Urban extension simulation method and device, equipment and storage medium
By combining high-order graph attention networks and vector cellular automata, the problem of insufficient urban expansion simulation accuracy is solved, precise capture of complex spatial relationships and adaptive simulation of urban expansion processes is achieved, and the scientific nature of urban planning and land use efficiency are improved.
Patent Information
- Application Number
- CN202510400783.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-01
- Publication Date
- 2025-08-22
AI Technical Summary
The existing urban expansion simulation methods have insufficient accuracy, especially the machine learning model has weak interpretability and relies on a large amount of labeled data. The combination of traditional methods and machine learning has not yet effectively improved the simulation ability of spatial interactions, and lacks the capture of complex spatial relationships.
Combining the advanced graph attention network and vector cellular automata, the interaction target is selected through neighborhood configuration, features are extracted, graph structure is constructed, and a self-attention mechanism is introduced, the interaction weight is calculated, and the vector cellular automata is integrated for urban expansion simulation to capture spatial heterogeneity and remote dependencies.
It improves the accuracy of urban expansion simulation, can adaptively learn complex models, simulate expansion dynamics under different policies and environmental conditions, provide scientific decision-making basis, and optimize urban planning and land use efficiency.
Smart Images

Figure CN120524550A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of urban simulation technology, and in particular to a method, device, equipment and storage medium for simulating urban expansion. Background Art
[0002] Currently, traditional methods for simulating urban expansion include statistical regression, system dynamics, cellular automata (CA), and multi-agent systems (MAS). Specifically, statistical regression uses historical data to establish statistical relationships and analyze the drivers of urban expansion (such as population growth, economic indicators, and policies). For example, multiple linear regression and logistic regression can be used to quantify the impact of drivers on urban expansion and predict future land use changes based on land type transition probabilities. However, these methods lack spatial explicitness and are difficult to simulate complex spatial interactions. System dynamics treats cities as complex systems composed of subsystems such as population, economy, and environment, analyzing the dynamic relationships between variables through feedback loops. Cellular automata divide space into grid cells and simulate urban expansion based on neighborhood states and transition rules (such as suitability and constraints). Multi-agent systems drive urban expansion from the bottom up by simulating the decision-making and interactions of "agents" such as residents, governments, and businesses.
[0003] Traditional methods have room for further improvement in the accuracy of urban expansion simulation. With the advancement of big data and computing power, machine learning models have been introduced into urban expansion simulation. Machine learning models include:
[0004] Random forest, support vector machine: for land conversion probability prediction.
[0005] Convolutional Neural Network (CNN): Extract spatial features from remote sensing images and predict urban expansion hotspots.
[0006] Generative Adversarial Networks (GANs): Generating high-resolution future urban morphology scenarios.
[0007] The aforementioned machine learning models suffer from weak interpretability and rely on large amounts of labeled data. Therefore, combining traditional methods with machine learning is one of the current technical approaches to improve the accuracy of urban expansion simulations. Currently, existing technologies combine CA, MAS, machine learning, and traditional statistical methods to address the shortcomings of single models. For example, the CA-MAS coupled model: CA simulates spatial expansion and MAS simulates agent decision-making; deep learning-driven CA: neural networks replace manually defined transformation rules. However, the combination of vector cellular automata (VCA) and high-order graph attention networks to achieve urban expansion simulation has not yet been reported in existing technologies. Summary of the Invention
[0008] The present application provides a city expansion simulation method, device, equipment and storage medium to solve the above technical problems.
[0009] In a first aspect, the present application provides a method for simulating urban expansion, comprising:
[0010] Selecting interactive targets based on neighborhood configuration; wherein the neighborhood configuration includes two interactive methods: buffer-based and graph-based to find neighborhood attributes of target plots. The neighborhood attributes include the size of the neighborhood radiation range, the number of plots, the radiation intensity, and the interactive influence of the plots;
[0011] Extracting features of the interactive target to obtain features of the interactive target; wherein the features of the interactive target include land use category, spatial variable features, and non-spatial variable features;
[0012] Constructing a graph structure based on the features of the interaction targets; wherein the graph structure includes edges and nodes, the nodes include the features of the interaction targets, and the edges include adjacency relationships between adjacent interaction targets;
[0013] Based on the graph structure, a high-order graph attention network is constructed. During the training phase of the high-order graph attention network, updated node feature representations are generated through weighted aggregation of neighboring nodes. The model is optimized using a cross-entropy loss function to output a suitability score for land use change of the plot. After the high-order graph attention network is constructed, a self-attention mechanism is introduced through the graph attention network to calculate the interaction weights between neighboring nodes to capture spatial heterogeneity and long-range dependencies.
[0014] A vector cellular automaton is integrated into the high-order graph attention network to obtain an HGAT-VCA model, and urban expansion simulation is implemented based on the HGAT-VCA model; wherein the vector cellular automaton is used to calculate the overall transition probability and determine the land use type.
[0015] In one possible design, the land use categories include cultivated land, forest land, grassland, water bodies, built-up land, and unused land; the spatial variable features are features constructed for each interactive target based on the attributes of the driver dataset of the target area, wherein the driver dataset attributes include neighborhood richness, terrain, and neighborhood accessibility; and the non-spatial variable features include geometric attributes of the interactive target, wherein the geometric attributes include the perimeter and / or area of the plot.
[0016] In one possible design, after building a high-order graph attention network, a self-attention mechanism is introduced through the graph attention network to calculate the interaction weights between neighboring nodes to capture spatial heterogeneity and long-range dependencies, including:
[0017] Select a node i in the graph structure as the central node, and calculate the attention coefficient e between the central node i and the neighbor node j by the following formulaij :
[0018]
[0019] Where: Represents the input features of the central node i in the lth layer; represents the input features of the neighbor node j in the lth layer, F is the feature dimension, represents the feature weight of the central node i in the lth layer, Represents the feature weight of the neighbor node j in the lth layer. When l = 0, the initial input of the high-order graph attention network is the constructed feature matrix X. The initial input in, Represents the first, second and Nth initial input features, through linear transformation Convert F-dimensional features into F'-dimensional features to enhance the expression ability and highlight the differences between nodes. ∥ represents the connection of features. It is a learnable vector used to calculate the attention coefficient. By activating the LeakyReLU function, the attention coefficient e is obtained. ij ;
[0020] Use the softmax function to normalize the attention coefficients between neighbor nodes and obtain the normalized attention coefficient a ij ;
[0021] Perform neighborhood aggregation and normalize the attention coefficient a ij Assigned to each neighbor node to update the characteristics of the central node, the aggregation result is expressed as:
[0022]
[0023] Where: represents the aggregation result, σ is the sigmoid activation function, and VN(i) is the number of neighbor nodes.
[0024] In one possible design, the self-attention mechanism introduced by the graph attention network is repeatedly executed multiple times in parallel, and the results of the self-attention mechanism repeated multiple times in parallel are combined to generate the final attention coefficient.
[0025] In a possible design, based on the vector cellular automaton, the overall transition probability is calculated using the following formula:
[0026]
[0027] Where: represents the overall transition probability of cell i changing to land use type s at time t+1, represents the transfer suitability of cell i to land use type s at time t, which is predicted by a high-order graph attention network; represents the constraint coefficient of cell i at time t. If cell i is developable, it is assigned a value of 1, otherwise it is assigned a value of 0, which is used to reflect geographical or physical constraints. RA is a random factor representing unknown disturbances, which is used to introduce randomness and simulate the impact of contingency and uncertainty in reality. The random factor is expressed as:
[0028] RA=1+(-lnγ) α
[0029] Where: γ is a random value between 0 and 1, α is a parameter from 1 to 10;
[0030] After calculating the overall transition probability of plot i changing to land use type s at time t+1, the land use status update is completed through the following steps:
[0031] Find the land use type with the highest conversion probability value among the transfer suitability of the node to all land use types S;
[0032] Based on the preset threshold, the highest conversion probability value is compared with the preset threshold to determine whether to perform a state transition. If the conversion probability value is higher than the preset threshold, the land use category in the node is converted to the corresponding land use type S; if the conversion probability value is lower than the preset threshold, the node remains in its current state.
[0033] The preset threshold is dynamically adjusted according to the historical urban land use data and the total changed area from time t to t+1 to ensure that the simulation results can reflect the actual land use changes.
