Automatic river system selection method based on heterogeneous graph convolutional network
By constructing the heterogeneous map of the river gyrus and using the RGCN model for information aggregation, the problems of low accuracy and poor topological connectivity of river gyrus selection in the existing technology are solved, and efficient automatic selection of river gyruses is achieved.
Patent Information
- Application Number
- CN202510508055.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-22
- Publication Date
- 2025-07-18
AI Technical Summary
The existing galaxy selection method has insufficient flexibility and generalization capabilities when dealing with complex and dynamically changing galaxy structures. Traditional graph convolutional neural networks have limited ability to express and aggregate heterogeneous graphs, resulting in low selection accuracy and poor topological connectivity.
The heterogeneous diagram of the river galaxy is constructed, the RGCN model is used to double aggregate the graph information, the deep-level features and selection rules of the river galaxy are excavated, and the geometric and topological features of the river section are extracted, and the node classification is performed in combination with the knowledge of the river space is realized to realize the automatic selection of the river galaxy.
The accuracy of river selection was improved, with the accuracy, recall and F1 value exceeding 92%, maintaining the topological connectivity and morphological similarity of the river network, and solving the problem of automatic selection of river systems.
Smart Images
Figure CN120339842A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of river system selection, and more specifically to an automatic river system selection method based on a heterogeneous graph convolutional network. Background Art
[0002] As the "skeleton" element on the map, the river system has always been the key object of map generalization research. Its generalization mainly includes processes such as river system selection and simplification. River system selection is a map generalization process in which, when transforming from a large-scale map to a small-scale map, due to space limitations, important rivers are preferentially selected and retained, while other rivers are discarded. As the core link of river system generalization, river system selection is the key to ensuring map accuracy and readability, and at the same time, it is also a high-dimensional complex decision-making problem.
[0003] Currently, river system selection methods mainly include three methods: selection based on indicators and knowledge derivation, selection based on machine learning, and selection based on deep learning.
[0004] The river system selection method based on indicators and knowledge derivation constructs a selection model by mining geometric features related to river system selection and using mathematical functions and knowledge rules. Although such methods improve the rationality of selection, due to their high dependence on pre-set expert knowledge, they show deficiencies such as poor flexibility and generalization ability when dealing with complex and dynamically changing river system structures.
[0005] The river system selection method based on machine learning constructs a river system selection model by combining indicators reflecting the importance degree of the river system and using machine learning methods such as K-nearest neighbor, random forest, and support vector machine. Such methods have promoted the development of automatic river system generalization. However, traditional machine learning methods perform better when dealing with regular Euclidean data (such as images), while for river system vector data with complex structural features, their capabilities are insufficient, resulting in obvious limitations of such river system selection methods.
[0006] Deep learning-based methods extract geometric and topological features of river systems as graph node attributes and use graph neural network models for node classification to achieve automatic selection of river systems. Such methods have good generalization ability and adaptability, but there are still the following deficiencies: First, their research objects are mainly homogeneous graphs (graphs containing only single types of nodes and edges), only considering the unified connection method between nodes, and failing to fully explore different types of river section connection relationships, resulting in insufficient learning ability of deep learning models for river system structures and prone to misjudgment or omission when selecting important river sections. Second, such methods use traditional graph convolutional neural networks such as GCN (Graph Convolutional Network) and GraphSAGE (Graph Sample and Aggregate) to process homogeneous graphs, with limited ability to express and aggregate graph information of complex river systems, resulting in problems such as low selection accuracy and poor topological connectivity of selection results in the model.
[0007] Compared with homogeneous graphs, heterogeneous graphs can represent different types of nodes and edges and can meet the diverse expression of relationships between nodes. Therefore, the present invention constructs a heterogeneous graph of river systems with multiple edge types through different river section connection relationships, providing more comprehensive information support for river system selection, enhancing the perception and learning ability of deep learning models for complex river system structures, and improving the accuracy of river system selection. RGCN (Relational Graph Convolutional Networks) is a heterogeneous graph convolutional network improved based on GCN. This model adds an aggregation process for multi-type node and edge relationships, can assign different weight matrices to efficiently process heterogeneous graph information of river systems, deepen the learning ability of the model for river system spatial structures, features, and river system selection expression rules, and improve selection accuracy and topological connectivity of selection results.
