A method for urban and rural space partitioning and demarcation based on reinforcement learning

Abstract

This invention relates to the fields of land planning management and computer-aided decision-making technology, specifically a method for urban and rural spatial zoning based on reinforcement learning. The method includes: collecting multi-source urban and rural geospatial data and constructing an adjacency graph of management nodes; aggregating features using a graph attention mechanism and outputting a topological state vector; modeling zoning as a Markov decision process; defining a planning adjustment instruction set and introducing a management constraint shielding unit to filter illegal instructions; constructing a composite reward function integrating carbon emission reduction, facility accessibility, and social equity; using a proximal policy optimization algorithm to drive the planning execution unit to perform policy search; and refining the generated preliminary scheme using a patch growth model. This invention solves the problems of slow response, multi-objective misalignment, and morphological fragmentation in spatial zoning under complex constraints.

Description

Technical Field

[0001] This invention relates to the field of land planning management and computer-aided decision-making technology, specifically a method for urban and rural spatial zoning based on reinforcement learning. Background Technology

[0002] In contemporary urban and rural planning systems, spatial zoning is the equilibrium point in the interplay of social, economic, and ecological interests. With the development of computational intelligence, urban and rural spatial research is shifting from descriptive simulation to generative decision-making, using sequential decisions based on the interaction between planning implementation units and the environment to explore optimal spatial structures.

[0003] Current technologies primarily rely on overlay analysis of geographic information systems, suitability assessments, and human experience-based judgment. When dealing with dynamic and nonlinear urban-rural evolution processes, they suffer from drawbacks such as slow response times and limited scientific support dimensions. While some studies have introduced genetic algorithms or linear regression benchmarks, they often face challenges such as high computational costs, low search efficiency, and imbalances in multi-objectives when solving large-scale urban-rural zoning optimization problems. Furthermore, traditional cellular automata models often use individual pixels as units, neglecting the topological attributes of land parcels in real-world development, making it difficult to simulate organic growth processes and leading to fragmented morphology and sporadic sprawl in spatial zoning results. Current deep learning models lack effective spatial state encoding capabilities when dealing with irregular geographic geometric features, and their black-box mechanism results in a lack of interpretability for zoning instructions, making it difficult to output compliant and efficient layout schemes under complex administrative constraints.

[0004] This invention aims to address the problem that traditional geographic information system (GIS) methods, which employ raster or vector overlays, struggle to effectively capture the complex adjacency and functional relationships between land parcels, roads, and intersections. By coupling graph neural networks with a near-end policy optimization algorithm, it achieves automated generation of spatial zoning schemes and optimization of physical boundaries.

[0005] To address this, a method for urban and rural spatial zoning based on reinforcement learning is proposed. Summary of the Invention

[0006] The purpose of this invention is to provide a method for urban and rural spatial zoning based on reinforcement learning. By integrating multi-source spatiotemporal data and constructing a sequential decision model, it can achieve automated optimization and compliant generation of urban and rural spatial layout under complex management constraints.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A method for urban-rural spatial zoning based on reinforcement learning includes: Collect multi-source urban and rural geospatial data, define land parcels, road segments, and intersections as management nodes, establish edge connections based on spatial adjacency relationships, and construct an urban and rural adjacency graph; use graph attention mechanism to aggregate the features of each management node in the urban and rural adjacency graph, and output a high-dimensional topological state vector; The urban and rural spatial zoning process is modeled as a Markov decision process, and the high-dimensional topological state vector is used as the input for the planning execution unit to perceive the environment. The planning adjustment instruction set is defined as the change of land parcel type and the adjustment of development intensity parameters. A management constraint shielding unit is introduced to filter illegal action instructions in real time according to the preset spatial control red line and technical criteria. Construct a composite reward function that integrates carbon emission reduction, facility accessibility optimization value, and social equity indicators to transform spatial layout targets into quantitative reward signals that guide model iteration; The high-dimensional topological state vector is used to drive the near-end policy optimization algorithm to search for policies, so that the policy loss and evaluation loss reach the preset convergence threshold. The preliminary zoning scheme generated by the near-end policy optimization algorithm is input into the patch growth model. The patch morphology is corrected by combining historical growth characteristics with a feedback mechanism to generate an urban and rural spatial zoning scheme that conforms to the physical actual boundary.

[0008] Preferably, the specific steps for constructing the urban-rural adjacency map include: The process involves acquiring multi-source urban and rural geospatial data and extracting the geometric center coordinates and boundary shape information of plots, road segments, and intersections. Each plot, road segment, and intersection is then set as an independent management node, and a unique identifier is assigned to each management node. The physical adjacency between management nodes is determined by calculating the length of the common boundary between two adjacent management nodes and determining that the Euclidean distance between the geometric centers of the two management nodes is within a preset threshold. The distribution density of points of interest, the functional category of regions of interest, and traffic flow data are mapped to the corresponding management nodes. Based on the physical adjacency, entity edges are established between the management nodes, and logical association edges are established between the management nodes based on functional complementarity and traffic accessibility, generating an urban and rural adjacency map representing spatial association characteristics.

[0009] Preferably, the specific steps for outputting the high-dimensional topological state vector are as follows: The geometric center coordinates of the management node are obtained and associated with the distribution density of points of interest and traffic flow data. The geometric attributes, functional attributes, and traffic flow attributes of the management node are concatenated into an initial feature vector. A linear transformation matrix is ​​used to perform feature mapping on adjacent management nodes, and a graph attention mechanism is used to calculate the association weight between the central management node and neighboring management nodes. The association weight represents the influence of the neighboring management nodes on the spatial layout of the central management node. Based on the association weight, the features of the neighboring management nodes are weighted and summed, and the feature representation of the central management node is updated by combining a nonlinear activation function. Multiple independently running graph attention mechanisms are executed in parallel, and the resulting feature vectors are concatenated and averaged to eliminate the bias of the single feature extraction process. The fused global features are extracted, mapped into a fixed-dimensional vector, and the high-dimensional topological state vector is generated.