[0034] In one possible design, after obtaining the HGAT-VCA model, the method further includes:
[0035] The HGAT-VCA model is evaluated for accuracy using evaluation indicators, wherein the evaluation indicators include fuzzy overlap, producer accuracy, user accuracy and / or overall accuracy.
[0036] In a second aspect, the present application provides an urban expansion simulation device, comprising:
[0037] A target selection module is configured to select an interactive target based on a neighborhood configuration; wherein the neighborhood configuration includes two interactive methods: a buffer-based method and a graph-based method to find neighborhood attributes of the target plot. The neighborhood attributes include the size of the neighborhood radiation range, the number of plots, the radiation intensity, and the interactive influence of the plots.
[0038] a feature extraction module configured to extract features from the interactive target to obtain features of the interactive target; wherein the features of the interactive target include land use categories, spatial variable features, and non-spatial variable features;
[0039] A graph structure construction module is configured to construct a graph structure based on the features of the interaction targets; wherein the graph structure includes edges and nodes, the nodes include the features of the interaction targets, and the edges include adjacency relationships between adjacent interaction targets;
[0040] A network construction module is configured to construct a high-order graph attention network based on the graph structure; wherein, during a training phase of constructing the high-order graph attention network, an updated node feature representation is generated by weighted aggregation of neighboring nodes, the model is optimized using a cross-entropy loss function, and a suitability score of land use change of a plot is output; after the high-order graph attention network is constructed, a self-attention mechanism is introduced through the graph attention network to calculate interaction weights between neighboring nodes to capture spatial heterogeneity and long-range dependencies;
[0041] An expansion simulation module is configured to integrate a vector cellular automaton into the high-order graph attention network to obtain an HGAT-VCA model, and implement urban expansion simulation based on the HGAT-VCA model; wherein the vector cellular automaton is used to calculate the overall transition probability and determine the land use type.
[0042] In a third aspect, an embodiment of the present application provides an electronic device comprising: at least one processor and a memory; the memory stores computer-executable instructions; the at least one processor executes the computer-executable instructions stored in the memory, so that the at least one processor executes the urban expansion simulation method described in the first aspect and various possible designs of the first aspect.
[0043] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, in which computer execution instructions are stored. When a processor executes the computer execution instructions, the urban expansion simulation method described in the first aspect and various possible designs of the first aspect is implemented.
[0044] In a fifth aspect, an embodiment of the present application provides a computer program product, including a computer program. When the computer program is executed by a processor, it implements the urban expansion simulation method described in the first aspect and various possible designs of the first aspect.
[0045] This application combines the advantages of the graph attention mechanism (GAT) and vector cellular automaton (VCA) in the graph neural network (GNN). The fusion of topological structure and spatial evolution mechanism not only improves the accuracy of urban expansion simulation, but also provides a valuable reference for the innovation and development of related theoretical models. It can simulate and analyze the expansion of urban construction land from a multi-scale and multi-level perspective, and can also reveal the complex process of urban expansion and the expansion path under different scenarios, enriching the existing theoretical models of urban expansion. In particular, the high-order graph attention network can process large-scale urban spatial data based on multi-order neighbor relationships, providing theoretical support for more accurate spatial analysis and expansion prediction. The combination of high-order graph attention network and vector cellular automaton provides a new idea for urban expansion research. Unlike traditional rule-based expansion models, this method combines deep learning and the dynamics of cellular automata, which can not only adaptively learn the complex patterns in the urban expansion process, but also simulate the expansion dynamics under different policy and environmental conditions, promoting the innovative development of urban expansion theory.
[0046] This application can accurately predict the expansion of construction land in the target urban area of the study area, provide scientific decision-making basis for local governments, and rationally plan the future development direction and spatial layout of the city, thereby improving land use efficiency and resource allocation efficiency. At the same time, urban expansion not only involves the spatial configuration of land use, but is also closely related to ecological environmental protection, transportation network construction, social resource allocation and other aspects. By simulating urban expansion paths under different scenarios, this application can provide urban planning departments with expansion characteristics and impact analysis under different policy scenarios, and help formulate urban expansion strategies that comply with the principles of sustainable development. In addition, the functional changes of each plot during the expansion process are predicted and optimized. Through simulation analysis of the expansion potential of different regions, this application can provide data support for land use planning, optimize urban functional layout, and improve the city's spatial efficiency and comprehensive service capabilities, especially in the demarcation of important functional areas such as transportation, commerce, and residence. BRIEF DESCRIPTION OF THE DRAWINGS
[0047] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.
[0048] Figure 1 A schematic diagram of the research area provided for the embodiments of this application;
[0049] Figure 2 A flow chart of a city expansion simulation method provided in an embodiment of the present application;
[0050] Figure 3 Schematic diagram of the feature extraction process of plots in the central urban area of the study area provided in the embodiment of this application;
[0051] Figure 4 A schematic diagram of graph topology construction provided in an embodiment of the present application;
[0052] Figure 5 A schematic diagram of establishing a neighborhood based on a buffer zone provided in an embodiment of the present application;
[0053] Figure 6 A schematic diagram of a topological neighborhood established based on a graph structure provided in an embodiment of the present application;
[0054] Figure 7 The graph structure of the central urban area of the study area provided in the embodiment of the present application is based on different neighborhood configurations;
[0055] Figure 8 Schematic diagram of the integration of the graph attention network and the vector cellular automaton model provided in an embodiment of the present application;
[0056] Figure 9 The HGAT model training loss function and accuracy function curve based on the 4th-order graph topology provided in the embodiment of the present application;
[0057] Figure 10 The actual land use distribution map of the central urban area of the study area in 2023 and the land use distribution map based on different models are provided in the embodiments of this application;
[0058] Figure 11 Provides visualization results of the attention coefficient of the selected plot for the embodiment of this application;
[0059] Figure 12 A schematic diagram of the structure of the urban expansion simulation device provided in an embodiment of the present application;
[0060] Figure 13 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application.
[0061] The above drawings illustrate specific embodiments of the present application, which will be described in more detail below. These drawings and the textual description are not intended to limit the scope of the present application in any way, but rather to illustrate the concepts of the present application to those skilled in the art by reference to specific embodiments. DETAILED DESCRIPTION
[0062] Exemplary embodiments will be described in detail herein, with examples illustrated in the accompanying drawings. In the following description, when referring to the drawings, identical numerals in different figures represent identical or similar elements, unless otherwise indicated. The embodiments described in the following exemplary embodiments are not intended to represent all embodiments consistent with the present application. Rather, they are merely examples of apparatus and methods consistent with certain aspects of the present application, as detailed in the appended claims.
[0063] In the technical solution of this application, the collection, storage, use, processing, transmission, provision and disclosure of information such as financial data, user data or medical image data involved shall comply with the provisions of relevant laws and regulations and shall not violate public order and good morals.
[0064] It should be noted that in the embodiments of the present application, certain software, components, models and other existing solutions in the industry may be mentioned. They should be regarded as exemplary. Their purpose is only to illustrate the feasibility of implementing the technical solution of the present application, but it does not mean that the applicant has or will necessarily use the solution.
[0065] It should be noted that in the embodiments of the present application, certain software, components, models and other existing solutions in the industry may be mentioned. They should be regarded as exemplary. Their purpose is only to illustrate the feasibility of implementing the technical solution of the present application, but it does not mean that the applicant has or will necessarily use the solution.
[0066] The following specific embodiments describe in detail the technical solution of the present application and how the technical solution of the present application solves the above-mentioned technical problems. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. The embodiments of the present application will be described below in conjunction with the accompanying drawings.
[0067] This embodiment of the application provides a method for simulating urban expansion. First, this embodiment provides an exemplary research area and data as an application scenario of the urban expansion simulation method. Figure 1 As shown in Figure 2, the central urban area of the selected study area is located between 106°12′36″-106°59′58″E longitude and 29°7′45″-30°7′22″N latitude, with an area of 5466km 2 , accounting for 6.6% of the total land area of the study area. These areas are the key areas for economic and social development in the study area and are among the fastest-growing areas in the study area.
[0068] Data sources: The datasets used in this example include urban plot data, remote sensing images and land use data, DEM, POIs, and road network vector data. The plot data is a nationwide plot-scale urban land use dataset provided by the paper "Mapping of Basic Urban Land Use Types in China: Preliminary Results in 2018" published in Science Bulletin in 2019 (Gong et al., 2019). The land use image data was obtained from the Institute of Geographic Sciences and Natural Resources Research, Chinese Academy of Sciences. The acquisition year and range include the study area in 2018 and the Yangtze River Basin in 2020 and 2023, with a resolution of 30 meters. In addition, DEM data was obtained from the National Basic Geographic Information Center with a resolution of 30 meters. Vector datasets, including POIs and road networks in 2015 and 2020, were downloaded from the Resources and Environmental Science and Data Center.
