Road network simplification method, device and equipment based on multi-branch graph neural network and storage medium

Abstract

The application provides a road network simplification method and device based on a multi-branch graph neural network, equipment and a storage medium, and relates to the technical field of road network simplification. The global topology, local neighborhood and attention weighted features of the road network are fused and extracted by the multi-branch graph neural network, and unified node embedding is generated through linear fusion. Then, the redundant nodes are identified based on unsupervised node importance evaluation and dynamic threshold optimization. A multi-criterion collaborative progressive pruning strategy is adopted to protect key nodes and implement dynamic connectivity repair. The method ensures geometric fidelity, topological consistency and semantic integrity while achieving efficient adaptive simplification of the road network.

Description

Method, apparatus, equipment, and storage medium for simplifying road networks based on multi-branch graph neural networks Technical Field

[0001] This application relates to the field of road network simplification technology, and in particular to a road network simplification method, apparatus, device and storage medium based on a multi-branch graph neural network. Background Technology

[0002] With the continuous development of smart city and digital twin construction, urban road networks, as a key spatial infrastructure, are increasingly in need of multi-scale, high-fidelity simplification technologies in applications such as autonomous driving navigation, urban traffic planning, and intelligent planning management. Currently, much publicly available road network information suffers from redundancy, excessive information volume, and chaotic classification, resulting in slow search speeds, numerous errors, and low decision-making efficiency.

[0003] Currently, road network simplification mainly employs geometric rule algorithms such as Douglas-Puk. While these algorithms can compress geometry, they easily disrupt topological connectivity. Existing graph neural network-based methods can learn node representations, but they mostly use single graph convolution or attention architectures, making it difficult to comprehensively capture the global topology, local structure, and semantic features of the road network.

[0004] Therefore, a road network intelligent simplification method that can integrate multi-dimensional features and has an adaptive fusion mechanism is needed to achieve unified optimization of geometric accuracy, topological integrity and semantic consistency. Summary of the Invention

[0005] To achieve unified optimization of geometric accuracy, topological integrity, and semantic consistency, this application provides a method, apparatus, device, and storage medium for simplifying road networks based on multi-branch graph neural networks.

[0006] Firstly, the road network simplification method based on a multi-branch graph neural network provided in this application adopts the following technical solution:

[0007] A method for simplifying road networks based on multi-branch graph neural networks includes:

[0008] The original road network data is preprocessed to construct a hierarchical graph structure, including extracting nodes, paths and spatial relationships, establishing a road topology map, and performing map projection transformation and road classification.

[0009] Global topological features, local neighborhood features, and attention-weighted features of the road network are extracted from the graph structure by parallel GCN, GraphSAGE, and GAT branches, respectively.

[0010] The global topological features learned by the GCN branch, the local neighborhood features generated by the GraphSAGE branch, and the attention-weighted features generated by the GAT branch are concatenated to generate a unified node embedding representation.

[0011] An unsupervised node importance evaluation system for graph autoencoders is constructed to calculate node importance scores in the node embedding representation and identify redundant nodes by combining a dynamic threshold optimization mechanism.

[0012] A progressive pruning strategy based on multiple criteria is adopted to protect topological nodes, high-level road nodes, and geographic entity centroid nodes, and to implement dynamic connectivity repair.

[0013] The geometric fidelity, topological consistency, and semantic integrity of the simplified results are quantified through a multi-dimensional quality assessment system.

[0014] An optimized road network is generated based on multi-scale pruning ratio parameters, and a simplified structure that meets application requirements is output.

[0015] Optionally, the road classification includes: obtaining the highway attribute value of the road and dividing the road into three functional levels according to predefined rules;

[0016] When the highway attribute value is motorway, trunk, or primary_link, the road level is 0, which corresponds to expressway, trunk road, and connecting road, and undertakes the main traffic flow function.

[0017] When the highway attribute value is secondary, tertiary or secondary_link, the road level is 1, corresponding to secondary arterial roads, tertiary roads and their connecting lines, and it undertakes the function of regional traffic connection.

[0018] When the highway attribute value is residential, service or living_link, the road level is 2, corresponding to residential area roads, service roads and living area connecting roads, and undertakes local traffic service functions;

[0019] The node road level is determined by the minimum value aggregation strategy, and centroid processing and DBSCAN spatial clustering optimization are performed on face-shaped geographic entities.

