Road network mode recognition method and system based on unsupervised graph representation learning

Abstract

The invention provides a road network mode recognition method and system based on unsupervised graph representation learning. The method comprises the following steps: firstly, modeling a road network through a spatial dual graph, and designing graph node features based on cognitive heuristic; then, a proposed graph neural network model is mainly based on GAE, the model can be trained in an unsupervised mode, and meanwhile, sub-graph isomorphic counting (SIC) is introduced in a road segment embedding learning stage and a global context attention mechanism (GCA) is introduced in a graph embedding generation stage, so that the representation performance of the model is enhanced; and finally, recognizing a road network mode by utilizing geometric similarity of graph-level embedding. Experimental results show that all indexes of the method in an experimental area are superior to those of a classical road network recognition method, the classification accuracy is improved by 12% or above, and the method is even slightly superior to a supervised baseline graph neural network method.

Description

Technical Field

[0001] The present invention relates to the technical field of road network pattern recognition, and in particular to a road network pattern recognition method and system based on unsupervised graph representation learning. Background Art

[0002] As a basic geographical element, understanding the spatial structure and morphological characteristics of the road network helps to explore the relationship between the earth's surface environment and human activities.

[0003] Classic road network pattern recognition is mainly divided into methods based on geometric calculation and graph theory. Among them, the geometric calculation method can be divided into two methods based on linear elements and surface elements according to the recognition unit. The linear element method recognizes the road network pattern from the perspective of the geometric characteristics of the road section. The surface element method recognizes the road network pattern through the spatial morphology of the road network mesh. The above-mentioned road network pattern recognition methods based on geometric calculation mainly rely on the shape characteristics of the road network, ignoring the topological connection properties of the road network as a spatial connection network. In this regard, graph theory is an effective modeling method that integrates the geometric shape characteristics and topological connection properties of the road network. In general, most traditional algorithms focus on describing the characteristics of specific patterns, that is, algorithms need to be designed separately for different patterns, and the generalization is not strong. In addition, road network pattern recognition is highly subjective and complex due to the influence of spatial perception and visual psychology, but traditional algorithms lack intelligent reasoning and decision-making processes, and it is difficult to reach the level of human recognition. Therefore, it is necessary to explore recognition methods with strong generalization and closer to human perception, so as to effectively recognize and classify various road network patterns.

[0004] In recent years, deep learning technology, especially Graph Neural Network (GNN), has made significant progress in processing graph data. Deep learning methods based on GNNs can automatically extract road network features without manually designing feature extractors. Compared with traditional algorithms, this method reduces manual intervention and improves the objectivity and accuracy of road network pattern recognition. However, it relies on a large amount of labeled training data, and the annotation of road network data is usually complex and expensive. The training and reasoning of the model also consume a lot of computing resources, especially when the data scale is large. In addition, the embedded representation generated by the GNNs model mainly serves specific tasks (such as pattern recognition), and may no longer be applicable or perform poorly when applied to new tasks, showing its lack of generalization ability and limited flexibility. Summary of the invention

[0005] In view of the problems that the existing GNNs-based road network feature extraction scheme relies on a large amount of labeled training data and the generated embedded representation has insufficient generalization ability, the present invention provides a road network pattern recognition method and system based on unsupervised representation learning.

[0006] In one aspect, the present invention provides a road network pattern recognition method based on unsupervised graph representation learning, comprising:

[0007] Step 1: Take the stroke sections in the target road network data as nodes of the graph, take the connection relationship between adjacent sections and the spatial relationship added for spatially similar sections as edges of the graph, construct a spatial dual graph of the target road network data, and generate an adjacency matrix of the target road network data based on the spatial dual graph;

[0008] Step 2: extracting the segment features of each stroke segment in the target road network data to generate a feature matrix of the target road network data; wherein the segment features include geometric attribute features and shape context SC attribute features of the segment;

[0009] Step 3: Based on the spatial dual graph, adjacency matrix and feature matrix of the target road network data, a preset graph neural network model is used to perform road graph representation learning to generate a graph-level embedding of the target road network data; wherein the graph neural network model includes a GAE model and a structure-aware graph representation learning module; the structure-aware graph representation learning module uses the node-level embedding output by the encoder in the GAE model as input to generate a graph-level embedding;

[0010] Step 4: Measure the similarity between the graph embedding of the target road network data and the graph embedding of multiple preset standard anchor points, and take the road network pattern of the standard anchor point with the highest similarity value as the road network pattern of the original road network data.

[0011] Furthermore, step 1 also includes: extracting spatially similar road sections and adding spatial relationships for the spatially similar road sections; wherein the spatially similar road sections refer to two road sections that meet preset spatial requirements but have no connection relationship; the preset spatial requirements refer to that the distance similarity and angle similarity of the two road sections are both greater than 0.

[0012] Furthermore, the distance similarity between two road sections is calculated according to the following formula: Similarity with angle

[0013]

[0014]

[0015] Among them, sp_dist(s i ,s j ) indicates calculating two road segments s i and j One-way Hausdorff distance between, ag_dist(s i ,s j ) indicates calculating two road segments s i andj The absolute angular distance between ds and δ as Indicates two preset distance thresholds.

[0016] Furthermore, the geometric attribute features of the road section include position features extracted from a macroscopic scale, connection features extracted from a mesoscopic scale, and length features, shape features, and angle features extracted from a microscopic scale.

[0017] Furthermore, the process of extracting the SC attribute features of the road section includes:

[0018] Establish a polar coordinate system at the midpoint of the target road section, and set the maximum radius of the polar coordinate system to r=γ×L(s); wherein L(s) is the length of the target road section, and γ is a scale factor;

[0019] In the polar coordinate system, p concentric circles are divided at equal distances, and q sectors are divided at equal angles to obtain p×q sectors, and the value of each sector is defined as the ratio of the total length of the road segments in the area to the total length of the road segments in the context range;

[0020] The values ​​of p×q sectors are represented as a vector, which is the SC attribute feature of the target road section.

[0021] Furthermore, a subgraph isomorphism counting SIC module is provided in the encoder, and the SIC module is used to generate structure identifiers for nodes in the space dual graph. The encoder performs neighborhood information transmission based on the structure identifier, adjacency matrix and feature matrix of the node to generate node-level embedding.

[0022] Furthermore, it also includes: training the encoder and decoder in the GAE model; during the training process, the decoder reconstructs the inter-node association based on the node-level embedding output by the encoder to generate a reconstructed adjacency matrix, and uses the following cross entropy loss function to calculate the distance between the reconstructed adjacency matrix A′ and the true adjacency matrix A:

[0023]

[0024] Where N represents the total number of nodes in the space-dual graph.