[0069] Combining the above research areas and data, the specific implementation principles and implementation effects of this urban expansion simulation method will be introduced in detail below.
[0070] Figure 2 This is a flow chart of the urban expansion simulation method provided in the embodiment of the present application. Figure 2 As shown, the city expansion simulation method can be implemented based on an electronic terminal, and the city expansion simulation method includes the following steps S10 to S50.
[0071] S10. Selecting an interactive target based on neighborhood configuration; wherein the neighborhood configuration includes two interactive methods: buffer-based and graph-topology-based to find the neighborhood attributes of the target plot. The neighborhood attributes include the size of the neighborhood radiation range and the number of plots, as well as the radiation intensity and the interactive influence of the plots.
[0072] The selection of the interaction target mainly depends on the configuration of the neighborhood, which determines which entities will be neighbors of the target entity and may interact with it. Neighborhood configuration can be roughly divided into two categories: buffer-based and topology-based (e.g. Figure 5 、 Figure 6 As shown in Figure 2, the neighborhood attributes of the research plots are found using two interactive methods: buffer-based and graph-based. Specifically, the neighborhood radiation range size (number of plots) and radiation intensity (interactive influence of plots) are used.
[0073] In this embodiment, a buffer zone was generated based on the boundary of the study plot, and the step size of the buffer distance was set to 5 meters. First, using the arcpy and os libraries in ArcGIS, a series of boundary-based buffer zones were generated for each plot in the Python 2.7 environment. The distance of the buffer zone was gradually expanded, increasing by 5 meters each time. Next, for each generated buffer zone, a spatial overlay analysis was performed to collect all plots that intersected with the buffer zone, and these plots were regarded as neighbors of the study plot. Specifically, the arcpy.Intersect_analysis method was used to perform intersection operations on plots at different buffer distances to determine which plots intersected with the buffer zone. For each intersection result, the ID of the central plot and the ID of its intersecting neighboring plots were recorded. Finally, the ID information of all these neighbor node pairs was organized and saved as CSV format files to facilitate subsequent data processing and analysis. This process provides detailed spatial information for adjacency modeling and supports further urban expansion simulation and analysis.
[0074] The process of constructing topologically based multi-order neighbor relationships utilized the Python 2.7 environment included with ArcGIS, and combined with the arcpy and os libraries in PyCharm for data processing and result storage. A maximum buffer distance of 60 meters ensured complete coverage of street gaps, thereby comprehensively capturing the neighborhood relationships of each parcel. This process involved multiple steps, including buffer calculation, neighbor relationship extraction, and data recording and storage, as described below. First, for each parcel in the input buffer parcel vector data, its neighborhood scope included first-order neighbors that shared a boundary or vertex with it. This was then expanded to include second-order neighbors that were connected to first-order neighbors but not directly connected to the central parcel. This process was then recursively repeated to obtain third-order and higher-order neighbors. This recursive process relied on dynamic buffer overlay and spatial overlay analysis, combined with topological relationship filtering rules, to ensure that the generated neighborhood relationships had a clear order hierarchy. In practice, when traversing all parcels, each parcel was treated as a "center parcel" or "center node," and its neighbor relationships were calculated and recorded one by one. To ensure data structure and clarity, each record contains the following fields: the center parcel (or center node) ID, the neighborhood order, and a list of neighboring parcels (or neighboring nodes) of the corresponding order. All records are ultimately organized into a CSV file with the headers center_ID, order, and neighbor_ID to facilitate subsequent model calls.
[0075] S20. Extract features of the interactive target to obtain features of the interactive target; wherein the features of the interactive target include land use categories, spatial variable features, and non-spatial variable features.
[0076] For interactive target feature extraction, the land use type corresponding to the centroid of each parcel was extracted from the grid cells of the corresponding land use map. Six land use categories were identified, including cultivated land, forest land, grassland, water bodies, built-up land, and unused land. These six land use types were used as the primary land use category attributes for the parcels. Simultaneously, the driver factor dataset collected in the study area was combined with the parcel data to construct 16 features (M = 16) for each parcel, including two nonspatial variables and 14 spatial variables. The specific parcel feature attributes are shown in Table 1. For the neighborhood richness variable, the ArcGIS kernel density tool was used to calculate the area ratio of the variable in different neighborhood states, i.e., the density value, for variables X1-X6. Furthermore, geometric attributes of the study parcels were considered, including the parcel area X7 and perimeter X8. Topographic spatial variables, including the study area elevation X9 and slope X10, were then used to measure the parcel's suitability for development. Finally, the road network and point of interest (POIs) were used to calculate the accessibility of the study parcels (X11-X16). To avoid scale effects, Min-Max normalization is used to map all values of plot features to [0, 1].
[0077] Table 1. Land plot feature attributes used in this example
[0078]
[0079] By extracting driving factor characteristics by plot, the continuous variables in raster data are combined with the discrete units of vector plots to provide accurate input data for subsequent model training and analysis. This process effectively integrates spatial information and attribute characteristics and is an important basic step in urban expansion simulation. Figure 3 Specific steps are as follows: ensure that the projection of the driving factor raster data is consistent with the plot vector data, resample the inconsistent raster data and unify the raster resolution to 30m; calculate the average value of the raster pixel value within the plot to represent the overall level of the driving factor; associate the extracted eigenvalues with the plot vector data, use the unique identification field (ID) of the plot as the primary key, store the extracted eigenvalues in the plot attribute table, and output the results in the form of a CSV file or an attribute table of the plot vector file, which contains the driving factor eigenvalues of each plot; check and verify the extracted eigenvalues to ensure that the extraction is complete.
[0080] S30. Construct a graph structure based on the features of the interaction targets; wherein the graph structure includes edges and nodes, the nodes include the features of the interaction targets, and the edges include the adjacency relationship between adjacent interaction targets.
[0081] The ultimate goal of interactive modeling is to construct a graph structure. This graph structure contains two key pieces of information: edges and nodes, which are also important inputs for subsequent network models. Edges contain the connectivity of neighbors, that is, the adjacency between plots. Nodes contain spatial features, geometric features, or other attributes, that is, the various attribute characteristics of the plots. All data (topological CSV files, plot attribute data (ID) and geometric attributes, land use raster data, and driver factor raster data) are preprocessed and normalized. Node features are extracted and encoded, and edge topology information is added to the graph structure to construct a plot graph, resulting in G = (V, E, X, A), where each node represents a plot, and adjacent plots are represented by an edge connecting them.
[0082] Relying on the networkx library, pandas library, and matplotlib library of Python 3.6, combined with topology and feature matrices, a graph structure for simulating urban expansion was created. First, the CSV file storing topological information and the node feature matrix were read through the pandas library. These data contain the adjacency relationship of the plots and the land use attributes of each plot. Based on the topological information, an undirected graph (e.g., Figure 7 ), where nodes represent different plots and edges represent the adjacency relationship between plots.
[0083] S40. Based on the graph structure, a high-order graph attention network is constructed; wherein, in the training phase of constructing the high-order graph attention network, an updated node feature representation is generated by weighted aggregation of neighbor nodes, the model is optimized using the cross-entropy loss function, and the suitability score of the land use change of the plot is output. After the high-order graph attention network has been constructed, a self-attention mechanism is introduced through the graph attention network to calculate the interaction weights between neighbor nodes to capture spatial heterogeneity and long-range dependencies.
[0084] In this embodiment, based on the graph structure, a HGAT network (high-order graph attention network) is constructed. After the graph structure has been constructed, the self-attention mechanism is introduced through the graph attention network (GAT) to calculate the interaction weights (attention coefficients e) between neighbor nodes. ij ) to capture spatial heterogeneity and long-range dependencies. During the training phase, an updated node feature representation is generated through weighted aggregation of neighboring nodes. The model is optimized using the cross-entropy loss function to output the suitability score of land use change for each plot.
