A land use change simulation method based on graph convolutional neural network and vector cellular automaton

Abstract

The present invention discloses a land use change simulation method based on graph convolutional neural networks and vector cellular automata, comprising the following steps: S1, constructing a graph-structured cellular space, which includes constructing a plot adjacency matrix and a plot feature matrix; S2, extracting cellular transition rules based on a graph convolutional neural network, which includes constructing a graph convolutional neural network and calculating the probability of cellular population transition; S3, simulating historical land use results and evaluating the model; and S4, predicting the future land use pattern of the city. The present invention is directed to a land use simulation method in the current urban development context. By considering the impact of human travel behavior on land function transformation, the method also takes into account the characteristics of distant plots with spatial interactions during the process of exploring plot transformation rules, making the model closer to the real world and thus improving the accuracy of land use change simulation for plots within the city.

Description

Technical Field

[0001] The present invention relates to the technical field of land use change simulation, and in particular to a land use change simulation method based on graph convolutional neural network and vector cellular automaton. Background Art

[0002] Urban simulation involves simulating future urban development based on past urban development patterns and patterns. Urban simulation typically involves simulating population mobility and land use changes. A commonly used model is the cellular automaton model, which can simulate overall urban land use changes from a microscopic perspective. Existing urban simulation models primarily focus on urban expansion and sprawl, assessing the suitability of non-construction land units to determine their probability of conversion to urban land. However, within the current urban development model driven by high-quality and sustainable development strategies, while current urban simulation models have demonstrated good performance in predicting urban expansion, they struggle to simulate changes in land use within cities.

[0003] Existing technologies calculate accessibility and connectivity, and combine them with VCA (vector cellular automata) to simulate land use change, which is very effective in simulating urban expansion. Land development depends on the convenience of transportation. However, cities are currently in a stage of high-quality development. The transportation facilities and urban road networks on the built-up urban land are relatively complete, and the importance of accessibility has declined. In contrast, the mature transportation infrastructure in cities allows residents to travel long distances within the city. This large-scale human travel activity in the city has changed the land use of the city, which is not taken into account by the model.

[0004] At the same time, the existing technology also uses CNN (convolutional neural network) to aggregate and learn the driving factors of other cells in the neighborhood, but CNN itself has some limitations. First, the size of the convolution kernel is fixed, and the setting of regular convolution kernel makes the shape of the neighborhood also fixed as a regular rectangle, which makes the extraction of features in the neighborhood deviate from reality to a certain extent. Secondly, although the neural network is used to extract features in the neighborhood, the neighborhood effect is still explicitly defined in the transformation rule. This effect is also artificially set. However, it is well known that there is obvious spatial heterogeneity in land use changes. The same rule will obviously bring certain deviations when used on cells with different neighborhood configurations and different geographical locations. Therefore, this scheme proposes a land use change simulation method based on graph convolutional neural network and vector cellular automaton. Summary of the Invention

[0005] The purpose of the present invention is to provide a land use change simulation method based on graph convolutional neural network and vector cellular automaton to solve the problems raised in the above background technology.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a land use change simulation method based on graph convolutional neural network and vector cellular automaton, the simulation method comprising the following steps:

[0007] S1, constructing a graph-structured cellular space, wherein the graph-structured cellular space construction includes constructing a plot adjacency matrix and a plot feature matrix;

[0008] S2, cell transition rule extraction based on graph convolutional neural network, including graph convolutional neural network construction and cell overall transition probability calculation;

[0009] S3: Simulate historical land use results and evaluate the model. The simulation results are obtained based on the overall transition probability obtained in S2. For each plot, the class with the highest transition probability is taken as the conversion result. The result is used as the attribute of the node in the graph structure. When calculating the accuracy and viewing the simulation results, the predicted type is rewritten into the attribute table of the plot.

[0010] S4, predicts the future land use pattern of the city. The pattern prediction is based on the model trained by the transition probability calculation in S2, and continuous simulation is performed to predict the future urban land use pattern. The process is to take the land use data and plot-related features of the three years extracted in S1 and put them into the hidden Markov model to predict the land use demand in future years.