[0020] Optionally, the step of extracting global topological features, local neighborhood features, and attention-weighted features of the road network from the graph structure through parallel GCN, GraphSAGE, and GAT branches includes:

[0021] The GCN branch captures global topological dependencies through spectral graph convolution operations, and its propagation formula is as follows:

[0022] in and They represent the first The node feature matrix and trainable weight matrix of the layer; A is the original adjacency matrix. It is then the identity matrix; express The degree matrix;

[0023] The GraphSAGE branch uses mean aggregation of neighbor features, calculated as follows:

[0024]

[0025]

[0026] in Represents the set of neighboring nodes. and Representing neighboring nodes respectively and Feature vectors in the current layer; It is a learnable weight matrix; It is a non-linear activation function;

[0027] The GAT branch weights and aggregates neighbor features through an attention mechanism that integrates edge attributes; the attention coefficients... The calculation is as follows:

[0028]

[0029] in It is a learnable weight; and Represents a node and eigenvectors; The eigenvectors of the edges; This is a splicing operation.

[0030] Optionally, the importance assessment uses unsupervised link prediction reconstruction error, and its loss function is defined as:

[0031]

[0032] in Represents the set of actual edges in the graph. Size; and Representing nodes respectively and The final embedding vector learned; This indicates the size of the negative sample set.

[0033] Optionally, multi-criteria collaborative pruning includes a triple protection mechanism:

[0034] Protect topological nodes and prevent connectivity disruption;

[0035] High-grade road nodes must be retained to maintain the integrity of the transportation backbone;

[0036] Protect the centroid nodes of geographic entities to ensure regional semantic consistency;

[0037] After pruning, isolated node pairs are quickly located using spatial indexing and connected edges are added to restore global connectivity.

[0038] Optionally, multi-dimensional quality assessment includes:

[0039] Geometric fidelity is achieved using the Hausdorff distance metric. To measure the difference in geometric shape between the simplified road network and the original road network, let the set of all points in the original road network graph be . The set of all points in the simplified road network diagram is For any two points and ,definition For point and points The Euclidean distance in a two-dimensional plane is the Hausdorff distance. Defined as

[0040] ;

[0041] Topological consistency is achieved through the entropy of node degree distribution. and attribute entropy Quantification;

[0042] Node degree distribution entropy is used to measure the complexity of topological connectivity patterns in a road network. It is defined by the degree of a node in the network. Let be the number of edges connected to this node. Count the degree of all nodes in the network to obtain the probability distribution of node degrees. The degree of a node is exactly The probability of is calculated using the following formula:

[0043]

[0044] The degree values ​​of the nodes appearing in the summation traversal of the network are... ;

[0045] Attribute entropy is used to measure the completeness of the retention of semantic attribute information of a road network. Let A be a road attribute to be evaluated, and all its discrete values ​​are... ,set up This indicates that the value of this attribute is exactly 1. The information entropy of attribute A is the proportion of the number of roads to the total number of roads. Represented as:

[0046] .

[0047] Optionally, multi-scale modeling generates simplified road network sequences by setting different pruning ratio parameters, and selects the optimal simplification ratio based on the comprehensive evaluation results of geometric fidelity, topological consistency and semantic integrity.

[0048] Secondly, this application provides a road network simplification device based on a multi-branch graph neural network, comprising:

[0049] The data processing module is used to preprocess the raw road network data, construct a hierarchical graph structure, including extracting nodes, paths and spatial relationships, establishing a road topology map, and performing map projection transformation and road classification.

[0050] The feature extraction module is used to extract global topological features, local neighborhood features, and attention-weighted features of the road network from the graph structure through parallel GCN, GraphSAGE, and GAT branches, respectively.

[0051] The node representation module is used to concatenate the global topological features learned by the GCN branch, the local neighborhood features generated by the GraphSAGE branch, and the attention-weighted features generated by the GAT branch to generate a unified node embedding representation.

[0052] The scoring module is used to construct an unsupervised node importance evaluation system for graph autoencoders to calculate node importance scores in the node embedding representation and identify redundant nodes by combining a dynamic threshold optimization mechanism.

[0053] The repair module is used to protect topological nodes, high-level road nodes, and geographic entity centroid nodes by adopting a progressive pruning strategy based on multiple criteria, and to implement dynamic connectivity repair.

[0054] The quantification module is used to quantify the geometric fidelity, topological consistency, and semantic integrity of simplified results through a multi-dimensional quality assessment system.

[0055] The output module is used to generate an optimized road network based on multi-scale pruning ratio parameters, and output a simplified structure that meets application requirements.

[0056] Thirdly, this application provides a computer device, the device comprising: a memory and a processor, wherein the processor, when executing computer instructions stored in the memory, performs the method described above.

[0057] Fourthly, this application provides a computer-readable storage medium including instructions that, when executed on a computer, cause the computer to perform the method described above.