[0085] The suitability of land use change at the central node is affected by its own driving factors and the surrounding plots. The transfer probability of the central node aggregates the state characteristics of its neighboring nodes and its own state characteristics. If a neighboring node has a stronger interaction with the central node, its influence on the central node will be greater, thus occupying a higher weight in the aggregation. This weight is the attention coefficient calculated by the graph attention mechanism. The attention coefficient e between the central node i and the neighboring node j is ij It can be expressed as:
[0086]
[0087] Where: Represents the input features of the central node i in the lth layer; represents the input features of the neighbor node j in the lth layer, and F is the feature dimension. When l = 0, the initial input of HGAT is the constructed feature matrix X, that is, in, Then, by linear transformation (i.e. ), ) converts F-dimensional features into F'-dimensional features to enhance the expressive power and highlight the differences between nodes. In addition, ∥ represents the connection of features. Is a learnable vector used to calculate the attention coefficient. By activating the function "LeakyReLU", the attention coefficient e can be obtained ij Then, the “softmax” function is used to normalize the attention coefficients between neighbor nodes. Finally, the normalized attention coefficient a ij It can be regarded as the quantitative influence received by the central node i from the neighbor node j, so the attention coefficient a ij It can be expressed as:
[0088]
[0089] Next, neighborhood aggregation is performed and the normalized attention coefficient a is ij Assigned to each neighbor node to update the characteristics of the central node. The final aggregation result is given by Output, where σ is the sigmoid activation function.
[0090]
[0091] All the steps mentioned in the above formula constitute the actual implementation of a graph attention operation, but a single operation may not be able to fully detect diverse interactions and may ignore some important neighbors. Therefore, the graph attention mechanism is repeated multiple times in parallel, each time called a separate attention head. These similar graph attention operations are then combined to generate the final attention coefficient. This process is called a multi-head attention operation, such as Figure 8 shown.
[0092] In this embodiment, the HGAT model was established using the PyTorch framework, which mainly consists of two graph attention layers. The attention layer of the first graph contains 8 attention heads and 8 hidden layers, calculating the 8-dimensional features of each head, while the second graph contains one attention head, collecting 64-dimensional features of the node, and outputting 6 land use categories. During the HGAT model training process, the learning rate was set to 0.0001, the training period was 10000, the early stopping parameter was set to 100, and the regularization coefficient was set to L1. 80% of the sample data from 2018 to 2020 was used as training data, and the remaining 20% was used as validation data to evaluate the performance of the HGAT-VCA model. Then, the data from 20020 to 2023 was used to test the stability of the proposed model.
[0093] S50. Integrate the vector cellular automaton into the high-order graph attention network to obtain the HGAT-VCA model, and implement urban expansion simulation based on the HGAT-VCA model; wherein, the vector cellular automaton is used to calculate the overall transition probability and determine the land use type.
[0094] In a typical cellular automation (CA) model, the transition rules defining each cell's state typically include five components: the cell state at the current time step—the cell's initial characteristics or category information; land transfer suitability—reflecting the cell's adaptability to different land use types, typically calculated from external models or data; the states of neighboring cells—representing the characteristics of the spatial neighborhood directly or indirectly related to the cell, reflecting neighborhood effects; and change constraints—including constraints on land use transitions such as policies, plans, or geographic restrictions, as well as related random perturbations—used to simulate the impact of unknown or accidental factors on land use change in actual processes. The vector cellular automaton (VCA) mechanism works similarly, except that the CA's research object is changed to vector plots.
[0095] (1) Calculation of overall transition probability
[0096] In this embodiment, since the current cell state and the states of neighboring cells have been input into the proposed HGAT model, and the neighborhood dependencies are learned through the high-order graph attention mechanism, the transition probability OP in VCA can be expressed as:
[0097]
[0098] in, represents the transfer suitability of cell i to land use type s at time t, which is predicted by HGAT; represents the constraint coefficient of cell i at time t. If cell i is developable, it is assigned a value of 1, otherwise it is assigned a value of 0. It is used to reflect policy, geographical or physical constraints. RA is a random factor representing unknown disturbances, which is used to introduce randomness and simulate the impact of contingency and uncertainty in reality. The random factor can be expressed as:
[0099] RA=1+(-lnγ) α
[0100] Where γ is a random value between 0 and 1, and α is a parameter ranging from 1 to 10.
[0101] (2) Determination of land use type
[0102] When calculating the overall transition probability of plot i to land use type s at time t+1 Finally, use the following steps to complete the land use status update:
[0103] Select the highest probability value: Find the land use type with the highest probability among all land use types S of the land parcel's transfer suitability.
[0104] Comparison with Threshold: The highest transition probability is compared with a preset threshold to determine whether to transition. If the transition probability is higher than the threshold, the cell transitions to the corresponding land use type S; if it is lower than the threshold, the cell remains in its current state.
[0105] Dynamically adjust the threshold: The threshold is dynamically adjusted based on the historical urban land use data and the total changed area from time t to t+1 to ensure that the simulation results can reflect the actual land use changes.
[0106] After step S50, the urban expansion simulation method further includes step S60, accuracy verification and comparison of the integrated model.
[0107] Because urban expansion is a long-term process, and most urban land use remains constant over short periods of time, using only the confusion matrix to evaluate the model may lead to overestimation of simulation results. Therefore, to quantitatively assess the accuracy of land use change simulation results, this study introduced multiple evaluation metrics, including FoM (Fuzzy Overlap Metric), PA (Producer's Accuracy), UA (User's Accuracy), and OA (Overall Accuracy), focusing on the accuracy of land use change during the simulation. These metrics quantify the model's predictive performance from different perspectives and are calculated in conjunction with the confusion matrix. A FOM value greater than 2.6 is considered to indicate high model stability and accuracy.
[0108] The proposed HGAT-VCA model is compared with four classic frameworks: LR-VCA, SVM-VCA, RF-VCA, and MLP-VCA. These classic frameworks are constructed similarly to HGAT-VCA, but their transition suitability is calculated using LR (logistic regression), SVM (support vector machine), RF (random forest), and MLP (multi-layer perceptron), respectively. The specific settings are as follows: LR-VCA: The inverse of the regularization strength is set to 1.0, and an L2 regularization penalty is used. SVM-VCA: The radial basis function (RBF) is used, and the inverse of the regularization strength is also set to 1.0. RF-VCA: 100 decision trees are used, and entropy is used as the partitioning quality metric. MLP-VCA: Two hidden layers are used, with 32 and 64 neurons, respectively. The activation function is ReLU, and the learning rate is set to 0.001.
[0109] Based on the processing of the above steps and the application to the study area, the effect of the urban expansion simulation method in actual application scenarios will be described in detail below.
[0110] Regarding the interactive modeling results for urban expansion simulations, the primary purpose of interactive modeling is to enhance the ability to express spatial relationships in urban expansion simulations, revealing the complex interactions between adjacent and distant plots and significantly improving model accuracy. This lays a solid foundation for understanding urban expansion processes and developing rational urban planning schemes. Interactive modeling effectively expresses the spatial interactions between urban plots and quantifies the impact of different neighborhoods on target plots.
[0111] The buffer-based neighborhood determination, whether based on centroid or boundary, has been repeated many times to determine the buffer distance. From a quantitative point of view (Table 2), when the buffer distance (buffer) is 5m, there are no neighbor nodes (the influence of road width); traversing the plots, when buffer = 8m, there are more than 2000 pairs of neighbors; visual observation, there are still a considerable number of plot buffers that do not intersect, so the buffer distance is enlarged, when buffer = 30m, there are more than 6000 pairs of neighbors; continue to increase the buffer distance gradually with a step size of 5m, and find that the buffer distance is between 30-60m, and the number of neighbors increases steadily until buffer = 65m, the number of neighbor pairs is less than when buffer = 60m. The buffer result is confirmed by visual inspection, as shown in the figure below. Figure 5 , it can be seen that when buffer = 35m, almost all neighbors have a connection relationship, except for some plots near wider roads; when buffer = 55m, all adjacent plots in the study area have an adjacency relationship.
[0112] Table 2. The number of neighbors in topological relationships established based on different buffer distances
[0113]
[0114] Neighborhood configuration based on graph topology: Based on the construction of neighborhood topology of graph structure, the connection relationship between the central plot and the high-order plot is realized through the operation of the adjacency matrix. The core of high-order neighborhood expansion is to iteratively search the neighborhood of the neighborhood according to the original adjacency matrix A, and collect all the selected neighborhoods to establish a new high-order adjacency matrix. The calculation method is:
[0115]
[0116] k-order (k=1,2,3…) adjacency matrix A K , represents the neighborhood connected to the central plot through k hops. K This can be done by multiplying the matrix A by itself k-1 times. 1 =A,A 2 =A*A,..., like Figure 6 As shown, the first-order neighbors defined by the adjacency matrix A of plot 1 are plots 2-7, its second-order neighbors are plots 8-16, and its third-order neighbors are plots 17-28. Finally, the k-order adjacency matrix A is obtained. K , which is defined to include all neighbors from order 1 to order k, which will aggregate all adjacency matrices at each order.
