Key node vector data vectorization and spatial perception method based on graph neural network

Abstract

The invention discloses a key node vector data vectorization and spatial perception method based on a graph neural network, and belongs to the technical field of geographic information systems and artificial intelligence. Obtaining vector shp layer data and DEM elevation data of the current region; converting the vector data into graph structure data, determining positions of water system nodes and water conservancy facility nodes based on the vector data, and extracting corresponding elevation values from the DEM as node additional features; constructing a GNN graph structure model based on node attributes (including elevation values) and a spatial relationship, and generating a spatial embedding vector; storing the embedded vector into a vector database to create an index, and retrieving a related embedded vector from the vector database when the large language model needs spatial reasoning or answers spatial related questions; and inputting the embedded vector as context information into a large language model, assisting in reasoning and answering, and generating an accurate response. The efficient modeling and accurate analysis of the complex spatial relationship are realized, and the accuracy and efficiency of spatial data processing are improved.

Description

Technical Field

[0001] The present invention relates to the field of geographic information systems and artificial intelligence technologies, and more particularly to a method for vectorizing key node vector data and spatial perception based on graph neural networks. Background Art

[0002] In recent years, with the rapid development of big data technologies and large language models (LLMs), the trend toward intelligent spatial data has become increasingly pronounced. Geographic Information Systems (GIS), as a core technology for processing and analyzing geospatial data, play a vital role in environmental monitoring, water conservancy construction, hydrological forecasting, and water resources management. However, traditional spatial data processing methods, which rely heavily on two- or three-dimensional spatial analysis, struggle to efficiently and intelligently process complex spatial relationships and high-dimensional data, limiting their effectiveness in large-scale and complex application scenarios.

[0003] Digital Elevation Models (DEMs), as an important data source reflecting surface elevation changes, are widely used in fields such as flood inundation analysis, water system perception, and terrain analysis. However, existing methods often face challenges with large data volumes and low processing efficiency when fusing DEM and vector data for spatial analysis. The combination of high-resolution DEM and vector data requires significant computing resources and storage space, which not only increases data processing complexity but also limits real-time and scalability in practical applications.

[0004] Graph Neural Networks (GNNs), as deep learning models that can effectively process graph-structured data, have shown great potential in capturing spatial relationships and elevation information. GNNs transfer and aggregate information through the connections between nodes, capable of learning representations of nodes, edges, or entire graphs, making them suitable for modeling and analyzing complex spatial relationships. However, when directly combining high-resolution DEM data with vector data for GNNs, the sheer volume of data often exceeds the processing capabilities of conventional computing resources, limiting their widespread adoption and practical use.

[0005] In addition, vector databases have significant advantages in storing and retrieving high-dimensional vectors, enabling efficient similarity search and query operations. Through optimized index structures and efficient search algorithms, vector databases can quickly find similar vectors in large-scale high-dimensional data, greatly enhancing the speed and accuracy of data retrieval. However, existing technologies lack a method to effectively combine the spatial embedding vectors generated by GNNs with vector databases and large language models (LLMs) to further improve the analysis and understanding capabilities of spatial data.

[0006] Although existing technologies have made some progress in the respective application fields of GNNs, vector databases, and LLMs, there is a lack of a comprehensive spatial data processing method that can simplify elevation data processing, improve computational efficiency, and seamlessly integrate with vector databases and large language models while ensuring the accuracy of spatial analysis. Therefore, developing a spatial data vectorization and spatial perception method that integrates GNNs, vector databases, and LLMs can effectively reduce the complexity of data processing and computational resource requirements, while also enhancing the accuracy and intelligence level of spatial relationship understanding, which has important theoretical value and broad application prospects. Summary of the Invention

[0007] In view of this, the present invention provides a method for vectorizing and spatially perceiving key node vector data based on graph neural networks, which realizes efficient modeling and accurate analysis of complex spatial relationships through advanced deep learning technologies, thereby improving the accuracy and efficiency of spatial data processing.