[0058] In summary, this application extracts global topology, local neighborhood, and attention-weighted features of a road network through multi-branch graph neural network fusion, and generates a unified node embedding through linear fusion. Then, redundant nodes are identified based on unsupervised node importance assessment and dynamic threshold optimization. A multi-criteria collaborative progressive pruning strategy is employed to protect key nodes and implement dynamic connectivity repair. This method achieves efficient adaptive simplification of the road network while ensuring geometric fidelity, topological consistency, and semantic integrity. Attached Figure Description

[0059] Figure 1 is a schematic diagram of the computer device structure of the hardware operating environment involved in the embodiment of this application;

[0060] Figure 2 is a flowchart illustrating the first embodiment of the road network simplification method based on a multi-branch graph neural network of this application.

[0061] Figure 3 is a structural block diagram of the first embodiment of the road network simplification device based on a multi-branch graph neural network of this application. Detailed Implementation

[0062] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0063] Referring to Figure 1, which is a schematic diagram of the computer device structure of the hardware operating environment involved in the embodiment of this application.

[0064] As shown in Figure 1, the computer device may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen or an input unit such as a keyboard; optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wireless-Fidelity (Wi-Fi) interface). The memory 1005 may be high-speed random access memory (RAM) or stable non-volatile memory (NVM), such as a disk drive. The memory 1005 may also optionally be a storage device independent of the aforementioned processor 1001.

[0065] Those skilled in the art will understand that the structure shown in Figure 1 does not constitute a limitation on the computer device and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0066] As shown in Figure 1, the memory 1005, which serves as a storage medium, may include an operating system, a network communication module, a user interface module, and a road network simplification program based on a multi-branch graph neural network.

[0067] In the computer device shown in Figure 1, the network interface 1004 is mainly used for data communication with the network server; the user interface 1003 is mainly used for data interaction with the user; the processor 1001 and the memory 1005 in this application can be located in the computer device. The computer device calls the road network simplification program based on multi-branch graph neural network stored in the memory 1005 through the processor 1001 and executes the road network simplification method based on multi-branch graph neural network provided in the embodiment of this application.

[0068] This application provides a method for simplifying road networks based on multi-branch graph neural networks. Referring to Figure 2, which is a flowchart of the first embodiment of the method for simplifying road networks based on multi-branch graph neural networks in this application.

[0069] In this embodiment, the road network simplification method based on multi-branch graph neural networks includes the following steps:

[0070] Step S10: Preprocess the original road network data to construct a hierarchical graph structure, including extracting nodes, paths and spatial relationships, establishing a road topology map, and performing map projection transformation and road level classification.

[0071] It should be noted that the technical terms and their corresponding explanations in this embodiment are as follows:

[0072] GCN (Graph Convolutional Network): A graph neural network based on spectral graph theory, which is good at capturing local neighborhood features;

[0073] GraphSAGE: A graph neural network that supports inductive learning, dynamic graphs, and large-scale networks;

[0074] GAT (Graph Attention Network): Introduces an attention mechanism into graph networks, allowing the model to determine the importance of different neighboring nodes;

[0075] Hausdorff distance: primarily used to measure the similarity between two geometric shapes;

[0076] Road classification: Roads are generally classified into three levels: arterial roads, secondary arterial roads, and other branch roads.

[0077] In practice, data preprocessing and hierarchical graph structure construction extract three basic elements from the original OSM data: nodes, paths, and spatial relationships, and create a road map with a complete topological structure. ,in It is a set of nodes. It is an edge set. Then, the geographic coordinate system is converted to a projected coordinate system through map projection transformation to establish an accurate spatial calculation benchmark. Based on this, a three-level road function classification system is established, dividing the road network into three functional levels: arterial roads, secondary arterial roads, and local roads. Level 0: When the value of the highway attribute is motorway, trunk, or primary_link, the road level is 0. This type usually corresponds to major traffic corridors such as expressways, arterial roads, or arterial road connectors.

[0078] Level 1: When the highway attribute value is secondary, tertiary, or secondary_link, the road level is 1. This category corresponds to secondary roads, tertiary roads, and their connecting lines.

[0079] Level 2: When the highway attribute value is residential, service, or living_link, the road level is 2. This category corresponds to residential roads, service roads, and roads connecting residential areas.

[0080] Based on the above rules, roads are divided into three functional levels: Level 0 (highways, arterial roads and connecting roads) mainly handles the main traffic flow between cities or regions; Level 1 (secondary arterial roads, tertiary roads and their connecting roads) mainly handles the traffic connection and distribution functions within the region; and Level 2 (residential area roads, service roads and residential area connecting roads) mainly serve the local traffic within the block.

