Cross-scale mesh river system matching method and system based on matching pattern classification

Abstract

The invention relates to the technical field of geographic space data matching, in particular to a cross-scale netlike river system matching method and system based on matching pattern classification, large-scale netlike river system data to be matched is constructed into a graph structure, graph node labels are extracted according to netlike river system characteristic indexes, and the graph node labels are extracted to be matched; the netted river system characteristic indexes comprise the river reach grade, the river reach drainage basin density, the river reach connectivity, the maximum length of the river reach, the overall curvature of the river reach and the close relation between the river reach and the road; using a pre-trained matching mode classification model and performing classification identification on the large-scale netted river system data graph structure according to the graph node labels, and establishing a mapping relation between graph structure features and matching modes, and executing a matching strategy corresponding to the matching mode according to the matching mode of each river reach of the to-be-matched large-scale net-shaped river system and based on spatial proximity and geometric similarity so as to realize matching with the matched small-scale net-shaped river system. Different matching strategies are adopted for the river reach according to the relation between the netted river system features and the matching modes, the matching result of the same-name entities in different-scale data is obtained, and the netted river system matching accuracy and efficiency are improved.

Description

Technical Field

[0001] The present invention relates to the field of geographic spatial data matching technology, and in particular to a cross-scale network river system matching method and system based on matching pattern classification. Background Art

[0002] Geospatial data matching plays a key role in multi-scale spatial data integration, fusion, scale transformation and updating, and is the core technology for maintaining the currentness and accuracy of the established database. The earth is rich in various types of water resources, so the geospatial database contains a large amount of water system data. The network river system is formed by human transformation of the natural water system, and its morphological structure is significantly different from the natural river system. This river system structure is widely distributed in the Netherlands, Jiangsu, my country, and other places, mainly distributed in towns and their surrounding areas, and is of great significance for ensuring urban and rural water supply, flood control and drainage, and maintaining the ecological health of water resources. Matching mode refers to the quantitative correspondence between entities with the same name in databases of different scales. Entities with the same name refer to spatial entities that reflect the same feature in the real world in multiple data sources. Since entities with the same name have significant differences in geometric and semantic features at different scales, there are many matching modes between them, which can be divided into 1:0, 1:1, m:1, etc.

[0003] At present, there are relatively few studies on the matching of networked river systems, while the matching of linear elements such as road networks is more extensive. In fact, networked river systems and road networks are similar in morphological structure. They are both composed of a series of interconnected line segments, forming a complex network structure. Although there are some differences in details between networked river systems and road networks, for example, there are no complex spatial structures such as overpasses or roundabouts in networked river systems, but in general, the matching methods of networked river systems and road networks have something in common, and they can learn from each other. The matching research of linear elements such as road networks can be divided into two categories. One is to measure the feature similarity of two line elements, and then compare them with the set similarity threshold to complete the determination of the matching relationship. In terms of similarity feature selection, geometric, topological and semantic features are widely used. However, in the setting of similarity threshold, a lot of prior knowledge is required to obtain a suitable threshold. If it is not set properly, it may affect the matching results. The second is to regard the matching problem as a global optimization problem, that is, to find an optimal matching solution that minimizes the differences between all elements in the two data sets, rather than measuring the similarity between elements. The advantage of this method is that it overcomes the suboptimality and greediness of the previous matching method, and can use constraints to filter fuzzy matching to improve matching efficiency. The difficulty of matching geographical entities with different matching modes is different. However, the above-mentioned studies did not actively calculate the matching mode, but classified the successfully matched ones as simple matching modes (such as 1:0) according to the matching results obtained by their algorithms, and the unmatched ones were regarded as complex matching modes (such as m:1), which affected the accuracy of the matching. For this reason, there is an urgent need for a means of pre-determining the matching of entities with the same name to efficiently and accurately complete the matching process of the network river system. Summary of the invention

[0004] To this end, the present invention provides a cross-scale network river system matching method and system based on matching pattern classification. In view of the problems of great difficulty and unsatisfactory accuracy in matching geographic entities with different matching patterns in the existing geographic space matching process, different matching strategies are adopted for river sections according to the relationship between the network river system characteristics and the matching patterns, so as to obtain matching results of entities with the same name in data of different scales, thereby improving the accuracy and efficiency of network river system matching.

[0005] According to the design scheme provided by the present invention, on the one hand, a cross-scale network river system matching method based on matching pattern classification is provided, comprising:

[0006] The large-scale networked river system data to be matched is constructed into a graph structure, and graph node labels are extracted according to the networked river system characteristic indicators, wherein the networked river system characteristic indicators include river section grade, river section watershed density, river section connectivity, maximum length of river section, overall curvature of river section, and proximity between river section and road;

[0007] The pre-trained matching pattern classification model is used to classify and identify the graph structure of the large-scale networked river system data according to the graph node labels, and a mapping relationship between the graph structure features and the matching patterns is established. According to the matching patterns of each river section of the large-scale networked river system to be matched and based on spatial proximity and geometric similarity, the matching strategy corresponding to the matching pattern is executed to achieve matching with the small-scale networked river system to be matched.

[0008] As a cross-scale network river system matching method based on matching pattern classification of the present invention, further, the large-scale network river system data to be matched is constructed into a graph structure, including:

[0009] Preprocessing the large-scale networked river system data to remove abnormal data in the networked river system data, wherein the preprocessing includes deleting disordered lines, broken river section connections, and topological checking, wherein the abnormal data includes false nodes and / or self-intersecting edges;

[0010] The networked river system is divided into river sections according to its confluence nodes and level changes. Each river section is regarded as a graph node, and the connections between river sections are regarded as edges, so as to construct a dual graph of the networked river system.

