Land utilization planning method based on AI driving

Abstract

The invention discloses a land utilization planning method based on AI driving, relates to the technical field of land planning, and solves the technical problems that it is difficult to construct a regional heterogeneous graph and calculate spatial dependence intensity by using a heterogeneous graph attention network, and it is also difficult to construct a multi-agent reinforcement learning model with a planned region as an environment. Comprising the following steps: firstly, constructing a land feature tensor containing land configuration affinity and a development pressure evaluation value; constructing a regional heterogeneous graph by taking the land feature tensor as an initial feature of each node in the graph, calculating spatial dependence intensity among the nodes by using a heterogeneous graph attention network, and generating a land unit enhanced feature vector; the method comprises the following steps: firstly, obtaining feature vectors of different planning target agents, constructing a multi-agent reinforcement learning model based on the feature vectors, simulating games among different planning target agents, generating a Pareto optimal land utilization scheme set, and finally, integrating a cellular automaton to carry out multi-scene dynamic deduction, and receiving a user interaction instruction to generate a final planning scheme.

Description

Technical Field

[0001] This invention belongs to the field of land planning technology, specifically relating to an AI-driven land use planning method. Background Technology

[0002] Land use planning is a key means of coordinating regional development, ensuring ecological security, and promoting the sustainable use of resources.

[0003] Traditional land use planning methods mainly rely on expert experience, static evaluation models, and simple spatial analysis tools, which have the following shortcomings: Traditional methods are difficult to handle massive spatial data from multiple sources, multiple time periods, and multiple modalities, and cannot dynamically simulate and respond to complex socio-economic and natural environmental changes, resulting in poor adaptability of planning schemes; Conventional models are mostly based on simple neighborhood rules, which cannot effectively capture and quantify complex functional connections across geographical space, thus limiting the rationality of spatial layout and making it difficult to construct regional heterogeneous maps and use heterogeneous map attention networks to calculate spatial dependency strength; Planning often involves multiple conflicting goals such as economic development, ecological protection, and social equity, and traditional methods usually use linear weighting and other methods to make trade-offs, making it difficult to systematically explore and generate non-dominated Pareto optimal solution sets, resulting in limited decision support capabilities; Furthermore, they lack the ability to conduct long-term, multi-scenario dynamic simulations of candidate schemes, and also fail to provide decision-makers with intuitive and flexible interactive means to adjust and optimize schemes in real time.

[0004] Therefore, there is an urgent need for an innovative land use planning method that can deeply integrate multi-source data, accurately quantify complex spatial relationships and intelligently coordinate multi-objective conflicts, and support dynamic simulation and human-computer interaction. Summary of the Invention

[0005] This invention aims to solve at least one of the technical problems existing in the prior art; to this end, this invention proposes an AI-driven land use planning method to solve the following technical problem: Traditional land use planning methods mainly rely on expert experience, static evaluation models, and simple spatial analysis tools, which have the following shortcomings: Traditional methods struggle to handle massive spatial data from multiple sources, across multiple time periods, and with multiple modalities, and cannot dynamically simulate and respond to complex socio-economic and natural environmental changes, resulting in poor adaptability of planning schemes; conventional models are mostly based on simple neighborhood rules, failing to effectively capture and quantify complex functional connections across geographical spaces, thus limiting the rationality of spatial layout and making it difficult to construct regional heterogeneous maps and utilize heterogeneous map attention networks to calculate spatial dependency strength; planning often involves multiple conflicting goals such as economic development, ecological protection, and social equity, and traditional methods typically use linear weighting and other methods to compromise, making it difficult to systematically explore and generate non-dominated Pareto optimal solution sets, resulting in limited decision support capabilities; furthermore, they lack the ability to conduct long-term, multi-scenario dynamic simulations of candidate schemes, and also fail to provide decision-makers with intuitive and flexible interactive means to adjust and optimize schemes in real time.

[0006] To address the above problems, this invention provides an AI-driven land use planning method, comprising the following steps: S1: Collect multimodal data of the target area, perform spatiotemporal alignment and fusion processing, and stack the multidimensional feature values ​​of each land unit on its spatial coordinates to form a three-dimensional raster data structure, thereby constructing a land feature tensor containing land use configuration affinity evaluation value and development pressure evaluation value; S2: Using land units as nodes and various spatial relationships as edges, construct a regional heterogeneous graph. Calculate the spatial dependency strength between nodes through a heterogeneous graph attention network, and use the spatial dependency strength as a weight to perform weighted aggregation of the neighbor features of the target node. Then, fuse the aggregated neighborhood information with the target node's own features to generate an enhanced feature vector of land units with spatial context information. S3: Based on the enhanced feature vector of the land unit, construct a multi-agent reinforcement learning model with the planning area as the environment, and generate a Pareto optimal land use scheme set by simulating the game and decision-making between agents representing different planning objectives. S4: Output the Pareto optimal land use scheme set through a visual interface, integrate a cellular automata model to perform multi-scenario dynamic simulation of candidate schemes, calculate the land unit conversion probability, receive user interaction instructions to adjust and optimize weights, and generate the final planning scheme.