[0117] Using a buffer vector plot map with buffer=60m, we established graph topology data with k=4. After multiple verifications, we found that there are 278,277 neighbor pairs in the fourth order.
[0118] Feature engineering and adjacency matrix construction:
[0119] We comprehensively extracted and encoded the graph node features of 2,294 land parcels, generating a graph node feature matrix X and label one-hot encodings (for six land use categories) to describe the parcel attributes and classification information. Furthermore, to characterize the spatial relationships between parcels, we constructed an adjacency matrix A to reflect the topological connectivity between them. Some of the results from these data structures are shown in Tables 3 and 4, respectively.
[0120] The feature matrix X contains 16-dimensional attribute features for each parcel, including driving factors, geometric characteristics, and land use type, providing rich input data for the graph neural network model. Label one-hot encoding converts the parcel classification information into a format suitable for processing by machine learning and deep learning models. Each land use category corresponds to a set of one-hot vectors, ensuring that the classification task can accurately identify and learn the different characteristics of different types of parcels.
[0121] Table 3. Partial results of feature matrix
[0122] ID X1 X2 X3 ... X16 Level_1 Level_2 Level_3 Level_4 Level_5 Level_6 0 0.115761489 0.20628 0.012009608 ... 0.023162 0 0 1 0 0 0 1 0.99427253 0.198259 0.010408327 ... 0.023162 1 0 0 0 0 1 2 0.127501482 0.215042 0.022417935 ... 0.023162 0 0 1 0 0 0 3 0.2223261 0.072127 0.105684549 ... 0.0955 0 0 1 0 0 1 4 0.2223261 0.072127 0.090472378 ... 0.0955 0 0 1 0 0 1 5 0.222819611 0.09617 0.084067255 ... 0073245 1 0 0 0 0 0 ... ... ... ... ... ... ... ... ... ... ... 2292 0.041922208 0.344235 0.06004804 ... 0.046324 1 0 0 0 1 0 2293 0.046870455 0.312551 0.055244196 ... 0.032756 0 0 0 0 1 1
[0123] The adjacency matrix A shows the spatial adjacency relationship between plots in the form of a sparse matrix. Each element of the matrix A i,j Indicates whether plots i and j are directly adjacent, with 1 indicating adjacent and 0 indicating non-adjacent. This topological matrix effectively encodes spatial interaction information, laying the foundation for subsequent graph structure analysis and graph neural network training.
[0124] Table 4. Partial results of adjacency matrix
[0125] ID 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 ... 2293 0 1 1 1 0 0 0 0 0 0 0 0 0 0 0 0 ... 0 1 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 ... 0 2 1 0 1 0 0 0 0 0 0 0 0 0 0 0 0 ... 0 3 0 0 0 1 1 1 1 1 0 0 0 0 0 0 0 0 4 0 0 1 0 1 1 1 1 1 0 0 0 0 0 0 ... 0 5 0 0 0 1 1 1 0 1 0 0 0 0 0 0 0 ... 0 ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... 2293 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 ... 0
[0126] To prevent the number of nodes from being less than the number of plots, resulting in incomplete plot addition to the graph structure and a mismatch in the number of edges, we also verified the number of graph results. This included establishing a topologically based fourth-order neighborhood using a maximum buffer distance of 60 meters, ensuring the adjacency matrix was symmetric, and adding self-loop operations (unit matrix + adjacency matrix). Because the number of neighborhood plots determined based on the buffer varies depending on the buffer distance, the number of nodes based on the buffer was not verified. The verification results are shown in Table 5, which illustrates that the graph structure satisfies the equation: number of nodes + number of isolated nodes = number of plots. Furthermore, the adjacency matrix and vector plot graph were checked, both showing seven isolated nodes per plot, indicating that all nodes participated in the graph construction and no plots were missed. The number of edges (edges connecting a node with its neighbors) + self-loop edges (edges connecting a node with itself) / number of nodes equals the total number of edges, indicating that the number of edges generated multiple times matches, and the graph construction is complete.
[0127] Table 5 Graph structure verification
[0128]
[0129]
[0130] Graph structure construction and visualization:
[0131] In order to better reflect the mutual influence between plots, adjacent edges are added to the graph, and it is ensured that each plot node can establish a connection with other nodes in the graph. In addition, considering that self-loops may represent the influence of the plot on itself, a self-loop operation is specially added to each node. Then, according to the land use attributes of the plot, 6 different colors (red, orange, yellow, green and pink) are assigned to each node, representing different land use types within the city, residential land, commercial land, industrial land, public land and default color. This allows the node color to intuitively reflect the land use of each plot in the visualization of the graph. In order to further enhance the readability and visualization of the graph, the matplotlib library is used to render the graph, and the size of the node, the thickness of the edge and the color are adjusted to highlight the density of the adjacency relationship. Finally, through graphical display, such as Figure 7 , we can clearly see the spatial topological structure between the plots, the distribution of land use types and the connections between the plots.
[0132] Simulation of the urban built-up area in the central urban area of the study area in 2023:
[0133] Model training:
[0134] After training based on buffer neighborhood and high-order graph topology datasets, taking the 4th-order graph topology sample dataset as an example, the model loss converged to 0.2961 and the model accuracy gradually converged to 0.9128. The model performance is good, such as Figure 9 .
[0135] Comparison of simulation results with the baseline model:
[0136] The simulation results obtained by different models are shown in Table 6. The following observations can be made from Table 6:
[0137] The proposed HGAT-VCA model performs best among these VCA models across all comparison metrics. In particular, the HGAT-VCA model achieves the highest score on the Functionality of Mandate (FoM) metric, demonstrating the superiority of our approach in learning transition rules. HGAT-VCA effectively captures the complex spatial dependencies and nonlinear growth patterns of urban expansion, demonstrating superior accuracy and stability compared to other methods during simulation.
[0138] Table 6 Comparison of simulation results of baseline models
[0139] Model OA PA UA FoM LR-VCA 0.7942 0.2342 0.4653 0.1865 SVM-VCA 0.8015 0.2838 0.4734 0.2367 RF-VCA 0.9087 0.4127 0.6055 0.2553 MLP-VCA 0.9202 0.4245 0.5891 0.2874 HGAT-VCA 0.9485 0.5591 0.6829 0.3956
[0140] The LR-VCA model produced the worst simulation results, demonstrating that traditional linear regression (LR) methods are ineffective in predicting nonlinear urban growth. This result suggests that the LR model oversimplifies the process of land use change and fails to capture complex spatial interactions, resulting in weak predictive power, particularly when predicting the dynamics of urban expansion.
[0141] While RF-VCA, MLP-VCA, and SVM-VCA outperform LR-VCA, they still lag behind HGAT-VCA, further demonstrating the importance of deep learning-based methods for simulating dynamic urban growth. While RF-VCA, MLP-VCA, and SVM-VCA can effectively capture nonlinear relationships, they still fail to effectively handle complex spatial topology and long-range dependencies, resulting in lower simulation accuracy than HGAT-VCA.
[0142] To more intuitively demonstrate the differences between the models, we compared the simulation results with the real ground land use data and selected a small part of the three study areas for visual comparison. Figure 10As shown in the figure, the simulation errors are mainly concentrated at the edges of the accumulation areas, indicating that land use shift prediction in these areas is more difficult. However, it is worth noting that the simulation results of LR-VCA and SVM-VCA deviate significantly from the actual ground truth, showing obvious mispredictions. Although RF-VCA and MLP-VCA perform well, there are still many mismatched prediction patches, especially in the rectangular areas on the far left and top of the figure. The prediction errors in these areas are more concentrated, indicating that the prediction capabilities of these models are insufficient in these locations. In contrast, the simulation results of HGAT-VCA have the highest degree of match with the actual ground truth, indicating that HGAT-VCA can better understand the laws of urban expansion through high-order topological learning and attention mechanisms, and accurately capture the spatial dependence and dynamic changes in land use change. Therefore, HGAT-VCA demonstrates the most accurate simulation results, demonstrating its advantages in urban expansion simulation.