[0010] Preferably, the steps for modeling the urban-rural spatial zoning process as the Markov decision process are as follows: The high-dimensional topological state vector is defined as the state quantity at the current moment. The state quantity encompasses the type, development intensity, and spatial association characteristics of each management node in the urban-rural adjacency graph. A set of decision-making behaviors that the planning execution unit can execute in the current state is defined. The set of decision-making behaviors includes discrete actions to change the land parcel type and continuous actions to fine-tune the plot ratio parameter. The immediate benefit of the planning execution unit after executing the decision-making behavior is calculated based on the composite reward function, and the immediate benefit is fed back to the planning execution unit to guide the planning execution unit in strategy optimization. The node attributes of the urban-rural adjacency graph are updated according to the decision-making behavior output by the planning execution unit, and the graph attention mechanism is re-triggered to generate the state quantity for the next moment.

[0011] Preferably, the specific steps for filtering illegal action commands in real time according to preset space control red lines and technical criteria are as follows: Based on the preset spatial control red line range and topographic parameters, constraint attribute labels are assigned to each management node in the urban-rural adjacency map. When the planning execution unit outputs a decision instruction for a certain management node, the constraint attribute label of the management node is extracted, and it is determined whether the change of land parcel type and adjustment of development intensity involved in the decision instruction violate the preset ecological protection threshold. For illegal decision instructions that violate the preset threshold, the probability score corresponding to the illegal decision instruction is mapped to zero, and an action mask vector with the same dimension as the planning adjustment instruction set is generated. The action mask vector is multiplied element-wise with the original action probability distribution output by the planning execution unit to eliminate the selection probability of illegal decision instructions. The clipped planning adjustment instruction set is normalized, and compliant decision instructions that conform to the spatial control red line and technical criteria are resampled and generated.

[0012] Preferably, the design of the composite reward function, which transforms the spatial layout objective into a quantitative reward signal to guide model iteration, includes the following steps: The carbon emission reduction refers to the reduction in carbon emissions under the current spatial layout; by calculating the total traffic carbon emissions generated by the management node and the total carbon emissions generated by the ecosystem, the difference between the total traffic carbon emissions and the total carbon emissions is calculated relative to the historical baseline, and the reduction is mapped to a quantitative value; the facility accessibility optimization value refers to the optimized walking distance for residents to reach various public service facilities; the average walking distance from the management node to the nearest public service facility is calculated, and the average walking distance is converted into a quantitative value representing spatial convenience using a negative correlation function; the social equity index refers to the distribution balance of public service resources within different zoning units; the Gini coefficient is used to measure the differences in resource allocation within the zoning units, and the distribution balance is mapped to a quantitative value; the weight coefficients of each index are preset, and the carbon emission reduction, the facility accessibility optimization value, and the social equity index are weighted and summed to output a total score representing the overall benefit of the current spatial layout scheme, and the total score is fed back to the planning execution unit as a quantitative reward signal.

[0013] Preferably, the specific steps for using the near-end policy optimization algorithm to drive the planning execution unit to perform policy search and identify spatial layout patterns that meet the reward function maximization requirement are as follows: Within the planning execution unit, an actor network and a critic network are established. The actor network outputs the probability distribution of decision instructions based on the high-dimensional topological state vector at the current moment. The critic network evaluates the value of the current spatial state and outputs a state score. The planning execution unit interacts with the sequential decision-making environment in multiple rounds, recording and storing time-series data including state representations, decision instructions, immediate rewards, and state representations at the next moment. Based on the state scores and immediate rewards output by the critic network, the advantage estimate corresponding to each decision instruction is calculated to represent the degree to which the decision instruction is better than the current average level. The parameters of the actor network are updated using the advantage estimate, and a proportional pruning mechanism is introduced during the update process to limit the probability ratio of the new strategy to the old strategy within a preset range. The mean square error between the predicted score output by the critic network and the actual reward is calculated, and the parameters of the critic network are updated using a gradient descent algorithm until the loss value of the critic network converges to below a preset threshold. The actor network after training convergence is extracted, and the optimal set of instructions for changing land parcel types and adjusting development intensity is output for the entire urban and rural spatial area, forming an urban and rural spatial zoning scheme.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention constructs an urban-rural adjacency map containing land parcels, road segments, and intersection management nodes, and utilizes a graph attention mechanism to aggregate node features, effectively solving the problem of missing topological relationships in traditional geographic information systems when processing irregular geometric features. The high-dimensional topological state vector can accurately capture the heterogeneity and connectivity of urban and rural spaces, providing a higher-dimensional spatial operating foundation for planning execution units, thereby making the generated spatial layout more consistent with the complex real-world geographical environment.

[0015] 2. This invention models the urban and rural spatial delineation process as a Markov decision process. By introducing management constraint shielding units and action mask vectors, it achieves real-time filtering of illegal action commands. Compared with traditional manual judgment or static simulation methods, the compliant decision command generation mechanism ensures that the zoning scheme strictly adheres to technical principles such as spatial control red lines and ecological protection thresholds while exploring the optimal spatial structure, significantly improving the administrative feasibility and search efficiency of generative planning.