[0007] Preferably, in step S1, the land use allocation affinity evaluation value includes the following steps: The planning area is divided into Regular grid land units, in which, and These represent the number of rows and columns of the planning area grid, respectively. For each land unit, its position in... Standardized scores on each evaluation factor form a dimension. The fundamental feature tensor, where, For each feature channel, calculate the land use configuration affinity tensor. With development pressure tensor The two tensors are concatenated along the feature channel dimension to form a final tensor with dimension . The land feature tensor; The land use allocation affinity evaluation value is as follows: ; in, For the first The land unit for the first Land use type allocation affinity evaluation value, This represents the total number of evaluation factors. Indicates the first The weights of each evaluation factor Indicates based on the first The first land unit Evaluator characteristics of each evaluation factor Standardized scoring function; Among them, standardized scoring function Specifically: ; in, To achieve the highest score, To achieve the minimum score, The upper limit of the threshold, This is the lower limit of the threshold.

[0008] Preferably, in step S1, the development pressure assessment value includes the following steps: ; in, For the first Development pressure assessment value for each land unit For the first The straight-line distance from the center of each land unit to the nearest city center For the first The set of neighboring units of a land unit, This is the first distance attenuation parameter. For the first From the center of each land unit to neighboring units The straight-line distance This is the second distance attenuation parameter. As an indicator function, when neighboring cells for That is, the value is 1 when it is for construction land, and 0 otherwise. It is a constant.

[0009] Preferably, in step S2, the regional heterogeneous map includes multiple types of nodes and multiple types of edges, wherein the node types include land unit nodes, transportation hub nodes and ecological nodes; and the edge types include spatial proximity edges, transportation connection edges and ecological corridor edges.

[0010] Preferably, in step S2, the spatial dependence strength includes the following steps: The spatial dependency strength is dynamically generated through a heterogeneous graph attention network, specifically: The transformed features of the target node, the transformed features of the neighboring nodes, and the relationship identifier features corresponding to the current spatial relationship type are synthesized to generate a preliminary relationship-aware interaction signal; the relationship-aware interaction signal is mapped to a scalar; the original importance of the target node's neighbors under all similar relationships is compared, normalized by the Softmax function, and finally a standardized attention coefficient is output. Specifically, mapping the relationship-aware interaction signal to a scalar involves: predefining a corresponding relationship-specific knowledge base for each spatial relationship type m; generating a relationship-aware query vector based on the features of target node a, the features of neighbor node b, and the relationship type between them; inputting the relationship-aware query vector into the relationship-specific knowledge base corresponding to relationship type m; and outputting a scalar value through the relationship-specific knowledge base, representing the original importance score of target node a to neighbor node b under relationship type m.

[0011] Preferably, in step S3, the multi-agent reinforcement learning model includes the following steps: Multiple agents, each assigned a different reward function, corresponding to different planning objectives; A central coordinator is used to fuse action proposals from various agents and resolve conflicts based on a comprehensive policy function, wherein the comprehensive policy function is specifically: ; in, The optimal action to be executed, i.e., the comprehensive strategy, For the number of sub-strategies, For the first The weights of each agent, For the first An agent is in state The following land use actions Action value function, A cost function for penalizing conflicts in land use spatial layout; The action value function is approximated by a Q-network. Specifically, the state code, action code, and agent identity code are taken as inputs, and the three input vectors are concatenated and fed into a multilayer perceptron. After training, the network outputs a scalar, i.e. .

[0012] Preferably, the cost function for penalizing conflicts in land use spatial layout is calculated based on predefined spatial planning rules and obtained by weighted summation of multiple sub-cost functions, including neighborhood compatibility conflict cost, public service overload conflict cost, infrastructure support conflict cost, and policy red line violation conflict cost.

[0013] Preferably, the calculation of the land unit conversion probability includes the following steps: ; in, In the first The first land unit The probability of conversion between different land use types. As a regulating factor, This is the overall quality assessment value for the state-action pair. For the first The land unit for the first Land use type allocation affinity evaluation value, For the first Among the land units belonging to the first External pressures for land use type conversion This is a constraint factor.