[0143] Efficiency of high-order graph attention mechanisms.
[0144] To verify the efficiency of graph attention, a comparative experiment is conducted with the improved HGAT-VCA, which removes the graph attention operation and is named HGNN-VCA. The results are listed in Table 7.
[0145] Table 7 Efficiency verification of image attention
[0146]
[0147] As shown in Table 7, HGAT-VCA achieves higher accuracy than HGNN-VCA at the same topological order, indicating that considering differences in neighborhood interactions can improve prediction accuracy. Furthermore, it was found that both HGAT-VCA and HGNN-VCA achieve the best simulation results for second-order neighborhoods.
[0148] As can be seen from Table 7, as the neighborhood order expands from 1 to 2, FoM (shows a gradually increasing trend, which indicates that when only the nearest neighbors (K = 1) are considered, the model's ability to express spatial interactions may be limited. In other words, if only the first-order neighborhood is used, the model will underestimate the complexity of spatial interactions because it cannot fully integrate neighborhood information from a larger range. However, when the neighborhood order is expanded to a higher order (such as K = 2), more spatial information is introduced, and the model can capture more comprehensive spatial interaction characteristics, thereby improving the accuracy of land use change simulation.
[0149] However, the FoM trend decreases when the neighborhood is further expanded (e.g., K = 3 and K = 4). This is because as the neighborhood range expands, the model begins to introduce some "noisy neighbors" that are unrelated to the target node. The aggregation of features from these "noisy neighbors" makes it difficult for the model to effectively distinguish relevant from irrelevant information, thereby reducing simulation accuracy. This phenomenon shows that while considering higher-order neighborhoods can integrate more information, there is also a trade-off. Beyond this point, the impact of information redundancy and noise outweighs the information gain brought by higher-order neighborhoods.
[0150] Further analysis shows that the HGAT-VCA model has a more pronounced FoM change trend than the HGNN-VCA model, indicating that HGAT-VCA is more sensitive to the neighborhood range and is better able to find a suitable neighborhood configuration. This advantage can be attributed to the attention mechanism in the HGAT model: HGAT assigns a specific attention coefficient to each neighbor node and dynamically adjusts its weight in the feature aggregation process based on the node's relevance. This selective aggregation strategy significantly reduces the impact of "noisy neighbors," allowing the model to maintain high simulation performance even when expanded to higher-order neighborhoods (such as K = 3 and K = 4).
[0151] In contrast, the HGNN-VCA model uses a unified feature aggregation approach, treating all nodes in a neighborhood equally and failing to consider the differences between different nodes. Consequently, the performance of HGNN-VCA degrades significantly with the introduction of more "noisy neighbors." However, HGAT-VCA effectively suppresses the influence of these "noisy neighbors" through its attention mechanism, outperforming HGNN-VCA even with high-order neighborhoods. This difference further demonstrates HGAT-VCA's advantage in capturing spatial interactions.
[0152] In the process of further analyzing land use change, in addition to focusing on the impact of different-order neighborhoods on model performance, we also explored the spatial distribution of attention coefficients at the plot level. To better understand this phenomenon, we selected three plots with IDs 1787, 1438, and 1574 and their fourth-order neighbors, and extracted the attention coefficients of these neighbors for visualization, as shown in the figure below. Figure 11 shown.
[0153] By observing the "Attention Coefficients (a2, b2, c2)" shown in the figure, we find that as the neighborhood order increases, the attention coefficients of most neighbors show a significant attenuation effect. In other words, as the neighborhood expands, the spatial interaction between plots gradually weakens, and neighbors farther from the central plot receive lower attention coefficients.
[0154] Comparison with the "Actual Land Use in 2018 (a3, b3, c3)" further reveals that clustered plots (i.e., plots that are closer together) generally have higher attention coefficients than dispersed plots. This indicates that the spatial interactions between the central plot and its neighbors are more intensive, as reflected by the higher attention coefficients in the HGAT model. This finding confirms the influence of spatial clustering on neighborhood relationships: in spatially denser areas, interactions between plots are generally stronger, and the attention coefficients are relatively higher.
[0155] Furthermore, the spatial distribution of the "attention coefficient" in the figure reveals some unique plots located within the fourth-order neighbors of the central plot, yet possessing higher attention coefficients than their second- or third-order neighbors. In reality, these unique plots are often important urban public service facilities, such as commercial land, hospitals, schools, or administrative buildings. The presence of these facilities attracts a large number of residents to settle nearby, so even at a considerable distance from the central plot, their spatial interactions remain close. This phenomenon further validates the HGAT model's ability to accurately quantify the diversity of interactions between the central plot and its distant neighbors.
[0156] Efficiency of different neighborhood configurations.
[0157] To identify the most appropriate neighborhood configuration, the accuracy changes in HGAT-VCA based on graph topology order adjustment were compared with those based on buffer distance adjustment. The initial adjacency matrix A was derived from the intersection of buffers rather than from adjacency topology. A total of 22 neighborhood configurations were defined, including four topology-based neighborhoods, nine boundary-buffer-based neighborhoods, and nine centroid-based neighborhoods. The buffer distances were increased in steps of 21 meters from 21 to 210 meters. This buffer distance was derived from the average radius of the plots and the optimal buffer distance setting in Section 3.4.1. Specifically, the centroid distance between two adjacent plots should be approximately equal to the sum of the boundary distance between the adjacent plots and the radius of the two plots. The boundary distance between adjacent plots was set according to the minimum intersection distance (10 meters). The radius mode of the plots (20.9789 meters) was calculated based on the perimeter and area mode of the plots (91.54% of the total). However, in practical applications, data primarily originates from the interpretation of remote sensing imagery and exhibits a grid structure. The shortest distance between two unconnected patches is equal to the resolution of the remote sensing imagery. Therefore, when the buffer distance is less than the resolution, due to the limitations of the original data, the first buffer of a newly added patch has no practical effect or significance, and will only overestimate the expansion of all newly added patches. Therefore, it is necessary to set the first buffer distance greater than the resolution of the remote sensing imagery, setting it to 42 meters. These 22 different adjacency matrices A were input into the HGAT-VCA model for simulation, while the input model feature matrix remained unchanged. The simulation results are shown in Table 8.
[0158] Table 8 Simulation results under different neighborhood configurations
[0159]
[0160]
[0161] The results in Table 8 show that the buffer-based neighborhood configuration achieves optimal results when the buffer distance is approximately 105 meters, while the topology-based neighborhood configuration performs best when the topology order is second. To further understand the correlation between these two configurations, we calculated the average distances from the central plot to adjacent plots for different topological orders, which were 61.96 meters, 104.78 meters, 146.03 meters, and 188.88 meters, respectively. This calculation clearly demonstrates the clear correspondence between the three neighborhood configurations in Table 8. For example, the simulation accuracy of the neighborhood configurations with first-order, second-order, third-order, and fourth-order topologies is close to that of the neighborhood configurations with buffer distances of 63 meters, 105 meters, 147 meters, and 189 meters, respectively.
[0162] It is worth noting that for different neighborhood configurations, buffer-based strategies usually require empirical knowledge to determine the optimal neighborhood spacing and often require more repeated experiments to determine the appropriate buffer distance, which means more debugging time. In contrast, topology-based strategies only need to explore discrete topological orders and can determine the optimal topological order through very limited experiments.
[0163] Therefore, using the proposed HGAT model, a topology-based neighborhood configuration strategy offers significant advantages over buffer-based configuration strategies. Not only can it model complex spatial interactions more quickly and conveniently, but it also significantly improves the efficiency of parameter calibration during CA model construction due to the more flexible adjustment of the topological order and the reduced number of experiments. This topology-based strategy effectively improves model accuracy and training efficiency through more refined spatial relationship modeling and fewer parameter adjustments, while reducing the need for excessive reliance on empirical knowledge and debugging time.
[0164] In summary, the purpose of this application is to model spatial interactions during urban expansion. We propose an innovative VCA model based on a high-order graph attention network (HGAT), namely HGAT-VCA, to accurately simulate the urban expansion process. The main innovations of this application can be summarized as follows:
[0165] (1) By organizing irregular cadastral plots into a graph structure, a model combining HGAT and VCA, HGAT-VCA, is proposed, which can accurately model spatial interactions and simulate the urban expansion process.
[0166] (2) In HGAT, high-order neighborhood expansion is used to extract long-range dependencies, and a graph attention mechanism is introduced to adaptively select interaction targets and accurately quantify heterogeneous spatial interactions.