[0008] To achieve the above object, the present invention adopts the following technical solutions:

[0009] A method for vectorizing and spatially perceiving key node vector data based on graph neural networks, comprising the following steps:

[0010] Obtain the vector shp layer data and DEM elevation data of the current region;

[0011] Convert the vector shp layer data into graph structure data, extract the geographical entities therein as the nodes of the graph, and the spatial adjacency relationships as the edges of the graph;

[0012] Extract the elevation values corresponding to the positions of each node from the DEM data as additional features of the nodes, and jointly constitute the node input features of the graph neural network with the spatial positions and functional attributes;

[0013] Construct a graph neural network model, adopt a multi-layer graph structure propagation mechanism, and under the unsupervised training method, use the node feature reconstruction loss as the optimization target to gradually aggregate the neighbor node information and its elevation features layer by layer to generate node embedding vectors that can express spatial relationships and elevation structures; and

[0014] Store the embedding vectors in a vector database and build an index structure for them to support subsequent fast retrieval and spatial semantic association;

[0015] When receiving a user query, retrieve the embedding vectors related to the spatial semantics from the vector database based on the query intention, and jointly construct the semantic information of the vectors and the user question into a context input;

[0016] Utilize a large language model based on Transformer to encode and infer the above context input, and through the semantic understanding and generation ability of the language model, output an intelligent answer highly consistent with the spatial semantics, realizing automatic question answering and spatial reasoning for complex geospatial problems.

[0017] Optionally, convert the vector shp layer data into graph structure data, where geographical entities are used as the nodes of the graph and spatial relationships are used as the edges of the graph.

[0018] Optionally, the DEM elevation data includes water system nodes, lake edge nodes, and reservoir edge nodes.

[0019] Optionally, use an unsupervised learning method for the GNN graph structure model, with the node feature reconstruction loss as the optimization objective, to ensure that the embedding vectors can accurately reflect the spatial relationships and elevation features.

[0020] Optionally, through a multi-layer graph neural network, integrate the neighborhood information layer by layer to generate embedding vectors representing spatial features.

[0021] Optionally, it also includes embedding the DEM elevation data as an important continuous feature into the graph neural network model to improve the accuracy of the model in hydrological analysis and spatial perception tasks, specifically including:

[0022] Extract the elevation value corresponding to each node's position and jointly input it into the graph neural network model with the spatial adjacency structure;

[0023] Utilize the elevation feature to enhance the performance of spatial decision-making tasks such as flood inundation range prediction, river flow direction identification, and catchment area delineation;

[0024] By encoding the elevation difference into the attention mechanism of the adjacent edges, guide the model to focus on the water flow path and the trend of terrain changes.

[0025] Optionally, the fusion of elevation information and spatial relationship information is realized by using a multi-layer neural network structure, specifically:

[0026] In each layer of the graph neural network, splice the elevation feature of the node and the neighborhood structure feature to form a combined vector;

[0027] Feature weighted aggregation is performed through the attention mechanism or the graph convolution operator, and the elevation difference plays a regulatory role in the propagation weight;

[0028] The finally output node embedding vectors have both spatial topology and elevation continuity, improving the model's ability to express the spatial information of the watershed structure and water system connectivity.

[0029] As can be seen from the above technical solutions, compared with the prior art, the present invention provides a method for vectorizing key node vector data and spatial perception based on a graph neural network, having the following beneficial effects:

[0030] 1. The method of the present invention is specifically designed for application scenarios related to specific environments such as water conservancy construction, hydrological forecasting, and water resource management, and customizes the process of spatial perception and reasoning. Through optimized key node selection (including water system nodes, lake boundary points, and reservoir boundary points) and efficient data processing, the present invention can achieve efficient processing and intelligent understanding of complex spatial data in these fields, support more accurate decision-making and resource optimization, and improve the practicality and pertinence of the system.

[0031] 2. The present invention constructs an end-to-end spatial intelligent reasoning method by integrating a graph neural network (GNN), a vector database, and a large language model (LLM). It can efficiently process large-scale, heterogeneous, and high-dimensional spatial data, and realize automatic modeling of complex spatial relationships, semantic retrieval, and intelligent question-answering reasoning. Compared with traditional methods, the present invention not only significantly improves the efficiency and accuracy of spatial data processing, but also enhances the system's intelligent understanding and reasoning ability, has good scalability and real-time performance, and can be widely applied to fields such as smart water conservancy, smart city, environmental monitoring, and spatial cognition, having important application value and broad market prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only the embodiments of the present invention, and those of ordinary skill in the art can also obtain other drawings according to the provided drawings without creative efforts.