[0016] 3. This invention solves the problems of goal conflict and fragmented morphology in traditional zoning methods by constructing a composite reward function that integrates carbon emission reduction, facility accessibility, and social equity, and by coupling the near-end strategy optimization algorithm with a patch growth model. The method enables the model to automatically identify efficient spatial layout patterns, achieving significant reductions in carbon emissions and equalization of public services. Simultaneously, by correcting patch morphology, it ensures that the zoning scheme conforms to physical boundaries, guiding the formation of a compact and intensive urban growth pattern. Attached Figure Description

[0017] Figure 1 This is a flowchart of a method for urban and rural spatial zoning based on reinforcement learning proposed in this invention. Figure 2 This is a flowchart illustrating a method for urban and rural spatial zoning based on reinforcement learning proposed in this invention. Figure 3 This is a flowchart of the method for constructing urban-rural adjacency graphs proposed in this invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figures 1 to 3 This invention provides a method for urban and rural spatial zoning based on reinforcement learning, the technical solution of which is as follows: A method for urban-rural spatial zoning based on reinforcement learning, such as Figures 1-2 As shown, it includes: Collect multi-source urban and rural geospatial data, define land parcels, road segments, and intersections as management nodes, establish edge connections based on spatial adjacency relationships, and construct an urban and rural adjacency graph; use graph attention mechanism to aggregate the features of each management node in the urban and rural adjacency graph, and output a high-dimensional topological state vector; The urban and rural spatial zoning process is modeled as a Markov decision process, and the high-dimensional topological state vector is used as the input for the planning execution unit to perceive the environment. The planning adjustment instruction set is defined as the change of land parcel type and the adjustment of development intensity parameters. A management constraint shielding unit is introduced to filter illegal action instructions in real time according to the preset spatial control red line and technical criteria. Construct a composite reward function that integrates carbon emission reduction, facility accessibility optimization value, and social equity indicators to transform spatial layout targets into quantitative reward signals that guide model iteration; The high-dimensional topological state vector is used to drive the near-end policy optimization algorithm to search for policies, so that the policy loss and evaluation loss reach the preset convergence threshold. The preliminary zoning scheme generated by the near-end policy optimization algorithm is input into the patch growth model. The patch morphology is corrected by combining historical growth characteristics with a feedback mechanism to generate an urban and rural spatial zoning scheme that conforms to the physical actual boundary.

[0020] Furthermore, such as Figure 3 As shown, the specific steps for constructing the urban-rural adjacency map include: The process involves acquiring multi-source urban and rural geospatial data and extracting the geometric center coordinates and boundary shape information of plots, road segments, and intersections. Each plot, road segment, and intersection is then set as an independent management node, and a unique identifier is assigned to each management node. The physical adjacency between management nodes is determined by calculating the length of the common boundary between two adjacent management nodes and determining that the Euclidean distance between the geometric centers of the two management nodes is within a preset threshold. The distribution density of points of interest, the functional category of regions of interest, and traffic flow data are mapped to the corresponding management nodes. Based on the physical adjacency, entity edges are established between the management nodes, and logical association edges are established between the management nodes based on functional complementarity and traffic accessibility, generating an urban and rural adjacency map representing spatial association characteristics.

[0021] When determining the physical adjacency between the management nodes, the preset threshold The value is typically set based on 1.2 to 2 times the average radius of the land parcel. The specific determination formula is as follows: Calculate the geometric center coordinates of the two management nodes. and Euclidean distance between .like If they are adjacent, then it is determined that they have potential adjacency.

[0022] Through spatial topology overlay analysis, the common boundary between two adjacent management nodes is extracted. If the length of the common boundary... If they are directly physically adjacent, then it is determined that they have direct physical adjacency.

[0023] The specific process of mapping the distribution density of points of interest, the functional category of regions of interest, and traffic flow data to the management node is as follows: Using a kernel density estimation algorithm, with the geometric center of the management node as the center, the number of points of interest and their category weights within a search radius R are calculated to determine the locational functional strength of each node. Traffic flow data is then mapped to corresponding road segment nodes according to spatial inclusion relationships, and corresponding traffic utility coefficients are assigned based on road grade.

[0024] The specific logic for establishing logically related edges based on the aforementioned functional complementarity and traffic accessibility is as follows: Calculate the functional feature vector between the two management nodes. and cosine similarity .like If the nodes are located within a pre-defined complementary region (e.g., a high complementarity region between residential and commercial nodes), then functional association edges are established. The network travel time between the two management nodes is calculated using a shortest path algorithm (such as Dijkstra's algorithm). .like If the time is less than a preset time threshold (e.g., 15 minutes), then the two are determined to have strong accessibility and a traffic association edge is established.

[0025] This embodiment significantly enhances the accuracy of urban-rural spatial relationship representation by introducing quantitative judgment criteria and multi-source data fusion. Utilizing a dual judgment based on the average radius of land parcels and the length of common boundaries, combined with a kernel density estimation algorithm, it effectively solves the ambiguity in the topological identification of irregular geographical entities. Simultaneously, by constructing logically related edges using the cosine similarity operator and the shortest path algorithm, it deeply explores the functional synergy and spatiotemporal accessibility between nodes.