[0011] As a cross-scale network river system matching method based on matching pattern classification of the present invention, further, extracting graph node labels according to network river system characteristic indicators includes:

[0012] Divide the networked river system into hierarchical structures according to river width, and obtain the river section grade of each river section in the networked river system, wherein the river section grade is the level of the river section in the hierarchical structure of the networked river system;

[0013] A river section grid is established based on the connecting lines of the first and last points of the river section, and the river section watershed density is obtained based on the total length of the river section in the grid and the grid area;

[0014] The connectivity of the river section is obtained according to the connection between the river section nodes in the graph structure, and the smooth chain stroke of the river section is constructed according to the semantic constraints and angle constraints in the graph structure. The maximum length of the river section is obtained according to the length of the smooth chain stroke of the river section.

[0015] The overall curvature of the river section is obtained based on the length of the connecting line between the first and last points of the smooth chain stroke of the river section and the area of ​​the shape enclosed, and the average length and average area of ​​the smooth chain stroke in the networked river system; and the proximity relationship between the river section and the road is obtained based on the minimum distance between the midpoint of the river section and the road.

[0016] As a cross-scale network river system matching method based on matching pattern classification of the present invention, further, before using a pre-trained matching pattern classification model and classifying and identifying the large-scale network river system data graph structure according to the graph node labels, it also includes:

[0017] Acquire large-scale remote sensing image data and small-scale terrain data, pre-process the data and segment the river sections, and construct the large-scale remote sensing image data after pre-processing and segmentation into a graph structure, use manual and / or machine methods to identify the matching pattern of each river section in the large-scale remote sensing image data, use the network river system characteristic indicators of the large-scale remote sensing image data as graph node labels, and use the graph structure and graph node labels to construct a river section sample data set for graph neural network training;

[0018] A graph neural network model is constructed based on GraphSAGE, and the graph neural network model is trained according to the model loss function and the river section sample dataset. The trained graph neural network model is used as a pre-trained matching pattern classification model.

[0019] As a cross-scale network river system matching method based on matching pattern classification of the present invention, further, the graph neural network model includes: an input layer, a hidden layer and an output layer, wherein the input layer is used to receive graph structure and graph node label data, the hidden layer includes three GraphSAGE convolutional layers and one fully connected layer, and the output layer takes the hidden layer input as input and uses a softmax function to generate a node classification probability distribution.

[0020] As a cross-scale network river system matching method based on matching pattern classification of the present invention, further, a river section sample data set is used to train the graph neural network model, including:

[0021] The graph structure and graph node labels are used as model inputs. Each node in the graph structure is taken as a core node. The graph structure is divided into k layers from the inside to the outside, and sampling nodes are selected from neighboring nodes according to the number of samples. The features of the sampling nodes in the kth layer are aggregated to the nodes of the corresponding k-1 layer, and the features of the sampling nodes in the k-1 layer are aggregated to the core nodes. The sampling node features of the core nodes in the k-1 layer are aggregated into the overall features of the core nodes. The classification function is used to map the aggregated feature information to the matching pattern category probability, output the network river system matching pattern classification type, compare the model output with the true matching pattern, and calculate the loss function. According to the calculated loss function, the chain rule is used to propagate the gradient from the model output layer to the model input layer to update the model parameters according to the gradient.

[0022] As a cross-scale network river system matching method based on matching pattern classification of the present invention, further, according to the matching pattern of each river section of the large-scale network river system to be matched and based on spatial proximity and geometric similarity, a matching strategy corresponding to the matching pattern is executed, including:

[0023] When the matching mode is 1:1, obtain the candidate matching set of each target river section under the matching mode, measure the similarity between each target river section and the small-scale meshed river system section in the corresponding candidate matching set, and select the small-scale meshed river system section in the corresponding target river section matching item according to the similarity; if multiple target river sections match the same small-scale meshed river system section, the target river section with the greatest similarity is identified as the 1:1 matching mode, and the matching modes of other target river sections in the multi-hop target river section are classified into m:1; if there is a target river section that does not match the small-scale meshed river system section, the matching mode of the target river section is classified into 1:0;

[0024] When the matching mode is m:1, determine whether the target river section of this category is adjacent to other m:1 river sections. If there is no adjacent situation, the matching mode of the target river section is classified as 1:1, and the corresponding matching strategy is adopted; if there is an adjacent situation, obtain the candidate matching set of each target river section under the m:1 matching mode, merge the other x target river sections connected to the target river section, and perform river section matching based on the combination of connected target river sections. The matching mode of the target river section that has not been successfully matched is classified as 1:0;

[0025] When the matching mode is 1:0, the river sections misclassified by the matching mode are identified and their matching items are found. By obtaining the candidate matching set of each target river section under the 1:1 matching mode, the similarity of each target river section and the small-scale network river section in the corresponding candidate matching set is measured to obtain the target river section matching items based on the similarity.

[0026] As a cross-scale network river system matching method based on matching pattern classification of the present invention, further, a candidate matching set of each target river section under the matching pattern is obtained, including:

[0027] Construct a river section buffer, and include at least one corresponding small-scale networked river system section data in each target river section buffer of this category, and use the small-scale networked river system sections falling into each target river section buffer as the candidate matching set of each target river section under the corresponding matching mode.

[0028] As a cross-scale network river system matching method based on matching pattern classification of the present invention, further, river section matching is performed according to a combination of connected target river sections, including:

[0029] When x>1 and the x target river sections are merged into a y river section line of the same level, the y river section line is measured for similarity with the small-scale network river section in the candidate matching set, and then the line formed by merging x-1 connected river sections is calculated and measured for similarity with the small-scale network river section in the candidate matching set, and the cycle is repeated z times until xz is equal to 1. The similarities are recorded and compared, and the target river section combination with the greatest similarity is selected as the target river section that successfully matches the small-scale network river section, and the corresponding small-scale network river section is used as the matching item of the target river section.