[0011] Preferably, the plot adjacency matrix in S1 is constructed by first extracting data containing land changes from the data of two years using a spatial overlay analysis tool based on the collected land use vector data, and then cutting the larger plots into smaller ones. After the plots are cut, they are divided into two parts to construct the neighborhood of the plots, including a local neighborhood and a global buffer zone.

[0012] Preferably, the local neighborhood is to construct a fixed-distance boundary truncation buffer zone for each small plot, and the plots in the buffer zone are used as plots in the local neighborhood of the central plot;

[0013] The method for constructing the global buffer is to extract the trajectory data of the taxi into OD data after cleaning, and connect the plot where the O point and the D point of the OD data are located to form a global connection.

[0014] Preferably, the two neighborhoods are processed into a graph format that can be used by GCN, where the local neighborhood is to calculate the distance decay function between the central plot and other plots as the weight of the edge between the nodes in the two graphs. The distance decay function is expressed as:

[0015]

[0016] Among them, weight represents the calculated edge weight, distance represents the Euclidean distance between two plots, and d represents the radius of the buffer zone. The edges are then stored as local adjacency matrices. The global connection is to count the frequency of OD occurrences, and then these frequencies are transformed to between 0 and 1 using the standardized interquartile range and stored in the matrix as the edge weights.

[0017] Preferably, the formula for the frequency standardized interquartile range is:

[0018]

[0019] Among them, weight represents the calculated edge weight, X represents the statistical OD frequency data, median(X) represents the median, Q3 represents the third quartile, and Q1 represents the first quartile. The above formula is used to scale the frequency data to between 0 and 1 while reducing the impact of outliers.

[0020] Preferably, the plot feature matrix construction in S1 is firstly to collect and process data, including collecting population density data, POI data, OSM data and DEM of the study area as the features of the plot;

[0021] The collected population density data and DEM are in raster format. The population density and height of each plot can be obtained by using the zoning statistics tool in the GIS software.

[0022] The POI data needs to be classified for kernel density analysis and then partitioned into plot characteristics;

[0023] The OSM data is used to extract the road network. The tool for calculating the distance to the road network in the GIS software can be used to produce a raster data whose value is the distance from the pixel to the road network. Zonal statistics are then used to obtain the characteristics of the plot.

[0024] Preferably, the graph convolutional neural network in S2 is constructed based on semi-supervised classification and graph convolutional network, and the network consists of two convolutional layers, one fully connected layer and a classifier, wherein the convolutional layer aggregates the features of the node itself and the surrounding nodes in each convolution, and its formula is expressed as:

[0025]

[0026] Among them, H(l+1) is the feature of layer l+1, when H (l) When the median l is 0, H (l) is the input feature matrix, A is the input adjacency matrix, σ is the nonlinear activation function, and A is obtained by the local neighborhood relationship and global connection relationship between plots;

[0027] The A~ matrix is ​​equal to the adjacency matrix A plus the identity matrix I. The A~ matrix can be used to incorporate the characteristics of the plot itself into the convolution process;

[0028] D~ is the degree matrix of matrix A~, and W (l) is the weight matrix of the l-layer network. In the process of graph convolution, W (l) Constantly updated.

[0029] Preferably, in the process of convolution, the features aggregated by the local and global adjacency matrices are merged by means of key information matrix fusion. The rule is to take the maximum value of the information aggregated by the two matrices. If the input of the first convolution layer is a feature matrix of N*F, F is the dimension of the feature, and the output shape is a matrix of N*80, the input of the second convolution layer is the output of the first convolution layer, and the output is a matrix of N*256, and the adjacency matrix remains unchanged in the network;

[0030] After two convolutional layers, a fully connected layer is used to transform the N*256 matrix into an N*M matrix, where M is the number of sample categories. After passing through the softmax function, the values ​​in the matrix are the predicted probabilities of the nodes transforming into each category.

[0031] Preferably, the state transition of the cells in S2 depends on the transition rule of the cells, and the transition rule of the cells consists of three parts, which can be expressed as follows:

[0032]

[0033] in, is the suitability of land use conversion from plot i to type k at time t. This value is the probability output by the GCN model. is a restricted transformation factor. If the plot is located in the restricted transformation area, the value is 0, that is, the probability of overall transformation is 0, and the plot does not undergo type transformation. RA is a random factor used to simulate random land type changes in real life that are not considered by the model. The value is between 0 and 1.

