A regional update prediction method and system based on a multi-view graph neural network

Abstract

The application provides a regional update prediction method and system based on a multi-view graph neural network, which comprises the following steps: obtaining multi-dimensional planning data of each plot in a target region and a regional status map, each data type in the multi-dimensional planning data being a preset data type, and the regional status map containing the spatial relationship, current land use property, current update mode and plot attribute of each plot in the target region; performing feature extraction on each multi-dimensional planning data to construct a feature matrix of the target region; inputting the feature matrix and the regional status map into a preset update prediction model, so that the update prediction model generates a plurality of plot relationship graphs based on the spatial relationship, current land use property, current update mode and plot attribute of each plot in the region, and then comprehensively predicts according to each plot relationship graph to generate the update potential value and future update mode of each plot, thereby improving the accuracy and rationality of plot update prediction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of machine learning and land resource management, and particularly relates to a regional update prediction method and system based on a multi-view graph neural network. BACKGROUND

[0002] Although rapid urbanization has promoted economic and social progress, it has also brought a series of problems such as land use inefficiency, environmental degradation, and loss of urban characteristics, prompting cities to gradually enter a new stage of optimizing the stock. To achieve high-quality and sustainable development, it is urgent to revitalize the stock land through urban renewal, improve construction efficiency, and improve the ecological environment. In this process, urban renewal potential assessment can scientifically quantify the potential of the region, helping decision-makers identify land parcels in need of renewal and rationally allocate resources, thereby improving planning efficiency and decision-making accuracy. However, potential assessment alone cannot meet the diverse needs of urban renewal; different regions have different conditions and the needs of interest groups, meaning that a combination of demolition and reconstruction, micro-update, and protection and repair must be flexible to balance social, economic, and ecological multi-dimensional benefits and achieve truly high-quality and multi-dimensional urban renewal.

[0003] Existing research still has many shortcomings: first, the research scale is mostly at the city or street level, making it difficult to accurately depict land parcel differences and land ownership relationships; second, the evaluation indicators are often not comprehensive, which can lead to one-sided results; third, there is a clear preference for the demolition and reconstruction mode, ignoring other more flexible and sustainable update methods; fourth, traditional methods lack big data support, and weight allocation and result determination are easily influenced by subjectivity; fifth, the complex spatial relationships between land parcels in urban renewal have not been fully captured. SUMMARY

[0004] To solve the above technical problems, the present application provides a regional update prediction method and system based on a multi-view graph neural network, which more accurately captures the complex spatial and semantic relationships between land parcels and improves the accuracy and rationality of land parcel update prediction, providing data support for regional planning decisions.

[0005] In a first aspect, an embodiment of the present application provides a regional update prediction method based on a multi-view graph neural network, comprising:

[0006] Obtaining multi-dimensional planning data of each land parcel in a target region and a regional status map, each data type in the multi-dimensional planning data being a preset data type, and the regional status map containing the spatial relationship, current land use property, current update mode, and land parcel attribute of each land parcel in the target region;

[0007] Performing feature extraction on each of the multi-dimensional planning data to construct a feature matrix of the target region;

[0008] input the feature matrix and the regional status map into a preset update prediction model, so that the update prediction model generates a plurality of land parcel relation graphs of different adjacency relations based on the spatial relationship of each land parcel in the region, the current land use property and the current update mode, and the land parcel attribute, and then comprehensively predicts according to each land parcel relation graph to generate an update potential value and a future update mode of each land parcel;

[0009] The update prediction model is obtained by training an initial update prediction model based on a plurality of historical update cases, and the initial update prediction model is obtained by constructing a graph neural network model.

[0010] The embodiment of the present application provides a regional update prediction method based on a multi-view graph neural network. The multi-dimensional planning data and the regional status map are fused to generate a multi-relation graph, and a graph neural network model is used for prediction to realize dual prediction of the update potential and the update mode of each land parcel in the target region. On the one hand, the feature matrix extracted from the multi-dimensional planning data is conducive to the model to comprehensively evaluate the development potential and the development direction of the corresponding land parcel, thereby improving the prediction accuracy of the model. On the other hand, the feature matrix is combined with the regional status map to generate a plurality of land parcel relation graphs of different adjacency relations, so that the model can more finely capture the complex relationship between each land parcel, overcome the limitation of the traditional prediction method which only predicts a single land parcel, and realize overall prediction of the development of the target region. In summary, the embodiment of the present application can not only accurately quantify the potential of the land parcel, but also match the most suitable update mode for the land parcel according to the overall situation of the target region, thereby providing more scientific and effective decision support for the city to promote sustainable development in the aspects of differentiation and refinement.

[0011] Further, the feature extraction of each multi-dimensional planning data and the construction of the feature matrix of the target region include:

[0012] For the multi-dimensional planning data of each land parcel, according to each data type in the multi-dimensional planning data, the planning data of different data types are respectively input into the corresponding preset feature extraction MLP model, so that each feature extraction MLP model generates a corresponding feature vector, and then each feature vector is convolved and spliced to obtain the land parcel feature of the land parcel.

[0013] According to each land parcel feature, the feature matrix of the target region is constructed.

[0014] The embodiment of the present application provides a feature extraction method, which fuses planning data of each plot in multiple dimensions into a feature vector of the same dimension, so as to realize effective combination of multi-source data. In the specific extraction mode, the embodiment of the present application presets different feature extraction MLP models for different types of planning data, so as to realize targeted feature extraction. The multi-branch MLP architecture effectively solves the problem of inconsistent dimensions of multi-source data, improves the efficiency and accuracy of feature extraction, and enhances the model generalization ability. Finally, the output results of each feature extraction MLP model are spliced into an overall plot feature by the convolution splicing technology, so as to provide a data basis for subsequent model prediction.

[0015] In a possible implementation manner, the update prediction model generates a plurality of plot relationship graphs of different adjacency relationships based on the spatial relationship, the current land use property and the current update mode of each plot in the region and the plot attribute, and then generates the update potential value and the future update mode of each plot according to each plot relationship graph, including:

[0016] According to the spatial relationship of each plot and the feature matrix, a spatial autocorrelation graph of the target region is generated, and a spatial feature matrix of the target region is extracted from the spatial autocorrelation graph;

[0017] According to the current land use property of each plot and the feature matrix, a functional heterogeneity graph of the target region is generated, and a functional feature matrix of the target region is extracted from the functional heterogeneity graph in combination with the spatial autocorrelation graph;

[0018] According to the current update mode of each plot and the feature matrix, a mode heterogeneity graph of the target region is generated, and a mode feature matrix of the target region is extracted from the mode heterogeneity graph in combination with the spatial autocorrelation graph;

[0019] According to the plot attribute of each plot and the feature matrix, a semantic similarity graph of the target region is generated, and a semantic feature matrix of the target region is extracted from the semantic similarity graph;

[0020] The spatial feature matrix, the functional feature matrix, the mode feature matrix and the semantic feature matrix are spliced to obtain the fusion feature of the target region;

[0021] The fusion feature is input into a preset double-layer output layer, so that the double-layer output layer generates the update potential value and the future update mode of each plot.