[0030] In another aspect, the present invention further provides a cross-scale network river system matching system based on matching pattern classification, comprising: a graph construction module and a graph matching module, wherein:

[0031] A graph construction module is used to construct the large-scale network river system data to be matched into a graph structure, and extract graph node labels according to the network river system characteristic indicators, wherein the network river system characteristic indicators include river section grade, river section watershed density, river section connectivity, river section maximum length, river section overall curvature and river section-road proximity;

[0032] The graph matching module is used to classify and identify the graph structure of large-scale networked river system data by using the pre-trained matching pattern classification model and according to the graph node labels, establish a mapping relationship between the graph structure features and the matching pattern, and execute the matching strategy corresponding to the matching pattern based on the matching pattern of each river section of the large-scale networked river system to be matched and based on spatial proximity and geometric similarity, so as to achieve matching with the small-scale networked river system to be matched.

[0033] Beneficial effects of the present invention:

[0034] The present invention uses graph neural networks to identify the matching patterns of large-scale networked river systems, and adopts corresponding matching strategies for different matching patterns to obtain more accurate matching results. Further experiments and related analysis show that: the scheme of this case can effectively identify the matching patterns of large-scale river sections, especially for 1:1 and 1:0 river sections, with a high classification accuracy; the matching strategies designed for different matching patterns can make full use of the advantages of various matching methods, so that the overall matching accuracy reaches 95%; by adopting different matching orders for river sections with different matching patterns, the uncertainty factors caused by artificially set thresholds can be effectively avoided, and it has a good application prospect in geospatial data matching. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a schematic diagram of the cross-scale network river system matching process based on matching pattern classification in the embodiment;

[0036] Figure 2It is a schematic diagram of the construction of the graph structure of the network river system in the embodiment;

[0037] Figure 3 It is a schematic diagram of the relationship between the cartographic synthesis and matching mode of the network river system in the embodiment;

[0038] Figure 4 It is a schematic diagram to illustrate the characteristic items of the river section in the embodiment;

[0039] Figure 5 This is a schematic diagram of the sample set test area in the embodiment;

[0040] Figure 6 It is a schematic diagram of the neighbor sampling and aggregation process in the embodiment;

[0041] Figure 7 It is a schematic diagram of a matching pattern recognition framework based on GraphSAGE in the embodiment;

[0042] Figure 8 It is a schematic diagram of the principle of the cross-scale network river system matching algorithm in the embodiment;

[0043] Fig. 9 This is an example of a conflict in matching patterns in the embodiment;

[0044] Fig.10 Schematic diagram of the matching pattern classification results of the test set in the embodiment;

[0045] Fig.11 Schematic diagram of the matching result of the network river system in the embodiment. DETAILED DESCRIPTION

[0046] In order to make the purpose, technical solutions and advantages of the present invention clearer and more understandable, the present invention is further described in detail below in conjunction with the accompanying drawings and technical solutions.

[0047] Cross-scale network river system matching is an important part of river system data integration, fusion and update. For network river systems with different matching modes, adopting differentiated matching methods is an effective means to deal with the morphological differences of multi-source data. However, there are few studies on the identification of network river system matching modes, and there is also a lack of research solutions that integrate the advantages of various matching methods. Figure 1 As shown, a cross-scale network river system matching method based on matching pattern classification is provided, comprising:

[0048] S101. Construct the large-scale networked river system data to be matched into a graph structure, and extract graph node labels based on the networked river system characteristic indicators, wherein the networked river system characteristic indicators include river section grade, river section watershed density, river section connectivity, maximum length of a river section, overall curvature of a river section, and proximity between a river section and a road.

[0049] Specifically, the large-scale network river system data to be matched is constructed into a graph structure, which can be designed to include:

[0050] Preprocessing the large-scale networked river system data to remove abnormal data in the networked river system data, wherein the preprocessing includes deleting disordered lines, broken river section connections, and topological checking, wherein the abnormal data includes false nodes and / or self-intersecting edges;

[0051] The networked river system is divided into river sections according to its confluence nodes and level changes. Each river section is regarded as a graph node, and the connections between river sections are regarded as edges, so as to construct a dual graph of the networked river system.

[0052] After completing the preprocessing of the large-scale network river system, the graph structure is constructed. The preprocessing mainly includes operations such as deleting disordered lines, connecting links, and topological checking. The network river system is divided into river sections at the confluence nodes and level change points, such as Figure 2 By constructing the dual graph of the network river system, each river section is regarded as a graph node, and the connection between river sections is regarded as an edge, as shown in Figure 2 As shown in (c) in the figure. In the process of dividing the river system into river sections, some simple 1:1 matches will evolve into m:1, but only a small number of m:n situations will occur. This is because the situation where one large-scale data corresponds to n small-scale data (i.e., 1:n) is not common. For example, there are 12,043 experimental data, and only 29 1:n data exist. Further analysis of these 1:n river sections found that their original data structure may not require segmentation, such as the absence of grade changes or confluence points. For m:1 river sections, the matching strategy adopted is to merge multiple connected m:1 river sections into one line for similarity calculation. This processing method actually simplifies the complex matching of m:1 to 1:1, thereby reducing the impact of the segmentation method on the matching accuracy to a certain extent. Data from different sources have differences in data structure. The use of this segmentation method helps to achieve unification at the data structure level and reduce the complex matching patterns of m:1 or m:n caused by differences in data structure.

[0053] In the embodiment of this case, the matching patterns between networked river systems of different scales can be used as node annotations. After preprocessing the two data sets, a manual + machine method is used to identify the matching patterns of each river section in the large-scale data. Since the cartographic comprehensive interoperation between networked river systems of different scales is mostly selected, there are fewer m:1 and m:n matching patterns in large-scale networked river systems. Considering that changes in the geographical environment are the main reason for the generation of the above-mentioned matching patterns, these matching patterns cannot be identified based on geographic entity features alone. In view of this, in the embodiment of this case, river sections with matching patterns of m:1 and m:n can be processed separately. On the premise of ignoring changes in the geographical environment and only considering cartographic comprehensive operations, it is judged whether these river sections are more in line with the matching patterns of 1:0, 1:1 or m:1. It should be pointed out that in the embodiment of this case, when identifying the matching patterns of ditches, it is limited to the three cases of 1:0, 1:1 and m:1.