[0034] Preferably, the model evaluation in S3 selects the evaluation indicators commonly used in cellular automata, including the quality factor FOM, the generator accuracy PA and the user accuracy UA, and the three evaluation indicator formulas are expressed as follows:

[0035]

[0036] Among them, A represents the error in plot area caused by the actual transformation of plot function while the simulation result remains unchanged; B represents the plot area where the transformation actually occurs and the simulation result is the same as the actual one; C represents the plot area error caused by the actual transformation and the simulation also occurring, but the type of transformation in the simulation does not match the actual one; D represents the plot area error caused by the transformation in the simulation but without actual transformation.

[0037] The technical effects and advantages of the present invention are as follows:

[0038] (1) The present invention is a land use simulation method for the current urban development context. By considering the impact of human travel behavior on land function transformation, the method also takes into account the characteristics of distant and spatially interactive plots in the process of exploring the transformation rules of plots, making the model closer to the real world, thereby improving the accuracy of land use change simulation for plots within the city;

[0039] (2) The neighborhood effect of the cells of the present invention is no longer explicitly expressed in the transition rules of the vector cellular automaton. Instead, the adjacency relationship is expressed through the edges of the graph structure, and the magnitude of the influence is expressed through the weights on the edges. During the graph convolution process, the neighborhood effect is calculated during the feature convergence process. At the same time, the addition of two convolutional layers enables the model to even calculate the interactions between plots in a larger range. At the same time, learning using a neural network is more in line with the data situation than explicit formula definitions, which makes the simulation accuracy higher. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 This is a flow chart of the simulation method based on graph convolutional neural network and vector cellular automaton of the present invention. DETAILED DESCRIPTION

[0041] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0042] The present invention provides Figure 1 The land use change simulation method based on graph convolutional neural network and vector cellular automation is shown in the figure. The simulation method includes the following steps:

[0043] S1, constructing a graph-structured cellular space, which includes the construction of the plot adjacency matrix and the plot feature matrix;

[0044] Specifically, the plot adjacency matrix in S1 is constructed based on the collected land use vector data. The data of two years are extracted using the spatial overlay analysis tool to include the data of land changes (equivalent to the land use data of two years on the plot). Then the larger plots are cut into smaller pieces. This step is because the plots are simulated for their own changes after being cut into smaller pieces. At the end of the simulation, they are spliced ​​together to simulate the shape and area changes of the larger overall plot.

[0045] After the plots are divided, the neighborhood of the plots is constructed in two parts, including the local neighborhood and the global buffer.

[0046] It's important to note that the neighborhood effect of cells is no longer explicitly expressed in the transition rules of the vector cellular automaton. Instead, adjacency relationships are expressed through edges in the graph structure, and the magnitude of influence is expressed through edge weights. During graph convolution, neighborhood effects are calculated during feature aggregation. The addition of two convolutional layers enables the model to calculate interactions between plots over a wider range. Learning using neural networks is more closely aligned with the data than using explicit formulas, resulting in higher simulation accuracy.

[0047] Furthermore, the local neighborhood is to construct a fixed distance boundary truncation buffer zone for each small plot. The shape of the buffer zone is expanded along the shape of the original plot and is truncated when encountering roads and rivers. The plots in the buffer zone are used as the plots in the local neighborhood of the central plot.

[0048] The method for constructing the global buffer is to clean the taxi trajectory data and extract it into OD data, and then connect the O point of the OD data with the plot where the D point is located to form a global connection.

[0049] Furthermore, the two neighborhoods are processed into a graph format that can be used by GCN, where the local neighborhood is to calculate the distance decay function between the central plot and other plots as the weight of the edge between the nodes in the two graphs. The distance decay function is expressed as:

[0050]

[0051] Among them, weight represents the calculated edge weight, distance represents the Euclidean distance between two plots, and d represents the radius of the buffer zone. The edges are then stored as local adjacency matrices. The global connection is to count the frequency of OD occurrences, and then these frequencies are transformed to between 0 and 1 using the standardized interquartile range and stored in the matrix as the edge weights.