[0081] In road network modeling, the node level of an intersection is usually determined by the highest level of all roads it connects to; this is known as the "minimum aggregation strategy," ensuring that a node matches the function of the most important road passing through it. Furthermore, for isal geographic features such as buildings, water bodies, and parks, this embodiment first extracts their geometric centroids as representative nodes and uses the DBSCAN spatial clustering algorithm to optimize high-density node groups, thereby identifying and simplifying the core road network structure.

[0082] The road node classification is determined using a minimum value aggregation strategy, prioritizing the protection of arterial road nodes. Simultaneously, isometric geographic entities such as buildings, water bodies, and parks undergo centroid processing, and their geometric center coordinates are calculated to simplify the data. Complex polygon boundary nodes are then optimized using the DBSCAN spatial clustering algorithm. The clustering parameter eps is automatically calculated based on the data spatial range, ensuring the adaptability of the processing. The formulas for calculating the geometric center and DBSCAN clustering optimization are as follows:

[0083] Geometric center calculation:

[0084]

[0085] DBSCAN clustering optimization:

[0086]

[0087]

[0088] in This represents the total number of boundary nodes for a isometric geographic entity. and Indicates the first The x and y coordinates of each boundary node in the projected coordinate system The length of the diagonal of the outer rectangle representing the entire road network map data range.

[0089] Step S20: Extract the global topological features, local neighborhood features, and attention-weighted features of the road network from the graph structure by using parallel GCN, GraphSAGE, and GAT branches respectively.

[0090] In practical implementation, multi-branch graph neural network feature extraction utilizes three parallel branches—GCN, GraphSAGE, and GAT—to extract global topological features, local neighborhood features, and attention-weighted features of roads, respectively, achieving collaborative learning of multi-dimensional features of the road network. The GCN branch constructs global topological relationships between nodes through multi-layer graph convolution, and its calculation process follows the hierarchical propagation formula of graph convolutional networks:

[0091]

[0092] in and They represent the first The node feature matrix and trainable weight matrix of the layer; A is the original adjacency matrix. It is then the identity matrix; express The degree matrix.

[0093] The GraphSAGE branch is used to learn the local neighborhood features of a node. It employs a fixed neighborhood aggregation mechanism, aggregating the features of all neighboring nodes for each node using the mean. Then, it transforms and fuses the aggregated neighborhood features with the node's own features. Finally, it uses activation functions and Dropout regularization. This process captures the feature representation of a node within its local topology. Its forward propagation process is represented as follows:

[0094]

[0095]

[0096] in Represents the set of neighboring nodes. and Representing neighboring nodes respectively and Feature vectors in the current layer; It is a learnable weight matrix; It is a non-linear activation function.

[0097] The GAT branch employs a hierarchical attention mechanism from single-head to multi-head to achieve differentiated feature propagation, mainly including three stages: edge attribute mapping, attention coefficient calculation, and weighted feature aggregation. It calculates the importance of connections between nodes using trainable attention weights and introduces edge attribute features as topological priors. The attention coefficients are normalized using a softmax function, and finally, attention-enhanced node representations are generated based on weighted summation. This method achieves differentiated weighted aggregation of neighborhood node features. The specific calculation method is as follows:

[0098] Linear transformation and input projection:

[0099]

[0100] In the formula For a trainable weight matrix, and This represents the original feature vector of the node.

[0101] Attention coefficient calculation for integrated edge attributes:

[0102]

[0103] in For trainable attention weight vectors, For the edge The attribute feature vector.

[0104] Attention weight normalization, weighted summation, and multi-head output concatenation calculation:

[0105]

[0106] In the formula Represents a node The neighborhood group, For the total number of attention heads, Indicates the activation function; Indicates the first In each attention head, the node For nodes Attention weights; For the first Linear transformation weight matrix for each attention head; For nodes The input feature vector.

[0107] Step S30: Concatenate the global topological features learned by the GCN branch, the local neighborhood features generated by the GraphSAGE branch, and the attention-weighted features generated by the GAT branch to generate a unified node embedding representation.

[0108] In its implementation, this step achieves dynamic adaptive fusion of multi-branch features. The global topological features learned by the GCN branch, the local neighborhood features generated by the GraphSAGE branch, and the attention-weighted features generated by the GAT branch are intelligently integrated through a dynamic concatenation mechanism to form a feature representation that integrates multi-dimensional information. Unlike traditional simple concatenation, this method introduces a dynamic weight allocation mechanism to adaptively evaluate and fuse the contributions of each branch feature, achieving a transformation from "static concatenation" to "dynamic adaptive fusion."