[0025] Furthermore, the process of generating node-level embedding specifically includes:

[0026] Define a subgraph collection For a subgraph collection For each subgraph H in the space dual graph G, first find its isomorphic subgraph G in the space dual graph G S , for G S Each node in Find the orbital Orb of its mapping f(v) in subgraph H H (f(v)), Orb H (*) represents the preset orbital mapping function, where the orbit refers to the position or role of a node in H; then, the position of each node v in different isomorphic subgraphs G is calculated. S The track times, and obtain the structural characteristics of node v on the i-th track of subgraph H represents the i-th track in subgraph H; f(v) represents the node mapping function that maps node v to the node in subgraph H; finally, the combined subgraph set The count of different subgraph structures and tracks in the node is used to obtain the structure identifier of the node in, It is subgraph H i The number of mid-tracks;

[0027] The adjacency matrix, feature matrix, and structure identifier are updated according to the following propagation rules between adjacent convolutional layers:

[0028]

[0029] in, A is the adjacency matrix, is the identity matrix, yes The degree matrix of is the convolutional signal matrix of the lth layer, representing the features of each node in the lth layer, m l is the dimension of the node feature at layer l; the input layer H 0 =M, represents the feature matrix, d1 represents the dimension of the node's geometric attribute feature, and d2 represents the dimension of the node's SC attribute feature; X v is the structure identifier of the node, and is the parameter training weight matrix and common bias of the lth layer, σ(·) is the nonlinear activation function, and N represents the total number of nodes in the space dual graph G.

[0030] Furthermore, the structure-aware graph representation learning module takes the node-level embedding output by the encoder in the GAE model as input to generate graph-level embedding, specifically including:

[0031]

[0032] Among them, E S represents graph-level embedding, Z represents the node-level embedding output by the encoder, is the embedding of the nth node, N represents the total number of nodes; is a learnable weight matrix, and tanh(·) is an activation function; represents the attention weight of the node, and σ(·) is a nonlinear activation function.

[0033] On the other hand, the present invention provides a road network pattern recognition system based on unsupervised graph representation learning, comprising:

[0034] A spatial dual graph construction module is used to construct a spatial dual graph of the target road network data by taking the stroke road segments in the target road network data as the nodes of the graph, and taking the connection relationship between adjacent road segments and the spatial relationship added for spatially similar road segments as the edges of the graph;

[0035] An adjacency matrix generation module, used to generate an adjacency matrix of target road network data according to the space dual graph;

[0036] A feature matrix generation module is used to extract the segment features of each stroke segment in the target road network data to generate a feature matrix of the target road network data; wherein the segment features include geometric attribute features and shape context SC attribute features of the segment;

[0037] A graph-level embedding generation module, which is used to perform road graph representation learning based on the spatial dual graph, adjacency matrix and feature matrix of the target road network data using a preset graph neural network model to generate a graph-level embedding of the target road network data; wherein the graph neural network model includes a GAE model and a structure-aware graph representation learning module; the structure-aware graph representation learning module uses the node-level embedding output by the encoder in the GAE model as input to generate a graph-level embedding;

[0038] The road network pattern recognition module is used to measure the similarity between the graph embedding of the target road network data and the graph embedding of multiple preset standard anchor points, and take the road network pattern to which the standard anchor point with the highest similarity value belongs as the road network pattern of the original road network data.

[0039] Beneficial effects of the present invention:

[0040] (1) The road network pattern recognition method and system of the present invention, in order to solve the problem of insufficient spatiality of road network modeling, introduces spatial edges to improve the road graph construction, and obtains a spatial dual graph, based on which the embedded representation of road network data can be better learned. At the same time, the graph-level embedding of road network data is learned based on the Graph Autoencoder (GAE) model, without relying on a large amount of labeled training data, and the generated graph-level embedding does not depend on specific downstream tasks.

[0041] (2) When extracting the initial node features of the road graph, the present invention follows the Gestalt theory and extracts visual variables such as position, connection, length, shape and angle from macro, meso and micro perspectives as the geometric features of the road segment; at the same time, the SC attribute features of the node are extracted to describe the spatial structure pattern of the neighborhood of the target road segment; the geometric attributes of each road segment and the SC attributes formed by its neighborhood are integrated as the features of the corresponding graph nodes, so that a more effective road graph embedding representation can be obtained.

[0042] (3) In the encoder’s road segment embedding learning stage, the present invention introduces subgraph isomorphism counting (SIC) to enhance the model’s expressiveness; in the graph-level embedding generation stage, the global context attention mechanism (GCA) is added to improve the representation of road network data. Through SIC and GCA, the model can better capture and understand the local structure and global topological relationship in the road network, and then generate a more representative, robust and structure-aware road graph-level embedding representation, better capture the road network structure information, and significantly improve the performance of the road network pattern recognition task.

[0043] (4) The present invention uses graph-level embedded geometric similarity to realize road network pattern recognition, and conducts experiments on real road network data to compare the performance of classic road network pattern recognition algorithms and multiple graph neural network recognition algorithms. The experiments verify the effectiveness of the recognition method. Compared with the classic road network pattern recognition method, the accuracy of SUGAR pattern classification is improved by more than 12%. Compared with the supervised baseline graph neural network method, SUGAR can achieve comparable or even slightly better performance without the need for sample annotation.

[0044] (5) The dimensionality reduction visualization technology was used to evaluate the road graph embedding performance. The pattern recognition performance of graph neural network models using different modules (SIC and GCA) was analyzed from the two aspects of road network samples and graph embedding representation. It was found that the SUGAR model that used both SIC and GCA had clear boundaries between categories, clear clustering of regular patterns, and was not confused with irregular patterns, achieving the best classification effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 A schematic diagram of a process flow of a road network pattern recognition method based on unsupervised graph representation learning provided by an embodiment of the present invention;

[0046] Figure 2 A process for constructing a space dual graph provided by an embodiment of the present invention;

[0047] Figure 3 A schematic diagram of a calculation process of geometric attribute features of a road section provided in an embodiment of the present invention;

[0048] Figure 4 Examples of SC attributes of road sections with different road network modes provided by embodiments of the present invention;

[0049] Figure 5 The generation process of the adjacency matrix A and the feature matrix M provided in the embodiment of the present invention;

[0050] Figure 6 A graph representation learning process of a graph neural network model with GAE as the main framework provided in an embodiment of the present invention;

[0051] Figure 7 An example of a subgraph and its track used by the road network subgraph isomorphism counting SIC provided in an embodiment of the present invention;

[0052] Figure 8 Subgraph isomorphism counting provided by the embodiment of the present invention (taking triangle and chord cycle as examples);

[0053] Fig. 9 Road network pattern recognition based on graph embedding geometric similarity provided by an embodiment of the present invention;