[0052] It should be noted that the GCN described above is a graph convolutional neural network, a deep learning model for processing graph-structured data. Unlike traditional CNNs, which focus on processing grid-structured data (such as images), GCN is specifically designed to learn and infer the relationships and features between nodes in graph data. The core idea of ​​GCN is to use the graph's adjacency matrix and node feature matrix to update the node representation in a manner similar to convolution operations, thereby capturing the local structure and global information between nodes. Compared to traditional CNNs, GCN needs to process graph structures of uncertain size and irregular connections, so its design takes into account the connection between nodes and feature propagation.

[0053] It should be noted that in the simulation of urban land use, global connections are added to the vector cellular machine. The calculation method is to perform spatial queries on the OD points to determine the location of the plot. The strength of the connection is calculated using the standardized interquartile range method, and together with the neighborhood influence calculated using the distance decay function, the key information matrix is ​​used to fuse the proximity relationship of the cells in the graph space, so that the VCA model can simultaneously calculate and learn the two types of connections between cells, thereby improving the accuracy of the simulation.

[0054] Furthermore, the formula for the frequency-standardized interquartile range is expressed as:

[0055]

[0056] Among them, weight represents the calculated edge weight, X represents the statistical OD frequency data, median(X) represents the median, Q3 represents the third quartile, and Q1 represents the first quartile. The above formula is used to scale the frequency data to between 0 and 1 while reducing the impact of outliers.

[0057] Specifically, the construction of the plot feature matrix in S1 first involves collecting and processing data, including collecting population density data, POI data, OSM data, and DEM of the study area as the characteristics of the plot;

[0058] The collected population density data and DEM are in raster format. The population density and height of each plot can be obtained using the zoning statistics tool in the GIS software.

[0059] It should be noted that the zoning statistics tools in GIS (Geographic Information System) software are tools used to perform statistical analysis on geospatial data. These tools can help users group and analyze data based on geographic areas (such as administrative divisions, land parcels, etc.), thereby revealing the relationships, trends, and patterns between geospatial data. The zoning statistics tools provided in common GIS software include: spatial query, which uses spatial location information to query data within a specified area, such as querying the population size and land use conditions within a specific geographic area. Spatial buffer analysis, which creates a buffer based on a given point, line, or surface and performs statistical analysis on the data within the buffer, such as analyzing the population density within a certain distance from a certain location. Spatial statistical analysis, which analyzes geospatial data through statistical methods such as spatial patterns and spatial associations to reveal the spatial distribution characteristics of geographic data. Spatial interpolation analysis, which infers the values ​​of unknown areas based on the distribution of spatial data points, such as interpolating the temperature data of a known station to derive the temperature distribution of the entire area.

[0060] POI data needs to be classified for kernel density analysis. For example, kernel density analysis of recreational POIs can output a raster data with the value of how many POI points of this type are near the pixel, and then perform zoning statistics to obtain plot characteristics.

[0061] OSM data is used to extract the road network. The tool for calculating the distance to the road network in the GIS software can be used to produce a raster data with the value of the distance from the pixel to the road network. Zonal statistics are then used to obtain the characteristics of the plot.

[0062] It should be noted that in GIS software, tools for calculating the distance to the road network are usually used to determine the distance from a geospatial data point to the nearest road or path. Such tools include: Network analysis tools. Typical GIS software (such as ArcGIS, QGIS, etc.) provides a network analysis tool set, which includes functions such as calculating the shortest path, the fastest path, and driving time. Users can load road network data into GIS and then use network analysis tools to calculate the distance or optimal path from a geographic data point to the nearest road network element (such as a road or path). Distance measurement tools. GIS software usually provides basic distance measurement tools, such as the distance from a point to a line, the distance from a point to a polygon, etc. These tools can quickly calculate the Euclidean distance or other distance metrics from a geospatial data point to the nearest road network element. Interpolation analysis tools. Some GIS software may also provide interpolation analysis functions to estimate the distance of an unknown location by measuring the distance of data points on a known road network.

[0063] Furthermore, after the characteristics of each plot have been obtained, in order to facilitate the input of the subsequent algorithm, the data is uniformly processed into a normalized matrix for storage. If the study area has N cells and the characteristics have F dimensions, then the matrix shape should be N*F.