[0109] This dynamic fusion mechanism calculates the importance weights of each branch in real time using a lightweight attention module, and performs weighted fusion and dimensionality reduction based on these weights. It can dynamically adjust the fusion strategy according to the specific structural characteristics of the road network, thereby generating node embedding representations with uniform dimensions and high information density. This method not only enhances the discriminative ability of node representations but also improves the model's cognitive level and adaptability to diverse road network structures, providing crucial technical support for the entire optimization process.

[0110] Step S40: Construct an unsupervised node importance evaluation system for the graph autoencoder to calculate node importance scores in the node embedding representation, and identify redundant nodes by combining a dynamic threshold optimization mechanism.

[0111] It should be noted that the constructed unsupervised node importance evaluation system based on graph autoencoders utilizes an encoder-decoder structure to learn low-dimensional embedding representations of nodes. The encoder can extract node features using a multi-branch graph neural network, and the decoder reconstructs the network topology connections based on the inner product of node representations. The reconstruction loss function considers the actual characteristics of the road network, as shown in the following equation. Here, the positive sample edge set E mainly comes from the real connections of the original road network, while the negative sample edge set N is randomly sampled. This function allows the model to learn the inherent connection patterns in the network, thus enabling the evaluation of the importance of each node in maintaining the global topology.

[0112]

[0113] in Represents the set of actual edges in the graph. Size; and Representing nodes respectively and The final embedding vector learned; This indicates the size of the negative sample set.

[0114] Calculating the L2 norm of a node's representation vector yields its importance score, as shown in the formula below. In the feature space, nodes with larger norms are generally located at the topological center or key connection points of the network, containing richer structural information. For road networks, these nodes mostly correspond to traffic hubs or important intersections; their removal would impact network connectivity.

[0115]

[0116] To further improve the accuracy of pruning, a dynamic threshold optimization mechanism is introduced to identify redundant nodes.

[0117]

[0118] in This represents the optimal parameter, typically a threshold or weight within the range [0,1]. This usually refers to the F1 score; This usually refers to accuracy, expressed within a threshold. The proportion of samples predicted to be positive that are actually positive. Recall rate, representing the percentage of time elapsed after reaching a certain threshold. Below, the proportion of all truly positive samples that are correctly predicted as positive.

[0119] The pruning effect under different thresholds was evaluated on the validation set. By maximizing the F1 score to balance precision and recall, the key structural features of the network were preserved to the maximum extent while removing redundant nodes. This data-driven threshold selection method overcomes the limitations of traditional methods that rely on empirical settings, and provides a reliable importance ranking basis for subsequent multi-criteria collaborative pruning.

[0120] Step S50: Employ a progressive pruning strategy based on multiple criteria to protect topological nodes, high-level road nodes, and geographic entity centroid nodes, and implement dynamic connectivity repair.

[0121] It's important to note the multi-criteria collaborative progressive pruning and topology repair approach. First, a triple protection mechanism is established, using graph theory to identify key nodes in the road network. Removing these nodes would disrupt network connectivity, so they must be protected. Nodes of main roads are forcibly retained according to the road hierarchy to ensure the integrity of the traffic function skeleton. All geographic entity centroid nodes are protected to maintain the integrity of regional semantic information. During the pruning decision-making process, node importance scores are jointly evaluated with the triple protection mechanism, removing only nodes that do not meet any protection conditions and have low importance scores. After pruning, repair processing is performed on potentially isolated connected components. When constructing a spatial index to quickly find nearest neighbor pairs, new connecting edges are added in the middle to ensure the network maintains global connectivity after processing. This repair process avoids the network breakage problem common in traditional methods.

[0122] Step S60: Quantify the geometric fidelity, topological consistency and semantic integrity of the simplified results through a multi-dimensional quality assessment system.

[0123] Understandably, a multi-dimensional quality assessment system is needed. An evaluation framework integrating geometric, topological, and semantic features is constructed to provide a quantitative evaluation standard for road network optimization effectiveness. Geometric fidelity is achieved using Hausdorff distance. This measure the geometric differences between the simplified and original road networks. Let the set of points in the original road network graph be denoted as . The set of all points in the simplified road network diagram is For any two points and ,definition For point and points The Euclidean distance in a two-dimensional plane. Then the Hausdorff distance. Defined as:

[0124] ;

[0125] Topological consistency is achieved through the entropy of node degree distribution. and attribute entropy Conduct a quantitative assessment:

[0126] Node degree distribution entropy: Primarily used to measure the complexity of topological connectivity patterns in a road network. Defines the degree of a node in the network. Let be the number of edges connected to this node. Count the degree of all nodes in the network to obtain the probability distribution of node degrees. Let be... The degree of a node is exactly The probability of [the outcome] is calculated using the following formula:

[0127]

[0128] The degree values ​​of the nodes appearing in the summation traversal of the network are... Before and after simplification The closer the values ​​are, the better the topological consistency is maintained.