[0054] According to the characteristic indicators of the network river system, the graph node labels can be extracted and designed to include:

[0055] Divide the networked river system into hierarchical structures according to river width, and obtain the river section grade of each river section in the networked river system, wherein the river section grade is the level of the river section in the hierarchical structure of the networked river system;

[0056] A river section grid is established based on the connecting lines of the first and last points of the river section, and the river section watershed density is obtained based on the total length of the river section in the grid and the grid area;

[0057] The connectivity of the river section is obtained according to the connection between the river section nodes in the graph structure, and the smooth chain stroke of the river section is constructed according to the semantic constraints and angle constraints in the graph structure. The maximum length of the river section is obtained according to the length of the smooth chain stroke of the river section.

[0058] The overall curvature of the river section is obtained based on the length of the connecting line of the first and last points of the smooth chain stroke of the river section and the area of ​​the shape enclosed, and the average length and average area of ​​the smooth chain stroke in the network river system; and the proximity relationship with the road is obtained based on the minimum distance between the midpoint of the river section and the road.

[0059] River segment generalization is an important factor that leads to different matching patterns between large and small scale data. For networked river systems, selection is the main consideration for generalization operators, such as Figure 3As shown. Therefore, in order to accurately identify the matching pattern of each river section in a large-scale networked river system, the key feature items adopted by the networked river system in the selection operation must be fully considered. Currently, a comprehensive model for mapping a networked river system with a wide range of applications can be constructed based on features such as river section length, river section grade, watershed density, river section connectivity, and networked river system spatial structure. This model is not limited to a networked river system with a certain specific distribution feature, and is applied in actual production. In the embodiment of this case, the above-mentioned feature items are sorted and expanded, and the following feature indicators are specified:

[0060] Feature 1: Level of the river section Le. It refers to the level of the river section in the river system hierarchy. It is usually divided into three levels based on the river width. It can be converted into a 3D vector using one-hot encoding.

[0061] Feature 2: River basin density D. Refers to the density of river sections within a certain range. First, a grid is constructed based on the connecting lines of the beginning and end points of the river section. The direction of the grid is consistent with the connecting lines, such as Figure 4 As shown in (a) in formula (1).

[0062] D=Len / S (1)

[0063] Where Len is the total length of the river section in the grid, that is, the sum of L1, L2, L3, and L4 in the figure, and S is the grid area.

[0064] Feature 3: Connectivity of river segments C R . Refers to the connectivity of the river section entity, which is calculated based on the dual graph of the river section and can reflect the accessibility of the river section, see formula (2).

[0065]

[0066] Where δ j Indicates the connection status between the node corresponding to the river section and node j. If they are directly connected, it is 1, otherwise it is 0.

[0067] Feature 4: Maximum length of the river section L Max Refers to the length of the stroke corresponding to the river section. The larger the length value, the more important it is in the overall spatial structure. The river section stroke is constructed through semantic constraints and angle constraints; the semantic constraints are set to "same level" and the angle constraints are set to be greater than "135°".

[0068] Feature 5: The overall curvature of the river section Cu. It refers to the curvature of the stroke corresponding to the river section, which can reflect the overall flow velocity of the river section.

[0069] like Figure 4 The calculation process shown in (b) in is shown in formula (3).

[0070]

[0071] Where W is the length of the line connecting the first and last points of the stroke, M is the area of ​​the shape enclosed by the stroke and the line connecting the first and last points (M is not 0), and L avg is the average length of the stroke in the network river system, M avg It is the average stroke area in the network river system; when M=0, set the value to 1.

[0072] Feature 6: The proximity between river sections and roads Cr. In cartographic synthesis, river sections close to roads often need to be shifted. In the case of low importance of rivers, they may even be considered for deletion. In this case, a place in the Netherlands was selected as the test area. Since the network of rivers on both sides of the roads in the Netherlands are often similar to the roads and close to them, the proximity between river sections and roads was also considered. Through the analysis and experiment of the matching data, when the minimum distance between the midpoint of the river section and the road is less than 10m, the two are considered to be close and assigned a value of 1, and 0 in other cases.

[0073] S102. Utilize the pre-trained matching pattern classification model and classify and identify the graph structure of the large-scale networked river system data according to the graph node labels, establish a mapping relationship between the graph structure features and the matching pattern, and execute the matching strategy corresponding to the matching pattern according to the matching pattern of each river section of the large-scale networked river system to be matched based on spatial proximity and geometric similarity, so as to achieve matching with the small-scale networked river system to be matched.

[0074] The model training process before classification and recognition can be designed to include:

[0075] Acquire large-scale remote sensing image data and small-scale terrain data, pre-process the data and segment the river sections, and construct the large-scale remote sensing image data after pre-processing and segmentation into a graph structure, use manual and / or machine methods to identify the matching pattern of each river section in the large-scale remote sensing image data, use the network river system characteristic indicators of the large-scale remote sensing image data as graph node labels, and use the graph structure and graph node labels to construct a river section sample data set for graph neural network training;

[0076] A graph neural network model is constructed based on GraphSAGE, and the graph neural network model is trained according to the model loss function and the river section sample dataset. The trained graph neural network model is used as a pre-trained matching pattern classification model.

[0077] Among them, the graph neural network model can be designed to include: input layer, hidden layer and output layer, wherein the input layer is used to receive graph structure and graph node label data, the hidden layer contains three GraphSAGE convolutional layers and one fully connected layer, and the output layer takes the hidden layer input as input and uses the softmax function to generate node classification probability distribution.