[0022] The embodiment of the application provides a model-based regional update prediction method, corresponding spatial autocorrelation graphs, function heterogeneity graphs, mode heterogeneity graphs and semantic similarity graphs are generated according to the spatial relationship of each plot, the current land use property, the current update mode and the plot attribute, and further, each corresponding feature matrix is extracted, and various potential correlation relationships between plots are fully mined. The spatial autocorrelation graph is used to solve the problem that the traditional method ignores the geographical proximity effect, and the spatial analysis accuracy is improved; the function heterogeneity graph is used to quantify the land use property difference, and supports differentiated update mode recommendation; the mode heterogeneity graph is used to model the historical update mode correlation, and the time sequence continuity of the prediction result is enhanced; the semantic similarity graph is used to mine the plot attribute implicit correlation, and the prediction robustness in a complex scene is improved. Finally, the multi-path feature matrix is aggregated or fused, the update potential value and the future update mode of each plot are generated through a double-layer output layer, and thus the update condition of each plot is accurately predicted, and the accuracy and rationality of plot update prediction are improved.

[0023] Further, the spatial autocorrelation graph of the target region is generated according to the spatial relationship of each plot and the feature matrix, and the spatial feature matrix of the target region is extracted from the spatial autocorrelation graph, including:

[0024] The Euclidean distance between each plot is calculated according to the spatial relationship of each plot;

[0025] The adjacency matrix of the target region is constructed using a Gaussian decay function according to each Euclidean distance, as the spatial autocorrelation graph;

[0026] The spatial autocorrelation graph and the feature matrix are subjected to a graph convolution operation to generate the spatial feature matrix of the target region.

[0027] The embodiment of the application provides a spatial autocorrelation graph generation method. In urban space, adjacent or close plots often have similar environmental conditions or development levels (such as similar building ages, population densities and land values in the same area). Therefore, the embodiment of the application defines the adjacency relationship based on geographical distance, calculates the Euclidean distance between different plots, and uses a Gaussian decay function to describe the adjacency matrix, thereby realizing the generation of the spatial autocorrelation graph. The nonlinear distance decay modeling of the Gaussian decay function is more in line with the real space interaction law and is superior to the traditional fixed threshold method. Finally, the spatial autocorrelation graph and the feature matrix are subjected to a graph convolution operation to generate the spatial feature matrix of the target region, thereby highlighting the similarity capture of plots close in physical space and improving the accuracy and rationality of subsequent plot update prediction.

[0028] Further, the function heterogeneity graph of the target area is generated according to the current land use property of each plot and the feature matrix, and the function feature matrix of the target area is extracted from the function heterogeneity graph in combination with the spatial autocorrelation graph, including:

[0029] The current land use property of each plot and the feature matrix are input into a preset function heterogeneity MLP model, so that the function heterogeneity MLP model is mapped to generate a corresponding embedding vector, and then the embedding vectors are combined to construct the function heterogeneity graph.

[0030] The function heterogeneity graph, the spatial autocorrelation graph and the feature matrix are subjected to a multi-order neighborhood graph convolution operation to generate a function feature matrix of the target area.

[0031] The embodiment of the application provides a function heterogeneity graph generation method, which maps the current land use property of each plot and the feature matrix into a corresponding embedding vector through a preset function heterogeneity MLP model. The embedding vector mapping can solve the semantic discrete problem of land use property and realize quantitative analysis of each plot. Then, on the basis of the spatial autocorrelation graph, the function heterogeneity graph constructed by the embedding vectors and the feature matrix are introduced to perform a multi-order neighborhood graph convolution operation, so that different plots have different embedding vectors to reflect their functional differences, thereby different convolution weights are given to the "same type" or "different type" plots during information transmission, and the accuracy and rationality of plot update prediction are improved.

[0032] Further, the mode heterogeneity graph of the target area is generated according to the current update mode of each plot and the feature matrix, and the mode feature matrix of the target area is extracted from the mode heterogeneity graph in combination with the spatial autocorrelation graph, including:

[0033] The current update mode of each plot and the feature matrix are input into a preset mode heterogeneity MLP model, so that the mode heterogeneity MLP model is mapped to generate a corresponding embedding vector, and then the embedding vectors are combined to construct the mode heterogeneity graph.

[0034] The mode heterogeneity graph, the spatial autocorrelation graph and the feature matrix are subjected to a multi-order neighborhood graph convolution operation to generate a mode feature matrix of the target area.

[0035] Further, the semantic similarity graph of the target area is generated according to the plot attribute of each plot and the feature matrix, and the semantic feature matrix of the target area is extracted from the semantic similarity graph, including:

[0036] inputting the plot attributes of each plot and the feature matrix into a preset semantic MLP model, so that the semantic MLP model calculates the cosine similarity between each plot according to the plot attributes of each plot, and generates a semantic similarity graph of the target region in combination with the feature matrix;

[0037] performing a graph convolution operation on the semantic similarity graph and the feature matrix to generate a semantic feature matrix of the target region.

[0038] The embodiment of the application provides a semantic similarity graph generation method. Even if the geographical distance is far away, if two plots are highly similar in the attribute space, their update needs or potential may be similar. Therefore, the embodiment of the application measures the cosine similarity between plots in a high-dimensional feature space, calculates a similarity matrix through a semantic MLP model, and generates a semantic similarity graph of the target region. The cosine similarity calculation breaks through the limitation of the traditional Euclidean distance, fully excavates the semantic similarity relationship between plots, and improves the recognition ability of the model for non-spatial correlation.

[0039] In a possible implementation manner, the training of the initial update prediction model based on the plurality of historical update cases to obtain the update prediction model comprises:

[0040] The initial update prediction model is constructed based on a graph neural network model and comprises a spatial autocorrelation graph convolution network, a functional heterogeneity graph convolution network, a pattern heterogeneity graph convolution network, a semantic similarity graph convolution network, a feature splicing layer and a double-layer output layer.

[0041] A plurality of historical update cases are acquired, and a training data set is constructed based on the plot update mode changes in each of the historical update cases.

[0042] According to the training data set, the initial update prediction model is trained using a preset Softmax function and a cross-entropy loss function, the learning parameters in the initial update prediction model are determined, and then the update prediction model is obtained.

[0043] The embodiment of the application provides a training method of an update prediction model. The basic structure of the update prediction model is determined based on the generation process of each relationship graph, and then an initial update prediction model is constructed through a plurality of different parameter graph convolution networks, a feature splicing layer and an output layer. In the training process, an end-to-end training method based on cross-entropy loss is used, and a Softmax function is used to realize multi-mode probability output, so that the model can simultaneously predict the update potential and the update mode of the plot, and provide data support for regional planning decisions.