[0129] Attribute entropy: Used to measure the completeness of the retention of semantic attribute information in a road network. Let the attribute of a road (or node) to be evaluated be A, and all its discrete values ​​be... .set up This indicates that the value of this attribute is exactly 1. The proportion of the number of roads (or nodes) to the total number of roads (or nodes). Then the information entropy of attribute A. Defined as:

[0130]

[0131] Before and after simplification The closer the values ​​are, the more completely the semantic information of the attribute is preserved.

[0132] Step S70: Generate an optimized road network based on the multi-scale pruning ratio parameters, and output a simplified structure that meets the application requirements.

[0133] In practical implementation, multi-scale modeling and optimization iterations are employed. Different pruning ratio parameters are set to construct multi-scale spatial representation sequences of the road network. During quality assessment, independent quantitative analyses are performed from three dimensions: geometric fidelity, topological consistency, and semantic integrity, recording the performance of the model at each scale across these dimensions. After comparing the results, the simplification ratio that performs best in terms of geometric fidelity is selected, completing the model output for the current iteration.

[0134] Understandably, this embodiment provides an intelligent simplification method for road networks based on multi-branch graph neural network fusion. It takes the original road network data as input and uses graph autoencoders—GCN, GraphSAGE, and GAT modules stacked in parallel—to extract road topology, geometry, and multi-level semantic features from the road structure. The GCN module captures the global topological features of the graph structure, while the GraphSAGE module uses a mean aggregation mechanism to learn the local neighborhood structure information of nodes. The GAT module relies on an attention mechanism and combines edge attribute weights to assign differentiated weights to adjacent nodes, enhancing the expressive power of important connections. The feature vectors output by each module are concatenated and input into a linear fusion module for feature fusion, generating a unified node representation with strong expressive power.

[0135] During the node importance assessment phase, nodes are ranked based on the reconstruction error of link prediction. A progressive node removal process is carried out using an intelligent method that prioritizes the preservation of key topology points, main roads, and functional centroids. After each pruning operation, a dynamic connectivity restoration mechanism is used to reassess the network structure to ensure that the overall connectivity and semantic consistency of the road network can be maintained while removing redundant nodes.

[0136] The final output simplified road network preserves the overall orientation and local details of the original road network geometrically using Hausdorff distance evaluation; ensures connectivity and path reachability from a topological perspective; and maintains road hierarchy distribution and regional functional characteristics semantically. This method achieves end-to-end automated simplification of road networks, overcoming the shortcomings of traditional geometric methods that destroy connectivity, rule-based methods that rely on manual thresholding, and the insufficient feature learning representation of existing single deep learning models. While ensuring a high simplification rate, it also enhances the practical value of simplified road networks, providing high-quality simplified road data for fields such as mapmaking, navigation services, and spatial analysis.

[0137] It should be noted that the core technical features of this embodiment include the following:

[0138] 1. Multi-branch graph neural network feature extraction module: A graph autoencoder consisting of parallel stacked GCN, GraphSAGE, and GAT modules with the Transformer module is used to extract road topology, geometry, and multi-level semantic features from the original road network.

[0139] 2. Gated Adaptive Feature Fusion Module: The feature vectors output from each branch are concatenated and then generated into a unified node representation with strong expressive power through a gated adaptive linear fusion module, using dynamic weight allocation and feature fusion.

[0140] 3. An unsupervised node importance evaluation system based on graph autoencoders: The importance of nodes is ranked by utilizing the reconstruction error of link prediction and the norm of node representation, and a dynamic threshold optimization mechanism is introduced.

[0141] 4. Multi-criteria collaborative progressive pruning and topology repair mechanism: Establish a triple protection mechanism (protecting topological key points, prioritizing the retention of main paths, and maintaining functional centroids), perform progressive node removal by combining node importance scores, and use a dynamic connectivity restoration mechanism after pruning to repair any isolated connected components that may be generated.

[0142] This embodiment extracts global topology, local neighborhood, and attention-weighted features of the road network through multi-branch graph neural network fusion, and generates a unified node embedding through linear fusion. Then, redundant nodes are identified based on unsupervised node importance assessment and dynamic threshold optimization. A multi-criteria collaborative progressive pruning strategy is adopted to protect key nodes and implement dynamic connectivity repair. This method achieves efficient adaptive simplification of the road network while ensuring geometric fidelity, topological consistency, and semantic integrity.

[0143] Referring to Figure 3, which is a structural block diagram of the first embodiment of the road network simplification device based on a multi-branch graph neural network of this application.