[0078] The graph neural network model is trained using the river section sample dataset and can be designed to include:

[0079] The graph structure and graph node labels are used as model inputs. Each node in the graph structure is taken as a core node. The graph structure is divided into k layers from the inside to the outside, and sampling nodes are selected from neighboring nodes according to the number of samples. The features of the sampling nodes in the kth layer are aggregated to the nodes of the corresponding k-1 layer, and the features of the sampling nodes in the k-1 layer are aggregated to the core nodes. The sampling node features of the core nodes in the k-1 layer are aggregated into the overall features of the core nodes. The classification function is used to map the aggregated feature information to the matching pattern category probability, output the network river system matching pattern classification type, compare the model output with the true matching pattern, and calculate the loss function. According to the calculated loss function, the chain rule is used to propagate the gradient from the model output layer to the model input layer to update the model parameters according to the gradient.

[0080] The Leeuwarden area in the Netherlands was selected as the experimental area. Data A was extracted from remote sensing images with a resolution of 1 meter. These images were taken in 2022 and extracted by a private company. They are considered large-scale data in the study. Data B comes from the Netherlands Topographic Database (Basisregistratie Topografie, BRT), belongs to the TOP50NL category, and was produced in 2019. Its scale is 1:25,000 to 1:75,000 and is considered small-scale data.

[0081] Before the experiment, the data was processed. First, the unordered lines without connection relationships in the experimental area were deleted, the broken river sections were connected, and the center lines of the planar network river system were extracted as experimental objects. Then, the river system data was topologically checked to ensure that there were no errors such as false nodes or self-intersections in the data. Finally, the entire network river system was divided into river sections at the intersection points and grade transition points of the river systems, and 12,043 large-scale river section data and 6,404 small-scale river section data were obtained. In the experiment, the 2,075 river section data in the lower right corner of the area were used as the test set, and the remaining river sections were used as training sets and validation sets (the quantity ratio was divided into 3:1), of which the non-test set data accounted for 82.8%, the test set data accounted for 17.2%, and the experimental area was as follows: Figure 5As shown. After the corresponding matching processing, the matching pattern of data A corresponding to data B is obtained. Specifically, there are 3650 data with matching pattern 1:0, 4937 data with matching pattern 1:1, and 3456 data with matching pattern m:1 (including 4 m:n and 29 1:m conversions). At the same time, a dual graph based on river sections and a node sample library are constructed. In this graph structure, there are a total of 12043 nodes and 45084 edges, which will be used as the input layer for subsequent training and testing operations.

[0082] Since the regional distribution density of the networked river system is high, the river sections are straight and short, and the network structure is relatively obvious. Therefore, compared with traditional machine learning methods, deep learning methods based on graph structures are more applicable. Graph neural networks pay more attention to the connection relationship between nodes and can adaptively learn the representation of each node, so that they have stronger generalization performance in tasks such as classification and prediction. GraphSAGE extracts the node neighborhood information of the graph through neighbor sampling and aggregation functions. It is an inductive spatial domain learning framework that can represent nodes on large-scale graphs. The innovation of GraphSAGE lies in sampling. The sampling idea is in line with the production process of networked river system data, that is, in the process of cartographic synthesis, key neighbor information is considered instead of all. At present, the method based on GraphSAGE has achieved certain results in river morphology recognition and river selection. At the same time, considering that there are many nodes in the graph structure constructed in this article, the neighborhood characteristics of the nodes need to be taken into account when matching pattern recognition. The convolution operation of GraphSAGE can be designed to mainly include three important steps: neighbor node sampling, aggregation, and prediction. Figure 6 Take this as an example to illustrate.

[0083] First, for each node v in the graph, take it as the core and subdivide the overall graph structure into k levels from the inside out. Generally, the k value is set to 2. Figure 6 In (a), the dotted lines indicate the level of neighbor nodes. Next, according to the set sampling number S, a corresponding number of nodes are randomly selected from the neighbor nodes. Figure 6 As shown in (a), the sampling number is set to 4, the unselected nodes are presented in blank form, and the arrows indicate the sampling direction of the nodes.

[0084] Then, the aggregation function shown in formula (4) is used to aggregate the neighbor nodes of node v. When aggregating, the sampling node features of the kth layer are first aggregated to the corresponding k-1th layer nodes, and then the sampling node features of the k-1th layer are aggregated to the central node v. Among them, the node feature of the k-1th layer is, the set of neighbor nodes of the kth layer is N(v), and for any node u∈N(v), its feature is expressed as Aggr represents the aggregation function, and you can choose the mean function or the maximum function, etc. It represents the feature of node v after aggregation of neighboring node features at layer k, that is, the feature representation obtained by aggregating node u.

[0085]

[0086] However, it is also necessary to construct the overall feature of node v based on the features of its neighbor nodes at layer k-1. Specifically, according to formula (5), To represent the features of node v after aggregation of neighbor node features at layer k, Represents the features of node v at the k-1th layer. By connecting the two using the Concat function, merging them using the learnable matrix W, and then mapping them through the nonlinear function σ, we finally get the feature representation of node v that aggregates the features of neighboring nodes. Concat is an operation that concatenates two or more tensors in a certain dimension, which can improve the performance of the model. In addition, before training begins, the learnable matrix W needs to be initialized. During the training process, the learnable matrix will be continuously adjusted according to the selected loss function and optimization algorithm until it converges to a better value as the learning result.

[0087]

[0088] Finally, the aggregated information is used as the input of the next layer of network to predict the matching pattern of the network river system, such as Figure 6 As shown in (c) in .

[0089] In the GraphSAGE framework, supervised parameter learning is similar to traditional deep learning models, and the loss function can be directly selected according to the task requirements. In this embodiment, the cross entropy loss function shown in formula (6) can be used, where y is the true label, is the predicted probability value. First, the output of the model is compared with the true label and the loss function is calculated. Next, based on the calculated loss function, the gradient is propagated from the output layer to the input layer using the chain rule. In GraphSAGE, the backpropagation process is used to calculate the gradient of each parameter with respect to the loss function. Finally, the model parameters are updated using the calculated parameter gradients and an optimization algorithm such as the Adam optimizer. By updating the parameters along the negative direction of the gradient, the model can be gradually optimized to improve accuracy and performance.