[0167] (3) With the support of HGAT, it is found that by simply adjusting the discrete topological order, it is possible to generate accuracy results similar to those of the traditional buffer-based neighborhood configuration, which significantly improves the efficiency of model calibration.
[0168] (4) The performance of the HGAT-VCA model was verified using a real plot-based land use dataset. The results showed that HGAT-VCA outperformed four classic VCA models and achieved the highest simulation accuracy (e.g., the FoM improvement was approximately 40.9%).
[0169] Figure 12 This is a schematic diagram of the structure of the urban expansion simulation device provided in the embodiment of the present application. This embodiment of the present application also provides an urban expansion simulation device. Figure 12 As shown, the urban expansion simulation device includes:
[0170] Target selection module 1501 is configured to select an interactive target based on a neighborhood configuration. The neighborhood configuration includes two interactive methods: a buffer-based method and a graph-based method to find neighborhood attributes of a target plot. The neighborhood attributes include the size of the neighborhood radiation range, the number of plots, the radiation intensity, and the interactive influence of the plots.
[0171] A feature extraction module 1502 is configured to extract features from the interactive target to obtain features of the interactive target; wherein the features of the interactive target include land use categories, spatial variable features, and non-spatial variable features;
[0172] A graph structure construction module 1503 is configured to construct a graph structure based on the features of the interaction targets; wherein the graph structure includes edges and nodes, the nodes include the features of the interaction targets, and the edges include the adjacency relationship between adjacent interaction targets;
[0173] The network construction module 1504 is configured to construct a high-order graph attention network based on the graph structure; wherein, during the training phase of constructing the high-order graph attention network, an updated node feature representation is generated by weighted aggregation of neighboring nodes, the model is optimized using a cross-entropy loss function, and a suitability score for land use change of the land parcel is output. After the high-order graph attention network is constructed, a self-attention mechanism is introduced into the graph attention network to calculate the interaction weights between neighboring nodes to capture spatial heterogeneity and long-range dependencies.
[0174] The expansion simulation module 1505 is configured to integrate a vector cellular automaton into the high-order graph attention network to obtain an HGAT-VCA model, and implement urban expansion simulation based on the HGAT-VCA model; wherein the vector cellular automaton is used to calculate the overall transition probability and determine the land use type.
[0175] In some embodiments, the land use categories include cultivated land, forest land, grassland, water bodies, built-up land and unused land; the spatial variable features are features constructed for each interactive target based on the attributes of the driving factor dataset of the target area, and the driving factor dataset attributes include neighborhood richness, terrain and neighborhood accessibility; the non-spatial variable features include geometric attributes of the interactive target, and the geometric attributes include the perimeter and / or area of the plot.
[0176] In some embodiments, the network construction module is further configured to:
[0177] Select a node i in the graph structure as the central node, and calculate the attention coefficient e between the central node i and the neighbor node j by the following formula ij :
[0178]
[0179] Where: Represents the input features of the central node i in the lth layer; represents the input features of the neighbor node j in the lth layer, F is the feature dimension, represents the feature weight of the central node i in the lth layer, Represents the feature weight of the neighbor node j in the lth layer. When l = 0, the initial input of the high-order graph attention network is the constructed feature matrix X. The initial input in, Represents the first, second and Nth initial input features, through linear transformation Convert F-dimensional features into F'-dimensional features to enhance the expression ability and highlight the differences between nodes. ∥ represents the connection of features. It is a learnable vector used to calculate the attention coefficient. By activating the LeakyReLU function, the attention coefficient e is obtained. ij ;
[0180] Use the softmax function to normalize the attention coefficients between neighbor nodes and obtain the normalized attention coefficient a ij ;
[0181] Perform neighborhood aggregation and normalize the attention coefficient a ij Assigned to each neighbor node to update the characteristics of the central node, the aggregation result is expressed as:
[0182]
[0183] Where: represents the aggregation result, σ is the sigmoid activation function, and VN(i) is the number of neighbor nodes.
[0184] In some embodiments, the network building module is further configured to repeatedly execute the self-attention mechanism introduced by the graph attention network in a parallel manner multiple times, combine the results of the self-attention mechanism repeatedly executed in parallel multiple times, and generate a final attention coefficient.
[0185] In some embodiments, the extended simulation module is further configured to calculate the overall transition probability based on the vector cellular automaton using the following formula:
[0186]
[0187] Where: represents the overall transition probability of cell i changing to land use type s at time t+1, represents the transfer suitability of cell i to land use type s at time t, which is predicted by a high-order graph attention network; represents the constraint coefficient of cell i at time t. If cell i is developable, it is assigned a value of 1, otherwise it is assigned a value of 0, which is used to reflect geographical or physical constraints. RA is a random factor representing unknown disturbances, which is used to introduce randomness and simulate the impact of contingency and uncertainty in reality. The random factor is expressed as:
[0188] RA=1+(-lnγ) α
[0189] Where: γ is a random value between 0 and 1, α is a parameter from 1 to 10;
[0190] After calculating the overall transition probability of plot i changing to land use type s at time t+1, the land use status update is completed through the following steps:
[0191] Find the land use type with the highest conversion probability value among the transfer suitability of the node to all land use types S;
[0192] Based on the preset threshold, the highest conversion probability value is compared with the preset threshold to determine whether to perform a state transition. If the conversion probability value is higher than the preset threshold, the land use category in the node is converted to the corresponding land use type S; if the conversion probability value is lower than the preset threshold, the node remains in its current state.
[0193] The preset threshold is dynamically adjusted according to the historical urban land use data and the total changed area from time t to t+1 to ensure that the simulation results can reflect the actual land use changes.
[0194] In some embodiments, the urban expansion simulation device further includes an accuracy assessment module, which is configured to perform accuracy assessment on the HGAT-VCA model using evaluation indicators; wherein the evaluation indicators include fuzzy overlap, producer accuracy, user accuracy and / or overall accuracy.
[0195] The urban expansion simulation device provided in the embodiment of the present application can be used to implement the technical solution of the urban expansion simulation method in the above embodiment. Its implementation principle and technical effects are similar and will not be repeated here.
[0196] Figure 13 This is a schematic diagram of the structure of the electronic device provided in the embodiment of the present application. Figure 13 As shown, the electronic device may include: a processor 161 and a memory 162 , wherein the processor 161 and the memory 162 may communicate with each other; illustratively, the processor 161 and the memory 162 communicate with each other via a communication bus 163 .
[0197] The processor 161 executes the computer-executable instructions stored in the memory 162, so that the processor 161 implements the solutions in the above embodiments. The processor 161 can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.
[0198] The communication bus 163 may be a peripheral component interconnect (PCI) bus or an extended industry standard architecture (EISA) bus, etc. The system bus may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, only one thick line is used in the figure, but this does not mean that there is only one bus or one type of bus. The transceiver is used to implement communication between the database access device and other computers (such as clients, read-write libraries, and read-only libraries). The memory may include random access memory (RAM) and may also include non-volatile memory.
[0199] The electronic device provided in the embodiment of the present application may be the terminal device of the above embodiment.
[0200] An embodiment of the present application further provides a computer-readable storage medium, in which computer instructions are stored. When the computer instructions are executed on a computer, the computer executes the technical solution of the urban expansion simulation method of the above embodiment.
[0201] An embodiment of the present application also provides a computer program product, which includes a computer program stored in a computer-readable storage medium. At least one processor can read the computer program from the computer-readable storage medium. When at least one processor executes the computer program, it can implement the technical solution of the urban expansion simulation method in the above embodiment.
[0202] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of modules is only a logical function division. In actual implementation, there may be other division methods, such as multiple modules can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interface, device or module, which can be electrical, mechanical or other forms.
[0203] Modules described as separate components may or may not be physically separate, and components shown as modules may or may not be physical units, that is, they may be located in one place or distributed across multiple network elements. Some or all of these modules may be selected to implement the solution of this embodiment based on actual needs.
[0204] In addition, the functional modules in the various embodiments of the present application may be integrated into a single processing unit, or each module may exist physically separately, or two or more modules may be integrated into a single unit. The above-mentioned modules may be implemented in the form of hardware or hardware plus software functional units.
[0205] The above-mentioned integrated module implemented in the form of a software functional module can be stored in a computer-readable storage medium. The above-mentioned software functional module is stored in a storage medium and includes a number of instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) or a processor to perform some steps of the methods of various embodiments of the present application.