[0064] S2, cell transition rule extraction based on graph convolutional neural network, including graph convolutional neural network construction and cell overall transition probability calculation;

[0065] Specifically, the graph convolutional neural network in S2 is based on semi-supervised classification with graph convolutional networks. The network consists of two convolutional layers, one fully connected layer, and a classifier. The convolutional layer aggregates the features of the node itself and the surrounding nodes during each convolution. Its formula is expressed as:

[0066]

[0067] Among them, H (l+1) is the feature of layer l+1, when H (l) When the median l is 0, H (l) is the input feature matrix, A is the input adjacency matrix, σ is the nonlinear activation function, and A is obtained by the local neighborhood relationship and global connection relationship between plots;

[0068] The matrix is ​​equal to the adjacency matrix A plus the identity matrix I. The A~ matrix can also incorporate the characteristics of the plot itself into the convolution process;

[0069] is a matrix The degree matrix of (l) is the weight matrix of the l-layer network. In the process of graph convolution, W (l) Constantly updated.

[0070] Furthermore, in the convolution process, the features aggregated by the local and global adjacency matrices are merged by means of key information matrix fusion. The rule is to take the maximum value of the information aggregated by the two matrices. If the input of the first convolution layer is the feature matrix of N*F, F is the dimension of the feature, and the output shape is a matrix of N*80, the input of the second convolution layer is the output of the first convolution layer, and the output is a matrix of N*256, and the adjacency matrix remains unchanged in the network;

[0071] After two convolutional layers, a fully connected layer is used to transform the N*256 matrix into an N*M matrix, where M is the number of sample categories. After passing through the softmax function, the values ​​in the matrix are the predicted probabilities of the nodes transforming into each category.

[0072] It should be noted that the Softmax function is a commonly used activation function, particularly suitable for multi-classification problems. It compresses a K-dimensional vector (usually the raw scores output by a neural network) into a K-dimensional probability distribution. The output of the softmax function is a probability distribution where the probability values ​​of all categories are between 0 and 1, and their sum is 1.

[0073] Specifically, the state transition of cells in S2 depends on the transition rule of the cells, and the transition rule of the cells consists of three parts, which can be expressed as follows:

[0074]

[0075] in, is the suitability of land use conversion from plot i to type k at time t. This value is the probability output by the GCN model. is a restricted transformation factor. If the plot is located in the restricted transformation area, the value is 0, that is, the probability of overall transformation is 0, and the plot does not undergo type transformation. RA is a random factor used to simulate random land type changes in real life that are not considered by the model. The value is between 0 and 1.

[0076] S3: Simulate historical land use results and evaluate the model. The simulation results are obtained based on the overall transition probability obtained in S2. For each plot, the class with the highest transition probability is taken as the conversion result. The result is used as the attribute of the node in the graph structure. When calculating the accuracy and viewing the simulation results, the predicted type is rewritten into the attribute table of the plot.

[0077] Specifically, the model evaluation in S3 selects the evaluation indicators commonly used in cellular automata, including the quality factor FOM, generator accuracy PA and user accuracy UA, and the three evaluation indicator formulas are expressed as follows:

[0078]

[0079] Among them, A represents the error in plot area caused by the actual transformation of plot function while the simulation result remains unchanged; B represents the plot area where the transformation actually occurs and the simulation result is the same as the actual one; C represents the plot area error caused by the actual transformation and the simulation also occurring, but the type of transformation in the simulation does not match the actual one; D represents the plot area error caused by the transformation in the simulation but without actual transformation.

[0080] It should be noted that the model analysis section mainly analyzes the impact of three types of parameters on the results. The first is the parameters of the GCN model, which changes the following parameters respectively: the number of convolutional layers, the input and output dimensions of the convolutional layers, the learning rate of the model, the number of epochs of model learning, the regularization method of the model, the data set division method of the model, and the dropout ratio. The contribution of driving factors to accuracy is determined through ablation experiments. A certain dimension of features in all plots is deleted, and the change in accuracy is calculated to derive the impact of adding or not adding the feature on the results, which indirectly reflects the contribution of the feature to land use change. Neighborhood configuration refers to calculating the change in accuracy by adjusting the size of the local neighborhood. The neighborhood starts from 100m and goes to 700m to calculate the changes in three evaluation indicators and overall accuracy.