[0014] Preferably, in step S4, outputting the Pareto optimal land use scheme through a visualization interface includes the following steps: The integrated cellular automata model performs long-term dynamic simulations of each candidate solution in the Pareto optimal solution set under multiple scenarios, simulating its spatial evolution over the next 20 to 30 years, and automatically generating a report comparing the advantages and disadvantages of each solution in key indicators. The system receives interactive commands from users through the visual interface. These commands include weight adjustment commands and scheme stitching commands, and both commands trigger a compliance verification process upon execution. Specifically, in the weight adjustment command, the user adjusts the target weights of the multi-agent reinforcement learning model using a slider. The system then generates a set of approximate Pareto frontier schemes quickly through transfer learning based on a pre-trained policy distillation model. In the scheme stitching command, the user selects several candidate schemes from the Pareto optimal scheme set and specifies the local retention region for each scheme. The system automatically verifies the schemes based on predefined spatial compatibility rules. For stitching operations that do not meet the compatibility rules, the system provides real-time warnings about the conflict and prohibits saving. For stitching schemes that pass the verification, the system automatically generates a stitching scheme boundary connection report. After comprehensively evaluating the dynamic simulation results and comparison reports, users can select one of the regenerated Pareto frontier schemes or user-defined stitching schemes as the final planning scheme.

[0015] The beneficial effects of this invention are: This invention employs multimodal data spatiotemporal fusion and three-dimensional land feature tensor construction to uniformly represent the multidimensional characteristics of land use allocation affinity and development pressure on spatial coordinates. This enables a refined and quantitative assessment of the land endowment of the planning area, providing a high-quality and high-dimensional data foundation for subsequent intelligent decision-making, and allowing planning schemes to more accurately reflect the underlying land conditions. This invention constructs a regional heterogeneous map and uses a heterogeneous map attention network to calculate the spatial dependency strength. It also introduces a relation-specific knowledge base to dynamically learn complex spatial interaction patterns, breaking through the limitations of traditional neighborhood models. This invention can accurately capture and quantify long-range spatial dependencies such as transportation and ecology, thereby generating a land use layout that is more functionally related and holistically coordinated in space. This invention constructs a multi-agent reinforcement learning model with the planning area as the environment and designs a comprehensive policy function that includes a conflict penalty cost function. By simulating the game between different target agents, it automatically explores and generates a set of Pareto optimal solutions, rather than a single compromise solution. At the same time, the cost function effectively avoids spatial layout conflicts, providing decision-makers with rich, high-quality and feasible alternatives, and significantly improving the global optimality and operability of the planning. This invention uses an integrated cellular automata model to perform dynamic simulations across multiple scenarios and provides interactive instructions such as weight adjustment and scheme stitching, as well as a compliance verification process. The system automatically performs compliance verification to ensure the feasibility of user-defined schemes, ultimately forming a final planning scheme that combines scientific rigor, foresight, and user intent, thus achieving a leap from machine output to human-machine collaborative decision-making. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0017] 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.

[0018] Please see Figure 1 As shown, this invention is an AI-driven land use planning method, comprising the following steps: S1: Collect multimodal data of the target area, perform spatiotemporal alignment and fusion processing, and stack the multidimensional feature values ​​of each land unit on its spatial coordinates to form a three-dimensional raster data structure, thereby constructing a land feature tensor containing land use configuration affinity evaluation value and development pressure evaluation value; S2: Using land units as nodes and various spatial relationships as edges, construct a regional heterogeneous graph. Calculate the spatial dependency strength between nodes through a heterogeneous graph attention network, and use the spatial dependency strength as a weight to perform weighted aggregation of the neighbor features of the target node. Then, fuse the aggregated neighborhood information with the target node's own features to generate an enhanced feature vector of land units with spatial context information. S3: Based on the enhanced feature vector of the land unit, construct a multi-agent reinforcement learning model with the planning area as the environment, and generate a Pareto optimal land use scheme set by simulating the game and decision-making between agents representing different planning objectives. S4: Output the Pareto optimal land use scheme set through a visual interface, integrate a cellular automata model to perform multi-scenario dynamic simulation of candidate schemes, calculate the land unit conversion probability, receive user interaction instructions to adjust and optimize weights, and generate the final planning scheme.

[0019] Specifically, assuming the planning objective is to achieve a macro-level distribution at the regional level, then land use types will be... Defined as the following categories: =1= Farmland =2= Forest land, =3=Grassland =4= Construction land, =5= Water area =6=Unused land.