[0054] Fig.10 The road network pattern recognition results provided by the embodiments of the present invention (including some correct recognition examples and all incorrect recognitions). DETAILED DESCRIPTION

[0055] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution in the embodiment of the present invention will be clearly described below in conjunction with the drawings in the embodiment of the present invention. Obviously, the described embodiment is a part of the embodiment of the present invention, not all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0056] As a basic geographical element, the road network studies its vector representation method and its application in pattern recognition, which can realize the analysis and understanding of the spatial structure of the road network. Traditional road network pattern recognition methods mainly rely on the geometric characteristics and topological properties of the road network, but such methods are computationally complex and lack the ability of intelligent reasoning, and their performance degrades under complex road network structures. In recent years, recognition methods based on graph neural networks have become a research hotspot. Such methods reduce manual intervention, improve the objectivity and accuracy of recognition, but rely on a large amount of labeled data. Moreover, the above methods are algorithms designed for specific tasks and have limited generalization capabilities. Based on this, the present invention mainly studies the road network vector representation method and its application in road network pattern recognition, maps the complex road network structure into a road network structure pattern vector representation suitable for analysis and calculation, realizes the analysis and understanding of the spatial structure of the road network, and provides a computational method for the information representation processing of the digital twin of the earth system.

[0057] The present invention uses graph neural networks to learn the representation of road network data, and infers the similarity relationship between spatial structures and topological connections between different road network data based on graph-level embedding, thereby realizing pattern recognition of road networks. The basic process of the method of the present invention is as follows: Figure 1 As shown, it mainly includes 4 steps:

[0058] S101: Road graph model construction.

[0059] Specifically, the present invention aims to use graph embedding to measure the geometric similarity of road networks for pattern recognition. Given that undirected graphs are more suitable for expressing the symmetry and interconnectivity of road networks, thereby effectively capturing the geometric similarities between road networks, the embodiments of the present invention select undirected graphs to construct a road graph model.

[0060] There are two common ways to construct road graphs: one is to use nodes and edges to represent road segment endpoints and road segments; the other is to use the idea of ​​dual graphs, using road segments as graph nodes and connection relationships as edges. In the first modeling method, nodes only express the connection relationship between road segments, and it is difficult to directly consider the attributes of road segments, but these attributes are particularly important when expressing the spatial pattern of the road network. In contrast, the dual graph modeling method using road segments as nodes has the advantage of representing the attribute characteristics of road segments and their topological relationships. The construction process of the dual graph includes the following key steps: Given a road network ( Figure 2 .a), first, filter out noise data such as small length and isolated points through preprocessing and perform stroke segment extraction ( Figure 2 .b); Then, each stroke segment is regarded as a node in the dual graph, and the node is usually located at the midpoint of the stroke segment; Finally, in the dual graph, the road segments connected by the original road network are connected to the corresponding nodes through edges ( Figure 2 .c).

[0061] The road network dual graph construction focuses on the topological structure of the graph, which allows the embedding of the target node to be learned by aggregating the embeddings of neighboring nodes when applying the GNNs model on the road graph. However, this approach ignores the spatial structure of the road network. For example, Figure 2 .d, although s1 and s2 are close in space and have similar directions, they are Figure 2 .e is cut into two parts in the dual graph of the road network; similarly, although s3 and s4 are similar in position and angle, there is no directly connected edge in the dual graph, and effective information transmission cannot be achieved. To solve this problem, the embodiment of the present invention adds a spatial relationship for spatially similar road sections. For example, in Figure 2 In .f, a red line segment is added between s1 and s2, and a red line segment is added between s3 and s4.

[0062] Based on the above content, the process of constructing the spatial dual graph can be summarized as follows: extract stroke segments from the road network data and use them to construct a graph model that describes the relationship between segments. The stroke segments in the road network data are used as nodes of the graph, and the connection relationship between adjacent segments and the spatial relationship added for spatially similar segments are used as edges of the graph to construct the spatial dual graph of the road network data. Next, the adjacency matrix of the road network data is generated based on the spatial dual graph.

[0063] It should be noted that, for the convenience of subsequent description, the edges corresponding to the connection relationship are called topological edges, and the edges corresponding to the spatial relationship are called spatial edges. Different from the dual graph that only considers the topological connection relationship of the road segment, the spatial dual graph ( Figure 2 .g) Contains two types of undirected edges: topological edges and spatial edges. Topological edges and spatial edges are represented by solid lines (blue) and dashed lines (red), respectively. Among them, spatial edges capture the spatial similarity between road segments, helping the model to identify road segments that are spatially close but not directly connected in the original road network. This modeling method enables the model to effectively capture those spatial similarities that cannot be directly expressed in the traditional dual graph model, optimizes the integration and transmission of information between graph nodes, and provides a way to fully understand the road network through topological and spatial structures, thereby improving the accuracy of road network pattern recognition.

[0064] S102: Extracting road map features.

[0065] Specifically, the geometric attribute features of the stroke segments in the road network data are extracted, and the shape context SC attribute features of the stroke segments are introduced as the neighborhood features of the segments. The segment geometry and its neighborhood features are integrated to form a feature matrix as the initial feature description of the road graph nodes.

[0066] S103: Road graph representation learning.

[0067] Specifically, based on the spatial dual graph, adjacency matrix and feature matrix of the road network data, a preset graph neural network model is used to perform road graph representation learning to generate a graph-level embedding of the target road network data. The graph neural network model includes a GAE model and a structure-aware graph representation learning module; the structure-aware graph representation learning module uses the node-level embedding output by the encoder in the GAE model as input to generate a graph-level embedding. GAE refers to Graph Autoencoder, which is a graph autoencoder model for graph data and a neural network model for unsupervised learning. Node-level embedding and graph-level embedding are two embedding methods in graph embedding technology in this field.

[0068] S104: Road network pattern recognition based on geometric similarity.

[0069] Specifically, for different road network data, different graph-level embeddings are generated respectively through the above steps S101 to S103, and the geometric similarity between different road network data is calculated using the graph-level embeddings, and finally different road network patterns are classified according to the similarity to achieve multi-pattern recognition of road networks. Specifically, the graph embedding of the target road network data is measured for similarity with the graph embedding of multiple preset standard anchor points, and the road network pattern to which the standard anchor point with the highest similarity value belongs is taken as the road network pattern of the original road network data.

[0070] The road network pattern recognition method provided by the embodiment of the present invention introduces spatial edges to improve the road graph construction to address the problem of insufficient spatiality in road network modeling, thereby obtaining a spatial dual graph. Based on the spatial dual graph, the embedded representation of road network data can be better learned. At the same time, the graph-level embedding of road network data is learned based on the GAE model, without relying on a large amount of labeled training data, and the generated graph-level embedding does not depend on specific downstream tasks. Finally, the present invention realizes the pattern recognition of road networks by utilizing the geometric similarity of graph-level embeddings.