[0044] In a second aspect, the embodiments of the present application provide a regional update prediction system based on a multi-view graph neural network, comprising an acquisition module, a feature extraction module, and a prediction module.

[0045] The acquisition module is configured to acquire multi-dimensional planning data of each plot in the target region and a regional status map, each data type in the multi-dimensional planning data being a preset data type, and the regional status map containing spatial relationships, current land use properties, current update modes, and plot attributes of each plot in the target region.

[0046] The feature extraction module is configured to perform feature extraction on each of the multi-dimensional planning data to construct a feature matrix of the target region.

[0047] The prediction module is configured to input the feature matrix and the regional status map into a preset update prediction model, so that the update prediction model generates a plurality of plot relationship graphs of different adjacency relationships based on the spatial relationships, current land use properties, and current update modes of each plot in the region as well as the plot attributes, and then performs comprehensive prediction according to each of the plot relationship graphs to generate an update potential value and a future update mode of each plot.

[0048] The update prediction model is obtained by training an initial update prediction model based on a plurality of historical update cases, and the initial update prediction model is obtained by constructing a graph neural network model.

[0049] In a possible implementation manner, the update prediction model generates a plurality of plot relationship graphs of different adjacency relationships based on the spatial relationships, current land use properties, and current update modes of each plot in the region as well as the plot attributes, and then performs comprehensive prediction according to each of the plot relationship graphs to generate an update potential value and a future update mode of each plot, comprising:

[0050] Generating a spatial autocorrelation graph of the target region according to the spatial relationships of each plot and the feature matrix, and extracting a spatial feature matrix of the target region from the spatial autocorrelation graph;

[0051] Generating a functional heterogeneity graph of the target region according to the current land use properties of each plot and the feature matrix, and extracting a functional feature matrix of the target region from the functional heterogeneity graph in combination with the spatial autocorrelation graph;

[0052] Generating a mode heterogeneity graph of the target region according to the current update modes of each plot and the feature matrix, and extracting a mode feature matrix of the target region from the mode heterogeneity graph in combination with the spatial autocorrelation graph;

[0053] According to the plot attribute of each plot and the feature matrix, a semantic similarity graph of the target region is generated, and a semantic feature matrix of the target region is extracted from the semantic similarity graph;

[0054] The spatial feature matrix, the functional feature matrix, the pattern feature matrix, and the semantic feature matrix are spliced to obtain a fusion feature of the target region.

[0055] The fusion feature is input into a preset double-layer output layer, so that the double-layer output layer generates an updated potential value and a future update mode of each plot. BRIEF DESCRIPTION OF DRAWINGS

[0056] Figure 1 A flowchart of a regional update prediction method based on a multi-view graph neural network provided by an embodiment of the present application;

[0057] Figure 2 A flowchart of feature extraction by an MLP model in a regional update prediction method based on a multi-view graph neural network provided by an embodiment of the present application;

[0058] Figure 3 A flowchart of generating a plot relationship graph in a regional update prediction method based on a multi-view graph neural network provided by an embodiment of the present application;

[0059] Figure 4 A flowchart of feature splicing and output of a prediction result in a regional update prediction method based on a multi-view graph neural network provided by an embodiment of the present application;

[0060] Figure 5 A structural diagram of a regional update prediction system based on a multi-view graph neural network provided by an embodiment of the present application. DETAILED DESCRIPTION

[0061] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0062] It should be noted that the step numbers in the text are only for the convenience of explaining the specific embodiments, and do not serve as a limitation on the execution sequence of the steps. In the description of the present application, the terms “first” and “second” are used for descriptive purposes only, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features limited by “first” and “second” can explicitly or implicitly include one or more of the features.

[0063] Embodiment One:

[0064] As shown in the embodiment one provides a multi-view graph neural network-based regional update prediction method, comprising steps S1-S3: Figure 1

[0065] Step S1, obtaining multi-dimensional planning data of each plot in the target region and the regional status map, each data type in the multi-dimensional planning data is a preset data type, and the regional status map contains the spatial relationship, current land use property, current update mode and plot attribute of each plot in the target region;

[0066] Step S2, feature extraction is performed on each of the multi-dimensional planning data to construct a feature matrix of the target region;

[0067] Step S3, inputting the feature matrix and the regional status map into a preset update prediction model, so that the update prediction model generates a plurality of plot relationship graphs with different adjacency relationships based on the spatial relationship, current land use property and current update mode of each plot in the region, and plot attribute, and then comprehensively predicts according to each of the plot relationship graphs to generate an update potential value and a future update mode of each plot;

[0068] Wherein, the update prediction model is obtained by training an initial update prediction model based on a plurality of historical update cases, and the initial update prediction model is obtained by constructing a graph neural network model.

[0069] The embodiment of the present application provides a multi-view graph neural network-based regional update prediction method, which generates a multi-relationship graph by fusing multi-dimensional planning data and regional status map, and uses a graph neural network model for prediction, realizing the dual prediction of the update potential and update mode of each plot in the target region. On the one hand, the feature matrix extracted from the multi-dimensional planning data is conducive to the model's comprehensive evaluation of the development potential and development direction of the corresponding plot, improving the prediction accuracy of the model. On the other hand, the combination of the feature matrix and the regional status map generates a plurality of plot relationship graphs with different adjacency relationships, enabling the model to more finely capture the complex relationships between each plot, overcoming the limitations of traditional prediction methods that only predict a single plot, and realizing the overall prediction of the development of the target region. In summary, the embodiment of the present application can not only accurately quantify the plot potential, but also match the most suitable update mode for the plot according to the overall situation of the target region, providing more scientific and effective decision support for the sustainable development of the city in the differentiation and refinement level.

[0070] In a preferred embodiment, in step S1, according to the urban renewal demand, the multi-dimensional planning data is used to evaluate seven indicators of the plot, and the seven indicators and the corresponding multi-dimensional planning data are respectively:​

[0071] Livable and workable index (X liv ): assessed by housing prices, population age structure, building density, plot ratio, etc.

[0072] Economic development index (X eco ): assessed by employment density, job-housing balance, etc.

[0073] Ecological and energy-saving index (X env ): assessed by 5-minute accessible green space, green space coverage, etc.

[0074] Community vitality index (X act ): assessed by nighttime light intensity, POI density, population flow, etc.

[0075] Safety resilience index (X res ): assessed by building age, 5-minute accessible rate of fire protection, etc.

[0076] Facility index (X fac ): assessed by traffic accessibility, public facility convenience, etc.

[0077] Cultural style (X cul ): assessed by green view rate, aesthetics, etc.