[0020] In one embodiment of the present invention, step S1, which involves setting the land use affinity evaluation value, includes the following steps: The planning area is divided into Regular grid land units, in which, and These represent the number of rows and columns of the planning area grid, respectively. For each land unit, its position in... Standardized scores on each evaluation factor form a dimension. The fundamental feature tensor, where, For each feature channel, calculate the land use configuration affinity tensor. With development pressure tensor The two tensors are concatenated along the feature channel dimension to form a final tensor with dimension . The land feature tensor; The land use allocation affinity evaluation value is as follows: ; in, For the first The land unit for the first Land use type allocation affinity evaluation value, This represents the total number of evaluation factors. Indicates the first The weights of each evaluation factor Indicates based on the first The first land unit Evaluator characteristics of each evaluation factor Standardized scoring function; Among them, standardized scoring function Specifically: ; in, To achieve the highest score, To achieve the minimum score, The upper limit of the threshold, This is the lower limit of the threshold.

[0021] Specifically, multimodal data of the target area is collected, including physical geographic data, socioeconomic and locational data, and constraint and policy data, and the planning area is divided into... Regular grid land units, in which, and These represent the number of rows and columns of the planning area grid, respectively. For each land unit, its advantages in terms of topography, location, transportation, ecology, socio-economic factors, and policy constraints are calculated. Standardized scores on each evaluation factor form a dimension. The fundamental feature tensor, where, Given the total number of feature channels, based on the aforementioned basic feature vector, and using the calculation formula for land use configuration affinity evaluation value, the approximation of each land unit for all features is calculated in parallel. Suitability evaluation values ​​for different land use types were used, and the results were organized into a single dimension. A land use affinity tensor is generated, where each pixel (i,j) is a k-dimensional vector representing the suitability of the unit for different land use types. Based on the aforementioned basic feature tensor, the development pressure evaluation value of each land unit is calculated in parallel using the development pressure calculation formula, and the results are organized into a k-dimensional vector. The development pressure tensor, and the land use allocation affinity tensor. With the aforementioned development pressure tensor The components are concatenated along the feature channel dimension to form a final dimension of... The land feature tensor; Among them, the land use allocation affinity tensor is a The development pressure tensor represents the degree of matching and compatibility between the natural endowments and socioeconomic attributes of a specific land unit and a certain type of land use; the development pressure tensor is a... The tensor quantifies the potential pressure from urban expansion and development at each location; in, and These are the preset maximum and minimum scores, respectively. To ensure the robustness of the model, their values ​​are usually set within a small buffer zone, for example... Set to 0.9, It is set to 0.1, and its specific value and the size of the buffer can be adjusted according to actual data and expert experience; and The upper and lower limits of the threshold are set based on expert knowledge or statistical standards. In the calculation of the land use allocation affinity evaluation value, Indicates the first The weights of each evaluation factor are determined using the analytic hierarchy process (AHP) or the entropy weight method.

[0022] In one embodiment of the present invention, step S1, developing a pressure assessment value, includes the following steps: ; in, For the first Development pressure assessment value for each land unit For the first The straight-line distance from the center of each land unit to the nearest city center For the first The set of neighboring units of a land unit, This is the first distance attenuation parameter. For the first From the center of each land unit to neighboring units The straight-line distance This is the second distance attenuation parameter. As an indicator function, when neighboring cells for That is, the value is 1 when it is for construction land, and 0 otherwise. It is a constant.

[0023] Specifically, The first distance attenuation parameter quantifies the rate at which development pressure decreases with increasing distance from the city center. A larger value indicates that the city center's radiating influence decreases faster with distance, and its influence is more concentrated around the center; a smaller value indicates that the radiating influence can extend to a more distant area. The second distance attenuation parameter quantifies the rate at which development pressure decreases with increasing distance from surrounding existing built-up areas. A larger value indicates that the sprawl effect of the built-up area is highly localized; a smaller value indicates that the influence range of the built-up area is wider. The calculation of the two parameters involves selecting land use maps of the target area at two time points (T0-T1), extracting all land units that were converted from non-construction land to construction land between T0 and T1, and denoting this set as DE. Simultaneously, the land units that remained unchanged as non-construction land are grouped into a set UDE. An objective function is then constructed to find a pair of... This leads to the development pressure calculated based on the distribution of the city center and built-up area in year T0, and the assumption that... Given certain parameters, the objective function (probability value) measures the probability that a randomly selected development unit has higher development pressure than a randomly selected undeveloped unit (approximated by calculating the Mann-Whitney U statistic). This probability value is obtained by calculating the difference between the mean of the target group (DE) and the reference group (UDE) relative to the joint volatility. It is quantified by dividing the difference between the two group means by the standardized dispersion of the entire data set. and For continuous parameters, grid search or optimization algorithms are used to find the optimal parameter in the parameter space.