[0071] Based on the above embodiments, the present invention provides a process for extracting spatially similar road segments. First, spatially similar road segments are defined as two road segments that meet preset spatial requirements but are not connected; the preset spatial requirements mean that the distance similarity and angle similarity of the two road segments are both greater than 0.

[0072] As an implementation method, the distance similarity between two road segments is calculated according to the following formula: Similarity with angle

[0073]

[0074] Among them, sp_dist(s i ,s j ) indicates calculating two road segments s i and jOne-way Hausdorff distance ag_dist(s i ,s j ) indicates calculating two road segments s i and j The absolute angular distance between i .radian-s j .radian|), δ ds and δ as Represents two preset distance thresholds, which respectively set the distance similarity to zero when the spatial distance and angular distance are too large. Similarity with angle The value of is set to 0 (because cosπ / 2=0).

[0075] In order to introduce spatial neighborhood information into the aggregation process, the embodiment of the present invention further proposes a spatial similarity adjacency matrix, which records the similarity between road sections according to the spatial attributes of the road sections (such as position, angle, etc.). The matrix elements The value of is shown as follows:

[0076]

[0077] It can be understood that the matrix element A s ij =1 means that when the road section s i and j When the spatial similarity between them meets the preset spatial requirements (i.e., the spatial distance is close and the spatial direction is similar), a spatial edge is added to the dual graph, such as Figure 2 .f.1; if the spatial similarity requirement is not met, or only one of the spatial distance relationship and spatial direction relationship is met, no spatial edge connection is added (A s ij =0), such as Figure 2 .f.2.

[0078] On the basis of the above embodiments, after the road graph is constructed, its space dual graph is represented as G=V,A t ,A s , where V = {v0,v1,…v N} is a set of nodes, N represents the total number of nodes in G; It is a topological adjacency matrix, which represents the topological connection relationship between nodes. If there is a connection v i and v j The edge ij , then A t ij =1; otherwise A t ij= 0, indicating that there is no such edge. Spatial similarity adjacency matrix Represents the spatial similarity between road segments. When the spatial distance and angular distance between two road segments are less than a certain threshold, A s ij =1, otherwise A s ij =0.

[0079] In the road network space dual graph, each node represents a road segment and carries a series of descriptive features. Since the road network pattern is affected by the morphological characteristics of the road and the connection mode of the neighboring roads, relying solely on the geometric features of a single road segment is not enough for accurate pattern recognition. Therefore, in order to obtain a more effective road graph embedding representation, the embodiment of the present invention extracts the geometric attributes of each road segment and the SC attributes formed by its neighborhood as the features of the corresponding graph node.

[0080] (1) Geometric properties of a single stroke segment

[0081] The embodiment of the present invention follows the Gestalt theory and extracts visual variables such as position, connection, length, shape and angle as the geometric features of the road section from the macro, meso and micro perspectives. Specifically, for each graph node, the visual variable method is used to extract the road section attributes through specific visual features to realize the holistic perception principle in the Gestalt theory, thereby obtaining better road network spatial cognition and graphical representation. Therefore, node features are extracted from three levels: macro, meso and micro: the macro scale involves the overall layout and pattern of the entire road network, the meso scale focuses on the connection between road sections and the hierarchical relationship of different road grades, and the micro scale focuses on the specific features of a single road section. Based on these different scales, in Table 1, the embodiment of the present invention lists 6 different geometric attribute features extracted from aspects such as position, connection relationship, length, shape and angle, forming a feature matrix Calculation method: Figure 3 shown. Figure 3 In the figure, point O is the center coordinate, which represents the arithmetic mean of the midpoints of all road sections. p ,y p ) is the coordinate of the midpoint of the road segment, l and w are the length and width of the minimum circumscribed matrix, and P1 and P2 are the two endpoints of the longest side of the minimum circumscribed matrix.

[0082] Table 1. Six road segment features extracted from global, meso and micro perspectives

[0083]

[0084] (2) SC attributes of road segment neighborhood

[0085] The embodiment of the present invention uses the SC descriptor to describe the spatial structure pattern of the neighborhood of the target road segment. First, a polar coordinate system is established at the midpoint of the road segment, with a maximum radius of r = γ × L (s), where L (s) is the length of the target road segment (red); γ is a proportional factor that determines the receptive field of the road segment context feature. Then, p concentric circles are divided at equal distances, and q sectors are divided at equal angles to obtain p × q sectors. The value of each sector is defined as the ratio of the total length of the road segment in the area to the total length of the road segment in the context range; for example, the value V of the ith sector is i Calculated as Finally, the values ​​of p×q sectors are represented as a vector to describe the SC attribute of the road segment. For a graph with n nodes, the neighborhood features are obtained through the SC descriptor

[0086] Figure 4 The SC descriptions of four road segments of four road network patterns (grid, complex interchange, parallel and irregular patterns) are shown, with parameters p, q and γ taken as 4, 8 and 1.20 respectively. It can be found that the SC descriptions of road segments with the same road network pattern are more similar.

[0087] Figure 5 The overall process of constructing the adjacency matrix A and feature matrix M of road network data is given. Given a road network, first build its spatial dual graph and extract the topological adjacency matrix A t And the spatial similarity matrix A s , the two are added to form the final adjacency matrix A; then, the geometric features M of the node segments are extracted f and SC attribute feature M sc , the 6 types of road segment features and the 32-dimensional SC attribute features are combined into a 38-dimensional vector, which is used as the input feature vector M of the corresponding graph node. Finally, the road network graph is represented as the adjacency matrix A and the feature matrix M for subsequent analysis.

[0088] Based on the above embodiments, Figure 6 As shown, the graph neural network model provided by the embodiment of the present invention is a structure-aware road network graph representation learning model (SUbgraph-aware Graph Autoencoder of Road network, SUGAR) that does not rely on specific downstream tasks. The graph representation learning process is as follows: the encoder uses the subgraph isomorphism counting SIC module as the core and the road graph (M, A) as input to calculate the node-level embedding Where N is the total number of nodes, m represents the dimension of the node features, and the nth row of Z is the embedding of node n. The decoder takes the node-level embedding Z as input to reconstruct the adjacency matrix The goal is to minimize the difference between the reconstructed adjacency matrix A′ and the original adjacency matrix A. In the graph representation learning phase, a simple mean readout function is used to convert the node-level embedding Z into a simple intermediate graph-level embedding Then GCA is used to calculate an attention weight for each node, and finally the structure-aware graph-level embedding is obtained by weighted average. It is used for cosine similarity calculation. The functions of each module will be introduced in detail below.