[0033] Figure 1 It is a schematic diagram of the overall process of the present invention;

[0034] Figure 2 It is a schematic diagram of converting the vector shp layer data of the present invention into a graph structure;

[0035] Figure 3 It is a schematic diagram of the graph neural network model structure of the present invention;

[0036] Figure 4 Schematic diagram of the embedding vector storage and retrieval of the present invention;

[0037] Figure 5 Flowchart of the integration of the LLM and the vector database of the present invention. Specific embodiments

[0038] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0039] Embodiment 1

[0040] See Figures 1-5 , the embodiment of the present invention discloses a method for vectorizing key node vector data and spatial perception based on a graph neural network, including the following steps:

[0041] Step 1: Obtain the vector shp layer data and DEM elevation data of the current area;

[0042] Step 2: Convert the vector shp layer data into graph structure data, determine the positions of water system nodes and water conservancy facility nodes based on the vector shp layer data, and extract the elevation values at the corresponding positions from the DEM data as node additional features;

[0043] Step 3: Construct a GNN graph structure model based on node attributes (including elevation values) and spatial relationships to generate embedding vectors representing spatial features;

[0044] Step 4: Store the embedding vectors in the vector database and create an index. When the large language model needs to perform spatial reasoning or answer spatial-related questions, retrieve the relevant embedding vectors from the vector database;

[0045] Step 5: Input the retrieved spatial embedding vectors into the large language model as context information to assist it in performing spatial reasoning and question answering, so as to generate an accurate response.

[0046] Figure 1 Schematic diagram of the overall process of the present invention, showing the overall implementation steps of the method for vectorizing key node vector data and spatial perception based on a graph neural network of the present invention, including links such as data input, data processing and modeling, spatial feature understanding, model prediction and decision-making, application deployment and feedback optimization.

[0047] Specifically include:

[0048] In the data input stage, the system receives water conservancy-related data, hydrological forecast data, and water resources management data respectively, as the basis for subsequent modeling and reasoning;

[0049] In the data processing and modeling stage, the input data is subjected to spatial structure conversion and elevation feature extraction, a graph structure data is constructed, and a graph neural network model is trained to generate spatial embedding vectors;

[0050] In the spatial feature understanding stage, the spatial relationship and elevation structure between nodes are deeply learned through the graph neural network model to form an embedding representation expressing complex geographical features;

[0051] In the model prediction and decision-making stage, spatial semantic retrieval and reasoning are carried out based on the embedding vectors to support intelligent decision-making in application scenarios such as water conservancy construction, hydrological forecasting, and water resources management;

[0052] In the application implementation stage, it is applied to the fields of water conservancy construction, hydrological forecasting, and water resources management respectively to generate intelligent application decision-making suggestions;

[0053] In the feedback and optimization stage, the model training and reasoning strategies are continuously optimized according to the application feedback to achieve system self-iteration and performance improvement.

[0054] Figure 2 This is a schematic diagram of converting the vector shp layer data of the present invention into graph structure data, showing the specific process of extracting nodes, edges, and node features based on spatial data. Specifically, it includes:

[0055] In the data source input stage, the system receives two types of original data: digital elevation model (DEM) elevation data and vector shp layer data, where the vector data includes pump station locations (Point), reservoir planar areas (Polygon), and river line segments (Line);

[0056] In the node extraction stage, spatial entities are extracted from the vector layer data as nodes, including key geographical objects such as pump stations, reservoirs, and rivers;

[0057] In the node feature extraction stage, according to the spatial position of the nodes, the corresponding elevation values are extracted from the DEM data as additional node features to form node attributes (such as node ID, type, elevation);

[0058] In the edge extraction stage, the connection edges between nodes are extracted according to the spatial connection relationship and the water flow direction relationship to construct the topological structure of the graph;

[0059] In the graph structure generation stage, a complete graph data structure including nodes (such as reservoir node ST001, river section node ST002, pump station node ST003) and their attributes (type, elevation) and edges (spatial connection relationship, water flow direction relationship) is formed.