[0008] In summary, in view of the deficiencies of existing research and the advantages of heterogeneous graphs and RGCN, the present invention proposes a method for automatic selection of river systems based on a heterogeneous graph convolutional network. Summary of the Invention
[0009] The present invention proposes a method for automatic selection of river systems based on a heterogeneous graph convolutional network. By combining river system spatial knowledge, a heterogeneous graph of river systems is constructed; then, using the dual aggregation mechanism of RGCN for graph information, the internal connection between deep features of river systems and river system selection rules is mined to achieve automatic selection of river systems. The present invention mainly includes three parts: extracting river system spatial features, constructing a heterogeneous graph of river systems, and establishing an RGCN river system selection model.
[0010] (1) Extracting river system spatial features: According to river system structured knowledge, extract the geometric features of river sections and river system spatial relationships;
[0011] (2) Construct a heterogeneous river network graph: Based on the river network vector data, the connection relationships of river sections, and the extracted spatial features of the river network, construct a heterogeneous river network graph containing three types of edges;
[0012] (3) Establish an RGCN river network selection model: Use the constructed heterogeneous river network graph and sample labels to train the RGCN model, learn the node and edge information of the heterogeneous river network graph, and perform graph node classification to achieve river network selection.
[0013] The river network automatic selection method proposed by the present invention has a relatively high selection accuracy, and indicators such as precision, recall, F1 value, and AUC all exceed 92%; in addition, this method also better maintains the topological connectivity and morphological similarity of the river network, and better solves the problem of river network automatic selection. Brief Description of the Drawings
[0014] 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 schematic diagrams of the present invention. For those of ordinary skill in the art, other drawings can be obtained according to the provided drawings without creative efforts.
[0015] Figure 1 It is a flowchart of the river network automatic selection method based on a heterogeneous graph convolutional network provided by the present invention.
[0016] Figure 2 It is a schematic diagram of the construction of the heterogeneous river network graph provided by the present invention.
[0017] Figure 3 It is a representation diagram of the feature matrix and adjacency matrix of the heterogeneous river network graph provided by the present invention.
[0018] Figure 4 It is a river network selection model diagram based on RGCN provided by the present invention.
[0019] Figure 5 It is a river network selection result diagram provided by the present invention.
[0020] Table 1 shows the calculation results of each accuracy evaluation index of the present invention
[0021] Detailed Embodiments
[0022] 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. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention.
[0023] An automatic river system selection method supported by a heterogeneous graph convolution network is disclosed in an embodiment of the present invention. The flowchart is as Figure 1 shown, and it includes the following steps:
[0024] 1. Extraction of river system spatial features:
[0025] The present invention extracts 10 river system spatial features, which include quantitative descriptions of river reaches and river network structures, and are divided into two categories: geometric features of river reaches and spatial relationships of river reaches for feature extraction.
[0026] Step1: The geometric features of the river reach describe the geometric shape and physical properties of the river reach. The present invention extracts 5 feature indicators including river reach length, river reach curvature, confluence angle, elongation ratio, and roundness ratio for extraction.
[0027] Step2: The spatial relationships of the river reaches describe the relationships between river reaches and their structural knowledge in the river network. The present invention selects 5 feature indicators including river distance, number of tributaries of the target river reach, symmetry of the number of left and right tributaries, Strahler coding, and Horton coding for extraction.
[0028] 2. Construction of the river system heterogeneous graph:
[0029] Step3: The edges of the river system heterogeneous graph are determined by the connection relationship information of the river reaches. The present invention classifies the connection relationships between river reaches into 3 types: co-flow connection, confluence connection, and tributary connection.
[0030] Step4: The present invention divides the river system into several single river reaches; then takes the midpoint of each river reach as a node and the connection relationship between river reaches as an edge to construct a river system homogeneous graph; finally, according to the 3 types of river reach connection relationships, the edges are classified to obtain the river system heterogeneous graph, as Figure 2 shown.
[0031] Step5: The present invention represents the river system heterogeneous graph as a feature matrix and an adjacency matrix, providing a data basis for subsequent graph operations, as Figure 3 shown. The model normalizes 10 river system feature indicators to construct the feature matrix of the river system heterogeneous graph. The adjacency matrix of the river system heterogeneous graph is constructed according to the type value of the edge.
[0032] 3. River system selection model based on RGCN:
[0033] The river system selection model based on RGCN consists of three parts: an input layer, an intermediate layer, and an output layer, as Figure 4 shown.
[0034] Step6: The model inputs the heterogeneous graph of the river system (represented as a feature matrix and an adjacency matrix) and the sample labels at the input layer.
[0035] Step7: The intermediate layer of the model is composed of several convolutional layers. Each convolutional layer consists of an RGCN convolutional network and a ReLU activation function.
