A surveying and mapping geographic information data management method and system based on cloud computing

Abstract

The present application discloses a surveying and mapping geographic information data management method and system based on cloud computing, which relates to the field of geographic information, including: constructing a semantic segmentation model based on the convolutional neural network Deep Lab, using the semantic segmentation model to extract features and semantically annotate the surveying and mapping geographic data set, and obtaining a geographic element set; using the R-tree spatial indexing method to spatially divide and index the geographic element set, and obtain a hierarchically organized geographic element tree; constructing a graph neural network model based on the graph convolutional network GCN, using the graph neural network model to model the topological relationship of the geographic element tree, and obtaining a spatiotemporal attribute graph containing geographic semantics and topological relationships; constructing a spatiotemporal prediction sequence model based on the spatiotemporal graph neural network STGCN, and using the spatiotemporal prediction sequence model to predict the evolution trend of the spatiotemporal attribute graph. In view of the difficulty in realizing dynamic prediction of geographic elements in the prior art, the present application predicts the spatiotemporal evolution trend of geographic elements and realizes dynamic prediction.

Description

Technical Field

[0001] The present application relates to the field of geographic information, and in particular to a surveying and mapping geographic information data management method and system based on cloud computing. Background Art

[0002] With the rapid development of modern surveying and mapping technology and the explosive growth of geographic information data, how to effectively manage and analyze massive, multi-source, and dynamically changing surveying and mapping geographic information data has become a major issue that needs to be solved in the field of geographic information science and engineering. Traditional surveying and mapping geographic information data management methods mainly use relational databases and file systems for static storage and retrieval, which is difficult to adapt to the current complex and changing geographical environment and application needs.

[0003] In this context, the rise of cloud computing technology has provided new ideas and means to solve the above challenges. Cloud computing platforms have the advantages of massive storage, high-performance computing, and on-demand services. They can achieve distributed storage, parallel processing, and elastic expansion of surveying and mapping geographic information data, greatly improving the efficiency and scalability of data management and analysis. In recent years, cloud GIS platforms represented by Google Maps and Baidu Maps have flourished, promoting the Internetization and popularization of geographic information services.

[0004] However, there are still many technical challenges and shortcomings in the process of using cloud computing platforms to manage surveying and mapping geographic information data. On the one hand, due to the characteristics of large data volume, diverse types, and uneven temporal and spatial distribution of surveying and mapping geographic data, how to achieve efficient extraction, organization and indexing of geographic elements in a cloud environment is an urgent problem to be solved. Traditional manual interpretation and regular processing methods are difficult to adapt to the processing needs of massive surveying and mapping data, and it is urgent to introduce intelligent and automated feature extraction and semantic interpretation technology.

[0005] On the other hand, there are complex semantic associations and topological relationships between geographic elements. How to build an association model that reflects the spatiotemporal evolution of geographic elements on the cloud platform is the key to achieving dynamic prediction of geographic elements. Traditional static topological relationship representation methods are difficult to describe the dynamic change process of geographic elements. It is urgent to introduce spatiotemporal data mining and machine learning technology to explore the evolution patterns and prediction laws of geographic elements in the spatiotemporal dimension.

[0006] At the same time, higher requirements are put forward for the dynamic monitoring and early warning forecast of geographic elements. Traditional surveying and mapping geographic information data management methods are difficult to achieve real-time prediction and scenario analysis of future development trends of geographic elements. It is urgent to introduce spatiotemporal big data analysis and deep learning technology to build a spatiotemporal intelligent model for dynamic prediction, and provide scientific decision-making support for applications such as disaster prevention and mitigation, and land planning. Summary of the invention

[0007] In view of the difficulty in realizing dynamic prediction of geographic elements in surveying and mapping geographic information in the prior art, the present application provides a surveying and mapping geographic information data management method and system based on cloud computing, which predicts the spatiotemporal evolution trend of geographic elements and realizes dynamic prediction by constructing a spatiotemporal topological relationship model.

[0008] The purpose of this application is achieved through the following technical solutions.

[0009] One aspect of the present application provides a surveying and mapping geographic information data management method based on cloud computing, including: S1, obtaining a surveying and mapping geographic data set; wherein the surveying and mapping geographic data set contains a variety of geographic elements; the geographic elements include buildings, roads, water systems and vegetation; S2, constructing a semantic segmentation model based on the convolutional neural network Deep Lab, using the semantic segmentation model to extract features and semantically annotate the surveying and mapping geographic data set to obtain a geographic element set; S3, using the R-tree spatial indexing method to spatially divide and index the geographic element set to obtain a hierarchically organized geographic element tree; S4, constructing a graph neural network model based on the graph convolutional network GCN, using the graph neural network model to model the topological relationship of the geographic element tree, and obtaining a spatiotemporal attribute graph containing geographic semantics and topological relationships; S5, constructing a spatiotemporal prediction sequence model based on the spatiotemporal graph neural network STGCN, and using the spatiotemporal prediction sequence model to predict the evolution trend of the spatiotemporal attribute graph.

[0010] Furthermore, S2 obtains a set of geographic features, including: S21, constructs a semantic segmentation model, the semantic segmentation model adopts an encoder-decoder architecture, and uses the convolutional neural network Deep Lab as the backbone network; wherein, the encoder extracts features from the surveying and mapping geographic data set through convolution and pooling operations to obtain a multi-scale semantic feature map; the decoder upsamples and fuses the multi-scale semantic feature map through deconvolution and jump connection operations to obtain a semantic segmentation map; S22, uses the connected domain analysis algorithm to perform pixel-level regional clustering on the semantic segmentation map, divides pixels with the same semantic label and spatially connected to each other into the same region, and forms an initial pixel region mask; S23, performs regional grouping on the initial pixel region mask, and divides the regions of different geographic features into different regions according to the predefined geographic feature category system. The pixel area masks are classified to obtain a pixel area mask set organized by geographic feature category; S24, according to the pixel area mask set organized by geographic feature category, a minimum enclosing rectangle algorithm is used to calculate the minimum enclosing rectangle bounding box of each pixel area mask, the minimum enclosing rectangle bounding box is used as the geometric boundary attribute of the geographic feature, and the geometric boundary attribute is associated with the corresponding semantic category label to obtain the corresponding geographic feature object; S25, each geographic feature object is organized into a geographic feature set, wherein each geographic feature object includes a semantic category label, a pixel area mask and a minimum enclosing rectangle bounding box.

[0011] Further, S24, obtaining the corresponding geographic feature object, includes: traversing the pixel area mask set, and using the minimum enclosing rectangle algorithm to calculate the minimum enclosing rectangle bounding box of each pixel area mask; using the position and size of the minimum enclosing rectangle bounding box as the geometric boundary attributes of the corresponding pixel area mask; associating the obtained geometric boundary attributes with the semantic category labels corresponding to the pixel area mask to construct a geographic feature object; wherein the geographic feature object includes the semantic category label, the pixel area mask and the geometric boundary attributes; according to the association result, organizing the geographic feature objects corresponding to each pixel area mask according to the semantic category labels to construct a geographic feature set.

[0012] Further, S3, a hierarchically organized geographic element tree is obtained, including: S31, according to the minimum bounding rectangle of the geographic element object, the R-tree spatial index method is used to spatially divide the geographic element set, and the geographic element objects are recursively aggregated into a nested tree structure to obtain a geographic element tree; wherein each node of the R-tree represents a geographic area, including the minimum bounding rectangle of all its child nodes; the leaf node represents a single geographic element object; by recursively aggregating the minimum bounding rectangle of the geographic element object from the bottom up, a tree-structured spatial index is formed to achieve hierarchical organization and rapid retrieval of geographic elements; S32, at each node of the geographic element tree, the attribute information of all geographic element objects in the geographic area corresponding to the corresponding node is stored, including semantic category labels, pixel area masks and minimum bounding rectangles; S33, at each node of the geographic element tree, a semantic attribute hash table is introduced, and the hash index method is used to index the corresponding node The semantic category labels of all geographic feature objects in the corresponding geographic area of ​​the point are indexed; wherein, a hash table is constructed with the semantic category label as the key and the set of geographic feature objects containing the semantic category label as the value; by associating the semantic attribute hash table with the R-tree node, the rapid positioning and aggregation of the semantic attributes of the geographic features in different spatial areas are realized, and the query and analysis efficiency of the semantic attributes are improved; S34, at each node of the geographic feature tree, a spatial topological relationship graph is introduced, and a graph data structure is used to represent the spatial topological relationship between all geographic feature objects in the geographic area corresponding to the corresponding node; wherein, a directed graph is constructed with the geographic feature objects as nodes and the spatial topological relationship as the edge; the spatial topological relationship includes adjacent, contained, crossed, etc., which can be annotated by the attributes of the edge; by associating the spatial topological relationship graph with the R-tree node, the efficient organization and query of the spatial topological relationship of the geographic features in different spatial areas are realized, and the graph algorithm is used to perform complex spatial pattern mining and association analysis.

[0013] The geographic feature tree uses the R-tree spatial indexing method to achieve spatial partitioning and indexing of geographic feature collections, forming a hierarchical tree-like organizational structure. At each node of the tree, by associating geographic feature attribute information, semantic attribute hash tables, and spatial topological relationship graphs, attribute storage, semantic retrieval, and topological query of geographic features at different spatial granularities are achieved. The hierarchical organizational structure of the R-tree node is combined with the semantic attribute hash table and spatial topological relationship graph to improve the efficiency of spatial retrieval, attribute analysis, and association mining of geographic feature data.