[0081] S4 predicts the future land use pattern of the city. This pattern prediction is based on the model trained by the transition probability calculation in S2, and continuous simulation is performed to predict the future urban land use pattern. The process is to take the land use data and plot-related features extracted from the three years in S1 and put them into the Hidden Markov Model to predict the land use demand in future years. For example, using data from 2009, 2014, and 2019 to predict the demand for various types of land use in 2025 and 2030, and using data from 2019 to predict land use change data for 2025, the land use demand is then used to select which plots to convert. For example, for type A land use, the predicted land use demand is 168 square kilometers. Then, among all plots, the plot with the highest probability of converting to type A land use is selected and converted until the land use demand is met. Other plots cannot be converted to type A land use.

[0082] It should be noted that the Hidden Markov Model (HMM) is a statistical model used to describe time series data. It assumes that the generation of observed data depends on a sequence of unobservable hidden states. In an HMM, the transitions between hidden states and the probability distribution of each hidden state generating observed data are key components of the model. This process involves: model training, which estimates the model parameters, including state transition probabilities and observation probabilities, from observed data. Decoding, which infers the most likely hidden state sequence given a sequence of observed data. Learning, which adjusts the model parameters based on the observed data to improve the model's fit to the data.

[0083] The above prediction process is actually a prediction method under the natural evolution scenario, because the prediction is made based on the original land use rules (the model learns the transformation rules of land use in the past). This prediction assumes that the rules of future land use changes are actually consistent with the past. Of course, this scenario is not universal, so scenarios are also selected based on the overall land space plan released by the city. This method selects two scenarios: ecological priority and people's livelihood security. One scenario emphasizes the protection of urban green spaces and the increase of urban green spaces, while the other emphasizes the construction of public service facilities such as parks, schools, and hospitals while also ensuring housing security. The establishment of the above scenarios only requires adjusting the land use demand output by the hidden Markov model to achieve the purpose. The simplest way is to directly modify the area value based on the published planning data.

[0084] Finally, it should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A land use change simulation method based on graph convolutional neural network and vector cellular automaton, characterized in that: The simulation method includes the following steps: S1, constructing a graph-structured cellular space, which includes constructing a plot adjacency matrix and a plot feature matrix. The plot adjacency matrix is ​​constructed by first extracting data containing land changes from two years of data using a spatial overlay analysis tool based on the collected land use vector data. Then, the larger plots are cut into smaller parts. After the plots are cut, they are divided into two parts to construct the neighborhood of the plots, including a local neighborhood and a global buffer zone. The global buffer is constructed by cleaning the taxi trajectory data and extracting it into OD data, and linking the plots where the O point and the D point of the OD data are located to form a global connection; The two neighborhoods are processed into a graph format that can be used by GCN, where the local neighborhood is to calculate the distance decay function between the central plot and other plots as the weight of the edge between the nodes in the two graphs. The distance decay function is expressed as: （1）； Among them, weight represents the calculated edge weight, distance represents the Euclidean distance between two plots, and d represents the radius of the buffer zone. The edges are then stored as local adjacency matrices. Global connections are calculated by counting the frequency of OD occurrences. These frequencies are then transformed to between 0 and 1 using the standardized interquartile range and stored in the matrix as edge weights. S2, cell transition rule extraction based on graph convolutional neural network, including graph convolutional neural network construction and cell overall transition probability calculation; S3: Simulate historical land use results and evaluate the model. The simulation results are obtained based on the overall transition probability obtained in S2. For each plot, the class with the highest transition probability is taken as the conversion result. The result is used as the attribute of the node in the graph structure. When calculating the accuracy and viewing the simulation results, the predicted type is rewritten into the attribute table of the plot. S4, predicts the future land use pattern of the city. The pattern prediction is based on the model trained by the transition probability calculation in S2, and continuous simulation is performed to predict the future urban land use pattern. The process is to take the land use data and plot-related features of the three years extracted in S1 and put them into the hidden Markov model to predict the land use demand in future years.

