.

[0059] Through the above steps, this invention achieves dynamic evolution and upgrading based on the three-dimensional spatial carrier and spatial value core. The optimal spatial suitability layout results output by this invention can be directly applied as core technical support in modern land spatial planning. Specifically, the dynamic optimization results can provide a scientific basis for the delineation of "three zones and three lines" under the background of "multi-plan integration," including guiding the precise location of agricultural production space, the flexible adjustment of urban construction and development boundaries, and the rigid early warning mechanism of ecological protection red lines. This invention transforms the research results of traditional static indicators into a practical and closed-loop intelligent planning tool, significantly improving the scientificity and practicality of land spatial planning.

[0060] To support the above-described method embodiments, this invention also provides a spatial adaptability optimization system based on three-dimensional spatial re-identification. This system is typically deployed on a cloud server or local workstation containing a high-performance graphics processor and large-capacity storage, processing massive spatial grid data through multi-threaded parallel computing. The system consists of five independent yet interconnected functional modules.

[0061] The Graph Network Spatial Recognition Module is responsible for performing spatiotemporal data parsing and graph structure construction. This module incorporates a data cleaning engine and a graph deep learning framework to acquire multi-source spatiotemporal data of the target region and construct a spatiotemporal semantic graph network. By calling the underlying graph convolutional neural network operators for forward reasoning, it finally outputs a three-dimensional spatial re-identification matrix containing the fuzzy membership degrees of each grid node belonging to production, living, and ecological spaces.

[0062] The mask constraint and set extraction module undertakes the calculation task of rigid constraints. This module receives the three-life space re-identification matrix as input, calls the carbon sink value accounting database, calculates the carbon sink opportunity cost of each grid cell, applies an ecological rigid constraint mask, and quickly eliminates prohibited construction areas through matrix bitwise operations to extract the set of potential production space cells that meet the safety bottom line.

[0063] The 3D suitability tensor construction module is the data fusion hub of the system. This module performs multidimensional tensor splicing operations, fusing multidimensional spatial assessments with baseline data to construct a 3D suitability tensor characterized by ecological constraints, value matching, and demand adaptation. Simultaneously, this module incorporates a scaling transformer to automatically normalize all tensor eigenvalues ​​to meet the input requirements of the neural network.

[0064] The game evolution optimization module is the decision-making brain of the entire system. This module integrates a reinforcement learning algorithm library and a cellular automaton engine. During runtime, the three-dimensional fitness tensor output by the tensor construction module is loaded into the GPU memory as the environment state, activating the pre-trained multi-agent game model. The output action value is converted into state transition probabilities in real time, which then drives the cellular automaton to perform thousands of parallel spatial evolution iterations, ultimately outputting the initial spatial fitness optimized layout and storing it in the spatial database.

[0065] The digital twin closed-loop feedback module is responsible for the system's adaptive correction. This module has an external application programming interface (API) that subscribes to and retrieves sensor data from the IoT monitoring platform in real time. Internally, the streaming data processing engine compares the ecological stress index in real time. Upon detecting anomalies, it generates a dynamic penalty factor using a nonlinear decay model and feeds the update command back to the value matching feature channel of the 3D suitability tensor construction module via the system bus. The data overwriting in the feature channel then triggers a hot restart of the game-theoretic optimization module, enabling the dynamic correction of the spatial layout map output.

[0066] The system module division in the above embodiments is merely a logical division for descriptive convenience. In actual industrial deployments, these modules can be integrated into the same large software package or distributed across different computing nodes using a microservice architecture for distributed processing. Those skilled in the art can optimize the internal functions of each module at the engineering level based on the actual data scale and computing power conditions. Such conventional engineering adjustments do not deviate from the core technical ideas of this invention. Through this method and system, the planning of production space is no longer a static blueprint on paper, but evolves into a dynamic intelligent control process with environmental perception capabilities, multi-party interest game capabilities, and self-correction capabilities.