[0090]

[0091] Given the advantages of GraphSAGE in processing large-scale graph structures and the consistency of its sampling ideas with the idea of ​​comprehensive cartography of networked river systems, this model has high applicability in processing networked river system data. The graph neural network model structure in the embodiment of this case, such as Figure 7 As shown in the figure, a GNN model with three GraphSAGE convolutional layers and one fully connected layer is shown. The classification model of the matching pattern mainly consists of an input layer, a hidden layer, and an output layer. The input of the model is the graph structure and the graph node labels, i.e., the feature items of the river section. The hidden layer contains three GraphSAGE convolutional layers and one fully connected layer. The input layer has 8 feature dimensions and the output layer has 3 dimensions. The output layer takes the output of the hidden layer as input and generates the probability distribution of the node classification as the output result. The number of neighbor sampling in the GraphSAGE convolutional layer is set to 5, and the aggregation function is set to mean; the Leaky Rectified Linear Unit (LeakyReLU) is used as the activation function to solve the gradient vanishing problem in the ReLU function, where negative_slope is set to 0.01. In the output layer, the node features are mapped to the category probabilities by applying the Softmax function, so as to obtain the classification results of the nodes, i.e., the matching pattern of the networked river system. After many experiments, the learning rate, regularization weight, dropout, and training times were set to 0.005, 5e-4, 0.3, and 1000, respectively, and the cross entropy loss function was selected as the loss function, and Adam was selected as the optimizer.

[0092] After completing the classification of matching patterns of large-scale river sections, according to different matching patterns, in the embodiments of this case, differentiated matching strategies and orders are also implemented based on spatial proximity and geometric similarity. Judging from the classification results, the classification accuracy of the 1:1 matching pattern is higher than that of the m:1 matching pattern. Starting from the 1:1 river section for matching, the obtained buffer threshold and similarity threshold are more accurate and can be better applied to other matching patterns. After adopting the corresponding matching strategy, a new matching pattern is obtained. When the current and subsequent matching patterns conflict, the matching patterns of the river section need to be re-divided and matched. Considering that when matching the 1:0 river section, there is no need to identify the m:1 river section (it is difficult to distinguish between the two), the 1:0 river section is matched last. Figure 8 As shown in the figure, ① Method 1 is used to match 1:1 river sections, and the river sections with changed matching patterns are classified into m:1 and 1:0 categories; ② the illogical m:1 river sections are classified into the 1:1 category for matching, and then method 2 is used to match the m:1 river sections, and the river sections that failed to match successfully are classified into the 1:0 category; ③ Method 3 is used to match the 1:0 river sections.

[0093] Among them, method 1 is used for 1:1 river section matching. First, a buffer is constructed for it. When setting the radius of the buffer zone, it is necessary to ensure that the buffer zone of each river section of this category contains at least one corresponding small-scale data. Next, the small-scale river section that falls within the buffer zone of the river section is regarded as the candidate matching set of the river section. Then, the similarity of the river section is measured with the small-scale data in the candidate set. Finally, the one with the best similarity that meets the conditions is selected as the matching item of the river section. In the similarity measurement: when the "mean direction" of the two river sections is less than 5° and their river section levels are consistent, if the means of their "shape similarity" and "distance similarity" are the largest, they are considered to have the greatest similarity.

[0094] After the above operations, the 1:1 river section matching results were successfully obtained. However, these results may conflict with the classification results of the initial matching mode, mainly in the following two aspects: Fig. 9 As shown:

[0095] 1. Multiple 1:1 river sections are matched with the same small-scale river section. Fig. 9 In (a), the large-scale river sections a and b, which were originally identified as 1:1, are matched with the same small-scale river section c after adopting the corresponding matching strategy, and the matching mode becomes 2:1. This conflict is mainly caused by the m:1 river section misclassified into this category sharing the same matching item with other river sections. To solve this problem, the similarity values ​​between the matching river sections can be compared. The large-scale river section with the greatest similarity will be identified as 1:1, and its matching item is the small-scale river section. Other large-scale river sections will be classified as m:1 and matched according to the corresponding method.

[0096] 2. A 1:1 river section does not match the small-scale river section. Fig. 9 In (b), the large-scale river section d, which was originally identified as 1:1, does not meet the similarity conditions with the candidate matching river section falling in its buffer zone, and the matching mode becomes 1:0. This conflict is mainly caused by the 1:0 river section misclassified into this category and unable to find a match. Therefore, the large-scale river section that causes this problem is classified as 1:0.

[0097] Before matching the river sections of the m:1 mode, preprocessing is required, that is, judging whether the river sections of this category are adjacent to other river sections of m:1. If a river section is found to be not adjacent, it means that the river section is likely to be misclassified. In this case, it is regarded as a 1:1 river section and the corresponding method is used to obtain the matching result.

[0098] After preprocessing, method 2 is used to match the m:1 river section. First, a buffer is constructed for it. Subsequently, the small-scale river section falling into the buffer is regarded as a candidate matching object for the river section. Then, x m:1 river sections with connectivity to the river section are obtained. When x>1 and these x river sections can be merged into a line y (same level), the similarity between line y and the small-scale river section is calculated. Next, the similarity between the line merged by x-1 connected river sections and the small-scale river section is calculated, and the cycle is repeated z times until xz equals 1. In each cycle, the size of the similarity is recorded and compared. Finally, the river section combination with the greatest similarity is selected, and it is considered that it is successfully matched with the small-scale river section. Similarity calculation is better applicable to situations where the length difference between line objects is large. When the "minimum direction change angle" of two river sections is less than 5° and their river section levels are consistent, when the "comprehensive median Hausdorff distance" of the two is the smallest, their similarity is considered to be the greatest. For those m:1 river sections that failed to match successfully with the small-scale river sections, they were regarded as 1:0 river sections and re-matched according to the 1:0 matching strategy to obtain the final matching results.