[0014] Further, S4, a spatiotemporal attribute graph containing geographic semantics and topological relationships is obtained, including: S41, using the nodes of the geographic element tree as graph nodes, and the parent-child relationships between the nodes of the geographic element tree as graph edges, to generate a geographic element graph with a tree topological structure; at each node of the geographic element graph, setting the attribute information of the geographic element object corresponding to the node, the attribute information including the semantic category label, the pixel area mask and the minimum circumscribed rectangular bounding box; S42, at each node of the geographic element graph, encoding the semantic attribute hash table into a semantic feature vector, and encoding the spatial topological relationship graph into a topological feature vector; using the attention mechanism to weightedly fuse the semantic feature vector and the topological feature vector to obtain a node feature vector; S43, using the node feature vector as the node feature of the graph convolutional network GCN, and using GCN to extract features from the geographic element graph of step S42 to obtain the node feature vector of the geographic element graph. Point embedding vector; S44, using the node embedding vector of the geographic feature map, construct the semantic topological relationship matrix between nodes, and by calculating the semantic similarity and spatial distance between the node embedding vectors, obtain the edge weight matrix representing the semantic similarity and topological connectivity between nodes; S45, using the node embedding vector and the semantic topological relationship matrix, construct a spatiotemporal attribute graph containing geographic semantics and topological relationships; wherein the nodes of the spatiotemporal attribute graph are geographic feature objects, the node features are node embedding vectors, the edges of the spatiotemporal attribute graph are the semantic topological relationships between geographic feature objects, and the edge weights are the semantic topological relationship matrix; timestamp attributes are set on nodes and edges to reflect the time evolution process of geographic features and semantic topological relationships; wherein the node feature vector is the basis of the node embedding vector: the node feature vector is the initial feature representation obtained by fusing the semantic attributes and spatial topological relationships of geographic features. The node feature vector contains the semantic information and topological structure information of geographic features, providing a basis for subsequent feature extraction and node embedding.

[0015] The node embedding vector is a high-level representation of the node feature vector: the node embedding vector is obtained by extracting the node feature vector through the graph convolutional network GCN. GCN transforms the node feature vector into a more abstract and high-level feature representation through convolution operations, capturing the local and global information of the node. Based on the node feature vector, the node embedding vector learns the semantic and topological features of the node in the geographic feature map.

[0016] The node embedding vector inherits and expands the information of the node feature vector: The node embedding vector inherits the semantic attributes and spatial topological relationship information in the node feature vector. Through the feature extraction process of GCN, the node embedding vector not only retains the original information of the node feature vector, but also learns the high-level features of the node in the graph structure. The node embedding vector expands the representation ability of the node feature vector and can better characterize the role and relationship of the node in the geographic feature map.

[0017] Further, S44, an edge weight matrix representing the semantic similarity and topological connectivity between nodes is obtained, including: obtaining a semantic similarity matrix representing the nodes by calculating the cosine similarity between the node embedding vectors of the geographic element map obtained in step S43; obtaining a spatial distance matrix representing the nodes by calculating the Euclidean distance between the center coordinates of the minimum circumscribed rectangular bounding box of the geographic element object corresponding to the node; normalizing the semantic similarity matrix and the spatial distance matrix by the softmax algorithm to obtain a semantic similarity weight matrix and a spatial distance weight matrix; utilizing the attention mechanism to weightedly fuse the semantic similarity weight matrix and the spatial distance weight matrix to obtain an edge weight matrix that fuses the semantic similarity and spatial distance; constructing a semantic topological relationship matrix according to the edge weight matrix as the edge attribute of the geographic element map.

[0018] Furthermore, an edge weight matrix integrating semantic similarity and spatial distance is obtained, including: setting a multi-head attention mechanism, each attention head corresponds to a different semantic and spatial scale, and learning the contribution of semantic similarity and spatial distance to edge weight at different scales; when calculating the semantic similarity weight matrix, setting the embedding vector of the semantic category label, and obtaining the semantic similarity weight of the fused node semantic feature and category semantic feature by calculating the attention weight between the node embedding vector obtained in step S43 and the embedding vector of the semantic category label; when calculating the spatial distance weight matrix, setting the spatial direction relationship and topological relationship of the geographic element, and calculating the attention weight between the geographic element object in step S41 The directional relationship and topological relationship between them are used to obtain the distance weight that integrates the spatial distance and directional topological relationship; the directional relationship is obtained by calculating the azimuth and angle between the center coordinates of the minimum circumscribed rectangular bounding box of the geographic element, which is used to reflect the directional distribution of the geographic element in space; the topological relationship is obtained by analyzing the intersection, inclusion or adjacency relationship between the minimum circumscribed rectangular bounding boxes of the geographic elements, which is used to reflect the topological connectivity of the geographic elements in space; the semantic similarity weights and spatial distance weights generated by different attention heads, as well as the semantic similarity weight matrix and the spatial distance weight matrix are weightedly fused through the attention mechanism to obtain the edge weight matrix that integrates the semantic similarity and spatial distance.

[0019] Further, S5, constructs a spatiotemporal prediction sequence model based on the spatiotemporal graph neural network STGCN, including: S51, constructs a spatiotemporal prediction sequence model based on the spatiotemporal graph neural network STGCN, wherein the input of the spatiotemporal prediction sequence model is the spatiotemporal attribute graph obtained in step S45, and the output is the predicted spatiotemporal attribute graph for multiple future moments; the spatiotemporal prediction sequence model adopts an encoder-decoder architecture, the encoder uses a graph convolution layer and a time convolution layer to encode the node features and edge features of the spatiotemporal attribute graph; the gated recurrent unit GRU is used as the backbone network for extracting temporal features; the decoder uses graph deconvolution and time deconvolution to decode the encoded features to generate a spatiotemporal attribute graph for a period of time in the future; in the graph convolution layer and the time convolution layer After the layer, a graph attention layer and a time attention layer are respectively set to adjust the importance between different nodes and different time steps; S52, based on the acquired surveying and mapping geographic data set, a training sample set is generated by a data enhancement method; S53, based on the training sample set, a classroom forcing method is used to train the spatiotemporal prediction sequence model, with the training samples as input and the real spatiotemporal attribute graph of a period of time in the future as the supervision signal, and the parameters of the spatiotemporal prediction sequence model are optimized by minimizing the difference between the predicted spatiotemporal attribute graph and the real spatiotemporal attribute graph; S54 uses the trained spatiotemporal prediction sequence model to perform spatiotemporal sequence prediction on the obtained spatiotemporal attribute graph, recursively generates predicted spatiotemporal attribute graphs for multiple moments in the future, and obtains the spatiotemporal evolution trend of geographic elements.

[0020] Furthermore, the mean square error loss function is used to minimize the difference between the predicted spatiotemporal attribute graph and the true spatiotemporal attribute graph; the expression of the mean square error loss function is: ,in, represents the value of the improved loss function, which takes into account spatial dependency and attribute diversity. T represents the number of predicted time steps, that is, the number of moments in the future. N represents the number of nodes in the spatiotemporal attribute graph. K represents the dimension of the node attribute, that is, the number of attribute types for each node. It represents the predicted value of the kth attribute of the i-th node at the t-th time step. It is the output result of the spatiotemporal prediction sequence model at the t-th time step. The true value of the kth attribute of the ith node at the tth time step. It is the node attribute value of the real spatiotemporal attribute graph at the tth time step. The spatial weight of the i-th node reflects the importance of the node in the spatial topology. It can be calculated based on indicators such as the node's degree and centrality, or learned adaptively through the attention mechanism. The attribute weight of the kth attribute reflects the importance of the attribute in the prediction task. It can be set based on factors such as the relevance of the attribute, the difficulty of prediction, etc., or learned adaptively through the attention mechanism.

[0021] This application introduces spatial weights and attribute weights . Spatial weights take into account the importance of nodes in the spatial topological structure, assign higher weights to key nodes, and emphasize the prediction accuracy of key areas. Attribute weights take into account the importance of different attributes in the prediction task, assign higher weights to key attributes, and emphasize the prediction accuracy of key attributes. By introducing spatial weights and attribute weights, the improved loss function can better capture the key information in the spatiotemporal attribute graph, highlight the role of key nodes and key attributes in the prediction task, and improve the prediction performance of the model. During the model training process, by minimizing the improved loss function, the model pays more attention to key nodes and key attributes, thereby obtaining more accurate prediction results.

[0022] Another aspect of the present application also provides a surveying and mapping geographic information data management system based on cloud computing, which is used to execute a surveying and mapping geographic information data management method based on cloud computing of the present application.

[0023] Compared with the prior art, the advantages of this application are:

[0024] By constructing a graph neural network model based on the graph convolutional network GCN, the geographic element tree is transformed into a geographic element graph with a tree topological structure, and the semantic features and topological features are fused on the graph nodes, and the embedded vector representation of the node is extracted through the graph convolution operation. On this basis, the semantic similarity and spatial distance between the node embedding vectors are calculated, and weighted fusion is performed through the attention mechanism to construct a semantic topological relationship matrix that reflects the semantic connection and spatial proximity between geographic elements. Unlike traditional static topological relationship representations, the method of this application can characterize the dynamic association of geographic elements in semantic space and physical space, providing a solid foundation for spatiotemporal evolution analysis.

[0025] By constructing a spatiotemporal prediction sequence model based on the spatiotemporal graph neural network STGCN, the spatiotemporal attribute graph is used as input, and the spatiotemporal features are extracted using graph convolution and time convolution. The encoder-decoder architecture and attention mechanism are combined to generate predicted spatiotemporal attribute graphs for multiple moments in the future. The spatial dependency weights and attribute diversity weights are introduced into the loss function to improve the spatial consistency and attribute expression ability of the prediction results. Compared with the traditional time series prediction model, the method of this application can fully explore the spatiotemporal correlation patterns between geographical elements and make long-term and dynamic trend predictions for complex geographical processes.

[0026] By building a semantic segmentation model based on the convolutional neural network Deep Lab, using the encoder-decoder architecture and jump connections, we have achieved multi-scale feature extraction and pixel-level semantic annotation of surveying and mapping geographic data sets, overcoming the limitations of manual feature extraction and regularization processing in traditional methods, and improving the accuracy and automation of geographic feature extraction. At the same time, using connected domain analysis and the minimum enclosing rectangle algorithm, we have achieved regional segmentation and vector representation of geographic features, reducing data redundancy and improving storage and retrieval efficiency.