[0144] As shown in Figure 3, the road network simplification device based on a multi-branch graph neural network proposed in this application includes:

[0145] The data processing module 10 is used to preprocess the original road network data, construct a hierarchical graph structure, including extracting nodes, paths and spatial relationships, establishing a road topology map, and performing map projection transformation and road level classification.

[0146] Feature extraction module 20 is used to extract global topological features, local neighborhood features and attention-weighted features of the road network in the graph structure through parallel GCN, GraphSAGE and GAT branches respectively;

[0147] The node representation module 30 is used to concatenate the global topological features learned by the GCN branch, the local neighborhood features generated by the GraphSAGE branch, and the attention-weighted features generated by the GAT branch to generate a unified node embedding representation.

[0148] The scoring module 40 is used to construct an unsupervised node importance evaluation system for the graph autoencoder, to calculate node importance scores in the node embedding representation, and to identify redundant nodes by combining a dynamic threshold optimization mechanism.

[0149] Repair module 50 is used to protect topological joints, high-level road nodes and geographic entity centroid nodes by adopting a progressive pruning strategy based on multiple criteria, and to implement dynamic connectivity repair.

[0150] Quantification module 60 is used to quantify the geometric fidelity, topological consistency and semantic integrity of simplified results through a multi-dimensional quality assessment system;

[0151] Output module 70 is used to generate an optimized road network based on multi-scale pruning ratio parameters and output a simplified structure that meets application requirements.

[0152] It should be understood that the above are merely illustrative examples and do not constitute any limitation on the technical solution of this application. In specific applications, those skilled in the art can make settings as needed, and this application does not impose any restrictions on this.

[0153] This embodiment extracts global topology, local neighborhood, and attention-weighted features of the road network through multi-branch graph neural network fusion, and generates a unified node embedding through linear fusion. Then, redundant nodes are identified based on unsupervised node importance assessment and dynamic threshold optimization. A multi-criteria collaborative progressive pruning strategy is adopted to protect key nodes and implement dynamic connectivity repair. This method achieves efficient adaptive simplification of the road network while ensuring geometric fidelity, topological consistency, and semantic integrity.

[0154] It should be noted that the workflow described above is merely illustrative and does not limit the scope of protection of this application. In practical applications, those skilled in the art can select some or all of it to achieve the purpose of this embodiment according to actual needs, and no restrictions are imposed here.

[0155] In addition, for technical details not described in detail in this embodiment, please refer to the method for simplifying road networks based on multi-branch graph neural networks provided in any embodiment of this application, which will not be repeated here.

[0156] Furthermore, it should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0157] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0158] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as read-only memory (ROM) / RAM, magnetic disk, optical disk), and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods of the various embodiments of this application. The above are only preferred embodiments of this application and do not limit the patent scope of this application. All equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for simplifying road networks based on multi-branch graph neural networks, characterized in that, include: The original road network data is preprocessed to construct a hierarchical graph structure, including extracting nodes, paths and spatial relationships, establishing a road topology graph, and performing map projection transformation and road level classification. Global topological features, local neighborhood features and attention-weighted features of the road network are extracted from the graph structure through parallel GCN, GraphSAGE and GAT branches, respectively. The global topological features learned by the GCN branch, the local neighborhood features generated by the GraphSAGE branch, and the attention-weighted features generated by the GAT branch are concatenated to generate a unified node embedding representation. An unsupervised node importance evaluation system for graph autoencoders is constructed to calculate node importance scores in the node embedding representation and identify redundant nodes by combining a dynamic threshold optimization mechanism. A progressive pruning strategy based on multiple criteria is adopted to protect topological key points, high-level road nodes, and geographic entity centroid nodes, and dynamic connectivity repair is implemented. A multi-dimensional quality assessment system quantifies the geometric fidelity, topological consistency, and semantic integrity of the simplified results. An optimized road network is generated based on multi-scale pruning ratio parameters, outputting a simplified structure that meets application requirements. The step of extracting global topological features, local neighborhood features, and attention-weighted features of the road network from the graph structure using parallel GCN, GraphSAGE, and GAT branches includes: the GCN branch captures global topological dependencies through spectral graph convolution operations, with the propagation formula as follows: in and They represent the first The node feature matrix and trainable weight matrix of the layer; A is the original adjacency matrix. It is then the identity matrix; express The degree matrix; the GraphSAGE branch uses mean aggregation of neighbor features, calculated using the following formula: in Represents the set of neighboring nodes. and Representing neighboring nodes respectively and Feature vectors in the current layer; It is a learnable weight matrix; It is a non-linear activation function; the GAT branch aggregates neighbor features through an attention mechanism that integrates edge attributes, and the attention coefficients are... The calculation is as follows: in It is a learnable weight; and Represents a node and eigenvectors; The eigenvectors of the edges; This is a splicing operation.