[0024] In one embodiment of the present invention, in step S2, the regional heterogeneous graph includes multiple types of nodes and multiple types of edges, wherein the node types include land unit nodes, transportation hub nodes and ecological nodes; and the edge types include spatial proximity edges, transportation connection edges and ecological corridor edges.

[0025] Specifically, the construction of the regional heterogeneous graph involves defining various types of nodes and edges based on the regular grid land units to construct the regional heterogeneous graph. Each land unit is treated as a node, and the feature vector at that location is extracted according to its spatial position in the land feature tensor, serving as the initial feature vector for that land unit node.

[0026] In one embodiment of the present invention, step S2, the spatial dependence strength, includes the following steps: The spatial dependency strength is dynamically generated through a heterogeneous graph attention network, specifically: The transformed features of the target node, the transformed features of the neighboring nodes, and the relationship identifier features corresponding to the current spatial relationship type are synthesized to generate a preliminary relationship-aware interaction signal; the relationship-aware interaction signal is mapped to a scalar; the original importance of the target node's neighbors under all similar relationships is compared, normalized by the Softmax function, and finally a standardized attention coefficient is output. Specifically, mapping the relationship-aware interaction signal to a scalar involves: predefining a corresponding relationship-specific knowledge base for each spatial relationship type m; generating a relationship-aware query vector based on the features of target node a, the features of neighbor node b, and the relationship type between them; inputting the relationship-aware query vector into the relationship-specific knowledge base corresponding to relationship type m; and outputting a scalar value through the relationship-specific knowledge base, representing the original importance score of target node a to neighbor node b under relationship type m.

[0027] Specifically, spatial relationship types refer to the channel types that connect two land unit nodes and assign them specific interactive semantics in a land use system. These include, but are not limited to, the following categories: spatial proximity relationships (semantic definition: based on direct physical adjacency, representing basic spatial diffusion and spillover effects; construction rules: based on von Neumann neighborhood or Moore neighborhood rules, connecting land units sharing common edges), transportation connectivity relationships (semantic definition: based on transportation infrastructure networks, representing the flow potential of economic factors such as population, capital, and goods; construction rules: connecting land units directly or indirectly connected by major transportation arteries, with connection strength weighted according to road grade or logistics intensity), and ecological corridor relationships (semantic definition: based on natural processes and ecological functions, representing the transmission and dependence of ecosystem services; construction rules: connecting land units located in the same watershed, the same migration corridor, or the same prevailing wind channel; for example, hydrological networks generated based on digital elevation models are the basis for constructing water system ecological relationships). The relation-specific knowledge base is a special data structure that encapsulates the domain knowledge of a specific type of spatial relation, either predefined or generated through training. Its implementation can be a key-value pair, a memory, or a lightweight neural network. The calculation of the original importance score involves initializing a knowledge base for each type of relation m. In this embodiment, a key-value pair memory is used as an example. This memory contains a set of keys (representing typical spatial interaction patterns) and a set of values ​​(representing the corresponding relation strength scores). The relation-aware interaction signal is used as a query vector, and its similarity (cosine similarity or dot product) with all keys in the knowledge base is calculated. The similarity is then converted into weights using the Softmax function. Subsequently, the original importance score is obtained by weighted summation of all values.

[0028] In one embodiment of the present invention, step S3 of the multi-agent reinforcement learning model includes the following steps: Multiple agents, each assigned a different reward function, corresponding to different planning objectives; A central coordinator is used to fuse action proposals from various agents and resolve conflicts based on a comprehensive policy function, wherein the comprehensive policy function is specifically: ; in, The optimal action to be executed, i.e., the comprehensive strategy, For the number of sub-strategies, For the first The weights of each agent, For the first An agent is in state The following land use actions Action value function, A cost function for penalizing conflicts in land use spatial layout; The action value function is approximated by a Q-network. Specifically, the state code, action code, and agent identity code are taken as inputs, and the three input vectors are concatenated and fed into a multilayer perceptron. After training, the network outputs a scalar, i.e. .

[0029] Specifically, the state encoding is a compact vector representation of the current land use pattern of the entire planning area, which can be processed by a graph neural network or a convolutional neural network. The action is encoded to obtain an action code for the agent. The proposed action can be represented by a vector, for example, an embedding vector, representing the land unit ID and the target land use type; the agent identity is encoded as a learnable vector to distinguish agents with different targets, enabling the network to differentiate the value judgments of different agents; and the training objective of the network is to make its output Q-value satisfy the Bellman equation. The function encapsulates the long-term strategic value of its specific goals (such as economy and ecology), which is the core basis for the agent's decision-making. After the model training converges, all network parameters are fixed, and the trained model is run through different weight combinations to complete a full planning process under each weight configuration. The series of land use schemes collected in this way constitutes the Pareto optimal land use scheme set.