[0027] By adopting the R-tree spatial indexing method, the geographic feature set is divided into a nested tree structure to form a geographic feature tree. The semantic attribute hash table and the spatial topological relationship diagram are introduced on each tree node to realize the semantic index and topological index of the geographic features. Compared with the traditional grid division and linear indexing, the method of the present application makes full use of the spatial proximity and topological association of the geographic features, supports fast spatial query and association analysis, and improves the efficiency of data organization and retrieval. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] The present application will be further described in the form of exemplary embodiments, which will be described in detail by the accompanying drawings. These embodiments are not restrictive, and in these embodiments, the same number represents the same structure, wherein:

[0029] Figure 1 is an exemplary flow chart of a method for managing surveying and mapping geographic information data based on cloud computing according to some embodiments of the present application;

[0030] Figure 2 is an exemplary flow chart of generating a set of geographic elements according to some embodiments of the present application;

[0031] Figure 3 is an exemplary flow chart of generating a geographic element tree according to some embodiments of the present application;

[0032] Figure 4 is an exemplary flow chart of constructing a spatiotemporal attribute graph according to some embodiments of the present application;

[0033] Figure 5 This is an exemplary flowchart of constructing a semantic topological relationship matrix according to some embodiments of the present application. DETAILED DESCRIPTION

[0034] The method and system provided in the embodiments of the present application are described in detail below with reference to the accompanying drawings.

[0035] like Figure 1As shown, S1, obtains a surveying and mapping geographic data set; wherein the surveying and mapping geographic data set contains a variety of geographic elements; the geographic elements include buildings, roads, water systems and vegetation; S2, constructs a semantic segmentation model based on the convolutional neural network Deep Lab, and uses the semantic segmentation model to extract features and semantically annotate the surveying and mapping geographic data set to obtain a geographic element set; S3, uses the R-tree spatial indexing method to spatially divide and index the geographic element set to obtain a hierarchically organized geographic element tree; S4, constructs a graph neural network model based on the graph convolutional network GCN, and uses the graph neural network model to model the topological relationship of the geographic element tree to obtain a spatiotemporal attribute graph containing geographic semantics and topological relationships; S5, constructs a spatiotemporal prediction sequence model based on the spatiotemporal graph neural network STGCN, and uses the spatiotemporal prediction sequence model to predict the evolution trend of the spatiotemporal attribute graph.

[0036] Specifically, S1 obtains surveying and mapping geographic data sets; uses multi-source sensors and surveying and mapping equipment, such as satellite remote sensing, aerial photography, and mobile measurement systems, to collect data for the target area. These devices can obtain high-resolution remote sensing images, point cloud data, vector maps, and other types of surveying and mapping geographic data. The collected raw surveying and mapping data are preprocessed, including data cleaning, noise removal, data format conversion, and other operations. Through preprocessing, the data quality is improved, outliers and inconsistencies are eliminated, and the data is converted into a standard format that is easy to store and analyze, such as GeoTIFF, Shapefile, LAS, etc. Due to the diverse sources of surveying and mapping geographic data, data fusion and integration are required. Using information such as spatial coordinate systems and timestamps, data from different sources and types are aligned and synchronized to form a unified space-time benchmark. Through data fusion, multi-source data such as remote sensing images, vector maps, and terrain data are superimposed and integrated to construct a complete surveying and mapping geographic data set. The preprocessed and fused surveying and mapping geographic data sets are stored in the distributed file system and NoSQL database of the cloud platform. Distributed file systems such as Hadoop HDFS and Apache Parquet are used to store large-scale remote sensing images and point cloud data; NoSQL databases such as MongoDB and Cassandra are used to store vector maps and attribute information. Through distributed storage, highly reliable, highly available and scalable storage of surveying and mapping geographic data is achieved.

[0037] The surveying and mapping geographic data sets are annotated with geographic elements to identify key geographic elements such as buildings, roads, water systems, and vegetation contained in the data. Remote sensing images and vector maps are interpreted and annotated by combining rule-based feature extraction algorithms with manual annotation. For buildings, building areas can be identified by extracting features such as building edges and shadows, combined with attributes such as shape, size, and texture; for roads, road networks can be extracted by extracting linear features, intersections, and other features, combined with road network topological relationships; for water systems, water system elements such as rivers and lakes can be identified by extracting water body boundaries, spectral features, and other features, combined with terrain undulation information; for vegetation, vegetation indexes, texture features, and other features can be extracted, combined with seasonal change information, to identify different types of vegetation areas. According to the type and attributes of geographic elements, they are organized into structured geographic element data sets. Using a vector data model, geographic elements such as buildings, roads, water systems, and vegetation are represented as vector elements such as points, lines, and surfaces, and their spatial coordinates, geometric shapes, topological relationships, and other information are recorded. At the same time, attribute tables and relational models are used to store attribute information of geographic elements, such as building height, road width, water system depth, vegetation type, etc. Through vectorization and attribute representation of geographic elements, a structured and semantic geographic element dataset is formed.

[0038] like Figure 2As shown in Figure 1, S2 builds a semantic segmentation model based on the convolutional neural network Deep Lab, and uses the semantic segmentation model to extract features and semantically annotate the surveying and mapping geographic data set to obtain a set of geographic features; first, a semantic segmentation model based on the convolutional neural network Deep Lab is built. The semantic segmentation model adopts an encoder-decoder architecture. The encoder extracts multi-scale semantic features from the input surveying and mapping geographic data, and then restores the spatial resolution of the feature map through the decoder to obtain a semantic segmentation map with the same size as the input data. The encoder is responsible for feature extraction and downsampling, and the decoder is responsible for feature recovery and upsampling. The two realize the fusion of features of different scales through jump connections. The convolutional neural network Deep Lab is used as the backbone network of the semantic segmentation model. Deep Lab is a deep convolutional neural network suitable for semantic segmentation tasks, with powerful feature extraction and semantic understanding capabilities. It uses technologies such as atrous convolution and spatial pyramid pooling to capture the semantic information of objects at different scales and improve segmentation accuracy. In the encoder part, features are extracted from the input surveying and mapping geographic data by recursively applying convolution and pooling operations. The convolution operation uses convolution kernels of different scales and receptive fields to extract local and global semantic features. By setting convolution kernels of different sizes (such as 3x3, 5x5, 7x7, etc.), semantic information at different scales can be captured. At the same time, by adjusting the step size and padding of the convolution kernel, the spatial resolution of the feature map can be controlled. The pooling operation uses maximum pooling or average pooling to downsample the feature map and reduce the spatial resolution. The pooling operation can reduce the size of the feature map while retaining significant semantic features, allowing the model to focus on more global information. By recursively applying convolution and pooling operations, the encoder generates semantic feature maps of multiple scales to represent semantic information at different levels of abstraction. In the decoder part, the multi-scale semantic feature maps generated by the encoder are upsampled and feature fused through deconvolution and skip connection operations. The deconvolution operation restores the low-resolution semantic feature map to the original resolution by learning the upsampling convolution kernel. The upsampling convolution kernel is learned through back propagation and can restore the spatial detail information of the feature map. The skip connection operation splices the high-resolution feature map in the encoder with the feature map of the corresponding layer in the decoder to achieve the fusion of features of different scales. Through the skip connection, the decoder can use the shallow feature information in the encoder to supplement the spatial details lost in the recovery process. The skip connection can improve the gradient propagation, accelerate the convergence of the model, and improve the segmentation accuracy. Through the forward propagation of the encoder-decoder, a semantic segmentation map with the same size as the input data is finally obtained. Each pixel point in the semantic segmentation map corresponds to a semantic category label, indicating which geographical feature category the pixel belongs to (such as buildings, roads, water systems, vegetation, etc.).The semantic segmentation map has the same spatial resolution as the original surveying and mapping geographic data image, achieving pixel-level semantic annotation. This application makes full use of the encoder-decoder architecture and the feature extraction capabilities of convolutional neural networks to automatically learn and extract multi-scale semantic features from surveying and mapping geographic data, and restore the spatial resolution of the feature map through deconvolution and jump connection operations to generate a high-precision semantic segmentation map.

[0039] Then perform pixel-level region clustering, scan each pixel in the semantic segmentation map, and extract its semantic category label. Assume that the size of the semantic segmentation map is HxW, with a total of N pixels. The semantic category label of each pixel is represented by an integer, such as building-1, road-2, water system-3, vegetation-4, etc. Create a region labeling map of size HxW, and initially mark all pixels as 0, indicating that they have not been visited. Starting from the upper left corner, traverse each pixel in the semantic segmentation map in row-first order. For the current pixel (i, j), check its semantic category label label (i, j). If the region of the current pixel is marked as 0, indicating that it has not been visited, it is used as the starting point of a new connected region and marked as a new region number. , initialize an empty queue queue, add the coordinates of the current pixel point (i, j) to the queue. When the queue is not empty, repeat the following steps: Take a pixel point coordinate (x, y) from the queue. Check each pixel point in the 8-neighborhood or 4-neighborhood of the pixel point (x, y) If the pixel The semantic category label of is the same as the current pixel, that is, , and its area mark is 0, then mark it as the current area number , and add its coordinates to the queue. Repeat the above steps until the queue is empty, indicating that the current connected region has been expanded. Continue to traverse the remaining unvisited pixels, repeat until all pixels are visited and assigned to the corresponding connected regions. Finally, a region labeling map is obtained, in which the value of each pixel represents the number of the connected region to which it belongs. For each connected region, a binary pixel region mask is generated. The size of the mask is the same as the semantic segmentation map, and the pixels belonging to the connected region are marked as 1, and the remaining pixels are marked as 0.

[0040] Region grouping, classify the initial pixel region mask according to a predefined system of geographic feature categories, such as buildings, roads, water systems, vegetation, etc. Create a dictionary , used to store the pixel region mask set corresponding to each semantic category. The key of the dictionary is the semantic category label, and the value is a list that stores all pixel region masks belonging to the category. Traverse each pixel region mask and extract its corresponding semantic category label label. Add the pixel region mask to the dictionary The process is repeated until all pixel region masks are classified and added to the corresponding set. Finally, a set of pixel region masks organized by geographic feature category is obtained, and each set contains all pixel region masks belonging to the same geographic feature category.