[0089] (1) Encoder

[0090] Neural networks can automatically identify repetitive patterns in data to understand complex structures. For example, when processing images, convolutional neural networks can identify different areas in the image (such as the facial features); or when processing text, they can identify different parts of a sentence (such as subject and predicate). However, existing GNNs cannot distinguish the different "structural roles" of nodes, that is, all nodes are treated in the same way. To this end, an embodiment of the present invention introduces a SIC module in the encoder, explicitly encodes the "structural role" as part of the message passing, and directly tells each node in the road graph the "structural role" to capture richer topological properties.

[0091] The road network spatial dual graph G is composed of a large number of subgraphs consisting of nodes and edges, and each node has a different "structural role". The "structural role" can be encoded into features by counting the number of occurrences of certain repeated subgraphs.

[0092] The embodiment of the present invention defines a set of small, connected, and recurring subgraphs for the road network pattern recognition task. They represent triangle, tailed-triangle, 4-cycle, chordal-cycle, 2-triangle and 4-clique respectively. Figure 7 shown.

[0093] For each subgraph First, find its isomorphic subgraph G in graph G. S For G S Each node in Find the orbital Orb of its mapping f(v) in H H (f(v)) to infer its role relative to H, Orb H (*) indicates the preset orbital mapping function, where orbit refers to the position or role of a node in H. Figure 7 As shown, nodes with the same color in the same subgraph are in the same orbit. Then, we calculate the SThe number of occurrences of a specific orbit in H is used to obtain its structural characteristics on the i-th orbit in H. in, represents the i-th track in H; f(v) represents the node mapping function that maps node v to the node in H. This paper uses the automorphism group Aut(H) function (i.e., the node transformation set that keeps the graph structure unchanged). Finally, by combining the subgraph set The counts of different subgraphs and tracks in the result are used to obtain the total feature vector The dimensions It is subgraph H i The number of tracks. Figure 8 Shows the use of triangle and chordal-cycle isomorphic subgraphs to count the structural identifiers X of the nodes v .

[0094] Although traditional GNNs can also distinguish nodes by introducing unique identifiers, when the order of nodes changes, the model output will also change, thus losing the permutation equivariance; while the structural identifier generated by SIC has permutation equivariance because the number of subgraphs is unchanged and changes in the order of nodes will not affect the structural identifier. In addition, using "structural roles" as identifiers is not necessarily unique, and multiple nodes can share the same structural identifier, which makes this method a good balance between uniqueness and generalization.

[0095] The encoder in the present invention can be implemented using a variety of networks, such as MPNN, GIN, etc. The embodiment of the present invention uses GCNN and introduces SIC, so the messages from neighboring nodes now also contain structural information, and each node v updates its state by combining its previous state with the neighborhood aggregate message. Each convolutional layer only processes the first-order neighborhood information, and the information transmission of multiple-order neighborhoods can be achieved by superimposing several convolutional layers. The propagation rules of the convolutional layer are as follows:

[0096]

[0097] in, Represents the adjacency matrix of the undirected graph G plus self-connection (that is, each vertex and itself plus an edge), is the identity matrix, yes The degree matrix of is the convolutional signal matrix of the lth layer, representing the features of each node in the lth layer, m l is the dimension of the node feature at layer l, the input layer H 0 =M, d1 represents the dimension of the node's geometric attribute feature, and d2 represents the dimension of the node's SC attribute feature.v Is the structure identifier of the node. and is the parameter training weight matrix and common bias of the lth layer. σ(·) is the nonlinear activation function. The hidden layer uses the linear rectifier unit function Relu(·), and the output layer uses the logistic sigmoid function.

[0098] (2) Decoder

[0099] During the training process of the SUGAR model, the node-level embedding Z uses the SC descriptor, which enhances the similarity between adjacent nodes and reduces the similarity between non-adjacent nodes. This embedding method makes adjacent nodes highly overlap in the embedding space, which helps to accurately reconstruct the association between nodes in the graph.

[0100] Specifically, if two graph nodes have a high degree of similarity in the embedding space, the possibility that there is an edge between them is relatively large. Therefore, the accuracy of the node-level embedding Z can be evaluated by comparing the distance between the reconstructed adjacency matrix A′ and the true adjacency matrix A. In this setting, the labels are not given externally, but directly derived from the structure of the graph (i.e., the adjacency matrix). This allows the model to learn complex relationships between nodes without external supervised labels, and optimizes the node-level embedding by minimizing the difference between the predicted and true adjacency matrices.

[0101] Furthermore, in this embodiment, the cross entropy loss function is used to calculate the distance between the reconstructed adjacency matrix A′ and the true adjacency matrix A:

[0102]

[0103] Where N is the total number of nodes. The logarithmic loss of the predicted connection probability is calculated for each pair of nodes, and the sum is averaged to form the loss value of the entire graph.

[0104] (3) Structure-aware graph representation learning module

[0105] The encoder output is a node-level embedding Z, but the geometric similarity measure between road networks requires a graph-level embedding. Typically, the conversion from node-level embedding to graph-level embedding is achieved through a readout function, such as average, maximum, or pooling.

[0106] The embodiment of the present invention uses an average value operator to embed the node level Convert to simple graph-level embedding Then the GCA module is introduced. The specific operation is as follows: First, a nonlinear transformation is performed on the graph-level embedding E. Get the global context in, is the embedding of the nth node; N represents the total number of nodes; is a learnable weight matrix, and its parameters are also updated by graph reconstruction to reweight and adjust the features of the global context vector; while tanh(·) provides the necessary nonlinearity, enabling the model to capture more complex graph structure features. The global context c provides the global structure and feature information of the graph adapted to the given similarity metric by learning the weight matrix. Then, we calculate an attention weight for each node based on c, and the calculation formula is It is represented as the inner product of node embedding and global context c, making its attention a n It can perceive the global context information. At the same time, the inner product result is processed using a nonlinear activation function σ(·) (the sigmoid function is used here) to ensure that the attention weight is in the range of (0,1). Finally, the resulting graph embedding is the weighted sum of node embeddings Z, i.e. The following formula summarizes the process of generating structure-aware graph-level embeddings based on GCA:

[0107]

[0108] The structure-aware road network graph representation learning model SUGAR proposed in an embodiment of the present invention, which is independent of specific downstream tasks, constructs a road graph by combining topological edges and spatial edges, uses an encoder-decoder architecture with SIC as the core to perform unsupervised learning of graph node embedding, and generates structure-aware graph-level embedding based on the attention mechanism of GCA.