[0030] In one embodiment of the present invention, the cost function for penalizing conflicts in land use spatial layout is calculated based on predefined spatial planning rules and obtained by weighted summation of multiple sub-cost functions, including neighborhood compatibility conflict cost, public service overload conflict cost, infrastructure support conflict cost, and policy red line violation conflict cost.

[0031] Specifically, the neighborhood compatibility conflict cost is used to penalize spatially adjacent but functionally incompatible land use layouts. First, a predefined... The matrix IM (Incompatibility Matrix) has the following values ​​for its elements: 1 indicates that land use type u and land use type v are spatially incompatible (e.g., a chemical plant and a residential area); 0 indicates that land use type u and land use type v are spatially compatible or unrelated. (This is used when performing actions.) In the newly formed state, each land unit and all its spatially adjacent neighboring units are traversed, specifically as follows: ,in, For nodes To the neighbor Spatial distance, Land use categories in the label space and A predefined similarity score (cosine similarity) between them; the public service overload conflict cost is used to penalize situations where regional public service demand exceeds facility capacity due to land use changes, for each type of public service facility. Definition, facilities Service capacity is denoted as After the action is taken, the total demand for the facility within the planned area will be [calculated / calculated]. The specific calculation of the conflict cost of public service overload is as follows: This formula calculates the relative overload rate, incurring a penalty only when demand exceeds capacity, and the more severe the overload, the higher the penalty. Infrastructure support conflict costs penalize situations where new land use patterns cause loads on transportation, municipal, and other infrastructure exceeding their carrying capacity. For each type of infrastructure... Define the carrying capacity threshold of the infrastructure. After the action is performed, the new load applied to the infrastructure The method of calculating load varies depending on the infrastructure. For example, for road intersections, load can be estimated based on the increased traffic volume caused by newly generated population and jobs. Similarly, penalties are only incurred when the load exceeds a threshold, with the severity of the penalty proportional to the degree of overload. The cost of violating policy red lines is the most punitive sub-cost, used to strictly prohibit the encroachment on areas protected by law or policy. After defining the red line area, the cost of violating policy red lines is calculated as follows: ,in, This refers to the total area of ​​the policy red line zone encroached upon by the new layout after the implementation of the action. The penalty coefficient per unit area is set to an extremely large value (e.g., several orders of magnitude higher than other sub-costs).

[0032] In one embodiment of the present invention, the calculation of the land unit conversion probability includes the following steps: ; in, In the first The first land unit The probability of conversion between different land use types. As a regulating factor, This is the overall quality assessment value for the state-action pair. For the first The land unit for the first Land use type allocation affinity evaluation value, For the first Among the land units belonging to the first External pressures for land use type conversion This is a constraint factor.

[0033] Specifically, It is an adjustment factor used to control the intensity of the impact of macro-strategic value on local transformation; it is an adjustable parameter greater than 0. For the first Among the land units belonging to the first External pressure for land use type conversion refers to the pressure exerted by the first The land use conditions of the surrounding area of ​​a land unit, which induce it to move towards the first External pressures for land use type conversion are traditionally addressed by calculating the first [unit / group] unit within a fixed-size neighborhood window. Quantified by the proportion of different land use types; The constraint factor is a value between 0 and 1, used to represent the first... The land unit is converted to the first The degree of restriction faced by a land use type is represented by a value closer to 1, indicating a greater restriction and a lower probability of conversion. For example, based on rigid constraint data such as ecological protection red lines, basic farmland, and steep slopes, a binary constraint layer is created for the first... The land unit and the first For land use types, if conversion is absolutely prohibited, set a very high constraint value (e.g., 0.99); if allowed, set it to 0.01.