[0041] Calculate the geometric boundary attributes, traverse the set of pixel area masks organized by geographic feature categories, and apply the minimum enclosing rectangle algorithm to each pixel area mask. For each pixel area mask, extract the coordinates of its boundary pixels. By scanning each pixel of the mask image, check whether there is a pixel with a value of 0 in its 4-neighborhood or 8-neighborhood. If so, mark the current pixel as a boundary point. Based on the coordinates of the boundary pixels, calculate the minimum rectangular bounding box that can completely contain the area. The coordinates of the upper left and lower right corners of the rectangle can be determined by finding the minimum and maximum horizontal and vertical coordinates of the boundary points. Record the position coordinates of the minimum enclosing rectangle and length and width (width, height). The position coordinates and size of the minimum enclosing rectangle are used as the geometric boundary attributes of the corresponding pixel region mask. Create a geographic feature object, which contains the following attributes: semantic category label label, indicating the category to which the geographic feature belongs. Pixel region mask mask, indicating the pixel-level representation of the geographic feature in the image. Geometric boundary attribute bbox, including the position coordinates and size of the minimum enclosing rectangle. Add the geographic feature object to the geographic feature collection of the corresponding semantic category. Repeat until all pixel region masks are processed and the corresponding geographic feature objects are generated.

[0042] Organize the geographic feature collection, group the geographic feature objects obtained by S24 according to the semantic category labels, and generate a geographic feature collection. Create a dictionary , used to store a list of geographic feature objects corresponding to each semantic category. The key of the dictionary is the semantic category label, and the value is a list that stores all geographic feature objects belonging to that category. Traverse each geographic feature object and extract its semantic category label label. Add the geographic feature object to the dictionary The process is repeated until all geographic feature objects are classified and added to the corresponding collection. Finally, a collection of geographic features organized by semantic category is obtained. Each collection contains all geographic feature objects belonging to the same geographic feature category, and each object contains attribute information such as semantic category label, pixel area mask, and minimum enclosing rectangular bounding box.

[0043] The present application realizes pixel-level regional clustering by using the connected domain analysis algorithm, and merges the pixels with the same semantic label and spatially interconnected in the semantic segmentation map into one region. Then, the initial pixel region mask is regionally grouped, and the pixel region mask is divided into different sets according to the predefined geographic element category system. Then, the geometric boundary attributes of each pixel region mask are calculated to generate the corresponding geographic element object. Finally, the geographic element objects are organized according to the semantic category label to form a geographic element set classified by the geographic element category. This data processing method makes full use of the pixel-level semantic information and spatial connectivity in the semantic segmentation results, and organizes discrete pixels into geographic element objects with semantic and geometric attributes through connected domain analysis and regional clustering. At the same time, by grouping geographic elements by semantic category, it is convenient to analyze and process different types of geographic elements in the future. Compared with traditional manual annotation and vectorization methods, this method realizes the automatic extraction and organization of geographic elements, greatly improving the efficiency and accuracy of data processing.

[0044] like Figure 3 As shown in S3, the R-tree spatial indexing method is used to spatially divide and index the geographic feature set to obtain a hierarchically organized geographic feature tree. First, the R-tree is used to spatially divide and index the geographic feature set. For each geographic feature set corresponding to each geographic feature category (such as buildings, roads, water systems, vegetation, etc.), each geographic feature object in it is traversed. For each geographic feature object, its minimum bounding rectangle (MBR) is extracted. The MBR is used to represent the approximate range of the geographic feature object in space, and is expressed by the upper left corner coordinates. and the lower right corner coordinates To define. The MBR of each geographic feature object can be calculated in the following way: If the geometry of the geographic feature object is represented as a pixel area mask, then traverse all pixel points in the mask, find the minimum and maximum x-coordinates and y-coordinates, and form the upper left corner and lower right corner coordinates of the MBR. If the geometry of the geographic feature object is represented as vector data (such as polygons, line segments, etc.), then traverse all vertices or control points, find the minimum and maximum x-coordinates and y-coordinates, and form the upper left corner and lower right corner coordinates of the MBR. Store the MBR of each geographic feature object in an array or list as input data for subsequent spatial indexing.

[0045] Construct an R-tree and select the R-tree as the spatial index method to spatially partition the geographic feature set. The R-tree is a multidimensional spatial index structure based on MBR, which can effectively support the rapid retrieval and query of spatial data. Recursively construct the R-tree to aggregate geographic feature objects into a nested tree structure. The specific steps are as follows: The MBRs of all geographic feature objects are used as leaf nodes of the R-tree, and each leaf node corresponds to a geographic feature object. The leaf node stores the identifier or pointer of the geographic feature object and its corresponding MBR. Spatially cluster the leaf nodes, and combine spatially adjacent leaf nodes into a higher-level node to form a non-leaf node of the R-tree. The MBR of a non-leaf node is the minimum circumscribed rectangle of the MBRs of all its child nodes. Spatial clustering can use different algorithms, such as k-means clustering, hierarchical clustering, etc. Here, k-means clustering is used as an example. Set the target number of clusters k, which represents the maximum number of child nodes for each non-leaf node. Randomly select k leaf nodes as the initial clustering centers. For each leaf node, calculate its distance to each cluster center (such as Euclidean distance, Manhattan distance, etc.), and assign it to the cluster with the nearest cluster center. For each cluster, calculate the minimum enclosing rectangle of the MBR of all its leaf nodes as the new cluster center of the cluster. Repeat the above two steps until the cluster center no longer changes or the maximum number of iterations is reached. Form each cluster into a non-leaf node, whose MBR is the minimum enclosing rectangle of the MBR of all leaf nodes of the cluster. Recursively perform spatial clustering and combination on non-leaf nodes until a root node is formed, that is, a complete R-tree structure is obtained. For each non-leaf node, use the MBR of all its child nodes (which may be leaf nodes or other non-leaf nodes) as input data and repeat the above clustering process. Recursively perform clustering and combination until a root node containing all geographic feature objects is obtained. In the process of building the R-tree, the balance and query efficiency of the tree are controlled by setting appropriate parameters (such as node capacity, split threshold, etc.). Node capacity refers to the maximum number of child nodes that each node can contain. It is generally set to the ratio of the disk page size to the node size to fully utilize the disk I / O performance. The split threshold indicates that when the number of child nodes of a node exceeds the node capacity, the node split operation is triggered. The split operation redistributes the child nodes of the node to two new nodes to maintain the balance of the tree. In the end, a hierarchically organized geographic feature tree is obtained. Each node of the tree corresponds to a geographic area and contains geographic feature objects in the area. The root node represents the entire geographic area, the leaf node represents a single geographic feature object, and the non-leaf node represents the division and aggregation of the geographic area. Each node stores the MBR of the corresponding geographic area, as well as information about the contained geographic feature objects or child nodes.

[0046] This application uses R-tree to perform spatial partitioning and indexing of geographic feature sets to obtain a hierarchically organized geographic feature tree. This method uses MBR to approximate geographic feature objects, and aggregates geographic feature objects into a nested tree structure through recursive spatial clustering and combination. By storing MBR information on each node, the spatial relationship between the query area and the geographic area corresponding to the node can be quickly determined, thereby reducing unnecessary node access and computational overhead. At the same time, the hierarchical organization of the R-tree can adapt to spatial query requirements of different scales and granularities. By recursively traversing the tree structure starting from the root node, the geographic area where the query area is located and the related geographic feature objects can be quickly located.

[0047] Secondly, store the attribute information. For each node of the geographic feature tree, obtain all the geographic feature objects in the area according to its corresponding geographic area. For each geographic feature object, extract its attribute information, including: semantic category label: indicates the category to which the geographic feature object belongs, such as buildings, roads, water systems, vegetation, etc. The semantic category label can be an integer value or a string, which is used to represent different types of geographic features. Pixel region mask: represents the pixel-level representation of the geographic feature object in the original image, represented by a binary mask. The pixel region mask is a two-dimensional array of the same size as the original image, in which the pixel values ​​belonging to the geographic feature object are 1, and the other pixel values ​​are 0. Minimum bounding rectangle (MBR): represents the approximate spatial range of the geographic feature object, represented by MBR. The MBR consists of the coordinates of the upper left corner. and the lower right corner coordinates Define the minimum bounding rectangle of a geographic feature object in space. Associate the extracted attribute information with the corresponding geographic feature object and store it in the corresponding node of the geographic feature tree. The association method can be to store the attribute information in the data structure of the geographic feature object, or to use a separate data structure (such as a dictionary or array) to store the attribute information and use the identifier of the geographic feature object as the key value.

[0048] For non-leaf nodes of the geographic element tree, their attribute information can be obtained by aggregating the attribute information of their child nodes. Specifically, the following strategies can be adopted: Count the number of each semantic category in the child node: traverse all child nodes of the non-leaf node, count the number of times each semantic category appears, and generate a semantic category-quantity mapping table. In this way, you can quickly understand the distribution of different types of geographic elements in the geographic area corresponding to the node. Calculate the union of the sub-node MBR: traverse all child nodes of the non-leaf node, obtain their MBRs, calculate the union of these MBRs, and obtain a minimum bounding rectangle that can contain all child nodes. The union MBR can be used as the MBR attribute of the non-leaf node, indicating the spatial range of the geographic area corresponding to the node. In this application, the attribute information of the geographic element object is stored in the corresponding node of the geographic element tree, and the attribute information of the non-leaf node is aggregated and calculated, so as to realize the hierarchical organization and management of the attribute information.

[0049] A semantic attribute hash index is constructed on each node of the geographic feature tree. A semantic attribute hash table is introduced on each node of the geographic feature tree to index the semantic category labels of all geographic feature objects in the geographic area corresponding to the node. The key of the hash table is the semantic category label, and the value is a list that stores the identifiers or pointers of all geographic feature objects belonging to the semantic category. For each geographic feature object, its semantic category label is extracted and inserted into the semantic attribute hash table of the corresponding node. The specific steps are as follows: Calculate the hash value of the semantic category label: Use a hash function to map the semantic category label to an integer hash value. Common hash functions include: Direct addressing method: Use the semantic category label directly as the index of the hash table, which is suitable for the case where the semantic category label is a continuous integer. Division remainder method: Convert the semantic category label to an integer, and then take the modulus of the size of the hash table to obtain the hash value. Square middle method: Convert the semantic category label to an integer, then square it, and take the middle digits as the hash value. Random number method: Multiply the semantic category label with a random number, and then take the modulus of the size of the hash table to obtain the hash value. Resolving hash conflicts: When different semantic category labels are mapped to the same hash value, a hash conflict occurs. Common methods for resolving hash conflicts are: Chained hashing: Store geographic feature objects with the same hash value in the same bucket in the form of a linked list. Open addressing: When a hash conflict occurs, search for the next empty bucket in the hash table according to a certain detection sequence until an empty bucket is found or the hash table is determined to be full.