2. The land use change simulation method based on graph convolutional neural network and vector cellular automaton according to claim 1, characterized in that: The local neighborhood is to construct a fixed distance boundary truncation buffer zone for each small plot, and the plots in the buffer zone are used as plots in the local neighborhood of the central plot.

3. The land use change simulation method based on graph convolutional neural network and vector cellular automaton according to claim 1, characterized in that: The formula for the frequency standardized interquartile range is: （2）； Among them, weight represents the weight of the calculated edge, Indicates the statistical OD frequency data, represents the median, represents the third quartile, Represents the first quartile. The above formula is used to scale the frequency data to between 0 and 1 while reducing the impact of outliers.

4. The land use change simulation method based on graph convolutional neural network and vector cellular automation according to claim 1, characterized in that: The construction of the plot feature matrix in S1 first involves collecting and processing data, including collecting population density data, POI data, OSM data, and DEM of the study area as the features of the plot; The collected population density data and DEM are in raster format. The population density and height of each plot can be obtained by using the zoning statistics tool in the GIS software. The POI data needs to be classified for kernel density analysis and then partitioned into plot characteristics; The OSM data is used to extract the road network. The tool for calculating the distance to the road network in the GIS software can be used to produce a raster data whose value is the distance from the pixel to the road network. Zonal statistics are then used to obtain the characteristics of the plot.

5. The land use change simulation method based on graph convolutional neural network and vector cellular automation according to claim 1, characterized in that: The graph convolutional neural network in S2 is constructed based on semi-supervised classification and graph convolutional networks. The network consists of two convolutional layers, one fully connected layer, and a classifier. The convolutional layer aggregates the features of the node itself and the surrounding nodes in each convolution. Its formula is expressed as: （3）； in, yes The characteristics of the layer, when median When it is 0, is the input feature matrix, A is the input adjacency matrix, is a nonlinear activation function, and A is obtained from the local neighborhood relationship and global connection relationship between plots; The matrix is ​​equal to the adjacency matrix A plus the identity matrix I, through The matrix can also incorporate the characteristics of the plot itself into the convolution process; is a matrix The degree matrix of for The weight matrix of the layer network, during the graph convolution process, Constantly updated.

6. The land use change simulation method based on graph convolutional neural network and vector cellular automaton according to claim 5, characterized in that: In the convolution process, the features aggregated by the local and global adjacency matrices are merged by means of key information matrix fusion. The rule is to take the maximum value of the information aggregated by the two matrices. If the input of the first convolution layer is an N*F feature matrix, F is the dimension of the feature, and the output shape is an N*80 matrix, the input of the second convolution layer is the output of the first convolution layer, and the output is an N*256 matrix, and the adjacency matrix remains unchanged in the network; After two convolutional layers, a fully connected layer is used to transform the N*256 matrix into an N*M matrix, where M is the number of sample categories. After passing through the softmax function, the values ​​in the matrix are the predicted probabilities of the nodes transforming into each category.

7. The land use change simulation method based on graph convolutional neural network and vector cellular automaton according to claim 1, characterized in that: The state transition of the cells in S2 depends on the transition rule of the cells, and the transition rule of the cells consists of three parts, which can be expressed as follows: （4）； in, is the suitability of land use conversion from plot i to type k at time t. This value is the probability output by the GCN model. is a factor that restricts the transformation. If the plot is located in the restricted transformation area, the value is 0, that is, the probability of the overall transformation is 0, and the plot does not undergo type transformation. It is a random factor used to simulate random land type changes in real life that are not considered by the model. Its value is between 0 and 1.

8. The land use change simulation method based on graph convolutional neural network and vector cellular automation according to claim 1, characterized in that: The model evaluation in S3 uses the commonly used evaluation indicators of cellular automata, including the quality factor FOM, the generator accuracy PA, and the user accuracy UA. The formulas of the three evaluation indicators are expressed as follows: （5）； （6）； （7）； Among them, A represents the error in plot area caused by the actual transformation of plot function while the simulation result remains unchanged; B represents the plot area where the transformation actually occurs and the simulation result is the same as the actual one; C represents the plot area error caused by the actual transformation and the simulation also occurring, but the type of transformation in the simulation does not match the actual one; D represents the plot area error caused by the transformation in the simulation but without actual transformation.

Images