[0036] Step8: The output layer takes the result of the intermediate layer as input and outputs the node classification probability through RGCN convolution and the Softmax function. The model selects the cross-entropy function as the loss function and updates the weight parameter matrix in RGCN by the gradient descent method to optimize the model.
[0037] Step9: Each node is classified according to the category to which the maximum probability value of the output result belongs, thereby completing the binary classification of graph node selection / discard. By connecting the river reaches represented by all the selected graph nodes, a small-scale river system is generated to achieve the automatic selection of the river system.
[0038] Step10: The model uses river system data of two scales, 1:24000 and 1:250000, collected from the USGS (https: / / apps.nationalmap.gov) river system vector database for training and testing. The test data is input into the trained model for river system selection, and the precision, recall rate, F1 value, and AUC of the selection result are calculated, as shown in Table 1. Analyze the above indicators and the selection result (as Figure 5 shown) to conduct a final evaluation of the quality of river system selection.
[0039] Step11: End.
[0040] The present invention proposes an automatic river system selection method based on a heterogeneous graph convolutional network. By combining river system spatial knowledge, a heterogeneous graph of the river system is constructed; then, using the dual aggregation mechanism of RGCN for graph information, the internal connection between the deep features of the river system and the river system selection rules is mined to achieve the automatic selection of the river system.
[0041] 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, and the general principles defined by this solution 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 by this solution, but will conform to the widest scope consistent with the principles and novel features disclosed by the present invention.
Claims
1. An automatic selection method for river systems based on heterogeneous graph convolutional network, which includes three parts: extraction of river system spatial features, construction of river system heterogeneous graph, and river system selection model based on RGCN. Extraction of river system spatial features: Step1: The geometric features of river reaches describe the geometric shape and physical properties of river reaches. Five feature indicators, namely river reach length, river reach curvature, confluence angle, elongation ratio, and roundness ratio, are extracted in this method. Step2: The spatial relationships of river reaches describe the relationships between river reaches and their structural knowledge in the river network. Five feature indicators, namely river distance, number of tributaries of the target river reach, symmetry of the number of left and right tributaries, Strahler coding, and Horton coding, are selected for extraction in this method. Construction of river system heterogeneous graph: Step3: The edges of the river system heterogeneous graph are determined by the connection relationship information of river reaches. The connection relationships between river reaches are divided into three types in this method: co-flow connection, confluence connection, and tributary connection. Step4: The river system is divided into several single river reaches; then each river reach (taking its midpoint) is used as a node, and the connection relationship between river reaches is used as an edge; finally, the edges are classified according to the three types of river reach connection relationships to obtain the river system heterogeneous graph. Step5: In this paper, the river system heterogeneous graph is represented as a feature matrix and an adjacency matrix, providing a data basis for subsequent graph operations. The feature matrix contains 10 river system spatial feature indicators (5 spatial relationship and 5 geometric feature indicators), and the adjacency matrix contains connection relationship information (co-flow, confluence, and tributary connection). River system selection model based on RGCN (input layer, intermediate layer, output layer): Step6: The model inputs the river system heterogeneous graph (represented as a feature matrix and an adjacency matrix) and the sample label at the input layer. Step7: The intermediate layer of the model is composed of several convolutional layers. Each convolutional layer consists of an RGCN convolutional network and a ReLU activation function. Step8: The output layer takes the result of the intermediate layer as input and outputs the node classification probability through RGCN convolution and Softmax function. The model selects the cross-entropy function as the loss function and updates the weight parameter matrix in RGCN by the gradient descent method to optimize the model. Step9: Each node is classified according to the category to which the maximum probability value of the output result belongs, thus completing the binary classification of selection / discard of graph nodes. By connecting the river reaches represented by all selected graph nodes, a small-scale river system is generated to realize the automatic selection of the river system. Step10: End.
2. The automatic river system selection method based on the heterogeneous graph convolutional network according to claim 1, wherein, In Steps S3 to S5, each river reach (taking its midpoint) is used as a node, and the connection relationship between river reaches is used as an edge; and the edges are classified according to the connection relationship between river reaches to obtain the river system heterogeneous graph.
3. The automatic selection method of river systems based on heterogeneous graph convolutional network according to claim 1 or claim 2, characterized in that In Steps S6 to S9, the river system heterogeneous graph is input into the RGCN model for information aggregation to obtain the classification result of graph nodes; and the automatic selection of the river system is realized based on the classification result.