[0206] It should be understood that the processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), or application-specific integrated circuits (ASICs). A general-purpose processor may be a microprocessor or any conventional processor. The steps of the method disclosed in the present invention may be directly implemented by a hardware processor or implemented by a combination of hardware and software modules in the processor.
[0207] The memory may include a high-speed RAM memory, and may also include non-volatile storage NVM, such as at least one disk memory, and may also be a USB flash drive, a mobile hard disk, a read-only memory, a magnetic disk or an optical disk.
[0208] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus. Buses can be classified into address buses, data buses, and control buses. For ease of illustration, the buses in the drawings of this application are not limited to just one bus or just one type of bus.
[0209] The storage medium may be implemented by any type of volatile or non-volatile memory device, or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The storage medium may be any available medium that can be accessed by a general-purpose or special-purpose computer.
[0210] An exemplary storage medium is coupled to a processor so that the processor can read information from the storage medium and write information to the storage medium. Of course, the storage medium can also be an integral part of the processor. The processor and the storage medium can be located in an application specific integrated circuit (ASIC). Of course, the processor and the storage medium can also exist as discrete components in an electronic control unit or a main control device.
[0211] Those skilled in the art will appreciate that all or part of the steps in the above-described method embodiments can be implemented using hardware associated with program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.
[0212] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some or all of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present application.
Claims
1. A method for simulating urban expansion, characterized in that: include: Selecting interactive targets based on neighborhood configuration; wherein the neighborhood configuration includes two interactive methods: buffer-based and graph-based to find neighborhood attributes of target plots. The neighborhood attributes include the size of the neighborhood radiation range, the number of plots, the radiation intensity, and the interactive influence of the plots; Extracting features of the interactive target to obtain features of the interactive target; wherein the features of the interactive target include land use category, spatial variable features, and non-spatial variable features; Constructing a graph structure based on the features of the interaction targets; wherein the graph structure includes edges and nodes, the nodes include the features of the interaction targets, and the edges include adjacency relationships between adjacent interaction targets; Based on the graph structure, a high-order graph attention network is constructed. During the training phase of the high-order graph attention network, updated node feature representations are generated through weighted aggregation of neighboring nodes. The model is optimized using a cross-entropy loss function to output a suitability score for land use change of the plot. After the high-order graph attention network is constructed, a self-attention mechanism is introduced through the graph attention network to calculate the interaction weights between neighboring nodes to capture spatial heterogeneity and long-range dependencies. A vector cellular automaton is integrated into the high-order graph attention network to obtain an HGAT-VCA model, and urban expansion simulation is implemented based on the HGAT-VCA model; wherein the vector cellular automaton is used to calculate the overall transition probability and determine the land use type.
2. The method according to claim 1, characterized in that The land use categories include cultivated land, forest land, grassland, water bodies, built-up land and unused land; the spatial variable features are features constructed for each interactive target based on the attributes of the driving factor dataset of the target area, and the driving factor dataset attributes include neighborhood richness, terrain and neighborhood accessibility; the non-spatial variable features include the geometric attributes of the interactive target, and the geometric attributes include the perimeter and / or area of the plot.
3. The method according to claim 1, characterized in that After building a high-order graph attention network, we introduce a self-attention mechanism through the graph attention network to calculate the interaction weights between neighboring nodes to capture spatial heterogeneity and long-range dependencies, including: Select a node i in the graph structure as the central node, and calculate the attention coefficient e between the central node i and the neighbor node j by the following formula ij : Where: Represents the input features of the central node i in the lth layer; represents the input features of the neighbor node j in the lth layer, F is the feature dimension, represents the feature weight of the central node i in the lth layer, Represents the feature weight of the neighbor node j in the lth layer. When l = 0, the initial input of the high-order graph attention network is the constructed feature matrix X. The initial input in, Represents the first, second and Nth initial input features, through linear transformation Convert F-dimensional features into F'-dimensional features to enhance the expression ability and highlight the differences between nodes. ∥ represents the connection of features. It is a learnable vector used to calculate the attention coefficient. By activating the LeakyReLU function, the attention coefficient e is obtained. ij ; Use the softmax function to normalize the attention coefficients between neighbor nodes and obtain the normalized attention coefficient a ij ; Perform neighborhood aggregation and normalize the attention coefficient a ij Assigned to each neighbor node to update the characteristics of the central node, the aggregation result is expressed as: Where: represents the aggregation result, σ is the sigmoid activation function, and VN(i) is the number of neighbor nodes.
4. The method according to claim 3, characterized in that The self-attention mechanism introduced by the graph attention network is repeatedly executed multiple times in parallel, and the results of the self-attention mechanism repeated multiple times in parallel are combined to generate the final attention coefficient.
5. The method according to claim 1, wherein Based on the vector cellular automaton, the overall transition probability is calculated using the following formula: Where: represents the overall transition probability of cell i changing to land use type s at time t+1, represents the transfer suitability of cell i to land use type s at time t, which is predicted by a high-order graph attention network; represents the constraint coefficient of cell i at time t. If cell i is developable, it is assigned a value of 1, otherwise it is assigned a value of 0, which is used to reflect geographical or physical constraints. RA is a random factor representing unknown disturbances, which is used to introduce randomness and simulate the impact of contingency and uncertainty in reality. The random factor is expressed as: RA=1+(-lnγ) α Where: γ is a random value between 0 and 1, α is a parameter from 1 to 10; After calculating the overall transition probability of plot i changing to land use type s at time t+1, the land use status update is completed through the following steps: Find the land use type with the highest conversion probability value among the transfer suitability of the node to all land use types S; Based on a preset threshold, the highest conversion probability value is compared with the preset threshold to determine whether to perform a state transition. If the conversion probability value is higher than the preset threshold, the land use category in the node is converted to the corresponding land use type S. If the conversion probability value is lower than the preset threshold, the node remains in its current state; The preset threshold is dynamically adjusted according to the historical urban land use data and the total changed area from time t to t+1 to ensure that the simulation results can reflect the actual land use changes.
6. The method according to claim 1, characterized in that After obtaining the HGAT-VCA model, the method further includes: The HGAT-VCA model is evaluated for accuracy using evaluation indicators, wherein the evaluation indicators include fuzzy overlap, producer accuracy, user accuracy and / or overall accuracy.
7. An urban expansion simulation device, characterized in that: include: A target selection module is configured to select an interactive target based on a neighborhood configuration; wherein the neighborhood configuration includes two interactive methods: a buffer-based method and a graph-based method to find neighborhood attributes of the target plot. The neighborhood attributes include the size of the neighborhood radiation range, the number of plots, the radiation intensity, and the interactive influence of the plots. a feature extraction module configured to extract features from the interactive target to obtain features of the interactive target; wherein the features of the interactive target include land use categories, spatial variable features, and non-spatial variable features; A graph structure construction module is configured to construct a graph structure based on the features of the interaction targets; wherein the graph structure includes edges and nodes, the nodes include the features of the interaction targets, and the edges include adjacency relationships between adjacent interaction targets; A network construction module is configured to construct a high-order graph attention network based on the graph structure; wherein, during a training phase of constructing the high-order graph attention network, an updated node feature representation is generated by weighted aggregation of neighboring nodes, the model is optimized using a cross-entropy loss function, and a suitability score of land use change of a plot is output; after the high-order graph attention network is constructed, a self-attention mechanism is introduced through the graph attention network to calculate interaction weights between neighboring nodes to capture spatial heterogeneity and long-range dependencies; An expansion simulation module is configured to integrate a vector cellular automaton into the high-order graph attention network to obtain an HGAT-VCA model, and implement urban expansion simulation based on the HGAT-VCA model; wherein the vector cellular automaton is used to calculate the overall transition probability and determine the land use type.
8. An electronic device, characterized in that: include: a processor, and a memory communicatively connected to the processor; The memory stores computer-executable instructions; The processor executes the computer-executable instructions stored in the memory to implement the method according to any one of claims 1 to 6.
9. A computer-readable storage medium, characterized in that The computer-readable storage medium stores computer-executable instructions, which are used to implement the method according to any one of claims 1 to 6 when executed by a processor.
10. A computer program product, characterized in that The invention comprises a computer program, which implements the method according to any one of claims 1 to 6 when the computer program is executed by a processor.
Citation Information
Cited By
Road network track k neighbor query method and system
CN120910131A