[0109] Based on the above embodiments, the embodiment of the present invention provides a geometric similarity measurement method based on road network graph level embedding. The road graph level embedding obtained by attention weighting is denoted as Because important nodes are given higher weights, the embedding representation maintains the main structure of the road network. The two road networks input in the test phase are denoted as E S1 and E S2 Therefore, the geometric similarity measure between road networks is regarded as the similarity between two high-dimensional vectors ( s im), the similarity is calculated by cosine similarity, and the calculation formula is as follows:

[0110]

[0111] After obtaining the similarity, the geometric similarity of graph embedding is used to identify the road network pattern. Specifically, four standard road network patterns are first selected as standard anchor points: grid pattern, parallel pattern, complex interchange pattern and radial pattern. Then the road network of the test set is measured with these four standard anchor points for geometric similarity, and four different similarity values ​​are obtained. The pattern corresponding to the highest similarity value is used as the classification of the road network, but if the highest similarity value is lower than 0.5, it is classified as an irregular pattern.

[0112] Furthermore, in order to reduce the bias that may be caused by a single anchor point and enhance the generalization ability of the model to the road network pattern, this embodiment selects four representative anchor points for each road network pattern to cover the typical features and changes of the category as much as possible. Correspondingly, the similarity measure is converted into the average similarity measure. The overall process is as follows Fig. 9 shown.

[0113] Corresponding to the above method, an embodiment of the present invention also provides a road network pattern recognition system based on unsupervised graph representation learning, including a space dual graph construction module, an adjacency matrix generation module, a feature matrix generation module, a graph-level embedding generation module and a road network pattern recognition module.

[0114] The spatial dual graph construction module is used to take the stroke sections in the target road network data as the nodes of the graph, and take the connection relationship between adjacent sections and the spatial relationship added for spatially similar sections as the edges of the graph to construct the spatial dual graph of the target road network data; the adjacency matrix generation module is used to generate the adjacency matrix of the target road network data according to the spatial dual graph; the feature matrix generation module is used to extract the section features of each stroke section in the target road network data to generate the feature matrix of the target road network data; wherein the section features include the geometric attribute features and the shape context SC attribute features of the section; the graph-level embedding generation module is used to generate the feature matrix of the target road network data based on the target road network data; The spatial dual graph, adjacency matrix and feature matrix of the marked road network data are used to perform road graph representation learning using a preset graph neural network model to generate a graph-level embedding of the target road network data; wherein the graph neural network model includes a GAE model and a structure-aware graph representation learning module; the structure-aware graph representation learning module uses the node-level embedding output by the encoder in the GAE model as input to generate a graph-level embedding; the road network pattern recognition module is used to measure the similarity between the graph embedding of the target road network data and the graph embedding of multiple preset standard anchor points, and the road network pattern to which the standard anchor point with the highest similarity value belongs is used as the road network pattern of the original road network data.

[0115] It should be noted that the road network pattern recognition system provided in the embodiment of the present invention is for implementing the above method. Its specific functions can be referred to the above method embodiments, which will not be described in detail here.

[0116] In order to verify the effectiveness of the solution of the present invention, the present invention also conducted relevant experiments.

[0117] 1. Experimental data and settings

[0118] (1) Experimental Dataset

[0119] The experimental data for road network pattern recognition comes from the road networks of 10 cities including Beijing, Xi'an, and Ansbach, Germany, downloaded from OpenStreetMap (OSM). The original road network was preprocessed by eliminating pseudo nodes and deleting short dangling line segments, and five road network patterns were collected: grid, complex interchange, radial, parallel, and irregular. 220 samples were made for each pattern, totaling 1,100 samples. The road networks of the first four patterns have typical spatial structures and road connectivity functions and appear frequently in road networks. Except for these four types of road network patterns, other road networks are marked as irregular patterns. In order to ensure that the model can be effectively evaluated on an independent dataset and to ensure sufficient training data, a ratio of 6:2:2 is used to divide the training set, validation set, and test set.

[0120] (2) Experimental setup

[0121] In the experiment, multiple SUGAR models were constructed by changing the model depth (i.e., the number of convolutional layers), named SUGAR-X, where X represents the model depth. Each convolutional layer contains 64 first-order Chebyshev polynomial kernels. The hidden layer uses the linear rectifier unit function Relu(·), the output layer uses the logical sigmoid function, and the optimization algorithm uses Adam. During the training process, after multiple tests, the hyperparameters such as batch size, learning rate, number of iterations, dropout value, and embedding dimension were set to 32, 0.001, 2000, 0.3, and 128, respectively. The training and validation process is unsupervised, and only the test set is labeled: y∈{0,1,2,3,4}, which represent grid, parallel, complex interchange, radial, and irregular patterns, respectively.

[0122] The SUGAR model is compared with four road network pattern recognition models. First, two rasterization models for complex interchange pattern recognition are compared, using AlexNet and [and GoogLeNet methods, generating 224×224 pixel images for each road network test sample as input, with a rendering width of 2 pixels. For the AlexNet model, the learning rate, batch size, maximum number of iterations, and dropout value are set to 0.02, 16, 10000, and 0.5, respectively. For the GoogLeNet model, these parameters are set to 0.02, 16, 20000, and 0.8. The Multi-Layer Perceptron (MLP) model is also compared. The MLP model contains a hidden layer of 64 neurons and uses Relu (·) as the activation function. Finally, it is also compared with the SOTA model GCNN-4 proposed for grid pattern recognition and complex overpass pattern recognition. The model contains 1 input layer, 4 graph convolution layers, and 1 output layer, using third-order Chebyshev polynomial convolution kernels, each convolution layer contains 64 kernels, and the maximum number of iterations is 2000. All of the above models randomly select 10% of the node labels for training.

[0123] In addition, the proposed SUGAR is compared with three other GAE models with ablation. The first model does not use SIC and GCA, but only uses the basic GCNN convolutional layer, which is called GCNN_GAE-3. The second model is GCA_GAE-3 using only GCA; the third model is SIC_GAE-3 using only SIC. All three models achieve the best pattern recognition effect in three convolutional layers, and the other parameter settings are exactly the same as the SUGAR model.

[0124] (3) Evaluation indicators

[0125] The experiment uses classification accuracy, that is, the ratio of correctly classified samples to the total sample size, to evaluate the pattern classification results. This is the most direct evaluation indicator, which calculates the ratio of the overall correctly classified road network patterns to the total road network samples.

[0126]

[0127] In addition, three quantitative metrics are used: precision, recall, and F1-score. For the results of each mode type, the three metrics are calculated as follows:

[0128]

[0129] For each mode category, precision is calculated as the ratio of the number of networks correctly predicted as that mode category to the total number of networks predicted as that mode category, while recall is calculated as the ratio of the number of networks correctly predicted as that mode category to the total number of networks actually in that mode category. TP, FP, and FN represent the number of true positives, false positives, and false negatives, respectively. Taking the results of the grid mode as an example, true positive results refer to network samples that have been manually marked and predicted as grid mode types. Network samples predicted as grid mode but manually marked as another type are considered false positive results. Network samples manually marked as grid mode but predicted as another type are considered false negative results. F1-score is the harmonic mean of precision and recall, and is an indicator that comprehensively considers precision and recall. F1-score can be calculated for each category, and the total F1-score can also be calculated as an evaluation of overall performance.