[0078] The acquisition method of the planning data is to pre-process and blockize multi-source heterogeneous information from POI, Baidu heat, nighttime light, buildings and planning, street view, Tencent location big data, obtain multi-dimensional planning data corresponding to each plot, and perform data preprocessing, and the specific steps include:

[0079] ① Coordinate unification and deduplication: project each data source to the same coordinate system, remove duplicate records and fill in missing values;

[0080] ② Plot aggregation: using kernel density analysis or spatial interpolation methods, the data originally in the form of grid or point features is aggregated into each plot boundary, forming a vector data table with unique number and multi-dimensional attributes;

[0081] ③ Street view image processing: use the API of Baidu Map (https: / / map.baidu.com / ) to grab street view images distributed along the road network, and considering that the dense distribution of shooting points can easily lead to unnecessary information redundancy, the shooting points also need to be sparsely pre-processed, selecting one from every three shooting points to form street view image data. Then use the semantic segmentation model (HRNetV2) to extract features such as buildings, vegetation, roads and sky from the street view image data;

[0082] (4) Population activity data: Tencent real-time user density data is obtained through the open API of Tencent "Easygo" service (https: / / heat.qq.com / ). Then, Tencent real-time user density (RTUD) is spatio-temporally aggregated to calculate the population distribution and activity intensity of each block on weekdays and holidays.

[0083] Further, in step S2, feature extraction is performed on each of the multi-dimensional planning data to construct a feature matrix of the target region, including:

[0084] For the multi-dimensional planning data of each block, according to each data type in the multi-dimensional planning data, the planning data of different data types is respectively input into a corresponding preset feature extraction MLP model, so that each feature extraction MLP model generates a corresponding feature vector, and then each feature vector is convoluted and spliced to obtain a block feature of the block.

[0085] According to each block feature, a feature matrix of the target region is constructed.

[0086] The feature extraction method provided in the embodiments of the present application fuses the planning data of each block in multiple dimensions into a feature vector of the same dimension, effectively combines multi-source data, and realizes targeted feature extraction. The multi-branch MLP architecture effectively solves the inconsistency problem of multi-source data dimensions, improves the efficiency and accuracy of feature extraction, enhances the model generalization ability. Finally, the output results of each feature extraction MLP model are spliced into a whole block feature by convolution splicing technology, providing a data basis for subsequent model prediction.

[0087] In a preferred embodiment, as shown in Figure 2 , for each block (such as V1-V3 in Figure 2 ), the multi-dimensional planning data of the block is divided into seven types of planning data according to the seven indicators, and then the seven types of planning data are respectively input into a preset MLP model (such as F livmlp model for livable and workable indicators, F ecomlp model for economic development indicators, etc.), and then each sub-vector is mapped to a hidden layer and spliced, and finally the fusion and dimension reduction are completed through a linear layer and an activation function (such as ReLU) to generate a unified feature representation matrix X i The specific formula of feature extraction is as follows:

[0088]

[0089] Wherein, i is the block number, W F , b FFor the learnable parameters of the linear layer, ReLU(·) is a non-linear activation function.

[0090] Taking the livability and workability index as an example, based on the multi-dimensional planning data, the average housing price, building density, plot ratio, and age structure category within the target plot are determined. This data is then input into the corresponding MLP model, using hidden layers to extract deep representations of these input features. The hidden layers map the input to a higher-level feature space using non-linear activation functions (such as ReLU). Finally, the MLP model outputs equal-length deep features, which will be used for subsequent spatial effects and semantic relationship modeling.

[0091] Furthermore, the training process of the feature extraction MLP model is as follows:

[0092] 1. Forward propagation: The data undergoes nonlinear transformation through multiple hidden layers to obtain the output of the MLP.

[0093] 2. Loss function: Mean Squared Error (MSE).

[0094] 3. Backpropagation: By calculating the gradient, the backpropagation algorithm can adjust each weight to minimize the loss.

[0095] 4. Optimization: Adam optimizer.

[0096] 5. Iterative training: Each epoch will traverse all training data once, going through forward propagation, loss calculation, backpropagation, and optimization update steps.

[0097] In one possible implementation, in step S3, the update prediction model generates several plot relationship diagrams with different adjacency relationships based on the spatial relationships of each plot in the region, the current land use nature and current update mode, and plot attributes. Then, based on these plot relationship diagrams, a comprehensive prediction is made to generate the update potential value and future update mode for each plot, including:

[0098] Based on the spatial relationships of each plot and the feature matrix, a spatial autocorrelation map of the target area is generated, and the spatial feature matrix of the target area is extracted from the spatial autocorrelation map.

[0099] Based on the current land use nature of each plot and the feature matrix, a functional heterogeneity map of the target area is generated, and the functional feature matrix of the target area is extracted from the functional heterogeneity map by combining the spatial autocorrelation map.

[0100] Based on the current update mode of each plot and the feature matrix, a pattern heterogeneity map of the target area is generated, and the pattern feature matrix of the target area is extracted from the pattern heterogeneity map by combining the spatial autocorrelation map.

[0101] Based on the land parcel attributes and the feature matrix, a semantic similarity map of the target area is generated, and the semantic feature matrix of the target area is extracted from the semantic similarity map.

[0102] The spatial feature matrix, functional feature matrix, pattern feature matrix, and semantic feature matrix are concatenated to obtain the fusion features of the target region.

[0103] The fusion features are input into a preset dual-layer output layer, so that the dual-layer output layer generates the update potential value and future update mode of each plot.

[0104] This application provides a model-based regional update prediction method. Based on the spatial relationships, current land use, current update mode, and land parcel attributes, it generates corresponding spatial autocorrelation maps, functional heterogeneity maps, pattern heterogeneity maps, and semantic similarity maps. Furthermore, it extracts the corresponding feature matrices to fully explore the various potential relationships between land parcels. Specifically, the spatial autocorrelation map addresses the problem of traditional methods neglecting the geographical proximity effect, improving the accuracy of spatial analysis; the functional heterogeneity map quantifies differences in land use, supporting differentiated update mode recommendations; the pattern heterogeneity map models historical update mode associations, enhancing the temporal continuity of prediction results; and the semantic similarity map uncovers implicit associations in land parcel attributes, improving prediction robustness in complex scenarios. Finally, the multi-path feature matrices are aggregated or fused, and a two-layer output layer generates the update potential value and future update mode for each land parcel, thereby obtaining accurate predictions of the update status of each land parcel and improving the accuracy and rationality of land parcel update prediction.

[0105] In a preferred embodiment, a simplified schematic diagram illustrating the process of generating a map of several plots with different adjacency relationships based on the spatial relationships, current land use, current update mode, and plot attributes of each plot in the region is shown below. Figure 3 As shown, in the graph learning framework, each "plot" is considered as a node v in the graph. i There are multiple potential relationships between the target node and surrounding plots, including geographical proximity (spatial autocorrelation), heterogeneity in land use function or ownership, or existing plot update patterns (spatial heterogeneity), and similarity in attribute features (semantic similarity). Therefore, to capture these relationships more comprehensively, this application proposes constructing three different types of adjacency relationships on the same node set V and performing graph convolution or attention learning on each, and then aggregating or fusing the multi-path features to obtain a more accurate characterization of the target node.