[0034] In one embodiment of the present invention, step S4, which involves outputting the Pareto optimal land use scheme through a visualization interface, includes the following steps: The integrated cellular automata model performs long-term dynamic simulations of each candidate solution in the Pareto optimal solution set under multiple scenarios, simulating its spatial evolution over the next 20 to 30 years, and automatically generating a report comparing the advantages and disadvantages of each solution in key indicators. The system receives interactive commands from users through the visual interface. These commands include weight adjustment commands and scheme stitching commands, and both commands trigger a compliance verification process upon execution. Specifically, in the weight adjustment command, the user adjusts the target weights of the multi-agent reinforcement learning model using a slider. The system then generates a set of approximate Pareto frontier schemes quickly through transfer learning based on a pre-trained policy distillation model. In the scheme stitching command, the user selects several candidate schemes from the Pareto optimal scheme set and specifies the local retention region for each scheme. The system automatically verifies the schemes based on predefined spatial compatibility rules. For stitching operations that do not meet the compatibility rules, the system provides real-time conflict reasons and prohibits saving. For stitching schemes that pass the verification, the system automatically generates a stitching scheme boundary connection report. After comprehensively evaluating the dynamic simulation results and comparison reports, users can select one of the regenerated Pareto frontier schemes or user-defined stitching schemes as the final planning scheme.

[0035] Specifically, for each candidate solution in the Pareto optimal solution set, the system uses it as the initial state of the cellular automata model and conducts iterative simulations over the next 20 to 30 years under various preset parameters, such as the baseline scenario, the high-speed development scenario, and the ecological protection scenario, to deduce its spatial evolution process. During the simulation, the system calculates and records a series of key performance indicators for each solution at different time points in real time, including economic indicators, ecological indicators, social indicators, etc. After the simulation is completed, the system automatically generates a multi-dimensional comparison report of multiple solutions. The report clearly shows the performance and advantages and disadvantages of each solution under different scenarios and at different time points in the form of a combination of charts and spatial maps. In the scheme stitching instruction, users can load two or more candidate schemes on the interface simultaneously and use the selection tool to specify the local areas they wish to retain on different schemes (e.g., retaining the industrial layout of scheme A, and retaining the ecological corridor and residential area of ​​scheme B). After the instruction is triggered, the system immediately starts the compliance verification process, including spatial compatibility verification and infrastructure connectivity verification. If the verification fails, the system will highlight the conflicting boundary locations in real time and pop up a prompt box explaining the reason for the conflict (e.g., industrial land and residential land are directly adjacent, violating compatibility rule X), and prohibit saving the stitching scheme; otherwise, the system will automatically perform the stitching operation, smooth the boundary, and generate a boundary connection report, detailing the boundary processing logic and results. The newly generated stitching scheme will be added as a new option to the candidate scheme list.

[0036] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. An AI-driven land use planning method, characterized in that, Includes the following steps: S1: Collect multimodal data of the target area, perform spatiotemporal alignment and fusion processing, and stack the multidimensional feature values ​​of each land unit on its spatial coordinates to form a three-dimensional raster data structure, thereby constructing a land feature tensor containing land use configuration affinity evaluation value and development pressure evaluation value; S2: Using land units as nodes and various spatial relationships as edges, construct a regional heterogeneous graph. Use the land feature tensor as the initial feature of each node in the graph and input it into the heterogeneous graph attention network. Calculate the spatial dependency strength between nodes and use the spatial dependency strength as a weight to perform weighted aggregation of the neighbor features of the target node. Fuse the aggregated neighborhood information with the target node's own features to generate an enhanced feature vector of land units with spatial context information. S3: Based on the enhanced feature vector of the land unit, construct a multi-agent reinforcement learning model with the planning area as the environment, and generate a Pareto optimal land use scheme set by simulating the game and decision-making between agents representing different planning objectives. S4: Output the Pareto optimal land use scheme set through a visual interface, integrate a cellular automata model to perform multi-scenario dynamic simulation of candidate schemes, calculate the land unit conversion probability, receive user interaction instructions to adjust and optimize weights, and generate the final planning scheme.

2. The AI-driven land use planning method according to claim 1, characterized in that, In step S1, the land use allocation affinity evaluation value includes the following steps: The planning area is divided into Regular grid land units, in which, and These represent the number of rows and columns of the planning area grid, respectively. For each land unit, its position is calculated... Standardized scores on each evaluation factor form a dimension. The fundamental feature tensor, where, For each feature channel, calculate the land use configuration affinity tensor. With development pressure tensor The two tensors are concatenated along the feature channel dimension to form a final tensor with dimension . The land feature tensor; The land use allocation affinity evaluation value is as follows: ； in, For the first The land unit for the first Land use type allocation affinity evaluation value, This represents the total number of evaluation factors. Indicates the first The weights of each evaluation factor Indicates based on the first The first land unit Evaluator characteristics of each evaluation factor Standardized scoring function; Among them, standardized scoring function Specifically: ； in, To achieve the highest score, To achieve the minimum score, The upper limit of the threshold, This is the lower limit of the threshold.