[0130] (II) Experimental results and analysis

[0131] Table 2 lists the pattern classification accuracy of SUGAR models with different depths and 7 comparison models in the test sample set.

[0132] Table 2 Classification accuracy of different models for test sample sets

[0133]

[0134] It can be found that the test accuracy of the SUGAR model is between 84.55% and 93.18%, which is significantly better than the AlexNet and GoogLeNet models. This is because the SUGAR model takes into account the connection relationship between road segments, while the gridding process makes it difficult for the AlexNet and GoogLeNet models to directly use this structured information, affecting the recognition performance. In addition, the SUGAR model is also better than the MLP model because the former effectively aggregates the information of nodes and their neighboring nodes through graph convolution, capturing the local and global structural features between nodes, while the MLP model can only process based on the independent features of each node and performs poorly when processing graph data. The SOTA model GCNN-4 achieved an accuracy of 89.55% on our test set, and the unsupervised training GCNN_GAE-3 model had an accuracy of 85.45%, a decrease of 4.10%. This decrease is due to the fact that the GCNN_GAE-3 model cannot directly use labels for optimization like supervised learning, but its performance still shows the potential of unsupervised learning in road network pattern recognition.

[0135] Comparing SUGAR models of different depths, the classification accuracy first increases and then decreases with increasing depth. There may be two reasons for this phenomenon: first, due to the limited sample size, deep models are prone to overfitting, resulting in poor performance on the test set; second, the continuous product effect of the back-propagation process may cause the gradient of the shallow convolutional layer to disappear or explode, affecting model learning. Therefore, although the overall accuracy of the SUGAR model is higher than that of the GCNN-4 model, its test accuracy is lower than that of the GCNN-4 model when the convolutional layer is 1 layer or more than 4 layers.

[0136] In addition, with the same number of convolutional layers, the performance of the SUGAR model is better than that of the GCNN_GAE, GCA_GAE, and SIC_GAE models. This result further verifies the importance of SIC and GCA. SIC enhances the model's ability to capture local structural features, while GCA improves the model's ability to understand the overall topological relationship. Specifically, SIC provides rich topological structure information for each graph node by identifying and counting the tracks of subgraphs, helping the model to more accurately describe and distinguish road network patterns. GCA dynamically adjusts the importance of nodes to ensure that the model focuses on the areas most relevant to the task, thereby improving the accuracy of classification and recognition.

[0137] As shown in Table 2, the SUGAR-3 model has the highest classification accuracy, reaching 93.18%. Therefore, we use it as the benchmark model for road network pattern recognition. Table 3 shows the confusion matrix and evaluation indicators of the recognition results of the test set. Among them, the grid, complex interchange and parallel patterns performed relatively well in terms of precision and recall, which were 93.62% and 100%, 97.62% and 93.18%, and 97.67% and 95.46%, respectively. This may be related to the unique spatial structure patterns of these three pattern types, such as the four-way vertical orthogonal distribution of the grid pattern, the two-way continuous distribution of the parallel pattern, and the central convergence distribution of the complex intersection, which makes these patterns easier to distinguish. There are also some misidentifications, which may be caused by confusion due to the structural similarities between different patterns, such as the diversity and complexity of the irregular pattern, which makes it easy to confuse with other patterns in some features.

[0138] Table 3 Confusion matrix and classification results evaluation of the SUGAR-3 model on the test dataset

[0139]

[0140] Fig.10The pattern recognition results of some correctly identified road network samples and all incorrectly identified samples are shown. In general, the model has a good recognition effect, but there are errors in the recognition of some patterns. Specifically, the model mistakenly identifies some parallel patterns (20, 21) and irregular patterns (30) as grid patterns. This may be because there are some small burrs when the road network samples are intercepted, which makes the samples have certain grid characteristics, resulting in misjudgment. In addition, the model mistakenly identifies multiple irregular patterns (32, 33, 34, 35) as radial patterns, because irregular patterns are mostly distributed in a divergent manner, and some samples may show characteristics similar to radial patterns. In some cases, the model identifies irregular pattern (30) as a grid pattern and irregular pattern (31) as a complex interchange pattern. This is because these samples have certain grid characteristics and complex interchange characteristics, respectively, which interfere with the model's judgment. It is worth noting that due to the subjective intention of the road network pattern, some ambiguous situations will occur during recognition. For example, although samples 13, 14, 26, 27, and 28 are classified as irregular patterns, they also have some characteristics of complex interchange patterns and radial patterns. Therefore, the samples have a high similarity with these patterns, but because they do not exceed the threshold, they are ultimately misclassified as irregular patterns. A similar situation also occurs in sample 12, which has the characteristics of both complex interchange patterns and radial patterns. Therefore, the similarity of these two patterns is very high during identification, resulting in misclassification. From the results, it can be seen that the model is more likely to identify radial patterns as irregular patterns, probably because the samples of the radial pattern themselves have some characteristics of irregular patterns. When the radial centers are not concentrated at one point, it is difficult for the model to classify correctly. The grid pattern is a pattern that is difficult to be misclassified because its characteristics are obvious and very different from the radial pattern and irregular pattern. Its vertical orthogonality can also be easily extracted through the direction of the road section.

[0141] The present invention is oriented to road network pattern recognition and proposes a recognition method based on geometric similarity graph representation learning. The method uses spatial dual graph modeling, cognitively inspired graph node feature extraction, introduces subgraph isomorphism count (SIC) in the segment embedding learning stage and global context attention mechanism (GCA) in the graph embedding generation stage, establishes a mapping between the input road graph and the output road network pattern, enhances the model's representation performance of the road network, and thus improves the effect of road network pattern recognition. Road network pattern recognition experiments were carried out using real data. The results show that the classification accuracy of the method of the present invention reaches 93.18%, which is higher than the grid-based AlexNet and GoogLeNet models, and also higher than the supervised MLP and GCNN models based on neural networks. Under the same parameter settings, it is also better than ablation models such as GCNN_GAE, SIC_GAE and GCA_GAE. Finally, through the comparative analysis of road network pattern samples and graph-level embedding representations, it can be found that the SUGAR model can effectively cluster the generated graph embeddings of the same pattern and the boundaries between different categories are relatively clear, verifying the effectiveness of SIC and GCA in improving recognition performance.