[0106] Further, the spatial self-correlation graph of the target region is generated according to the spatial relationship of each plot and the feature matrix, and the spatial feature matrix of the target region is extracted from the spatial self-correlation graph, including:

[0107] According to the spatial relationship of each plot, the Euclidean distance between each plot is calculated;

[0108] According to each of the Euclidean distances, a neighborhood matrix of the target region is constructed using a Gaussian decay function as the spatial self-correlation graph;

[0109] The spatial self-correlation graph and the feature matrix are subjected to a graph convolution operation to generate a spatial feature matrix of the target region.

[0110] The embodiment of the application provides a spatial self-correlation graph generation method. In urban space, adjacent plots or plots with a short distance often have similar environmental conditions or development levels (such as similar building ages, population densities, and land values in the same area). Therefore, the embodiment of the application defines the adjacency relationship based on geographical distance, calculates the Euclidean distance between different plots, and uses a Gaussian decay function to depict the adjacency matrix to realize the generation of the spatial self-correlation graph. The nonlinear distance decay modeling of the Gaussian decay function is more in line with the real space interaction law and is superior to the traditional fixed threshold method. Finally, the spatial self-correlation graph and the feature matrix are subjected to a graph convolution operation to generate a spatial feature matrix of the target region, highlight the similarity capture of plots close in physical space, and improve the accuracy and rationality of subsequent plot update prediction.

[0111] In a preferred embodiment, the adjacency relationship is defined based on geographical distance. The Euclidean distance between plot i and plot j is calculated, and a Gaussian decay function is used to depict the adjacency matrix AD. The specific formula is as follows:

[0112]

[0113] where d i,j represents the distance between node i and node j; σ D is a decay control parameter; is a distance threshold value calculated by a Top K function, where k is the number of selected nearest neighbors.

[0114] Let the node feature matrix be X∈R N×C , where N is the number of nodes and C is the feature dimension. Then in the graph convolution process, the graph convolution formula is:

[0115] Y A = F sa (X) = g(X, A D ) = ReLU(A D XθD +b D )

[0116] where A D is the above-mentioned spatial autocorrelation adjacency matrix (which has been normalized in a symmetric manner), D is a degree matrix, θ D and b D are learnable parameters representing weights and biases, respectively. X is the input feature matrix, and the corresponding output spatial feature matrix is Y A After superimposing multiple layers, the result is an updated embedding vector of each node under the geographical "neighborhood" relationship, highlighting the similarity capture of the physical space near the plot.

[0117] Further, the functional heterogeneity map of the target area is generated according to the current land use properties of each plot and the feature matrix, and the functional feature matrix of the target area is extracted from the functional heterogeneity map in combination with the spatial autocorrelation map, including:

[0118] The current land use properties of each plot and the feature matrix are input into a preset functional heterogeneity MLP model, so that the functional heterogeneity MLP model maps to generate each corresponding embedding vector, and then each embedding vector is combined to construct the functional heterogeneity map;

[0119] The functional heterogeneity map, the spatial autocorrelation map and the feature matrix are subjected to multi-order neighborhood graph convolution operation to generate the functional feature matrix of the target area.

[0120] The embodiment of the application provides a functional heterogeneity map generation method, which maps the current land use properties of each plot and the feature matrix into each corresponding embedding vector through a preset functional heterogeneity MLP model. The embedding vector mapping can solve the semantic discrete problem of land use properties and realize quantitative analysis of each plot. Then, on the basis of the spatial autocorrelation map, the functional heterogeneity map constructed by each embedding vector and the feature matrix are introduced to perform multi-order neighborhood graph convolution operation, so that different plots have different embedding vectors to reflect their functional differences, thereby different convolution weights are given to "similar" or "different" plots during information transmission, improving the accuracy and rationality of plot update prediction.

[0121] Further, the pattern heterogeneity map of the target area is generated according to the current update mode of each plot and the feature matrix, and the pattern feature matrix of the target area is extracted from the pattern heterogeneity map in combination with the spatial autocorrelation map, including:

[0122] inputting the current update mode of each plot and the feature matrix into a preset mode heterogeneity MLP model, so that the mode heterogeneity MLP model maps to generate a corresponding embedding vector of each plot, and then combining each embedding vector to construct the mode heterogeneity graph;

[0123] performing a multi-order neighborhood graph convolution operation on the mode heterogeneity graph, the spatial autocorrelation graph and the feature matrix to generate a mode feature matrix of the target region.

[0124] Within the same range, plots show obvious differences due to land use properties, ownership categories or planning functions, i.e. "heterogeneity". At the same distance, residential land and industrial land may be completely different in terms of value potential, update demand, etc. Therefore, in a preferred embodiment, we construct two networks to distinguish the learnable coefficients in different land class partitions or different update mode partitions. That is, a functional heterogeneity graph convolution network (H1GCN) and a mode heterogeneity graph convolution network (H2GCN)

[0125] The structures of the two networks are the same, except for the grouping. Take functional grouping as an example for introduction. First, grouping, define a grouping label (land use property) for each plot. There are r different groups (for example: residential, commercial, industrial, public management and service, etc.). Each node v i is mapped to an embedding vector E h (i)∈R d These vectors can be stacked into a matrix E h ∈R N×d where d is the embedding dimension.

[0126] The mapping is learned by a functional heterogeneity graph convolution network (HMLP):

[0127] E h =F hmlp (X)=W2(ReLU(W1X+b1))+b2

[0128] where W1, W2 are learnable parameters, and b1, b2 are bias terms.

[0129] After introducing the heterogeneity vector, the information transmission of each edge (i, j) is weighted and corrected in the graph convolution calculation, to generate a functional feature matrix Y H1 or a mode feature matrix Y H2 of the target region. Wherein, the multi-order neighborhood (q=0 to Q) convolution formula is as follows:

[0130]

[0131] denotes the adjacency matrix of the q-order neighborhood. and are learnable parameters of q-th order group. Different patches have different E h (i) to reflect its function or mode group, so as to give different convolution weights to "same kind" or "different kind" patches when information is transmitted.

[0132] Further, the semantic similarity graph of the target region is generated according to the patch attribute of each patch and the feature matrix, and the semantic feature matrix of the target region is extracted from the semantic similarity graph, comprising:

[0133] The patch attribute of each patch and the feature matrix are input into a preset semantic MLP model, so that the semantic MLP model calculates the cosine similarity between each patch according to the patch attribute of each patch, and generates a semantic similarity graph of the target region in combination with the feature matrix;

[0134] The semantic similarity graph and the feature matrix are subjected to graph convolution operation to generate a semantic feature matrix of the target region.