[0060] It shows that in the data preprocessing stage of the present invention, by efficiently extracting spatial entities and constructing a spatial topology graph structure, it lays a foundation for subsequent graph neural network modeling and spatial reasoning.

[0061] Figure 3 It is a schematic diagram of the graph neural network (GNN) model structure of the present invention, showing the modeling process of performing graph convolution processing based on node features and generating spatial embedding vectors. Specifically, it includes:

[0062] Input layer: The system receives node features as input, and the node features include information describing spatial terrain and functional characteristics such as geometric attributes (such as node type, spatial position), elevation value, catchment area, river length, etc.;

[0063] Hidden layer: Through the structure of a multi-layer graph convolutional network (GCN) or graph attention network (GAT), it aggregates neighbor node information and elevation features layer by layer to extract local and global spatial relationship features of the nodes;

[0064] Output layer: Generates node-level embedding vectors, and the embedding vectors encode the spatial position features, elevation change features of the nodes and their spatial relevance with surrounding nodes, which are used for subsequent vector retrieval and intelligent reasoning.

[0065] It shows how, in the modeling stage of the present invention, through the graph neural network structure, it realizes the joint modeling and characterization of complex spatial relationships and elevation structures, providing a basis for subsequent efficient spatial understanding and application decision-making.

[0066] Figure 4 It is a flowchart of the storage and retrieval of the embedding vectors of the present invention, showing the process of storing the spatial embedding vectors generated by the graph neural network (GNN) in a vector database and supporting fast similarity retrieval. Specifically, it includes:

[0067] In the stage of generating embedding vectors, through the training of the graph neural network model, node embedding vectors that can represent the spatial features and elevation characteristics of the nodes are obtained;

[0068] In the stage of uploading to the vector database, the generated embedding vectors are batch imported into the vector database for centralized management;

[0069] In the indexing and storage stage, based on efficient indexing algorithms (such as inverted index, hash index, tree structure index, etc.), indexes are established for the vector data to accelerate the subsequent retrieval process;

[0070] In the similarity search stage, according to the input query, the similarity between the embedding vectors (such as Euclidean distance, cosine similarity) is calculated to quickly locate the most relevant spatial feature vectors;

[0071] In the fast retrieval of relevant embedding vectors stage, a set of spatial embedding vectors most relevant to the user's query intention is retrieved from the vector database, providing context support for subsequent large language model reasoning.

[0072] Illustrates the efficient mechanism of the present invention in spatial vector data management and retrieval, ensuring fast response to large-scale spatial data and efficient reasoning support capabilities.

[0073] Figure 5 This is the integration flowchart of the large language model (LLM) and vector database of the present invention, showing how the LLM improves the process of spatial understanding and answer generation by retrieving spatial embedding vectors. Specifically, it includes:

[0074] In the stage of initiating a spatial-related query, when the large language model (LLM) receives a spatial-related question from the user (such as position relationship, flow direction judgment, etc.), it automatically initiates a query request;

[0075] In the stage of the vector database retrieving embedding vectors, according to the query intention, the spatial embedding vectors most relevant to the query are retrieved from the vector database;

[0076] In the context construction stage, the retrieved embedding vectors are injected as external knowledge to construct enhanced context information;

[0077] In the stage of inputting to the LLM to generate an answer, the context containing spatial features is input to the LLM, and the LLM performs reasoning and natural language answer generation based on the retrieved spatial semantic information;

[0078] Illustrates the innovative mechanism of the present invention in a spatial intelligent question-answering system. Through the seamless integration of the vector database and the large language model, it realizes the intelligent understanding and reasoning of complex spatial relationships, effectively improving the accuracy and professionalism of answers.