[0142] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A road network pattern recognition method based on unsupervised graph representation learning, characterized in that: include: Step 1: Take the stroke sections in the target road network data as nodes of the graph, take the connection relationship between adjacent sections and the spatial relationship added for spatially similar sections as edges of the graph, construct a spatial dual graph of the target road network data, and generate an adjacency matrix of the target road network data based on the spatial dual graph; Step 2: extracting the segment features of each stroke segment in the target road network data to generate a feature matrix of the target road network data; wherein the segment features include geometric attribute features and shape context SC attribute features of the segment; Step 3: Based on the spatial dual graph, adjacency matrix and feature matrix of the target road network data, a preset graph neural network model is used to perform road graph representation learning to generate a graph-level embedding of the target road network data; wherein the graph neural network model includes a GAE model and a structure-aware graph representation learning module; the structure-aware graph representation learning module uses the node-level embedding output by the encoder in the GAE model as input to generate a graph-level embedding; Step 4: Measure the similarity between the graph embedding of the target road network data and the graph embedding of multiple preset standard anchor points, and take the road network pattern of the standard anchor point with the highest similarity value as the road network pattern of the original road network data.

2. A road network pattern recognition method based on unsupervised graph representation learning according to claim 1, characterized in that: Step 1 also includes: extracting spatially similar road sections and adding spatial relationships for the spatially similar road sections; wherein the spatially similar road sections refer to two road sections that meet preset spatial requirements but have no connection relationship; the preset spatial requirements refer to that the distance similarity and angle similarity of the two road sections are both greater than 0.

3. A road network pattern recognition method based on unsupervised graph representation learning according to claim 2, characterized in that: The distance similarity between two road segments is calculated according to the following formula: Similarity with angle Among them, sp_dist(s i ,s j ) indicates calculating two road segments s i and j One-way Hausdorff distance between, ag_dist(s i ,s j ) indicates calculating two road segments s i and j The absolute angular distance between ds and δ as Indicates two preset distance thresholds.

4. The road network pattern recognition method based on unsupervised graph representation learning according to claim 1, characterized in that: The geometric attribute features of the road section include position features extracted from a macroscopic scale, connection features extracted from a mesoscopic scale, and length features, shape features, and angle features extracted from a microscopic scale.

5. The road network pattern recognition method based on unsupervised graph representation learning according to claim 1, characterized in that: The process of extracting the SC attribute features of the road section includes: Establish a polar coordinate system at the midpoint of the target road section, and set the maximum radius of the polar coordinate system to r=γ×L(s); wherein L(s) is the length of the target road section, and γ is a scale factor; In the polar coordinate system, p concentric circles are divided at equal distances, and q sectors are divided at equal angles to obtain p×q sectors, and the value of each sector is defined as the ratio of the total length of the road segments in the area to the total length of the road segments in the context range; The values ​​of p×q sectors are represented as a vector, which is the SC attribute feature of the target road section.

6. A road network pattern recognition method based on unsupervised graph representation learning according to claim 1, characterized in that: A subgraph isomorphism counting SIC module is provided in the encoder, and the SIC module is used to generate structure identifiers for nodes in the space dual graph. The encoder performs neighborhood information transmission based on the structure identifier, adjacency matrix and feature matrix of the node to generate node-level embedding.

7. The road network pattern recognition method based on unsupervised graph representation learning according to claim 1, characterized in that: The method further includes: training the encoder and decoder in the GAE model; during the training process, the decoder reconstructs the inter-node association based on the node-level embedding output by the encoder to generate a reconstructed adjacency matrix, and uses the following cross entropy loss function to calculate the distance between the reconstructed adjacency matrix A′ and the true adjacency matrix A: Where N represents the total number of nodes in the space-dual graph.

8. The road network pattern recognition method based on unsupervised graph representation learning according to claim 6, characterized in that: The process of generating node-level embedding specifically includes: Define a subgraph collection For a subgraph collection For each subgraph H in the space dual graph G, first find its isomorphic subgraph G in the space dual graph G S , for G S Each node in Find the orbital Orb of its mapping f(v) in subgraph H H (f(v)), Orb H (*) represents the preset orbital mapping function, where the orbit refers to the position or role of a node in H; then, the position of each node v in different isomorphic subgraphs G is calculated. S The track times, and obtain the structural characteristics of node v on the i-th track of subgraph H represents the i-th track in subgraph H; f(v) represents the node mapping function that maps node v to the node in subgraph H; finally, the combined subgraph set The count of different subgraph structures and tracks in the node is used to obtain the structure identifier of the node in, It is subgraph H i The number of mid-tracks; The adjacency matrix, feature matrix, and structure identifier are updated according to the following propagation rules between adjacent convolutional layers: in, A is the adjacency matrix, is the identity matrix, yes The degree matrix of is the convolutional signal matrix of the lth layer, representing the features of each node in the lth layer, m l is the dimension of the node feature at layer l; the input layer H 0 =M, represents the feature matrix, d1 represents the dimension of the node's geometric attribute feature, and d2 represents the dimension of the node's SC attribute feature; X v is the structure identifier of the node, and is the parameter training weight matrix and common bias of the lth layer, σ(·) is the nonlinear activation function, and N represents the total number of nodes in the space-dual graph G.

9. The road network pattern recognition method based on unsupervised graph representation learning according to claim 1, characterized in that: The structure-aware graph representation learning module takes the node-level embedding output by the encoder in the GAE model as input to generate graph-level embedding, specifically including: Among them, E S represents graph-level embedding, Z represents the node-level embedding output by the encoder, is the embedding of the nth node, N represents the total number of nodes; is a learnable weight matrix, and tanh(·) is an activation function; represents the attention weight of the node, and σ(·) is a nonlinear activation function.

10. A road network pattern recognition system based on unsupervised graph representation learning, characterized in that: include: A spatial dual graph construction module is used to construct a spatial dual graph of the target road network data by taking the stroke road segments in the target road network data as the nodes of the graph, and taking the connection relationship between adjacent road segments and the spatial relationship added for spatially similar road segments as the edges of the graph; An adjacency matrix generation module, used to generate an adjacency matrix of target road network data according to the space dual graph; A feature matrix generation module is used to extract the segment features of each stroke segment in the target road network data to generate a feature matrix of the target road network data; wherein the segment features include geometric attribute features and shape context SC attribute features of the segment; A graph-level embedding generation module, which is used to perform road graph representation learning based on the spatial dual graph, adjacency matrix and feature matrix of the target road network data using a preset graph neural network model to generate a graph-level embedding of the target road network data; wherein the graph neural network model includes a GAE model and a structure-aware graph representation learning module; the structure-aware graph representation learning module uses the node-level embedding output by the encoder in the GAE model as input to generate a graph-level embedding; The road network pattern recognition module is used to measure the similarity between the graph embedding of the target road network data and the graph embedding of multiple preset standard anchor points, and take the road network pattern to which the standard anchor point with the highest similarity value belongs as the road network pattern of the original road network data.

Images