[0135] The embodiment of the application provides a semantic similarity graph generation method. Even if the geographical distance is far away, if two patches are highly similar in the attribute space, their update needs or potential may be similar. Therefore, the embodiment of the application measures the cosine similarity between patches in a high-dimensional feature space, calculates a similarity matrix through a semantic MLP model, and generates a semantic similarity graph of the target region. The cosine similarity calculation breaks through the limitation of traditional Euclidean distance, fully excavates the semantic similarity relationship between each patch, and improves the recognition ability of the model to non-spatial correlation.

[0136] In a preferred embodiment, the cosine similarity between patches in a high-dimensional feature space can be measured based on the patch attribute (such as building age, POI distribution, population structure) of each patch, a similarity matrix is calculated through a small MLP (SMLP), and then it is regarded as the adjacency matrix A S of the "semantic similarity graph".

[0137] A S = F smlp (W3X + b3) + b4)

[0138] wherein, W3∈R C×C′ , W4∈R C′×N , b3∈R C′ , b4∈R N are learnable parameters; F smlp (·) represents a semantic MLP layer; ReLU(·) is an activation function; As represents the adjacency matrix of the semantic graph; σ(·) is a SoftMax function for normalization.

[0139] After the semantic similarity adjacency matrix AS is completed, a multi-layer GCN can also be used for information transmission and aggregation to generate a semantic feature matrix of the target area, and the specific formula is as follows:

[0140] Y s =F ss (X,A s )=g(X,A s )=ReLU(A s Xθ S +b s )

[0141] Where D is a degree matrix; θ S ∈R C×C′ , b s ∈R C′ are learnable parameters; the dimension of the input feature X is NxC, and the dimension of the output feature Ys is Nx C'. The far distance but similar feature plots can exchange information in the convolution process, avoiding the neglect phenomenon caused by pure geographical proximity.

[0142] In a preferred embodiment, as shown in Figure 4 After multi-source feature fusion and relationship graph construction are performed, a plurality of different feature matrices are generated based on a spatial autocorrelation graph convolution network (SA), a functional heterogeneity graph convolution network (SH1), a pattern heterogeneity graph convolution network (SH2), and a semantic similarity graph convolution network (SS). At this time, in order to aggregate various features of multiple nodes into the target plot, the embodiments of the present application use a feature concatenation (concat) method to concatenate the feature vectors output by different graphs in the channel dimension, and then input them into the subsequent MLP.

[0143]

[0144] Y t ∈R 1×C is the fusion feature of the target plot; Y A (i), Y H1 (i), Y H2 (i), and Y S (i) represent the features from the SA, SH, and SS modules, respectively.

[0145] Finally, the fusion feature is input into a preset double-layer output layer to obtain the updated potential value and the future update mode of the target plot. The design of the double-layer output layer is as follows:

[0146] The first layer outputs two types of 0 and 1, which respectively represent whether to update and the confidence of the update, that is, the size of the update potential. The category with the update of 1 is transmitted into the second classification task, and the second layer outputs the update mode of the preset category. In order to achieve the dual goals of "update potential prediction" and "update mode recommendation", multi-task learning is used to train the two output layers in the training process.

[0147] In a possible implementation manner, the training of the initial update prediction model based on a plurality of historical update cases to obtain the update prediction model comprises:

[0148] The initial update prediction model is obtained based on a graph neural network model construction, and the initial update prediction model comprises a spatial autocorrelation graph convolution network, a functional heterogeneity graph convolution network, a mode heterogeneity graph convolution network, a semantic similarity graph convolution network, a feature splicing layer and a double-layer output layer.

[0149] A plurality of historical update cases are obtained, and a training data set is constructed based on the change of the plot update mode in each of the historical update cases;

[0150] According to the training data set, the initial update prediction model is trained using a preset Softmax function and a cross-entropy loss function, the learning parameters in the initial update prediction model are determined, and then the update prediction model is obtained.

[0151] The embodiment of the application provides a training method of an update prediction model, determines the basic structure of the update prediction model based on the generation process of each relationship graph, and then constructs an initial update prediction model through a plurality of different parameter graph convolution networks, a feature splicing layer and an output layer. In the training process, the end-to-end training method based on the cross-entropy loss is used in the embodiment, the multi-mode probability output is realized through the Softmax function, so that the model can simultaneously predict the update potential and the update mode of the plot, and provide data support for regional planning decision.

[0152] In a preferred embodiment, the update potential prediction is regarded as a regression task. The potential prediction result is a continuous numerical value, and the range is 0-1. The update mode recommendation is regarded as a classification problem (such as the categories of "demolition and reconstruction", "micro-update" and "repair and protection"), and can be trained using SoftMax+cross-entropy loss. Specifically, in the training process, based on a certain historical update case, each learning parameter (including graph convolution weight, MLP weight and heterogeneity embedding) is supervised and optimized, and after the model converges, the final update prediction model is obtained.

[0153] Embodiment two:

[0154] As Figure 5As shown, correspondingly, embodiment two provides a regional update prediction system based on a multi-view graph neural network, comprising an acquisition module 10, a feature extraction module 20, and a prediction module 30;

[0155] The acquisition module 10 is configured to acquire multi-dimensional planning data of each plot in the target region and a regional status map, each data type in the multi-dimensional planning data is a preset data type, and the regional status map contains spatial relationships, current land use properties, current update modes, and plot attributes of each plot in the target region.

[0156] The feature extraction module 20 is configured to perform feature extraction on each of the multi-dimensional planning data and construct a feature matrix of the target region.

[0157] The prediction module 30 is configured to input the feature matrix and the regional status map into a preset update prediction model, so that the update prediction model generates a plurality of plot relationship graphs of different adjacency relationships based on the spatial relationships, current land use properties, current update modes, and plot attributes of each plot in the region, and then performs comprehensive prediction according to each of the plot relationship graphs to generate an update potential value and a future update mode of each plot.

[0158] The update prediction model is obtained by training an initial update prediction model based on a plurality of historical update cases, and the initial update prediction model is obtained by constructing a graph neural network model.

[0159] Further, the feature extraction module 20 performs feature extraction on each of the multi-dimensional planning data and constructs a feature matrix of the target region, including:

[0160] For the multi-dimensional planning data of each plot, different data types of planning data are input into corresponding preset feature extraction MLP models according to each data type in the multi-dimensional planning data, so that each feature extraction MLP model generates a corresponding feature vector, and then each feature vector is convolved and spliced to obtain a plot feature of the plot.

[0161] The feature matrix of the target region is constructed based on each of the plot features.