[0026] Furthermore, the specific steps for outputting the high-dimensional topological state vector are as follows: The geometric center coordinates of the management node are obtained and associated with the distribution density of points of interest and traffic flow data. The geometric attributes, functional attributes, and traffic flow attributes of the management node are concatenated into an initial feature vector. A linear transformation matrix is ​​used to perform feature mapping on adjacent management nodes, and a graph attention mechanism is used to calculate the association weight between the central management node and neighboring management nodes. The association weight represents the influence of the neighboring management nodes on the spatial layout of the central management node. Based on the association weight, the features of the neighboring management nodes are weighted and summed, and the feature representation of the central management node is updated by combining a nonlinear activation function. Multiple independently running graph attention mechanisms are executed in parallel, and the resulting feature vectors are concatenated and averaged to eliminate the bias of the single feature extraction process. The fused global features are extracted, mapped into a fixed-dimensional vector, and the high-dimensional topological state vector is generated.

[0027] The initial feature vector By management node It is composed of linearly concatenated multidimensional attributes: Geometric properties: include the coordinates of the geometric center. The dimension is 2.

[0028] Functional attributes: include The distribution density vector of the points of interest of the class, with dimension . .

[0029] Traffic attribute: Includes traffic flow scalar of associated road segments, with a dimension of 1.

[0030] The association weights between the central management node and its neighboring management nodes are calculated using the graph attention mechanism, specifically implemented using the following operator: First, the node features are augmented using a linear transformation matrix with shared parameters. Then, the center node is calculated using the vector dot product operator. Its neighboring nodes Original correlation .

[0031] Introducing nonlinear activation functions The data is processed, and the Softmax function is used to normalize all neighboring nodes of the central node to obtain the final association weights. Ensure that the sum of the weights of all neighborhoods equals 1.

[0032] The specific fusion method for executing multiple independently running graph attention mechanisms in parallel is as follows: In the hidden layer processing stage, the feature vectors output by K independent attention mechanisms are concatenated laterally to enhance the model's ability to capture spatially heterogeneous features; In the final feature aggregation stage, the output results of multiple heads are weighted and averaged, and the average value is used as the stable feature representation of the node in this iteration, thereby eliminating the single feature bias caused by random initialization.

[0033] The specific steps for extracting the fused global features and generating a high-dimensional topological state vector are as follows: A global average pooling operator is used to aggregate the spatial dimensions of the features of all management nodes in the graph, extracting a global feature vector that reflects the overall urban and rural layout characteristics. The aggregated features are then mapped to a preset fixed dimension (e.g., 128 or 256 dimensions) through one or more fully connected networks, and a linear rectified function is used for nonlinear mapping, ultimately generating the high-dimensional topological state vector, which serves as the state input for the Markov decision process.

[0034] This embodiment significantly improves the extraction accuracy of urban-rural spatial heterogeneity features through deep coupling of a multi-head graph attention mechanism and a global pooling operator. By utilizing linear transformation for dimensionality enhancement and normalized association weight operators, the spatial influence between neighborhood management nodes is accurately characterized, resolving the problem of fuzzy associations of irregular plot features. Through the splicing and averaging of the multi-head mechanism, random initialization bias is effectively eliminated, enhancing the robustness of the model representation. Combined with global average pooling and nonlinear mapping, complex topological attributes are transformed into high-dimensional state vectors, providing Markov decision-making with precise perceptual input containing global layout logic, greatly improving the scientific rigor and stability of zoning decisions.

[0035] Furthermore, the steps of modeling the urban-rural spatial zoning process as the aforementioned Markov decision process are as follows: The high-dimensional topological state vector is defined as the state quantity at the current moment. The state quantity encompasses the type, development intensity, and spatial association characteristics of each management node in the urban-rural adjacency graph. A set of decision-making behaviors that the planning execution unit can execute in the current state is defined. The set of decision-making behaviors includes discrete actions to change the land parcel type and continuous actions to fine-tune the plot ratio parameter. The immediate benefit of the planning execution unit after executing the decision-making behavior is calculated based on the composite reward function, and the immediate benefit is fed back to the planning execution unit to guide the planning execution unit in strategy optimization. The node attributes of the urban-rural adjacency graph are updated according to the decision-making behavior output by the planning execution unit, and the graph attention mechanism is re-triggered to generate the state quantity for the next moment.

[0036] The planning adjustment instruction set includes discrete actions for changes in land parcel type and continuous actions for fine-tuning of plot ratio parameters. Its quantification process is as follows: The land use categories in urban and rural planning are preset as a set of integer indices. When the planning execution unit outputs discrete instructions, it converts the type characteristics of the target node into corresponding one-hot codes by looking up the index table.

[0037] The fine-tuning command for the plot ratio parameter is generated by the output layer of the neural network, and its original value is in the range of [-1, 1]. Before being applied to the management node, it needs to be multiplied by a preset step size operator (e.g., 0.05 or 0.1) to ensure the smoothness of the development intensity adjustment.

[0038] The specific steps for updating the urban-rural adjacency map based on the specific instructions output by the planning execution unit are as follows: Extract the index of the management node pointed to by the instruction. If it is a type change instruction, directly replace the category field in the node's feature vector; if it is an intensity fine-tuning instruction, accumulate the scaled continuous displacement amount on the original volume ratio value. Reinsert the updated node vector into the global feature matrix X. Since the topological adjacency relationship of the spatial nodes remains stable during short-term adjustments, the adjacency matrix A remains unchanged.

[0039] The process of re-triggering the graph attention mechanism and generating the state variables for the next time step is as follows: The updated feature matrix The original adjacency matrix A is input to a preset graph attention layer, and the spatial association weights between nodes are recalculated using a multi-head attention mechanism. After nonlinear activation and linear mapping, the output is a high-dimensional topological state vector reflecting the latest layout changes, completing the state transition for one decision step.