[0050] Insert geographic feature object: According to the calculated hash value, insert the identifier or pointer of the geographic feature object into the corresponding bucket. If chain hashing is used, insert it into the linked list of buckets; if open addressing is used, insert it into the empty bucket detected. Through the semantic attribute hash table, geographic feature objects of a specific semantic category can be quickly retrieved and queried. Given a semantic category label, calculate its hash value, locate the corresponding bucket in the hash table, and then traverse the geographic feature object identifiers or pointers in the bucket to obtain all geographic feature objects belonging to the semantic category.

[0051] Time complexity analysis of semantic attribute hash index: Insert operation: The average time complexity is In the worst case, you may need to traverse the entire hash table, and the time complexity is , where n is the number of geographic feature objects. Query operation: The average time complexity is In the worst case, you may need to traverse the entire hash table, and the time complexity is . Construct a spatial topological relationship graph. At each node of the geographic feature tree, introduce a spatial topological relationship graph to represent the spatial topological relationship between all geographic feature objects in the geographic area corresponding to the node. The spatial topological relationship graph is represented by a graph data structure, in which nodes represent geographic feature objects and edges represent the spatial topological relationship between geographic feature objects, such as intersection, adjacency, and inclusion. Common spatial topological relationships include: Intersects: Two geographic feature objects have common points inside, but do not completely contain each other. Touches: Two geographic feature objects have common points on their boundaries, but no common points inside. Contains: One geographic feature object completely contains another geographic feature object. Within: One geographic feature object is completely contained by another geographic feature object. Overlaps: Two geographic feature objects partially overlap, but do not completely contain each other. Disjoint: Two geographic feature objects do not have any common points.

[0052] For each geographic feature object, analyze the spatial position relationship between it and other geographic feature objects, and determine the spatial topological relationship between them according to certain topological relationship judgment rules. The judgment rules can be based on the geometric attributes of the geographic feature objects, such as MBR, pixel area mask, etc. Common judgment methods are: MBR intersection judgment: compare whether the MBRs of two geographic feature objects intersect. If they intersect, they are considered to have an intersection relationship. Pixel area mask overlap judgment: perform logical AND operation on the pixel area masks of two geographic feature objects. If the result is not empty, they are considered to have an overlapping relationship. Boundary intersection judgment: calculate whether the boundaries of the two geographic feature objects have an intersection. If there is an intersection, they are considered to have an adjacent relationship. Containment relationship judgment: judge whether the pixel area mask of one geographic feature object completely contains the pixel area mask of another geographic feature object. If it does, it is considered to have a containment relationship. In the spatial topological relationship diagram, add nodes representing geographic feature objects, and establish directed edges or undirected edges between nodes according to the spatial topological relationship between them. For symmetrical relationships such as intersection, adjacency, and overlap, use undirected edges to connect two nodes. For asymmetric relationships such as containment and being contained, a directed edge is used to connect two nodes, and the direction of the directed edge indicates the direction of the relationship.

[0053] By traversing the spatial topological relationship graph, you can quickly query and analyze the spatial relationship between geographic feature objects. Common query and analysis operations include: Find adjacent geographic feature objects: For a given geographic feature object, traverse the edges connected to it to obtain adjacent geographic feature objects. Determine whether two geographic feature objects intersect: Find the nodes representing the two geographic feature objects in the spatial topological relationship graph, and determine whether there is an edge connection between them. Find the containment relationship: For a given geographic feature object, traverse the directed edges starting from the object to obtain the geographic feature objects contained by it. Analyze the transitivity of spatial relationships: Use graph traversal algorithms, such as depth-first search or breadth-first search, to analyze indirect relationships between geographic feature objects, such as whether there is a path between two geographic feature objects connected through other geographic feature objects. Time complexity analysis of spatial topological relationship graph: Constructing a spatial topological relationship graph: It is necessary to compare the spatial relationship between each pair of geographic feature objects, and the time complexity is , where n is the number of geographic feature objects. Query operation: By traversing the spatial topological relationship graph, the time complexity of the query is related to the number of edges in the graph. In the worst case, , where n is the number of geographic feature objects and m is the number of edges in the spatial topological relationship graph.

[0054] In this application, a semantic attribute hash index and a spatial topological relationship graph are constructed on each node of the geographic element tree, which can realize efficient retrieval and spatial relationship analysis of geographic element objects. The semantic attribute hash index indexes the geographic element objects according to semantic categories in the form of a hash table, so as to realize fast query and retrieval of geographic element objects of a specific category. The spatial topological relationship graph uses a graph data structure to represent the spatial topological relationship between geographic element objects. By traversing the graph, the spatial relationship between geographic element objects, such as adjacency, intersection, and inclusion, can be quickly analyzed. The combination of these two index structures provides efficient support for the retrieval and analysis of geographic element objects. Through the semantic attribute hash index, geographic element objects of a specific category can be quickly located; through the spatial topological relationship graph, the spatial relationship between these geographic element objects can be further analyzed to discover potential patterns and laws. This composite index structure is suitable for the organization and management of large-scale geographic element data, and improves the efficiency of data processing and analysis.

[0055] like Figure 4 As shown in S4, a graph neural network model based on the graph convolutional network GCN is constructed, and the graph neural network model is used to model the topological relationship of the geographic feature tree to obtain a spatiotemporal attribute graph containing geographic semantics and topological relationships. First, a geographic feature map is generated: the nodes of the geographic feature tree are used as graph nodes, and the attribute information of the nodes includes: semantic category labels: indicating the semantic category of the geographic feature object corresponding to the node, such as buildings, roads, water systems, vegetation, etc. Pixel area mask: indicating the pixel-level representation of the geographic feature object corresponding to the node in the image, represented by a binary mask. Minimum circumscribed rectangular bounding box: indicating the spatial range of the geographic feature object corresponding to the node, represented by the coordinates of the rectangular bounding box. The parent-child relationship between the nodes of the geographic feature tree is used as the graph edge to represent the topological relationship between the nodes. The generated geographic feature map has a tree-like topological structure, reflecting the hierarchical relationship between geographic feature objects.

[0056] Perform feature fusion and semantic feature vector encoding for each node in the geographic feature graph: For each node in the geographic feature graph, extract the semantic attribute hash table of the corresponding geographic feature object. According to the semantic category label in the semantic attribute hash table, assign a unique integer code to each semantic category. For example, "building" can be encoded as 0, "road" can be encoded as 1, "water system" can be encoded as 2, "vegetation" can be encoded as 3, and so on. For each node, generate a one-hot encoded vector according to the integer encoding of its semantic category label. The length of the one-hot encoded vector is equal to the total number of semantic categories. The position corresponding to the semantic category in the vector is 1, and the rest of the positions are 0. For example, if the semantic category label of a node is "building" and the corresponding integer code is 0, assuming that the total number of semantic categories is 4, the one-hot encoded vector of the node is [1, 0, 0, 0]. For another example, if the semantic category label of a node is "water system" and the corresponding integer code is 2, the one-hot encoded vector of the node is [0, 0, 1, 0]. The generated one-hot encoded vector is used as the semantic feature vector of the node to represent the semantic attributes of the node.

[0057] Topological feature vector encoding: For each node in the geographic feature map, analyze its connection relationship with other nodes in the spatial topological relationship map. Generate a connection relationship vector, the length of which is equal to the total number of nodes in the geographic feature map. For each node, traverse the other nodes connected to it in the spatial topological relationship map: if there is a connection relationship between the two nodes (such as adjacent, contained, etc.), set the corresponding position in the connection relationship vector to 1. If there is no connection relationship between the two nodes, set the corresponding position in the connection relationship vector to 0. Use the generated connection relationship vector as the topological feature vector of the node to represent the connection mode of the node in the spatial topological relationship map.

[0058] Attention mechanism fusion: Set an attention weight matrix to indicate the contribution of semantic features and topological features to node features. The dimension of the attention weight matrix is ​​(semantic feature vector length, topological feature vector length). For each node, multiply its semantic feature vector and topological feature vector by the attention weight matrix respectively to obtain weighted semantic features and topological features. Add the weighted semantic features and topological features to obtain the fused node feature vector. The attention weight matrix can be learned and optimized through training data to adapt to different geographic feature data sets. The following methods can be used to learn the attention weight matrix: use the attention weight matrix as a trainable parameter of the model and update it as the model is trained. Using supervised learning, calculate the loss function based on the node label information (such as semantic category, spatial position, etc.), and update the parameters of the attention weight matrix through the back propagation algorithm. Through multiple iterations and optimizations, the attention weight matrix can learn the importance of semantic features and topological features on different nodes and perform appropriate weighted fusion on them.

[0059] In this application, the semantic feature vector encoding process converts the semantic attributes of the node into a one-hot encoding vector, which represents the semantic category information of the node. The topological feature vector encoding process converts the connection relationship of the node in the spatial topological relationship graph into a connection relationship vector, which represents the topological structure information of the node. The semantic feature vector and the topological feature vector are weightedly fused through the attention mechanism to obtain a node feature vector that comprehensively considers the semantic attributes and topological structure. The introduction of the attention weight matrix enables the model to adaptively learn the importance of different features and dynamically adjust them to adapt to different geographic feature data sets.

[0060] like Figure 5 As shown, graph convolutional network feature extraction: The graph convolutional network GCN is used to extract features from the geographic feature map obtained in step S42. The graph convolutional network takes the node feature vector as input and extracts the high-level feature representation of the node through convolution and pooling operations. The convolution operation updates the node feature vector by aggregating the neighborhood information of the node; the pooling operation reduces the feature dimension by downsampling the node feature vector. After multiple layers of graph convolution and pooling operations, the node embedding vector of the geographic feature map is obtained, which represents the semantic and topological features of the node.