[0162] In one possible implementation, the update prediction model generates a plurality of plot relationship graphs of different adjacency relationships based on the spatial relationships, current land use properties, current update modes, and plot attributes of each plot in the region, and then performs comprehensive prediction according to each of the plot relationship graphs to generate an update potential value and a future update mode of each plot, including:

[0163] generate a spatial autocorrelation graph of the target region according to the spatial relationship of each plot and the feature matrix, and extract a spatial feature matrix of the target region from the spatial autocorrelation graph;

[0164] generate a functional heterogeneity graph of the target region according to the current land use property of each plot and the feature matrix, and extract a functional feature matrix of the target region from the functional heterogeneity graph in combination with the spatial autocorrelation graph;

[0165] generate a pattern heterogeneity graph of the target region according to the current update mode of each plot and the feature matrix, and extract a pattern feature matrix of the target region from the pattern heterogeneity graph in combination with the spatial autocorrelation graph;

[0166] generate a semantic similarity graph of the target region according to the plot attribute of each plot and the feature matrix, and extract a semantic feature matrix of the target region from the semantic similarity graph;

[0167] perform feature splicing on the spatial feature matrix, the functional feature matrix, the pattern feature matrix and the semantic feature matrix to obtain a fusion feature of the target region;

[0168] input the fusion feature into a preset double-layer output layer to enable the double-layer output layer to generate an update potential value and a future update mode of each plot.

[0169] Further, the generation of the spatial autocorrelation graph of the target region according to the spatial relationship of each plot and the feature matrix, and the extraction of the spatial feature matrix of the target region from the spatial autocorrelation graph, comprise:

[0170] According to the spatial relationship of each plot, the Euclidean distance between each plot is calculated;

[0171] According to each of the Euclidean distances, a neighborhood matrix of the target region is constructed using a Gaussian decay function as the spatial autocorrelation graph;

[0172] Perform graph convolution operation on the spatial autocorrelation graph and the feature matrix to generate a spatial feature matrix of the target region.

[0173] Further, the generation of the functional heterogeneity graph of the target region according to the current land use property of each plot and the feature matrix, and the extraction of the functional feature matrix of the target region from the functional heterogeneity graph in combination with the spatial autocorrelation graph, comprise:

[0174] input the current land use property of each plot and the feature matrix into a preset functional heterogeneity MLP model to enable the functional heterogeneity MLP model to map to generate each corresponding embedding vector, and then combine each of the embedding vectors to construct the functional heterogeneity graph;

[0175] performing a graph convolution operation on the functional heterogeneity graph, the spatial autocorrelation graph and the feature matrix in a multi-order neighborhood to generate a functional feature matrix of the target region.

[0176] Further, the mode heterogeneity graph of the target region is generated according to the current update mode of each plot and the feature matrix, and a mode feature matrix of the target region is extracted from the mode heterogeneity graph in combination with the spatial autocorrelation graph, including:

[0177] inputting the current update mode of each plot and the feature matrix into a preset mode heterogeneity MLP model, so that the mode heterogeneity MLP model maps to generate a corresponding embedding vector, and then combining each embedding vector to construct the mode heterogeneity graph;

[0178] performing a graph convolution operation on the mode heterogeneity graph, the spatial autocorrelation graph and the feature matrix in a multi-order neighborhood to generate a mode feature matrix of the target region.

[0179] Further, the semantic similarity graph of the target region is generated according to the plot attribute of each plot and the feature matrix, and a semantic feature matrix of the target region is extracted from the semantic similarity graph, including:

[0180] inputting the plot attribute of each plot and the feature matrix into a preset semantic MLP model, so that the semantic MLP model calculates the cosine similarity between each plot according to the plot attribute of each plot, and generates a semantic similarity graph of the target region in combination with the feature matrix;

[0181] performing a graph convolution operation on the semantic similarity graph and the feature matrix to generate a semantic feature matrix of the target region.

[0182] In a possible implementation manner, the initial update prediction model is trained based on a plurality of historical update cases to obtain the update prediction model, including:

[0183] The initial update prediction model is obtained based on a graph neural network model, and the initial update prediction model includes a spatial autocorrelation graph convolution network, a functional heterogeneity graph convolution network, a mode heterogeneity graph convolution network, a semantic similarity graph convolution network, a feature concatenation layer and a double-layer output layer.

[0184] A plurality of historical update cases are obtained, and a training data set is constructed based on the plot update mode change in each of the historical update cases;

[0185] According to the training data set, the initial update prediction model is trained using a preset Softmax function and a cross-entropy loss function, the parameters in the initial update prediction model are determined, and then the update prediction model is obtained.

[0186] The embodiment of the present application provides a regional update prediction system based on a multi-view graph neural network. The system generates multi-relation graphs by fusing multi-dimensional planning data and regional status graphs, and uses a graph neural network model for prediction, thereby realizing dual prediction of the update potential and update mode of each plot in the target region. On the one hand, the feature matrix extracted from the multi-dimensional planning data is conducive to the model's comprehensive evaluation of the development potential and development direction of the corresponding plot, thereby improving the prediction accuracy of the model. On the other hand, the combination of the feature matrix and the regional status graph generates a number of plot relation graphs with different adjacency relations, so that the model can more finely capture the complex relationships between plots, overcoming the limitations of traditional prediction methods that only predict a single plot, and realizing overall prediction of the development of the target region. In summary, the embodiment of the present application can not only accurately quantify the plot potential, but also match the most suitable update mode for the plot according to the overall situation of the target region, thereby providing more scientific and effective decision support for the sustainable development of the city at the differentiation and refinement level.

[0187] The more detailed working principle and step flow of the embodiment can be but not limited to the related description of embodiment one.

[0188] The above specific embodiments further illustrate the purpose, technical solutions and advantages of the present application. It should be understood that the above description is only for specific embodiments of the present application and is not intended to limit the protection scope of the present application. It is particularly pointed out that any modification, equivalent replacement, improvement, etc. made by those skilled in the art within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A regional update prediction method based on a multi-view graph neural network, characterized in that, The method comprises the following steps: acquiring multi-dimensional planning data of each plot in a target area and an area status map, each data type in the multi-dimensional planning data being a preset data type, the area status map containing spatial relationships, current land use properties, current update modes, and plot attributes of each plot in the target area; performing feature extraction on each of the multi-dimensional planning data to construct a feature matrix of the target area; inputting the feature matrix and the area status map into a preset update prediction model, so that the update prediction model generates a plurality of plot relationship graphs of different adjacency relationships based on the spatial relationships, current land use properties, current update modes, and plot attributes of each plot in the area, and then performs comprehensive prediction according to each of the plot relationship graphs to generate an update potential value and a future update mode of each plot, including: generating a spatial autocorrelation graph of the target area according to the spatial relationships of each plot and the feature matrix, and extracting a spatial feature matrix of the target area from the spatial autocorrelation graph; generating a functional heterogeneity graph of the target area according to the current land use properties of each plot and the feature matrix, and extracting a functional feature matrix of the target area from the functional heterogeneity graph in combination with the spatial autocorrelation graph; generating a mode heterogeneity graph of the target area according to the current update modes of each plot and the feature matrix, and extracting a mode feature matrix of the target area from the mode heterogeneity graph in combination with the spatial autocorrelation graph; generating a semantic similarity graph of the target area according to the plot attributes of each plot and the feature matrix, and extracting a semantic feature matrix of the target area from the semantic similarity graph; and performing feature splicing on the spatial feature matrix, the functional feature matrix, the mode feature matrix, and the semantic feature matrix to obtain a fusion feature of the target area; and inputting the fusion feature into a preset double-layer output layer so that the double-layer output layer generates an update potential value and a future update mode of each plot. The update prediction model is obtained by training an initial update prediction model based on a plurality of historical update cases, and the initial update prediction model is obtained by constructing a graph neural network model.