[0040] This embodiment models urban-rural zoning as a Markov decision process, achieving dynamic interaction and smooth evolution of planning decisions. Utilizing one-hot encoding and step-size operator quantization of the instruction set, it accurately maps land use changes and plot ratio fine-tuning, ensuring the stability of development intensity adjustments. Through cyclical triggering of local feature replacement and graph attention mechanisms, the model can capture spatial correlation changes caused by layout shifts in real time, achieving closed-loop feedback of state evolution. This sequential optimization mechanism replaces traditional static simulation methods, significantly enhancing the planning scheme's ability to perceive complex spatial logic, making the zoning results more scientific and dynamically adaptable.

[0041] Furthermore, the specific steps for filtering illegal action commands in real time based on preset space control red lines and technical criteria are as follows: Based on the preset spatial control red line range and topographic parameters, constraint attribute labels are assigned to each management node in the urban-rural adjacency map. When the planning execution unit outputs a decision instruction for a certain management node, the constraint attribute label of the management node is extracted, and it is determined whether the change of land parcel type and adjustment of development intensity involved in the decision instruction violate the preset ecological protection threshold. For illegal decision instructions that violate the preset threshold, the probability score corresponding to the illegal decision instruction is mapped to zero, and an action mask vector with the same dimension as the planning adjustment instruction set is generated. The action mask vector is multiplied element-wise with the original action probability distribution output by the planning execution unit to eliminate the selection probability of illegal decision instructions. The clipped planning adjustment instruction set is normalized, and compliant decision instructions that conform to the spatial control red line and technical criteria are resampled and generated.

[0042] The constraint attribute labels are defined as multidimensional vectors. ,in: The control zone code refers to the spatial control zone (e.g., prohibited construction zone, restricted construction zone, suitable construction zone) to which the management node belongs. This is a terrain constraint value, recording the average slope within the area where the management node is located; The setback distance threshold is used to record the spatial distance from the management node to the nearest ecologically sensitive element (such as water system or forest). This represents the upper limit of environmental capacity, recording the maximum allowable floor area ratio or development density threshold for this node.

[0043] To determine whether the decision instruction violates a preset ecological protection threshold, the following logical judgment rule is adopted: If the decision instruction involves changing the land parcel to urban construction land, and the corresponding... If the tag is a restricted area code, it is considered a violation; if the decision instruction involves a development intensity adjustment value... Or the slope of the area where the decision instruction applies. If the slope exceeds a preset threshold, it is considered a violation. The distance between the management node and the ecological element... If the span is less than the preset minimum buffer zone span, it is considered a violation.

[0044] For the planning adjustment instruction set containing both discrete and continuous actions, the specific steps for generating the action mask vector are as follows: based on the index position of the decision instruction, construct a 0-1 mask vector with the same dimension as the instruction probability distribution. For illegal instructions that are determined to violate the threshold, set their mask bit to 0 and set the compliant instructions to 1.

[0045] The original probability distribution vector is obtained by using the Hadamard product (element-wise multiplication). With mask vector Calculations are performed to obtain the corrected distribution. .

[0046] right conduct Norm normalization (i.e., dividing each element by its sum) ensures that the probability and redistribution of compliant actions are 1. Then, the Categorical distribution is used for resampling to output the final compliance decision instruction.

[0047] This embodiment significantly improves the compliance and search efficiency of planning decisions by introducing multi-dimensional constraint attribute labels and action masking mechanisms. It utilizes the Hadamard product operator to filter illegal instructions that violate spatial red lines and ecological thresholds in real time, addressing the technical shortcomings of reinforcement learning in complex control environments where it is prone to generating non-compliant solutions. Through norm normalization and resampling logic, it ensures the completeness of the decision space under legal constraints, achieving deep decoupling between management red lines and intelligent algorithms.

[0048] Furthermore, the design of the composite reward function, which transforms the spatial layout objective into a quantitative reward signal to guide model iteration, includes the following steps: The carbon emission reduction refers to the reduction in carbon emissions under the current spatial layout; by calculating the total traffic carbon emissions generated by the management node and the total carbon emissions generated by the ecosystem, the difference between the total traffic carbon emissions and the total carbon emissions is calculated relative to the historical baseline, and this reduction is mapped to a quantitative value; the facility accessibility optimization value refers to the optimized walking distance for residents to reach various public service facilities; the average walking distance from the management node to the nearest public service facility is calculated, and the average walking distance is converted into a quantitative value representing spatial convenience using a negative correlation function; the social equity index refers to the distribution balance of public service resources within different zoning units; the Gini coefficient is used to measure the differences in resource allocation within the zoning units, and the distribution balance is mapped to a quantitative value; weighting coefficients are preset for each index, and the carbon emission reduction, the facility accessibility optimization value, and the social equity index are weighted and summed to output a total score representing the overall benefit of the current spatial layout scheme, and this total score is fed back to the planning execution unit as a quantitative reward signal.

[0049] The carbon emission reduction is quantified by constructing a carbon metabolism balance model, and the specific steps are as follows: By combining traffic flow data from the management nodes with a preset emission factor per unit mileage, the traffic turnover emissions between each management node are calculated. Based on the land parcel type (e.g., forest, cultivated land, wetland) of the management node, the corresponding carbon sequestration rate coefficient per unit area is multiplied to obtain the total carbon sink within the node area. The carbon footprint value, obtained by subtracting the total carbon sink value from the total traffic emissions, is calculated and then compared with the carbon footprint under historical baseline conditions to obtain the quantified value.