[0099] When formulating the matching strategy for 1:0 river sections, the matching goal is to identify those river sections that are misclassified as other modes of this category and find matches. Considering that the m:1 river sections that are misclassified into this category are difficult to convert into 1:1 for matching, the main purpose is to identify the 1:1 river sections. Method 3 is used for 1:1 river section matching. First, the buffer radius is obtained, and the buffer is constructed. The small-scale river sections that fall within the buffer of the river section are regarded as the candidate matching set of the river section. Then, the similarity between the river section and the small-scale data in the candidate set is measured. If the similarity meets the conditions and the similarity is greater than the similarity mean of the 1:1 river section matching results, the match is considered successful and used as the matching result.

[0100] Furthermore, based on the above method, an embodiment of the present invention also provides a cross-scale network river system matching system based on matching pattern classification, comprising: a graph construction module and a graph matching module, wherein:

[0101] A graph construction module is used to construct the large-scale network river system data to be matched into a graph structure, and extract graph node labels according to the network river system characteristic indicators, wherein the network river system characteristic indicators include river section grade, river section watershed density, river section connectivity, river section maximum length, river section overall curvature and river section-road proximity;

[0102] The graph matching module is used to classify and identify the graph structure of large-scale networked river system data by using the pre-trained matching pattern classification model and according to the graph node labels, establish a mapping relationship between the graph structure features and the matching pattern, and execute the matching strategy corresponding to the matching pattern based on the matching pattern of each river section of the large-scale networked river system to be matched and based on spatial proximity and geometric similarity, so as to achieve matching with the small-scale networked river system to be matched.

[0103] In order to verify the effectiveness of this solution, the following is a further explanation based on experimental data:

[0104] Three machine learning methods, GCN, graph attention network (GAT) and random forest (RF), are selected to compare with the classification model in this case. In the GCN method, there are 3 GCNConv layers and one fully connected layer, and other hyperparameters are consistent with the model in this case. The parameter settings of the GAT method are the same as those of the GCN method. In these three methods, the embedding vector dimension is set to 64, 128, and 256 respectively. In the RF method, the number of decision trees is set to 50, 100, and 200. Table 1 shows the classification accuracy of the four models on the test data set under different parameters.

[0105] Table 1 Matching pattern classification accuracy of different machine learning methods

[0106]

[0107] From the above results, it can be seen that the classification model based on GraphSAGE in this case solution has excellent overall performance, and it performs best when the vector dimension is 128. Comparing the four models, the best accuracy of the GraphSAGE model in this case solution reached 91.5%, which has certain advantages over GCN, GAT and RF. The traditional machine learning-RF model also performed better than GCN and GAT, with the best accuracy reaching 84.9%. The GCN and GAT models performed generally, with the best accuracies of 76.7% and 74.7%. The GraphSAGE model can learn the node information that is not a first-order neighbor but close to it, and complete the iterative transmission task of information with the help of sampling and aggregation operations, which is in line with the mapping production process. The RF method performs well in dealing with complex nonlinear relationships, and its ability to fit data is relatively stable, thus leading GCN and GAT in performance. In addition, when allocating weights to first-order neighbor nodes, the attention mechanism of GAT may lead to unreasonable weights due to differences in sensitivity to different features, thereby affecting the overall performance.

[0108] After setting the embedding vector dimension to 128, the classification results of the test set were obtained, and the accuracy, precision, recall and F1 value of various matching patterns were calculated. Fig.10Specifically, the matching patterns of most river sections were correctly identified, especially the 1:0 matching pattern, with an accuracy of 96.3%. This may be because when transforming cross-scale data, the 1:0 river section only underwent a comprehensive operation of deletion, and this feature is easy to be identified by the graph neural network.

[0109] The FP and FN of each matching mode are statistically analyzed, as shown in Table 2. FP represents the number of samples predicted to belong to category a but actually belong to other categories, and the categories to which they actually belong are classified and counted; FN represents the number of samples predicted not to belong to category a but actually belong to category a, and the categories to which they are predicted are classified and counted. Category a can be one of the 1:0, 1:1, and m:1 matching modes. It can be observed from Table 2 that the 1:0 matching mode is easily misjudged as 1:1, the 1:1 matching mode is easily misjudged as m:1, and the m:1 matching mode is easily misjudged as 1:1, while the classification model is not easy to confuse the 1:0 and m:1 matching modes, which is also consistent with the human cognitive process.

[0110] Table 2 FP and FN of each matching mode

[0111]

[0112]

[0113] In order to evaluate the regional generalization performance of the matching pattern classification method, the Bolsward region of the Netherlands was selected as the experimental object, with a total of 1,198 river sections. By using the pre-trained graph neural network to classify the matching patterns of these river sections, the overall classification accuracy was 84.81%. Table 3 shows the accuracy, precision, recall and F1 value of various matching patterns in detail. These experimental results show that the matching model classification method has good generalization performance and can effectively identify matching patterns of river sections in different regions.

[0114] Table 3 Classification accuracy of matching patterns in the bolsward area

[0115]

[0116] Different matching strategies are adopted for river sections with different matching patterns, and finally the matching results of the test set data and other data sets are obtained. In the experiment, precision P and recall R are constructed to comprehensively test the matching effect, where C represents the correct number of matching results, M represents the number of matching pairs obtained by the matching algorithm, and A represents the number of matching pairs identified manually (including 1:m and m:n matching pairs in total).

[0117] As shown in Table 4, the overall matching rate of the test set reached 93% and the recall rate reached 94%. Among them, the matching accuracy of the 1:1 river section was the highest. In the classification matching process, the 1:1 river section that was misclassified into m:1 and 1:0 was effectively extracted, while the error correction difficulty of the actual m:1 river section predicted as 1:0 river section was relatively large, which reduced the matching accuracy of the 1:0 and 1:m categories. At the same time, the matching algorithm for the 1:1 river section is relatively mature and it is easier to obtain accurate matching results. For the m:1 river section, its matching accuracy is relatively low, mainly due to the high difficulty of matching this type of river section. At the same time, only partial matching results were obtained during the matching process. The above performance is due to the fact that different matching strategies are adopted for river sections with different matching modes in this case scheme, absorbing the advantages of multiple matching methods. At the same time, by setting the matching order, matching starts from the 1:1 river section, and the buffer threshold and similarity threshold of the matching mode are applied to other matching modes. Since the threshold has strong constraints, the reliability of the matching results is further improved. In addition, when the matching pattern obtained by the matching strategy is inconsistent with the matching pattern obtained by the graph neural network, a differentiated processing method is adopted to determine the final matching pattern, thereby improving the matching accuracy.