2. The regional update prediction method based on the multi-view graph neural network according to claim 1, wherein, The feature extraction on each of the multi-dimensional planning data to construct a feature matrix of the target area comprises: for the multi-dimensional planning data of each plot, different data types of planning data are input into corresponding preset feature extraction MLP models according to each data type in the multi-dimensional planning data, so that each feature extraction MLP model generates a corresponding feature vector, and then each feature vector is convoluted and spliced to obtain a plot feature of the plot; a feature matrix of the target area is constructed based on each of the plot features.

3. The regional update prediction method based on the multi-view graph neural network according to claim 1, wherein, The generation of a spatial autocorrelation graph of the target area according to the spatial relationships of each plot and the feature matrix, and the extraction of a spatial feature matrix of the target area from the spatial autocorrelation graph, comprises: calculating the Euclidean distances between each plot according to the spatial relationships of each plot; constructing an adjacency matrix of the target area using a Gaussian decay function according to each of the Euclidean distances, as the spatial autocorrelation graph. performing graph convolution operations on the spatial autocorrelation graph and the feature matrix to generate a spatial feature matrix of the target region.

4. The region updating prediction method based on multi-view graph neural network according to claim 1, wherein, The functional heterogeneity graph of the target region is generated according to the current land use properties of each plot and the feature matrix, and the functional feature matrix of the target region is extracted from the functional heterogeneity graph in combination with the spatial autocorrelation graph, including: The current land use properties of each plot and the feature matrix are input into a preset functional heterogeneity MLP model, so that the functional heterogeneity MLP model maps to generate each corresponding embedding vector, and then each embedding vector is combined to construct the functional heterogeneity graph; The functional feature matrix of the target region is generated by performing multi-order neighborhood graph convolution operations on the functional heterogeneity graph, the spatial autocorrelation graph and the feature matrix.

5. The region updating prediction method based on multi-view graph neural network according to claim 1, wherein, The mode heterogeneity graph of the target region is generated according to the current update mode of each plot and the feature matrix, and the mode feature matrix of the target region is extracted from the mode heterogeneity graph in combination with the spatial autocorrelation graph, including: The current update mode of each plot and the feature matrix are input into a preset mode heterogeneity MLP model, so that the mode heterogeneity MLP model maps to generate each corresponding embedding vector, and then each embedding vector is combined to construct the mode heterogeneity graph; The mode feature matrix of the target region is generated by performing multi-order neighborhood graph convolution operations on the mode heterogeneity graph, the spatial autocorrelation graph and the feature matrix.

6. The region updating prediction method based on multi-view graph neural network according to claim 1, wherein, The semantic similarity graph of the target region is generated according to the plot attributes of each plot and the feature matrix, and the semantic feature matrix of the target region is extracted from the semantic similarity graph, including: The plot attributes of each plot and the feature matrix are input into a preset semantic MLP model, so that the semantic MLP model calculates the cosine similarity between each plot according to the plot attributes of each plot, and generates a semantic similarity graph of the target region in combination with the feature matrix; The semantic feature matrix of the target region is generated by performing graph convolution operations on the semantic similarity graph and the feature matrix.

7. The region updating prediction method based on the multi-view graph neural network according to any one of claims 1-6, characterized in that, The update prediction model is obtained by training an initial update prediction model based on a plurality of historical update cases, including: The initial update prediction model is obtained by constructing based on a graph neural network model, and the initial update prediction model includes a spatial autocorrelation graph convolution network, a functional heterogeneity graph convolution network, a mode heterogeneity graph convolution network, a semantic similarity graph convolution network, a feature concatenation layer and a double-layer output layer; A plurality of historical update cases are obtained, and a training data set is constructed based on the plot update mode changes in each of the historical update cases; According to the training data set, the initial update prediction model is trained using a preset Softmax function and a cross-entropy loss function to determine each learnable parameter in the initial update prediction model, and then the update prediction model is obtained.

8. A region update prediction system based on multi-view graph neural network, characterized in that, It includes an acquisition module, a feature extraction module and a prediction module; The acquisition module is configured to acquire multi-dimensional planning data of each plot in the target region and a regional status map, each data type in the multi-dimensional planning data being a preset data type, and the regional status map including spatial relationships, current land use properties, current update modes, and plot attributes of each plot in the target region. The feature extraction module is configured to perform feature extraction on each of the multi-dimensional planning data to construct a feature matrix of the target region. The prediction module is configured to input the feature matrix and the regional status map into a preset update prediction model, so that the update prediction model generates a plurality of plot relationship graphs of different adjacency relationships based on the spatial relationships, current land use properties, current update modes, and plot attributes of each plot in the region, and then performs comprehensive prediction according to each of the plot relationship graphs to generate an update potential value and a future update mode of each plot, including: generating a spatial autocorrelation graph of the target region according to the spatial relationships of each plot and the feature matrix, and extracting a spatial feature matrix of the target region from the spatial autocorrelation graph; generating a functional heterogeneity graph of the target region according to the current land use properties of each plot and the feature matrix, and extracting a functional feature matrix of the target region from the functional heterogeneity graph in combination with the spatial autocorrelation graph; generating a mode heterogeneity graph of the target region according to the current update modes of each plot and the feature matrix, and extracting a mode feature matrix of the target region from the mode heterogeneity graph in combination with the spatial autocorrelation graph; generating a semantic similarity graph of the target region according to the plot attributes of each plot and the feature matrix, and extracting a semantic feature matrix of the target region from the semantic similarity graph; performing feature splicing on the spatial feature matrix, the functional feature matrix, the mode feature matrix, and the semantic feature matrix to obtain a fusion feature of the target region; and inputting the fusion feature into a preset double-layer output layer to cause the double-layer output layer to generate the update potential value and the future update mode of each plot. The update prediction model is obtained by training an initial update prediction model based on a plurality of historical update cases, and the initial update prediction model is obtained by constructing a graph neural network model.

Images