[0050] The Gini coefficient is used to measure the differences in resource allocation within the partitioned units. The specific calculation logic is as follows: The total amount of public service resources (such as the number of medical beds or educational places) within each partitioned unit is calculated. The partitioned units are then sorted according to their population size, and the per capita resource difference between units is calculated using the Lorenz curve cumulative distribution function. The calculated Gini coefficient G is then converted into a fairness reward. .when When the value approaches 0, it indicates that the resource allocation is absolutely average, and at this time the mapping value of the distribution balance reaches its maximum.

[0051] Because carbon emission reduction, distance values, and the Gini coefficient use different physical units, Min-Max normalization is needed to map each indicator to the [0, 1] interval before weighted summation. Preset weight coefficients for each indicator. And satisfy By adjusting the weighting coefficients, spatial layout optimization can be achieved under different planning orientations (such as low-carbon priority or fairness priority).

[0052] This embodiment achieves precise quantification and adaptive trade-offs in multi-dimensional urban and rural spatial management objectives by constructing a composite reward system integrating a carbon metabolism balance model, an accessibility decay function, and a Gini coefficient. By utilizing the Lorenz curve cumulative distribution function and carbon footprint difference calculation, it overcomes the technical bottleneck of simultaneously measuring ecological benefits and social equity in traditional planning. Combined with normalization processing and weight adjustment mechanisms, this reward function provides feedback signals with uniform magnitude and clear guidance, effectively guiding the model to identify the optimal layout pattern with low carbon emissions, convenient facilities, and balanced resources, significantly improving the comprehensive social benefits and environmental sustainability of generative solutions.

[0053] The historical growth characteristics are statistical probability features extracted from land use change data from remote sensing images of the city over the past 10 to 20 years: Growth probability matrix: Using multi-temporal land use data, the spatial transfer probability of different land parcel types is calculated, and the spatial probability distribution reflecting the expansion inertia of a specific city is extracted.

[0054] Neighborhood attraction coefficient: Based on historical data, the attraction weight of different types of patches to surrounding undeveloped plots is extracted and used to characterize the unique physical characteristics of "contiguous development" or "cluster expansion" of the city.

[0055] Shape index constraint: Extract the average compactness and fractal dimension of historical mature partitions as geometric target values ​​in the correction process to counteract the discrete point distribution that may occur in the reinforcement learning decision-making process.

[0056] The specific steps for revising the preliminary zoning scheme using a patch growth model are as follows: First, the growth probability matrix and neighborhood attraction coefficient are extracted from historical land use change data as historical growth characteristics. The preliminary scheme generated by the near-end strategy optimization algorithm is used as seed points input into the model, and a feedback mechanism is used for local morphological correction. This feedback mechanism, while maintaining the functional positioning of the plots, clusters isolated small patches into neighboring larger patches of the same type. If the area change caused by the correction exceeds a preset threshold, a reverse fine-tuning is triggered. During the correction process, a management constraint masking unit is simultaneously invoked, using action mask vectors to ensure that patch growth does not encroach on spatial control red lines such as prohibited construction zones. The correction effect is evaluated by calculating quantitative indicators such as compactness and edge density, ensuring that the final urban and rural spatial zoning scheme eliminates the morphological fragmentation problem and conforms to the actual physical boundaries.

[0057] Furthermore, the specific steps for using the near-end policy optimization algorithm to drive the planning execution unit to perform policy search and identify spatial layout patterns that meet the reward function maximization requirement are as follows: Within the planning execution unit, an actor network and a critic network are established. The actor network outputs the probability distribution of decision instructions based on the high-dimensional topological state vector at the current moment. The critic network evaluates the value of the current spatial state and outputs a state score. The planning execution unit interacts with the sequential decision-making environment in multiple rounds, recording and storing time-series data including state representations, decision instructions, immediate rewards, and state representations at the next moment. Based on the state scores and immediate rewards output by the critic network, the advantage estimate corresponding to each decision instruction is calculated to represent the degree to which the decision instruction is better than the current average level. The parameters of the actor network are updated using the advantage estimate, and a proportional pruning mechanism is introduced during the update process to limit the probability ratio of the new strategy to the old strategy within a preset range. The mean square error between the predicted score output by the critic network and the actual reward is calculated, and the parameters of the critic network are updated using a gradient descent algorithm until the loss value of the critic network converges to below a preset threshold. The actor network after training convergence is extracted, and the optimal set of instructions for changing land parcel types and adjusting development intensity is output for the entire urban and rural spatial area, forming an urban and rural spatial zoning scheme.

[0058] Both the actor network and the critic network employ a multilayer perceptron structure, specifically configured as follows: each network includes an input layer, two hidden layers, and an output layer; the hidden layers all use ReLU as the activation function, and the number of hidden layer nodes is preset to 256 or 512. The output layer of the actor network uses Softmax to handle discrete instruction probabilities and Tanh to handle continuous instructions; the output layer of the critic network does not have an activation function and directly outputs scalar state scores.

[0059] To ensure that the planning execution unit can identify the optimal layout pattern, the following training criteria are set: A synchronous parallel sampling method is employed, extracting a fixed number of trajectory segments (BatchSize of 64 or 128) from the experience storage in each round. The total loss function consists of a policy pruning loss, a mean squared error loss of the value function, and a weighted entropy regularization term, where the entropy regularization coefficient is set to 0.01 to encourage exploratory behavior of the agent in the early stages of training. Training is considered convergent when the average composite reward value fluctuates by less than 1% over 100 consecutive training cycles and the policy entropy value tends to stabilize.