[0079] Furthermore, in step one, vectorization of vector data: Convert vector shp layer data into graph structure data. Geographic entities (such as reservoirs, rivers, pumping stations, lake boundary points, reservoir boundary points, etc.) are used as nodes of the graph, and spatial relationships (such as adjacency, inclusion, connection) are used as edges of the graph. This step ensures that the topological relationships between geographic entities are accurately modeled, laying a foundation for subsequent graph neural network processing.

[0080] Extraction of key node elevation data: Only extract the elevation values of important spatial nodes such as key water system nodes, lake boundary points, and reservoir boundary points from Digital Elevation Model (DEM) data as additional features of the nodes. Through this selective extraction method, the data volume is significantly reduced, the processing efficiency is improved, and at the same time, the elevation information crucial for spatial analysis is retained for inundation analysis and water system perception.

[0081] Further, in step 2, graph structure construction: Based on the vector data and the extracted elevation data of key nodes, construct a graph structure suitable for GNN, including nodes, edges, and their features. This graph structure can effectively reflect the spatial relationship and elevation change between geographical entities, enhancing the model's perception ability of terrain and spatial position.

[0082] GNN model design and training: Design and train a GNN model (such as a graph convolutional network GCN or a graph attention network GAT) to capture the spatial relationship and elevation information between key nodes. Through multiple layers of graph neural networks, the model can integrate neighborhood information layer by layer to generate embedding vectors representing spatial features. During the training process, an unsupervised learning method is adopted, with the node feature reconstruction loss as the optimization objective, ensuring that the embedding vectors can accurately reflect the spatial relationship and elevation features.

[0083] Further, in step 4, embedding vector storage: Store the embedding vectors generated by GNN into a vector database to support efficient similarity search and query operations. The vector database ensures the efficient management and retrieval of large-scale spatial embedding vectors through an optimized index structure and an efficient search algorithm, improving the response speed and scalability of the system in practical applications.

[0084] Embedding vector retrieval: When the large language model needs to perform spatial reasoning or answer spatial-related questions, retrieve relevant embedding vectors from the vector database. Through efficient similarity search, quickly locate the spatial feature vectors related to the user query, providing a basis for subsequent context fusion.

[0085] Further, in step 5, context fusion and answer generation: Input the retrieved embedding vectors as context information into the large language model to enhance its spatial understanding ability and generate accurate natural language answers. Through this integrated method, the LLM can combine the high-dimensional features of spatial embedding vectors to achieve intelligent understanding and reasoning of complex spatial relationships, significantly improving its application performance in environment-related fields.

[0086] This embodiment also includes further applications of elevation data, as follows:

[0087] Flood inundation analysis and water system perception: Use the DEM elevation data of key nodes as additional features to enhance the performance of the GNN model in flood inundation analysis and water system perception. Through the fusion of elevation information, the model can more accurately predict the flood inundation range and depth, identify river flow directions, catchment areas, and drainage systems, supporting precise water resource management and environmental monitoring.

[0088] Multi-level Fusion: Through a multi-layer neural network structure, the elevation data and spatial relationship data are fused at multiple levels to further enhance the expression ability of the embedded vectors. The multi-level feature fusion can capture more complex and detailed spatial patterns, enhancing the model's comprehensive understanding of terrain and spatial distribution.

[0089] Apply the above method to environment-related fields such as water conservancy construction, hydrological forecasting, and water resource management to achieve efficient processing and intelligent understanding of key spatial data. Through efficient data processing and intelligent spatial perception, the present invention supports precise decision-making and resource optimization, improving the scientificity and effectiveness of environmental management.

[0090] Decision Support and Resource Optimization: Based on the generated spatial embedded vectors and intelligent answers of the LLM, the system can provide a scientific basis for water conservancy project design, water resource allocation, hydrological prediction, etc., optimize resource allocation, reduce environmental risks, and improve the overall management level.

[0091] Example 2

[0092] This example illustrates the understanding of spatial relationships in the application of the present invention to water conservancy construction:

[0093] S1: Data Preparation: Obtain the vector shp layer data of a certain area, including water conservancy facilities such as reservoirs, rivers, and pumping stations. Obtain the DEM elevation data of the corresponding area.