[0067] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A spatial adaptability optimization method based on three-dimensional space re-identification, characterized in that, The steps of this method include: Acquire multi-source spatiotemporal data of the target area, construct a spatiotemporal semantic graph network, and aggregate the spatial features of each grid node through a graph convolutional neural network to output a three-life space re-identification matrix containing the fuzzy membership degree of each grid node belonging to production, living and ecological spaces. Based on the three-life space re-identification matrix, the carbon sink opportunity cost of the target area is calculated and an ecological rigid constraint mask is applied to extract the set of potential production space units in the target area. By integrating spatial value assessment data of potential production area spatial units, industrial evolution demand data, and natural background data, a three-dimensional suitability tensor is constructed with spatial unit coordinates as the plane and ecological constraints, value matching, and demand adaptation as characteristic channels. The characteristic values ​​of each channel of the three-dimensional suitability tensor are then normalized. The three-dimensional suitability tensor is input into the multi-agent game model as the environmental state. The action value output by the multi-agent game model is transformed into the state transition probability, which drives the cellular automaton to perform spatial evolution and output the corresponding spatial adaptability optimization layout.

2. The spatial adaptability optimization method based on three-dimensional space re-identification according to claim 1, characterized in that, The output is a three-dimensional spatial re-identification matrix containing the fuzzy membership degrees of each grid node to production, living, and ecological spaces. Specifically: Define the spatiotemporal semantic graph network as ,in, For a set of grid nodes, Let be the set of edges. This is the edge weight matrix based on the intensity of material flow and information flow; The graph convolutional neural network aggregates node features and outputs arbitrary nodes. Fuzzy membership vectors belonging to production, living, and ecological spaces The calculation formula is as follows: in, Let D be the set of neighboring nodes of node i, and let D be the degree matrix. Let be the input feature vector of node j. The weight parameter matrix, For activation functions; Extracting satisfaction Node i serves as the fuzzy transition zone in the three-dimensional space, where... This is the preset confidence threshold.

3. The spatial adaptability optimization method based on three-dimensional space re-identification according to claim 2, characterized in that, The carbon sink opportunity cost of the target area is calculated and an ecologically rigid constraint mask is applied, specifically as follows: Define the carbon sink opportunity cost for any unit k in the set of potential production spatial units. The calculation formula is as follows: in, The amount of carbon sequestration per unit area that maintains the natural background state for unit k. This represents the estimated carbon sink after unit k is transformed into its production location. For the unit price equivalent in the carbon trading market, This is the amplification factor for ecological sensitivity. Construct an ecologically rigid constraint mask matrix M, if the region where cell k is located... If the cost exceeds a preset threshold or is within the legally mandated ecological red line, then the mask matrix elements... ,otherwise The potential set of spatial units is extracted by performing a Hadamard product operation on the spatial grid and the mask matrix M.

4. The spatial adaptability optimization method based on three-dimensional space re-identification according to claim 3, characterized in that, A three-dimensional suitability tensor is constructed, with spatial unit coordinates as the plane and ecological constraints, value matching, and demand adaptation as characteristic channels. Specifically: Constructing tensors ,in, These are the latitude and longitude grid coordinates of the spatial unit; The first channel of the tensor Fill as cells The ecological constraint index is obtained by weighted summation and normalization of soil erosion sensitivity and biodiversity maintenance index; The second channel of tensor Fill as cells The value matching index is obtained by subtracting the opportunity cost of carbon sink from the economic value of basic production. The third channel of tensor Fill as cells The demand matching index is obtained by combining the potential index of industrial scale contiguous area and the resource footprint matching degree.

5. The spatial adaptability optimization method based on three-dimensional space re-identification according to claim 4, characterized in that, The third channel of tensor Demand Fit Index The calculation formula, based on the composite industry compatibility assessment model, is as follows: in, and Units Soil physical and chemical suitability and agricultural water resource carrying capacity; and These are the solar thermal radiation resource index and the capacity margin of the grid access node, respectively. and These are the agricultural demand weight and the new energy access weight, respectively. As a factor for adjusting the spatial compatibility of composite industries, and .