[0060] This embodiment significantly improves the search efficiency and convergence stability of high-dimensional space decision-making through a two-layer perceptron architecture actor-commentator network and a proportional pruning mechanism. By utilizing generalized advantage estimation and entropy regularization, it addresses the problem of policy optimization easily getting trapped in local optima in complex urban and rural environments, enhancing the agent's global exploration capability. Synchronous parallel sampling and gradient descent updates ensure the model's spatiotemporal consistency when handling large-scale partitioning tasks, enabling the planning execution unit to accurately identify layout patterns that maximize multidimensional rewards, producing scientifically sound and logically robust urban and rural partitioning schemes, effectively avoiding the technical shortcomings of traditional reinforcement learning in terms of convergence performance and policy jitter.

[0061] Example 2 This embodiment uses the central urban area of ​​a megacity as a specific application scenario, focusing on how the present invention can achieve the spatial delineation goals of low carbon emissions and efficient facility allocation in high-density urbanized areas by adjusting reward weights and management constraints.

[0062] In view of the dense building characteristics of the central urban area, this embodiment performs the following operations when constructing the urban-rural adjacency map: Geometric information of 1200 independent plots of land, 450 road segments, and major transportation hubs within the urban area is acquired and set as independent management nodes. A preset threshold is then set. Calculate the Euclidean distance between the management nodes, which is 1.8 times the average radius of the plot. .

[0063] When the length of the common boundary between adjacent nodes When the center distance is within a threshold, nodes are considered direct physical adjacencies. Using a kernel density estimation algorithm with a search radius R of 400 meters, the intensity of office, commercial, and residential functions of each plot node is calculated, and traffic flow is mapped to road nodes. The cosine similarity of the functional feature vectors between nodes is calculated. Establish logical connections between nodes that have a high degree of complementarity between work and residence.

[0064] The zoning process is modeled as a Markov decision process, with parameter settings tailored to high-density scenarios: a graph attention mechanism is used to aggregate the geometric, functional, and traffic attributes of the management nodes, outputting a high-dimensional topological state vector incorporating global features. A planning adjustment instruction set is defined, where the step size operator for continuous actions is set to 0.05 to ensure the smoothness of development intensity adjustments in high-density areas. Constraint attribute labels are assigned to the management nodes within the historical preservation area and river protection line boundaries. If the planning adjustment instruction involves changing a restricted construction zone into urban construction land, the action mask vector maps the probability of the instruction to zero, thereby forcibly eliminating the risk of violation.

[0065] To drive the planning execution unit to identify low-carbon layout patterns, this embodiment performs the following quantitative calculations on the composite reward function: It uses a carbon metabolism balance model to calculate traffic turnover emissions and total ecosystem carbon emissions, calculates the reduction relative to historical baselines, and maps it to a quantitative value. It calculates the average walking distance from the plot to the nearest medical and educational facilities, converts this distance into a reward signal using a negative correlation function, and guides public service resources to concentrate in residential areas. It sets weight coefficients for each indicator, with carbon emission reduction having a specific weight. Accessibility weight of facilities Social fairness weight .

[0066] Within the planning and execution unit, the actor network and the critic network are established, with a preset hidden layer node count of 512. A proportional pruning mechanism is introduced during the update process. A synchronous parallel sampling method is used to extract trajectory segments with a batch size of 128 from the experience storage for iteration.

[0067] The generated preliminary zoning scheme is input into the patch growth model, and the patch edges are corrected using the historical land development and growth characteristics of the area to generate a zoning scheme that conforms to the actual physical boundaries.

[0068] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for urban-rural spatial zoning based on reinforcement learning, characterized in that, include: Collect multi-source urban and rural geospatial data, define land parcels, road segments, and intersections as management nodes, establish edge connections based on spatial adjacency relationships, and construct an urban and rural adjacency graph; use graph attention mechanism to aggregate the features of each management node in the urban and rural adjacency graph, and output a high-dimensional topological state vector; The urban and rural spatial zoning process is modeled as a Markov decision process, and the high-dimensional topological state vector is used as the input for the planning execution unit to perceive the environment. The planning adjustment instruction set is defined as land parcel type change and development intensity parameter adjustment. A management constraint shielding unit is introduced to filter illegal action instructions in real time based on the preset spatial control red line and technical criteria. Construct a composite reward function that integrates carbon emission reduction, facility accessibility optimization value, and social equity indicators to transform spatial layout targets into quantitative reward signals that guide model iteration; The high-dimensional topological state vector is used to drive the near-end policy optimization algorithm to search for policies, so that the policy loss and evaluation loss reach the preset convergence threshold. The preliminary zoning scheme generated by the near-end policy optimization algorithm is input into the patch growth model. The patch morphology is corrected by combining historical growth characteristics with a feedback mechanism to generate an urban and rural spatial zoning scheme that conforms to the physical actual boundary.

2. The method for urban-rural spatial zoning based on reinforcement learning according to claim 1, characterized in that, The specific steps for constructing the urban-rural adjacency map include: The process involves acquiring multi-source urban and rural geospatial data and extracting the geometric center coordinates and boundary shape information of plots, road segments, and intersections. Each plot, road segment, and intersection is then set as an independent management node, and a unique identifier is assigned to each management node. The physical adjacency between management nodes is determined by calculating the length of the common boundary between two adjacent management nodes and determining that the Euclidean distance between the geometric centers of the two management nodes is within a preset threshold. The distribution density of points of interest, the functional category of regions of interest, and traffic flow data are mapped to the corresponding management nodes. Based on the physical adjacency, entity edges are established between the management nodes, and logical association edges are established between the management nodes based on functional complementarity and traffic accessibility, generating an urban and rural adjacency map representing spatial association characteristics.