[0094] S2: Data Preprocessing: Use GIS tools to convert the vector shp layer data into graph structure data, define reservoirs, rivers, pumping stations, etc. as nodes, and define the adjacency relationship as edges.

[0095] Extract only the elevation values of key water system nodes and important water conservancy facility nodes from the DEM data as part of the node features for flood inundation analysis and water system perception.

[0096] S3: Graph Neural Network Modeling and Training: Construct a graph structure containing nodes and edges, and the node features include geometric attributes, elevation values, and functional attributes (such as reservoir capacity).

[0097] Design a GNN model containing two layers of graph convolutional networks.

[0098] Use unsupervised learning methods to train the GNN model to generate the embedded vectors of the nodes.

[0099] S4: Vector Database Storage:

[0100] Upload the trained embedded vectors to the vector database and create an index to support fast retrieval.

[0101] S5: Integration with Open-source Large Language Model:

[0102] Develop an interface. When the LLM receives a query related to space, such as "How far is a certain reservoir from the nearest pumping station?", the system retrieves relevant embedding vectors from the vector database.

[0103] Input the retrieved embedding vectors into the LLM as context information to generate accurate answers.

[0104] S6: System testing and verification:

[0105] Test the system with a series of space-related questions, such as location queries, distance calculations, boundary identifications, etc.

[0106] Evaluate the answer accuracy and spatial understanding ability of the LLM after integrating the embedding vectors.

[0107] Example 3

[0108] This example illustrates the spatial data analysis of the present invention applied to hydrological forecasting:

[0109] S1: Data preparation:

[0110] Obtain the vector shp layer data of a certain basin, including river channels, gauging stations, sluices, etc.

[0111] Obtain the DEM elevation data and historical hydrological data (such as rainfall, flow) of the corresponding area.

[0112] S2: Data preprocessing:

[0113] Convert the vector shp layer data into graph structure data, define river channels, gauging stations, sluices, etc. as nodes, and define the connection relationship as edges.

[0114] Extract only the elevation values of key water system nodes and important water conservancy facility nodes from the DEM data, and combine with historical hydrological data as part of the node features for inundation analysis and water system perception.

[0115] S3: Graph neural network modeling and training:

[0116] Construct a graph structure, and the node features include geometric attributes, elevation values, and hydrological features.

[0117] Design and train a GNN model to generate embedding vectors to capture hydrological spatial relationships.

[0118] S4: Vector database storage:

[0119] Store the embedding vectors in the vector database to support fast retrieval.

[0120] S5: Integration with open-source large language models:

[0121] When the LLM receives a query related to hydrological forecasting, such as "What is the current water level of a certain river section?", the system retrieves relevant embedding vectors from the vector database.

[0122] Input the embedding vectors as context information into the LLM to generate accurate hydrological forecasts.

[0123] S6: System testing and verification:

[0124] Test the system with hydrological-related questions and evaluate its performance in hydrological forecasting.

[0125] Compare the model predictions with the actual data to verify the accuracy and reliability of the system.

[0126] Example 4

[0127] This example illustrates the spatial optimization of the present invention applied to water resource management:

[0128] S1: Data preparation:

[0129] Obtain the vector shp layer data of a certain area, including water sources, water use points, pipe network facilities, etc.

[0130] Obtain the DEM elevation data and water use demand data of the corresponding area.

[0131] S2: Data preprocessing:

[0132] Convert the vector shp layer data into graph structure data, define water sources, water use points, pipe network facilities, etc. as nodes, and define the water supply relationship as edges.

[0133] Extract only the elevation values of key water system nodes and important water conservancy facility nodes from the DEM data, and combine them with the water use demand data as part of the node features for flood inundation analysis and water system perception.

[0134] S3: Graph neural network modeling and training:

[0135] Construct a graph structure, and the node features include geometric attributes, elevation values, and water use demands.

[0136] Design and train a GNN model to generate embedding vectors and optimize the water supply network.

[0137] S4: Vector database storage:

[0138] Store the embedding vectors in the vector database to support fast retrieval.

[0139] S5: Integration with open-source large language models:

[0140] When the LLM receives a query related to water resource management, such as "How to optimize the water supply network to reduce water loss?", the system retrieves relevant embedded vectors from the vector database.