[0061] Construct a semantic topological relationship matrix and calculate a semantic similarity matrix: Use the node embedding vectors of the geographic element map obtained in step S43 to calculate the semantic similarity between nodes. For each pair of nodes, calculate the cosine similarity between their embedding vectors to obtain a semantic similarity matrix. The calculation formula for cosine similarity is: , where A and B are the embedding vectors of the two nodes, and Respectively represent the modulus of the vector. The value range of cosine similarity is [-1, 1]. The larger the value, the closer the directions of the two vectors are, and the higher the semantic similarity.

[0062] Calculate the spatial distance matrix: For each pair of nodes, calculate the Euclidean distance between the center coordinates of the minimum bounding rectangle of the corresponding geographic feature objects to obtain a spatial distance matrix. The calculation formula of the Euclidean distance is: ,in and are the center coordinates of the two nodes. The Euclidean distance measures the straight-line distance between two coordinate points. The smaller the distance, the closer the spatial positions are and the higher the topological connectivity.

[0063] Normalization: Normalize the semantic similarity matrix and the spatial distance matrix respectively, and convert the elements in the matrix into probability distribution. Use the softmax function to normalize the matrix. The calculation formula of the softmax function is: ,in are the elements in the matrix, is the exponential sum of all elements in the matrix. The normalized semantic similarity matrix represents the semantic similarity weights between nodes, and the normalized spatial distance matrix represents the spatial distance weights between nodes.

[0064] Attention mechanism fusion: Set up a multi-head attention mechanism, where each attention head corresponds to a different semantic and spatial scale. When calculating the semantic similarity weight matrix, introduce the embedding vector of the semantic category label: for each semantic category, learn an embedding vector to represent the semantic features of the category. For each node, obtain the corresponding category embedding vector based on its semantic category label. Through the attention mechanism, calculate the attention weight between the node embedding vector and the category embedding vector, and obtain the semantic similarity weight of the fused node semantic features and category semantic features.

[0065] In this application, the spatial direction relationship and topological relationship of geographic elements are introduced when calculating the spatial distance weight matrix, and a semantic topological relationship matrix is ​​constructed to calculate the spatial direction relationship: for each pair of nodes in the geographic element graph, the azimuth and angle between their corresponding geographic elements are calculated to represent the spatial direction relationship. Azimuth calculation: for two geographic element objects A and B, the center coordinates of their minimum enclosing rectangular bounding box are calculated. and . Calculate the direction vector of vector AB ,in , . Calculate the azimuth angle θ, , where atan2 is the inverse tangent function, and the return value range is [-π, π]. Convert the azimuth angle θ to the degree system, , get the azimuth with the east direction as 0 degrees and rotated counterclockwise. Angle calculation: For two geographic feature objects A and B, calculate the coordinates of the four vertices of their minimum enclosing rectangular bounding box. Calculate the main direction vector of the bounding box , you can choose the long side or short side of the bounding box as the main direction vector. Calculate the vector and Angle , ,in represents the dot product of vectors, and represents the magnitude of the vector. Convert to degrees. , get the angle representing the relative direction between geographic feature objects.

[0066] Analyze topological relationships: For each pair of nodes in the geographic feature map, analyze the intersection, inclusion, and adjacency relationships between the minimum bounding rectangles of the corresponding geographic features to represent the topological relationship. Determine intersection relationships: For two geographic feature objects A and B, determine whether their minimum bounding rectangles intersect. You can use the coordinates of the lower left corner and upper right corner of the bounding box to make a judgment, that is, determine and Whether there is an overlap. If the two bounding boxes intersect, it is considered that geographic feature objects A and B have an intersection relationship, recorded as A∩B. Judgment of the inclusion relationship: For two geographic feature objects A and B, determine whether the minimum enclosing rectangular bounding box of one geographic feature completely contains the bounding box of the other geographic feature. If the bounding box of A completely contains the bounding box of B, it is considered that geographic feature object A contains B, recorded as A⊃B. If the bounding box of B completely contains the bounding box of A, it is considered that geographic feature object B contains A, recorded as B⊃A.

[0067] Judgment of adjacency relationship: For two geographic feature objects A and B, determine whether their minimum circumscribed rectangular bounding boxes have common edges or vertices. If the two bounding boxes have common edges or vertices but do not completely overlap, the geographic feature objects A and B are considered to have an adjacency relationship, recorded as A to B. Attention mechanism fusion: Through the attention mechanism, the spatial distance weight is calculated together with the attention weights of the directional relationship and the topological relationship to obtain the distance weight that integrates the spatial distance and directional topological relationship. Calculation of spatial distance weight: For each pair of nodes, the corresponding elements in the calculated spatial distance matrix are used as the initial spatial distance weight. Through the attention mechanism, the importance of the spatial distance weight is learned to obtain the adjusted spatial distance weight.

[0068] Calculation of directional relationship weights: For each pair of nodes, the calculated azimuth and angle are used to construct a directional relationship feature vector. Through the attention mechanism, the importance of the directional relationship feature vector is learned to obtain the directional relationship weight. Calculation of topological relationship weights: For each pair of nodes, the topological relationship feature vector is constructed using the analyzed intersection, inclusion, and adjacency relationships. Through the attention mechanism, the importance of the topological relationship feature vector is learned to obtain the topological relationship weight. Calculation of fusion weights: The spatial distance weight, directional relationship weight, and topological relationship weight are weighted and fused to obtain a fusion weight that comprehensively considers the spatial distance, directional relationship, and topological relationship. The calculation of fusion weights can use weighted average, weighted sum, and other methods, and the distribution of weights can be learned through the attention mechanism.

[0069] Multi-head attention fusion: Through the attention mechanism, the semantic similarity weights and spatial distance weights generated by different attention heads are weighted and fused to obtain the final edge weight matrix. Set multiple attention heads, each corresponding to different semantic and spatial scales, and learn the semantic similarity weights and spatial distance weights at different scales. For each attention head, calculate the semantic similarity weight and spatial distance weight respectively to obtain the attention head-specific edge weight matrix. Through the attention mechanism, learn the importance of different attention heads and obtain the weights of the attention heads. The edge weight matrices of different attention heads are weighted and fused to obtain the final edge weight matrix, which comprehensively considers the semantic similarity and spatial distance information at different scales.

[0070] Constructing the semantic topological relationship matrix: Based on the final edge weight matrix obtained, construct the semantic topological relationship matrix as the edge attribute of the geographic feature graph. The semantic topological relationship matrix is ​​an N×N matrix, where N is the number of nodes in the geographic feature graph. The rows and columns of the semantic topological relationship matrix correspond to the nodes in the geographic feature graph, and the elements in the matrix represent the semantic similarity and topological connectivity between the nodes. The elements in the semantic topological relationship matrix can directly use the corresponding values ​​in the edge weight matrix, or they can be normalized to map the weights to the range of [0, 1]. Through the above steps, the spatial direction relationship and topological relationship of the geographic elements can be introduced when calculating the spatial distance weight matrix, and the attention mechanism can be used to fuse them to obtain an edge weight matrix that comprehensively considers the spatial distance, direction relationship and topological relationship. At the same time, through the multi-head attention mechanism, the semantic similarity weights and spatial distance weights can be learned at different scales, and fused to obtain the final edge weight matrix. This method can capture the association characteristics of geographic elements at different semantic and spatial scales, and improve the accuracy and richness of relationship representation. Finally, a semantic topological relationship matrix is ​​constructed based on the edge weight matrix as the edge attribute of the geographic feature graph, which represents the semantic similarity and topological connectivity between nodes. The semantic topological relationship matrix provides important relationship information for subsequent geographic feature analysis and application. By introducing spatial direction relationships and topological relationships and fusing them using the attention mechanism, the spatial relationship between geographic features can be more comprehensively characterized, important spatial patterns such as direction, adjacency, and inclusion can be captured, and the quality and richness of the representation of geographic feature relationships can be improved.

[0071] Construct a spatiotemporal attribute graph containing geographic semantics and topological relationships, and construct nodes of the spatiotemporal attribute graph: the nodes of the spatiotemporal attribute graph are geographic feature objects, and each node represents a specific geographic entity, such as a building, road, river, etc. For each geographic feature object, the node embedding vector obtained in step S43 is used as the feature representation of the node. The node embedding vector is obtained by extracting features from the geographic feature graph using a graph convolutional network, and represents the semantic and topological features of the geographic feature. The node embedding vector can be a real-valued vector of a fixed length, such as a vector of length 256, and each element represents a specific semantic or topological attribute.

[0072] Construct the edges of the spatiotemporal attribute graph: The edges of the spatiotemporal attribute graph represent the semantic topological relationships between geographic feature objects, such as similarity, adjacency, inclusion, etc. For each pair of geographic feature objects, the corresponding elements in the semantic topological relationship matrix obtained in step S44 are used as the weights of the edges. The semantic topological relationship matrix is ​​obtained by calculating the semantic similarity and spatial distance between node embedding vectors, and combining the directional relationship and the topological relationship. The elements in the semantic topological relationship matrix represent the semantic similarity and topological connectivity between geographic feature objects. The larger the weight, the stronger the relationship.

[0073] Set the timestamp attribute of the node: For each geographic feature object, set the corresponding timestamp attribute for the node according to its status and changes in different time periods. The timestamp attribute can be a continuous time value, such as year, month, day, hour, minute, second, etc., or a discrete time segment number, such as T1, T2, T3, etc. Setting of continuous time value: For each geographic feature object, record its status and attribute value at different time points. The status can include the existence, shape, size, etc. of the geographic feature, and the attribute value can include the category, name, function, etc. of the geographic feature. Associate the time point with the corresponding status and attribute value to form time series data, such as (time 1, status 1, attribute value 1), (time 2, status 2, attribute value 2), etc. Attach the time series data to the node as the timestamp attribute of the node. Setting of discrete time segment number: Divide the time axis into time segments of fixed length, such as each time segment represents a month or a year. For each geographic feature object, record its status and attribute value in different time segments. Associate the time segment number with the corresponding state and attribute value to form time segment data, such as (T1, state 1, attribute value 1), (T2, state 2, attribute value 2), etc. Attach the time segment data to the node as the timestamp attribute of the node.