3. The method for urban-rural spatial zoning based on reinforcement learning according to claim 2, characterized in that, The specific steps for outputting the high-dimensional topological state vector are as follows: Obtain the geometric center coordinates of the management node, associate them with the distribution density of points of interest and traffic flow data, and concatenate the geometric attributes, functional attributes and traffic flow attributes of the management node into an initial feature vector; The adjacent management nodes are feature-mapped using a linear transformation matrix, and the association weight between the central management node and the neighboring management nodes is calculated using a graph attention mechanism. The association weight represents the influence of the neighboring management nodes on the spatial layout of the central management node. The features of the neighborhood management nodes are weighted and summed based on the association weights, and the feature representation of the central management node is updated by combining a nonlinear activation function. Multiple independently running graph attention mechanisms are executed in parallel, and the resulting feature vectors are concatenated and averaged to eliminate the bias of the single feature extraction process. The fused global features are extracted, mapped to fixed-dimensional vectors, and the high-dimensional topological state vector is generated.

4. The method for urban-rural spatial zoning based on reinforcement learning according to claim 1, characterized in that, The steps to model the urban-rural spatial zoning process as a Markov decision process are as follows: The high-dimensional topological state vector is defined as the state quantity at the current moment, which covers the type, development intensity, and spatial association characteristics of each management node in the urban-rural adjacency graph. Define the set of decision-making actions that the planning execution unit can execute in the current state. The set of decision-making actions includes discrete actions to change the land parcel type and continuous actions to fine-tune the plot ratio parameter. The immediate reward of the planning execution unit after executing the decision behavior is calculated based on the composite reward function. The immediate reward is fed back to the planning execution unit to guide the planning execution unit to optimize the strategy. The node attributes of the urban-rural adjacency graph are updated according to the decision behavior output by the planning execution unit, and the graph attention mechanism is re-triggered to generate the state variables for the next time step.

5. The method for urban-rural spatial zoning based on reinforcement learning according to claim 1, characterized in that, The specific steps for filtering illegal action commands in real time based on preset space control red lines and technical guidelines are as follows: Based on the preset spatial control red line range and topographic parameters, constraint attribute labels are assigned to each management node in the urban-rural adjacency map. When the planning execution unit outputs a decision instruction for a certain management node, the constraint attribute label of the management node is extracted, and it is determined whether the change of land parcel type and adjustment of development intensity involved in the decision instruction violate the preset ecological protection threshold. For illegal decision instructions that violate the preset threshold, the probability score corresponding to the illegal decision instruction is mapped to zero, and an action mask vector with the same dimension as the planning adjustment instruction set is generated. The action mask vector is multiplied element-wise with the original action probability distribution output by the planning execution unit to eliminate the selection probability of illegal decision instructions. The clipped planning adjustment instruction set is normalized, and compliant decision instructions that conform to the spatial control red line and technical criteria are resampled and generated.

6. The method for urban-rural spatial zoning based on reinforcement learning according to claim 1, characterized in that, The design of the composite reward function, which transforms spatial layout objectives into quantitative reward signals to guide model iteration, includes the following steps: The carbon emission reduction refers to the reduction in carbon emissions under the current spatial layout; by calculating the total traffic carbon emissions generated by the management nodes and the total carbon emissions generated by the ecosystem, the difference between the total traffic carbon emissions and the total carbon emissions is calculated relative to the historical baseline, and this reduction is mapped to a quantitative value; the facility accessibility optimization value refers to the optimized walking distance for residents to reach various public service facilities; the average walking distance from the management node to the nearest public service facility is calculated, and the average walking distance is converted into a quantitative value representing spatial convenience using a negative correlation function; the social equity index refers to the distribution balance of public service resources within different zoning units; the Gini coefficient is used to measure the differences in resource allocation within the zoning units, and the distribution balance is mapped to a quantitative value; the weighting coefficients of each index are preset, and the carbon emission reduction, the facility accessibility optimization value, and the social equity index are weighted and summed to output a total score representing the overall benefit of the current spatial layout scheme, and this total score is fed back to the planning execution unit as a quantitative reward signal.

7. The method for urban-rural spatial zoning based on reinforcement learning according to claim 1, characterized in that, The specific steps for using the near-end policy optimization algorithm to drive the planning execution unit to perform policy search and identify spatial layout patterns that meet the requirement of maximizing the reward function are as follows: Within the planning execution unit, an actor network and a critic network are established respectively. The actor network is used to output the probability distribution of decision instructions based on the high-dimensional topological state vector at the current moment. The critic network is used to evaluate the value of the current spatial state and output a state score. The planning execution unit is controlled to interact with the sequential decision environment in multiple rounds, and time-series data including state representation, decision instructions, immediate rewards, and state representation at the next moment are recorded and stored. Based on the state score output by the critic network and the immediate reward, the advantage estimate corresponding to each decision instruction is calculated to characterize the degree to which the decision instruction is better than the current average level. The parameters of the actor network are updated using the advantage estimate, and a proportional pruning mechanism is introduced during the update process to limit the probability ratio of the new strategy to the old strategy within a preset range. The mean square error between the predicted score output by the critic network and the actual return is calculated, and the parameters of the critic network are updated using a gradient descent algorithm until the loss value of the critic network converges to below a preset threshold. The actor network after training convergence is extracted, and the optimal set of instructions for changing land parcel types and adjusting development intensity is output for the entire urban and rural space, forming an urban and rural space zoning scheme.

Images