2. The method according to claim 1, characterized in that, The road classification includes: obtaining the highway attribute value of a road and dividing the road into three functional levels according to predefined rules; when the highway attribute value is motorway, trunk, or primary_link, the road level is 0, corresponding to expressways, arterial roads, and connecting roads, undertaking the main traffic flow function; when the highway attribute value is secondary, tertiary, or secondary_link, the road level is 1, corresponding to secondary arterial roads, tertiary roads, and their connecting roads, undertaking the regional traffic connection function; when the highway attribute value is residential, service, or living_link, the road level is 2, corresponding to residential area roads, service roads, and living area connecting roads, undertaking the local traffic service function; the node road level is determined through a minimum value aggregation strategy, and centroid processing and DBSCAN spatial clustering optimization are performed on face-shaped geographic entities.

3. The method according to claim 1, characterized in that, The importance assessment uses unsupervised link prediction reconstruction error, and its loss function is defined as: in Represents the set of actual edges in the graph. Size; and Representing nodes respectively and The final embedding vector learned; This indicates the size of the negative sample set.

4. The method according to claim 1, characterized in that, Multi-criteria collaborative pruning includes a triple protection mechanism: protecting topological nodes to avoid connectivity disruption; forcibly retaining high-level road nodes to maintain the integrity of the traffic backbone; Protect the centroid nodes of geographic entities to ensure regional semantic consistency; after pruning, quickly find isolated node pairs through spatial indexing and add connecting edges to restore global connectivity.

5. The method according to claim 1, characterized in that, Multi-dimensional quality assessment includes: geometric fidelity using the Hausdorff distance metric. To measure the difference in geometric shape between the simplified road network and the original road network, let the set of all points in the original road network graph be . The set of all points in the simplified road network diagram is For any two points and ,definition For point and points The Euclidean distance in a two-dimensional plane is the Hausdorff distance. Defined as Topological consistency is achieved through the entropy of node degree distribution. and attribute entropy Quantification; Node degree distribution entropy is used to measure the complexity of topological connectivity patterns in a road network, by defining the degree of a node in the network. Let be the number of edges connected to this node. Count the degree of all nodes in the network to obtain the probability distribution of node degrees. The degree of a node is exactly The probability of is calculated using the following formula: The degree values ​​of the nodes appearing in the summation traversal of the network are... Attribute entropy is used to measure the completeness of the retention of semantic attribute information in a road network. Let the attribute of a road to be evaluated be A, and all its discrete values ​​be... ,set up This indicates that the value of this attribute is exactly 1. The information entropy of attribute A is the proportion of the number of roads to the total number of roads. Represented as: 。 6. The method according to claim 1, characterized in that, Multi-scale modeling generates simplified road network sequences by setting different pruning ratio parameters, and selects the optimal simplification ratio based on the comprehensive evaluation results of geometric fidelity, topological consistency and semantic integrity.

7. A road network simplification device based on a multi-branch graph neural network, characterized in that, The method described in claim 1 comprises: a data processing module for preprocessing the original road network data, constructing a hierarchical graph structure, including extracting nodes, paths, and spatial relationships, establishing a road topology graph, and performing map projection transformation and road level classification; a feature extraction module for extracting global topological features, local neighborhood features, and attention-weighted features of the road network from the graph structure through parallel GCN, GraphSAGE, and GAT branches; and a node representation module for concatenating the global topological features learned by the GCN branch, the local neighborhood features generated by the GraphSAGE branch, and the attention-weighted features generated by the GAT branch. The system generates a unified node embedding representation; a scoring module is used to construct an unsupervised node importance evaluation system for graph autoencoders to calculate node importance scores in the node embedding representation and identify redundant nodes by combining a dynamic threshold optimization mechanism; a repair module is used to protect topological key nodes, high-level road nodes, and geographic entity centroid nodes by adopting a multi-criteria collaborative progressive pruning strategy and to implement dynamic connectivity repair; a quantification module is used to quantify the geometric fidelity, topological consistency, and semantic integrity of the simplified results through a multi-dimensional quality evaluation system; and an output module is used to generate an optimized road network based on multi-scale pruning ratio parameters and output a simplified structure that meets application requirements.

8. A computer device, characterized in that, The device includes a memory and a processor, wherein the processor, when executing computer instructions stored in the memory, performs the method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, Includes instructions that, when executed on a computer, cause the computer to perform the method as described in any one of claims 1 to 6.

Images