[0141] Input the embedded vectors as context information into the LLM to generate optimization suggestions.

[0142] S6: System testing and verification:

[0143] Test the system with water resource management-related questions to evaluate its performance in water supply optimization.

[0144] Field verify the feasibility and effectiveness of the optimization suggestions.

[0145] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. For the same or similar parts among the embodiments, reference can be made to each other. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple. For the relevant parts, reference can be made to the description in the method section.

[0146] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be obvious to those skilled in the art. The general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to these embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for vectorizing key node vector data and spatial perception based on graph neural network, characterized in that, It includes the following steps: Obtain the vector shp layer data and DEM elevation data of the current area; Convert the vector shp layer data into graph structure data, extract the geographical entities as the nodes of the graph, and the spatial adjacency relationship as the edges of the graph; Based on the vector shp layer data, determine the positions of water system nodes and water conservancy facility nodes, and extract the elevation values at the corresponding positions from the DEM data as additional features of the nodes, which together with the spatial positions and functional attributes constitute the node input features of the graph neural network; Construct a graph neural network model, adopt a multi-layer graph structure propagation mechanism, and under the unsupervised training method, use the node feature reconstruction loss as the optimization objective to aggregate the neighbor node information and its elevation features layer by layer to generate node embedding vectors that can express spatial relationships and elevation structures; Store the embedding vectors in a vector database and establish an index structure for it to support subsequent rapid retrieval and spatial semantic association; When receiving a user query, based on the query intention, retrieve the embedding vectors related to the spatial semantics from the vector database, and jointly construct the context input with the semantic information of the vectors and the user's question; Utilize a large language model based on Transformer to encode and reason the above context input, and through the semantic understanding and generation ability of the language model, output an intelligent answer highly consistent with the spatial semantics to achieve automatic question answering and spatial reasoning for complex geospatial problems.

2. The vectorization and spatial perception method for key node vector data based on graph neural network according to claim 1, characterized in that, Convert the vector shp layer data into graph structure data, where the geographical entities are the nodes of the graph and the spatial relationship is the edge of the graph.

3. A method for vectorizing key node vector data and spatial perception based on a graph neural network according to claim 1, characterized in that, The DEM elevation data includes water system nodes, lake edge nodes, and reservoir edge nodes.

4. A method for vectorizing key node vector data and spatial perception based on a graph neural network according to claim 1, characterized in that Adopt an unsupervised learning method for the GNN graph structure model, with the node feature reconstruction loss as the optimization objective, to ensure that the embedding vectors can accurately reflect spatial relationships and elevation features.

5. A key node vector data vectorization and spatial perception method based on a graph neural network according to claim 1, characterized in that Through a multi-layer graph neural network, integrate the neighborhood information layer by layer to generate embedding vectors representing spatial features.

6. A method for vectorizing key node vector data and spatial perception based on a graph neural network according to claim 1, characterized in that It also includes embedding the DEM elevation data as an important continuous feature into the graph neural network model to improve the accuracy of the model in hydrological analysis and spatial perception tasks, specifically including: Extract the elevation value at the corresponding position of each node and input it into the graph neural network model together with the spatial adjacency structure; Utilize the elevation feature to enhance the performance of spatial decision-making tasks such as flood inundation range prediction, river flow direction identification, and catchment area delineation; By encoding the elevation difference into the attention mechanism of the adjacent edges, guide the model to focus on the water flow path and the trend of terrain change.

7. A key node vector data vectorization and spatial perception method based on a graph neural network according to claim 1, characterized in that The fusion of elevation information and spatial relationship information is realized by using a multi-layer neural network structure, specifically: In each layer of the graph neural network, splice the elevation feature of the node with the neighborhood structure feature to form a combined vector; Conduct feature weighted aggregation through the attention mechanism or graph convolution operator, and the elevation difference plays a regulatory role in the propagation weight; The finally output node embedding vectors have both spatial topology and elevation continuity, improving the model's ability to express spatial information such as watershed structure and water system connectivity.

Images