[0074] Setting the timestamp attribute of the edge: For each semantic topological relationship edge, set the corresponding timestamp attribute for the edge according to its strength changes in different time periods. The timestamp attribute can be a continuous time value, such as year, month, day, hour, minute, second, etc., or a discrete time segment number, such as T1, T2, T3, etc. Setting continuous time values: For each semantic topological relationship edge, record its weight value at different time points. The weight value represents the strength of the semantic topological relationship between geographic feature objects, which can be obtained by calculating the similarity or distance between node embedding vectors. Associate the time point with the corresponding weight value to form time series data, such as (time 1, weight value 1), (time 2, weight value 2), etc. Attach the time series data to the edge as the timestamp attribute of the edge. Setting discrete time segment numbers: Divide the time axis into time segments of fixed length, such as each time segment represents a month or a year. For each semantic topological relationship edge, record its weight value in different time segments. Associate the time segment number with the corresponding weight value to form time segment data, such as (T1, weight value 1), (T2, weight value 2), etc. Attach the time slice data to the edge as the edge's timestamp attribute.

[0075] Construct a complete spatiotemporal attribute graph: combine the obtained nodes, the obtained edges, and the obtained timestamp attributes to form a complete spatiotemporal attribute graph. The nodes of the spatiotemporal attribute graph represent geographic feature objects, and the node features are node embedding vectors, which represent the semantic and topological features of the geographic features. The edges of the spatiotemporal attribute graph represent the semantic topological relationships between geographic feature objects, and the edge weights are semantic topological relationship matrices, which represent the semantic similarity and topological connectivity between geographic features. The timestamp attributes on the nodes and edges represent the states and changes of geographic features and semantic topological relationships in different time periods. This application uses node embedding vectors and semantic topological relationship matrices to construct a spatiotemporal attribute graph that contains geographic semantics and topological relationships. The spatiotemporal attribute graph not only captures the semantic and topological features of geographic features, but also describes the evolution of geographic features and semantic topological relationships over time through timestamp attributes.

[0076] S5, build a spatiotemporal prediction sequence model based on the spatiotemporal graph neural network STGCN, and use the spatiotemporal prediction sequence model to predict the evolution trend of the spatiotemporal attribute graph. First, build the spatiotemporal prediction sequence model: Input data: The input of the spatiotemporal prediction sequence model is the spatiotemporal attribute graph obtained in step S45, including node features, edge features, and timestamp attributes. Model architecture: The encoder-decoder architecture is adopted. The encoder is responsible for extracting the features of the spatiotemporal attribute graph, and the decoder is responsible for generating predicted spatiotemporal attribute graphs at multiple moments in the future. Encoder: Graph convolution layer: Perform convolution operations on the node features of the spatiotemporal attribute graph to extract the local features of the nodes. Temporal convolution layer: Perform convolution operations on the edge features of the spatiotemporal attribute graph to extract the features of the time series. Gated recurrent unit (GRU): GRU is used as the backbone network for extracting temporal features to capture the long-term dependencies of the spatiotemporal attribute graph. Graph attention layer: The graph attention layer is set after the graph convolution layer. The importance between different nodes is adjusted through the attention mechanism, and the node features are adaptively aggregated. Temporal attention layer: The temporal attention layer is set after the temporal convolution layer. The attention mechanism adjusts the importance between different time steps and adaptively aggregates temporal features. Decoder: Graph deconvolution layer: Deconvolution is performed on the features extracted by the encoder to generate node features for multiple future moments. Temporal deconvolution layer: Deconvolution is performed on the features extracted by the encoder to generate edge features for multiple future moments. The output of the decoder is a predicted spatiotemporal attribute graph for multiple future moments, including predicted node features and edge features.

[0077] Generate a training sample set and collect remote sensing image data, such as satellite images and aerial images, covering different time periods of the study area. Collect geographic information system (GIS) data, such as land use classification maps, road network maps, building distribution maps, etc., to reflect the distribution and attribute information of geographical elements in the study area. Ensure that remote sensing image data and GIS data are aligned in time and space, and perform necessary preprocessing, such as geometric correction and image fusion. Generate training samples through data enhancement methods: Time slicing: Divide the spatiotemporal attribute map into time segments of fixed length at equal intervals along the time axis, such as each time segment contains 5 consecutive time steps. Each time segment is used as a training sample. The input of the sample is the spatiotemporal attribute map of the first 4 time steps, and the output of the sample is the spatiotemporal attribute map of the 5th time step. Move the time segment on the time axis by sliding the window to generate multiple training samples. Spatial sampling: Sample the spatiotemporal attribute map spatially and select sub-maps in different areas as training samples. Use random sampling or sampling methods based on geographic location, such as grid sampling and radial sampling, to ensure that the sampled sub-maps cover different parts of the study area. The size of the subgraph can be adjusted according to the receptive field of the model and the computing resources, such as selecting a 64x64 or 128x128 subgraph. Attribute noise: Add random noise to the attribute values ​​of nodes and edges to increase the robustness of the model. For continuous attributes, such as temperature and humidity, add Gaussian noise or uniform noise, and the noise amplitude can be controlled within a certain range. For discrete attributes, such as land use type and road grade, randomly change the attribute value with a certain probability to simulate the uncertainty of the attribute. Rotation and flipping: Rotate and flip the spatiotemporal attribute graph to increase data diversity. Randomly select a rotation angle, such as 90 degrees, 180 degrees, 270 degrees, etc., to rotate the spatiotemporal attribute graph. Randomly select a flip direction, such as horizontal flip, vertical flip, etc., to flip the spatiotemporal attribute graph. Divide the generated training samples into input and output: For each training sample, take the spatiotemporal attribute graph of the first 4 time steps as input, and take the spatiotemporal attribute graph of the 5th time step as output. Form training sample pairs, each sample pair contains the input historical spatiotemporal attribute graph and the output future spatiotemporal attribute graph. All training sample pairs are combined into a training sample set for model training.

[0078] Training spatiotemporal prediction sequence model: The model is trained using the teacher forcing method, and the training sample set is divided into batches, each batch containing multiple training sample pairs. For each batch, the input historical spatiotemporal attribute graph is fed into the model to obtain the model's predicted output. The predicted output of the model is compared with the true future spatiotemporal attribute graph to calculate the loss function. Loss function: The mean squared error (MSE) loss function is used to measure the difference between the predicted spatiotemporal attribute graph and the true spatiotemporal attribute graph. The MSE loss function calculates the squared difference between the predicted value and the true value, and then takes the average. The goal is to minimize the MSE loss function so that the predicted spatiotemporal attribute graph is as close to the true spatiotemporal attribute graph as possible.

[0079] Improvement of loss function: Introduce spatial weight matrix, assign a spatial weight to each node according to the importance of the node in the spatiotemporal attribute graph. Importance can be measured by indicators such as node degree, centrality, connectivity, etc., and weights can be set by expert knowledge or data statistics. When calculating the loss function, weight the prediction errors of different nodes, give higher weights to important nodes, and improve the prediction accuracy of key nodes. Introduce attribute weight matrix: assign an attribute weight to each attribute according to the importance of different attributes in the prediction task. Importance can be measured by factors such as the degree of influence of the attribute on the prediction result and the scarcity of the attribute, and weights can be set by expert knowledge or data analysis. When calculating the loss function, weight the prediction errors of different attributes, give higher weights to key attributes, and improve the prediction accuracy of key attributes.

[0080] The improved loss function expression is as follows ,in, represents the value of the improved loss function, which takes into account spatial dependency and attribute diversity. T represents the number of predicted time steps, that is, the number of moments in the future. N represents the number of nodes in the spatiotemporal attribute graph. K represents the dimension of the node attribute, that is, the number of attribute types for each node. It represents the predicted value of the kth attribute of the i-th node at the t-th time step. It is the output result of the spatiotemporal prediction sequence model at the t-th time step. The true value of the kth attribute of the ith node at the tth time step. It is the node attribute value of the real spatiotemporal attribute graph at the tth time step. The spatial weight of the i-th node reflects the importance of the node in the spatial topology. It can be calculated based on indicators such as the node's degree and centrality, or learned adaptively through the attention mechanism. The attribute weight of the kth attribute reflects the importance of the attribute in the prediction task. It can be set based on factors such as the relevance of the attribute, the difficulty of prediction, etc., or learned adaptively through the attention mechanism.

[0081] Model optimization: Use optimization algorithms, such as the Adam optimizer, to optimize model parameters. The gradient of the loss function to the model parameters is calculated through the back-propagation algorithm. Based on the gradient information, the model parameters are updated to minimize the loss function. Iterate multiple training rounds until the model converges or reaches the predetermined performance index. Through the above steps, a rich set of training samples can be generated, and the spatiotemporal prediction sequence model can be trained using the teacher forcing method and the improved loss function. Data augmentation technology helps to improve the generalization ability and robustness of the model, while the introduction of spatial weight matrix and attribute weight matrix can improve the prediction accuracy of key nodes and attributes.

[0082] The model training process continuously adjusts the model parameters through the optimization algorithm to minimize the difference between the predicted spatiotemporal attribute graph and the real spatiotemporal attribute graph. The trained model can be used to predict the spatiotemporal sequence of the new spatiotemporal attribute graph, recursively generate prediction results for multiple moments in the future, and reveal the spatiotemporal evolution trend of geographic elements. This application achieves modeling and prediction of complex spatiotemporal dynamic processes by integrating the semantic, topological and temporal features of geographic elements.