6. The spatial adaptability optimization method based on three-dimensional space re-identification according to claim 5, characterized in that, The three-dimensional suitability tensor is used as the environmental state input into the multi-agent game model, specifically as follows: The developers and protectors of the production area are instantiated as intelligent agents in a game of mutual competition, and the state space of each agent is a three-dimensional suitability tensor T. Define the local joint reward function for an agent developing intelligence at any origin at time t. The calculation formula is as follows: in, The set of potential production space units occupied by the agent at time t; , , Units The three corresponding channel eigenvalues ​​in the tensor; As a positive incentive weight, This is the penalty coefficient for crossing ecological boundaries. This is an ecological sensitivity parameter.

7. The spatial adaptability optimization method based on three-dimensional space re-identification according to claim 6, characterized in that, The action value output by the multi-agent game model is converted into state transition probabilities, and the state transition probabilities of its cellular automaton are... The calculation formula is: in, The action value Q function output by the agent, which transforms unit i into place type j. Represents the set of neighboring nodes of cell i. The proportion of cells whose internal state is j. As an indicator function, when the state of cell k... hour ,otherwise , This is the global spatial suitability control function, where element i corresponds to the mask matrix element. hour, Otherwise, it is 1.

8. The spatial adaptability optimization method based on three-dimensional space re-identification according to claim 7, characterized in that, The corresponding space-adaptive optimized layout is output, specifically as follows: Extracting the eigenvectors of potential production space units that have evolved to a stable state in the three-dimensional suitability tensor ; Define the positive ideal solution vector for space adaptation. Calculate the weighted Euclidean distance ,in, These are the weighting coefficients for each dimension; Based on weighted Euclidean distance The spatial units are arranged in ascending order, and based on the allocation ratio of the total control indicators for the target area, the first suitable scale development zone, the second suitable characteristic compatibility zone, and the third suitable restricted retreat zone are delineated in sequence.

9. The spatial adaptability optimization method based on three-dimensional space re-identification according to claim 8, characterized in that, Also includes: Acquire real-time ecological monitoring data from the Internet of Things, generate dynamic penalty factors through a nonlinear decay model, and feed the dynamic penalty factors back to the value matching feature channel of the three-dimensional suitability tensor to update feature values, thereby triggering the re-evolution of spatial layout and completing the dynamic closed-loop management of spatial adaptability. Specifically, a dynamic penalty factor is generated through a nonlinear decay model, and this dynamic penalty factor is fed back into the value matching feature channel of the three-dimensional suitability tensor to update the feature values. Computing unit Ecological stress index at time t ; If unit The ecological stress index in continuous If the preset ecological tolerance threshold is exceeded within a monitoring cycle, a dynamic penalty factor for that unit will be triggered and calculated. The calculation formula is as follows: in, For unit The dynamic penalty factor at the current time t. As a penalty decay rate constant, For the summation index variable of historical monitoring times, The set number of continuous monitoring cycles, For unit At historical monitoring moments Ecological stress index. This is the ecological tolerance threshold; Apply a dynamic penalty factor to the second channel of the 3D fitness tensor, updating the local tensor value. This is to suppress the game expansion probability of the unit in the spatial evolution model.

10. A spatial adaptability optimization system based on three-dimensional space re-identification, characterized in that, include: The graph network spatial recognition module is used to acquire multi-source spatiotemporal data of the target area and construct a spatiotemporal semantic graph network. It outputs a three-life space re-identification matrix containing the fuzzy membership degree of each grid node belonging to production, living and ecological spaces through a graph convolutional neural network. The mask constraint and set extraction module is used to calculate the opportunity cost of carbon sinks and apply an ecologically rigid constraint mask to extract a set of potential production site spatial units; The 3D suitability tensor construction module is used to fuse multidimensional data, construct a 3D suitability tensor with ecological constraints, value matching and demand adaptation as feature channels, and normalize the feature values. The game evolution optimization module is used to input the three-dimensional suitability tensor as the environmental state into the multi-agent game model, and convert the action value into the state transition probability to drive the spatial evolution of the cellular automaton to output the adaptive optimized layout. The digital twin closed-loop feedback module is used to generate dynamic penalty factors based on real-time ecological monitoring data and feed back to update the value matching feature channels of the three-dimensional suitability tensor to trigger dynamic correction of spatial layout.

Images