3. The AI-driven land use planning method according to claim 1, characterized in that, In step S1, the development pressure assessment value includes the following steps: ； in, For the first Development pressure assessment value for each land unit For the first The straight-line distance from the center of each land unit to the nearest city center For the first The set of neighboring units of a land unit, This is the first distance attenuation parameter. For the first From the center of each land unit to neighboring units The straight-line distance This is the second distance attenuation parameter. As an indicator function, when neighboring cells for That is, the value is 1 when land is designated for construction purposes, and 0 otherwise. It is a constant.

4. The AI-driven land use planning method according to claim 1, characterized in that, In step S2, the regional heterogeneous map contains multiple types of nodes and multiple types of edges. The node types include land unit nodes, transportation hub nodes, and ecological nodes; the edge types include spatial proximity edges, transportation connection edges, and ecological corridor edges.

5. The AI-driven land use planning method according to claim 1, characterized in that, In step S2, the spatial dependence strength includes the following steps: The spatial dependency strength is dynamically generated through a heterogeneous graph attention network, specifically: The transformed features of the target node, the transformed features of the neighboring nodes, and the relationship identifier features corresponding to the current spatial relationship type are synthesized to generate a preliminary relationship-aware interaction signal; the relationship-aware interaction signal is mapped to a scalar; the original importance of the target node's neighbors under all similar relationships is compared, normalized by the Softmax function, and finally a standardized attention coefficient is output. Specifically, mapping the relationship-aware interaction signal to a scalar involves: predefining a corresponding relationship-specific knowledge base for each spatial relationship type m; generating a relationship-aware query vector based on the features of target node a, the features of neighbor node b, and the relationship type between them; inputting the relationship-aware query vector into the relationship-specific knowledge base corresponding to relationship type m; and outputting a scalar value through the relationship-specific knowledge base, representing the original importance score of target node a to neighbor node b under relationship type m.

6. The AI-driven land use planning method according to claim 1, characterized in that, In step S3, the multi-agent reinforcement learning model includes the following steps: Multiple agents, each assigned a different reward function, corresponding to different planning objectives; A central coordinator is used to fuse action proposals from various agents and resolve conflicts based on a comprehensive policy function, wherein the comprehensive policy function is specifically: ； in, The optimal action to be executed, i.e., the comprehensive strategy, For the number of sub-strategies, For the first The weights of each agent, For the first An agent is in state The following land use actions Action value function, A cost function for penalizing conflicts in land use spatial layout; The action value function is approximated by a Q-network. Specifically, the state code, action code, and agent identity code are taken as inputs, and the three input vectors are concatenated and fed into a multilayer perceptron. After training, the network outputs a scalar, i.e. .

7. The AI-driven land use planning method according to claim 6, characterized in that, The cost function for penalizing conflicts in land use spatial layout is calculated based on predefined spatial planning rules and obtained by weighted summation of multiple sub-cost functions, including neighborhood compatibility conflict cost, public service overload conflict cost, infrastructure support conflict cost, and policy red line violation conflict cost.

8. The AI-driven land use planning method according to claim 1, characterized in that, The calculation of the land unit conversion probability includes the following steps: ； in, In the first The first land unit The probability of conversion between different land use types. As a regulating factor, This is the overall quality assessment value for the state-action pair. For the first The land unit for the first Land use type allocation affinity evaluation value, For the first Among the land units belonging to the first External pressures for land use type conversion This is a constraint factor.

9. The AI-driven land use planning method according to claim 1, characterized in that, In step S4, the Pareto optimal land use scheme is output through a visualization interface, including the following steps: The integrated cellular automata model performs long-term dynamic simulations of each candidate solution in the Pareto optimal solution set under multiple scenarios, simulating its spatial evolution over the next 20 to 30 years, and automatically generating a report comparing the advantages and disadvantages of each solution in key indicators. The system receives interactive commands from users through the visual interface. These commands include weight adjustment commands and scheme stitching commands, and both commands trigger a compliance verification process upon execution. Specifically, in the weight adjustment command, the user adjusts the target weights of the multi-agent reinforcement learning model using a slider. The system then generates a set of approximate Pareto frontier schemes quickly through transfer learning based on a pre-trained policy distillation model. In the scheme stitching command, the user selects several candidate schemes from the Pareto optimal scheme set and specifies the local retention region for each scheme. The system automatically verifies the schemes based on predefined spatial compatibility rules. For stitching operations that do not meet the compatibility rules, the system provides real-time warnings about the conflict and prohibits saving. For stitching schemes that pass the verification, the system automatically generates a stitching scheme boundary connection report. After comprehensively evaluating the dynamic simulation results and comparison reports, users can select one of the regenerated Pareto frontier schemes or user-defined stitching schemes as the final planning scheme.

Images