Claims

1. A surveying and mapping geographic information data management method based on cloud computing, characterized in that: include: S1, obtaining a surveying and mapping geographic data set; wherein the surveying and mapping geographic data set includes a variety of geographic elements; the geographic elements include buildings, roads, water systems and vegetation; S2, build a semantic segmentation model based on the convolutional neural network Deep Lab, use the semantic segmentation model to extract features and semantically annotate the surveying and mapping geographic data set to obtain a set of geographic features; S3, using the R-tree spatial indexing method to spatially partition and index the geographic feature set, and obtain a hierarchically organized geographic feature tree; S4, build a graph neural network model based on graph convolutional network GCN, use the graph neural network model to model the topological relationship of the geographic feature tree, and obtain a spatiotemporal attribute graph containing geographic semantics and topological relationships; S5, build a spatiotemporal prediction sequence model based on the spatiotemporal graph neural network STGCN, and use the spatiotemporal prediction sequence model to predict the evolution trend of the spatiotemporal attribute graph; S2, obtains a set of geographic elements, including: S21, build a semantic segmentation model. The semantic segmentation model adopts an encoder-decoder architecture and uses the convolutional neural network Deep Lab as the backbone network; The encoder extracts features from the surveying and mapping geographic data set through convolution and pooling operations to obtain a multi-scale semantic feature map; The decoder upsamples and fuses the multi-scale semantic feature maps through deconvolution and skip connection operations to obtain a semantic segmentation map. S22, using a connected domain analysis algorithm to perform pixel-level region clustering on the semantic segmentation map, dividing pixels with the same semantic label and spatially connected to each other into the same region, and forming an initial pixel region mask; S23, performing regional grouping on the initial pixel region masks, and classifying the pixel region masks of different geographic elements according to a predefined geographic element category system to obtain a pixel region mask set organized by geographic element category; S24, according to the pixel region mask set organized by geographic feature category, a minimum bounding rectangle algorithm is used to calculate the minimum bounding rectangle bounding box of each pixel region mask, the minimum bounding rectangle bounding box is used as the geometric boundary attribute of the geographic feature, and the geometric boundary attribute is associated with the corresponding semantic category label to obtain the corresponding geographic feature object; S25, organizing each geographic feature object into a geographic feature set, wherein each geographic feature object includes a semantic category label, a pixel region mask, and a minimum circumscribed rectangular bounding box.

2. The method for managing surveying and mapping geographic information data based on cloud computing according to claim 1, characterized in that: S24, obtaining the corresponding geographic element object, including: Traverse the pixel region mask set and use the minimum bounding rectangle algorithm to calculate the minimum bounding rectangle bounding box of each pixel region mask; The position and size of the minimum bounding rectangle are used as the geometric boundary attributes of the corresponding pixel area mask; Associating the obtained geometric boundary attributes with the semantic category labels corresponding to the pixel region masks to construct geographic feature objects; wherein the geographic feature objects include the semantic category labels, the pixel region masks and the geometric boundary attributes; According to the association results, the geographic feature objects corresponding to each pixel area mask are organized according to the semantic category labels to construct a geographic feature set.

3. The method for managing surveying and mapping geographic information data based on cloud computing according to claim 1, characterized in that: S3, obtains a hierarchically organized geographical element tree, including: S31, spatially partitioning the geographic feature set using an R-tree spatial indexing method according to the minimum circumscribed rectangular bounding box of the geographic feature object, recursively aggregating the geographic feature objects into a nested tree structure, and obtaining a geographic feature tree; S32, storing, at each node of the geographic element tree, attribute information of all geographic element objects in the geographic area corresponding to the corresponding node, including semantic category labels, pixel area masks, and minimum circumscribed rectangular bounding boxes; S33, introducing a semantic attribute hash table at each node of the geographic element tree, and using a hash indexing method to index the semantic category labels of all geographic element objects in the geographic area corresponding to the corresponding node; S34, introducing a spatial topological relationship graph at each node of the geographic element tree, and using a graph data structure to represent the spatial topological relationship between all geographic element objects in the geographic area corresponding to the corresponding node.

4. The method for managing surveying and mapping geographic information data based on cloud computing according to claim 3, characterized in that: S4, obtains a spatiotemporal attribute graph containing geographic semantics and topological relationships, including: S41, using the nodes of the geographic element tree as graph nodes and the parent-child relationships between the nodes of the geographic element tree as graph edges, generating a geographic element graph with a tree topological structure; setting attribute information of the geographic element object corresponding to the node on each node of the geographic element graph, the attribute information including a semantic category label, a pixel region mask and a minimum circumscribed rectangular bounding box; S42, at each node of the geographic feature map, encode the semantic attribute hash table into a semantic feature vector, and encode the spatial topological relationship map into a topological feature vector; use the attention mechanism to weightedly fuse the semantic feature vector and the topological feature vector to obtain a node feature vector; S43, using the node feature vector as the node feature of the graph convolutional network GCN, and using the GCN to extract features from the geographic element graph of step S42 to obtain a node embedding vector of the geographic element graph; S44, using the node embedding vectors of the geographic element graph, constructing a semantic topological relationship matrix between nodes, and obtaining an edge weight matrix representing the semantic similarity and topological connectivity between nodes by calculating the semantic similarity and spatial distance between the node embedding vectors; S45, using node embedding vectors and semantic topological relationship matrix, construct a spatiotemporal attribute graph containing geographic semantics and topological relationships; wherein the nodes of the spatiotemporal attribute graph are geographic feature objects, the node features are node embedding vectors, the edges of the spatiotemporal attribute graph are semantic topological relationships between geographic feature objects, and the edge weights are the semantic topological relationship matrix; timestamp attributes are set on nodes and edges to reflect the temporal evolution of geographic features and semantic topological relationships.

5. The method for managing surveying and mapping geographic information data based on cloud computing according to claim 4, characterized in that: S44, obtaining an edge weight matrix representing the semantic similarity and topological connectivity between nodes, including: By calculating the cosine similarity between the node embedding vectors of the geographic element graph obtained in step S43, a semantic similarity matrix representing the nodes is obtained; By calculating the Euclidean distance between the center coordinates of the minimum circumscribed rectangular bounding box of the geographic feature object corresponding to the node embedding vector, a spatial distance matrix representing the nodes is obtained; The semantic similarity matrix and the spatial distance matrix are normalized by the softmax algorithm to obtain the semantic similarity weight matrix and the spatial distance weight matrix; Using the attention mechanism, the semantic similarity weight matrix and the spatial distance weight matrix are weighted and fused to obtain the edge weight matrix that integrates semantic similarity and spatial distance. A semantic topological relationship matrix is ​​constructed based on the edge weight matrix as the edge attribute of the geographic feature map.

6. The method for managing surveying and mapping geographic information data based on cloud computing according to claim 5, characterized in that: The edge weight matrix that integrates semantic similarity and spatial distance is obtained, including: Set up a multi-head attention mechanism, where each attention head corresponds to a different semantic and spatial scale, and learn the contribution of semantic similarity and spatial distance to edge weight at different scales; When calculating the semantic similarity weight matrix, the embedding vector of the semantic category label is set, and the attention weight between the node embedding vector obtained in step S43 and the embedding vector of the semantic category label is calculated to obtain the semantic similarity weight of the fused node semantic feature and the category semantic feature; When calculating the spatial distance weight matrix, the spatial direction relationship and topological relationship of the geographic elements are set, and the distance weight integrating the spatial distance and the direction topological relationship is obtained by calculating the direction relationship and topological relationship between the geographic element objects in step S41; The directional relationship is obtained by calculating the azimuth and angle between the center coordinates of the minimum circumscribed rectangular bounding box of the geographic element, which is used to reflect the directional distribution of the geographic element in space; The topological relationship is obtained by analyzing the intersection, inclusion or adjacency relationship between the minimum circumscribed rectangular bounding boxes of geographic elements, and is used to reflect the topological connectivity of geographic elements in space; The semantic similarity weights and spatial distance weights generated by different attention heads, as well as the semantic similarity weight matrix and spatial distance weight matrix are weightedly fused through the attention mechanism to obtain the edge weight matrix that integrates semantic similarity and spatial distance.

7. The method for managing surveying and mapping geographic information data based on cloud computing according to any one of claims 1 to 6, characterized in that: S5, build a spatiotemporal prediction sequence model based on the spatiotemporal graph neural network STGCN, including: S51, constructing a spatiotemporal prediction sequence model based on the spatiotemporal graph neural network STGCN, wherein the input of the spatiotemporal prediction sequence model is the spatiotemporal attribute graph obtained in step S45, and the output is the predicted spatiotemporal attribute graph at multiple moments in the future; The spatiotemporal prediction sequence model adopts an encoder-decoder architecture. The encoder uses graph convolutional layers and temporal convolutional layers to encode the node features and edge features of the spatiotemporal attribute graph. The gated recurrent unit GRU is used as the backbone network for extracting temporal features; The decoder uses graph deconvolution and temporal deconvolution to decode the encoded features and generate a spatiotemporal attribute graph for a period of time in the future; The graph attention layer and the temporal attention layer are set after the graph convolution layer and the temporal convolution layer respectively to adjust the importance between different nodes and different time steps; S52, generating a training sample set by a data enhancement method according to the acquired surveying and mapping geographic data set; S53, according to the training sample set, a classroom forcing method is used to train the spatiotemporal prediction sequence model, with the training sample as input and the real spatiotemporal attribute graph of a period of time in the future as the supervision signal, and the parameters of the spatiotemporal prediction sequence model are optimized by minimizing the difference between the predicted spatiotemporal attribute graph and the real spatiotemporal attribute graph; S54, using the trained spatiotemporal prediction sequence model, perform spatiotemporal sequence prediction on the obtained spatiotemporal attribute graph, recursively generate predicted spatiotemporal attribute graphs for multiple moments in the future, and obtain the spatiotemporal evolution trend of the geographic elements.

8. The method for managing surveying and mapping geographic information data based on cloud computing according to claim 7, characterized in that: The mean square error loss function is used to minimize the difference between the predicted spatiotemporal attribute map and the true spatiotemporal attribute map; The expression of the mean square error loss function is: in, L Improved represents the value of the improved loss function; T represents the number of predicted time steps; N represents the number of nodes in the spatiotemporal attribute graph; K represents the dimension of the node attribute; represents the predicted value of the kth attribute of the i-th node at the t-th time step; represents the true value of the kth attribute of the i-th node at the t-th time step; represents the spatial weight of the i-th node, reflecting the importance of the node in the spatial topological structure; It represents the attribute weight of the k-th attribute, reflecting the importance of the attribute in the prediction task.

9. A surveying and mapping geographic information data management system based on cloud computing, characterized in that: include: At least one processing unit; used to execute instructions to implement the cloud computing-based surveying and mapping geographic information data management method as described in any one of claims 1 to 8.

Images