[0118] Table 4 Statistical results of evaluation indicators for different matching modes

[0119]

[0120] Fig.11 (a) in the figure shows the matching results of the river sections in the test set. In the matching process, there are mainly the following errors and deficiencies: First, the matching results of the m:1 river section are incomplete, and only partial matching results can be obtained. Fig.11 As shown in (b) in the figure. These m:1 river sections have significant morphological differences from the corresponding small-scale river sections, which results in poor accuracy of current matching methods when processing these river sections. Since they cannot be converted into 1:1 similarity metrics, effective matching strategies need to be developed for these specific m:1 river sections. Second, there are cases of mismatching, such as Fig.11 This type of river section is often surrounded by other river sections with similar geometric and semantic features, which increases the difficulty of correct matching and affects the final matching results. Third, the current method is still unable to effectively match 1:m and m:n river sections, such as Fig.11 As shown in (d), how to deal with this type of river section will be the next research direction.

[0121] In order to verify the effect of the idea based on matching pattern classification in this case in improving the matching accuracy, the experimental results are detailed in Table 5. The method (a) converts the m:1 river section into 1:1 for matching, and the matching accuracy of the 1:1 river section is high. The method (b) obtains the matching result by calculating the feature item. The design of the feature item effectively considers the local and global features, and the matching effect is better for the m:1 river section. The experimental results show that this case is better than the other two methods and can better take into account their algorithm advantages. At the same time, the solution of this case sets the matching order to avoid the manual setting of the buffer threshold and similarity threshold.

[0122] Table 5 Statistical results of evaluation indicators of different matching methods

[0123]

[0124] This strategy is used to match the data of the Bolsward area in the Netherlands, and the matching results are shown in Table 6. As can be seen from Table 6, this solution still has a certain degree of reliability in other areas, and the overall matching accuracy is 91.8%. Among them, the matching accuracy of the 1:1 river section reached 99.4%, and the matching accuracy of the m:1 river section was lower, which is basically consistent with the above experimental results.

[0125] Table 6 Matching accuracy of Bolsward area

[0126]

[0127] The above theoretical analysis and experimental data show that this scheme can effectively improve the matching accuracy and recall rate of networked river systems, and has good theoretical and application value.

[0128] Unless otherwise specifically stated, the relative steps, numerical expressions and values ​​of the components and steps set forth in these embodiments do not limit the scope of the present invention.

[0129] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0130] The units and method steps of each example described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in the above description according to function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. A person of ordinary skill in the art may use different methods to implement the described functions for each specific application, but such implementation is not considered to be beyond the scope of the present invention.

[0131] Those skilled in the art will appreciate that all or part of the steps in the above method can be completed by instructing related hardware through a program, and the program can be stored in a computer-readable storage medium, such as a read-only memory, a disk or an optical disk. Optionally, all or part of the steps in the above embodiment can also be implemented using one or more integrated circuits, and accordingly, each module / unit in the above embodiment can be implemented in the form of hardware or in the form of software function modules. The present invention is not limited to any specific form of combination of hardware and software.

[0132] Finally, it should be noted that the above-described embodiments are only specific implementations of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The protection scope of the present invention is not limited thereto. Although the present invention is described in detail with reference to the above-described embodiments, ordinary technicians in the field should understand that any technician familiar with the technical field can still modify the technical solutions recorded in the above-described embodiments within the technical scope disclosed by the present invention, or can easily think of changes, or make equivalent replacements for some of the technical features therein; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A cross-scale network river system matching method based on matching pattern classification, characterized in that: Include: The large-scale networked river system data to be matched is constructed into a graph structure, and graph node labels are extracted according to the networked river system characteristic indicators, wherein the networked river system characteristic indicators include river section grade, river section watershed density, river section connectivity, maximum length of river section, overall curvature of river section, and proximity between river section and road; The pre-trained matching pattern classification model is used to classify and identify the graph structure of the large-scale networked river system data according to the graph node labels, and a mapping relationship between the graph structure features and the matching patterns is established. According to the matching patterns of each river section of the large-scale networked river system to be matched and based on spatial proximity and geometric similarity, the matching strategy corresponding to the matching pattern is executed to achieve matching with the small-scale networked river system to be matched.

2. The cross-scale network river system matching method based on matching pattern classification according to claim 1 is characterized in that: The large-scale network river system data to be matched is constructed into a graph structure, including: Preprocessing the large-scale networked river system data to remove abnormal data in the networked river system data, wherein the preprocessing includes deleting disordered lines, broken river section connections, and topological checking, wherein the abnormal data includes false nodes and / or self-intersecting edges; The networked river system is divided into river sections according to its confluence nodes and level changes. Each river section is regarded as a graph node, and the connections between river sections are regarded as edges, so as to construct a dual graph of the networked river system.

3. The cross-scale network river system matching method based on matching pattern classification according to claim 1 is characterized in that: Extract graph node labels based on network river system characteristic indicators, including: Divide the networked river system into hierarchical structures according to river width, and obtain the river section grade of each river section in the networked river system, wherein the river section grade is the level of the river section in the hierarchical structure of the networked river system; A river section grid is established based on the connecting lines of the first and last points of the river section, and the river section watershed density is obtained based on the total length of the river section in the grid and the grid area; The connectivity of the river section is obtained according to the connection between the river section nodes in the graph structure, and the smooth chain stroke of the river section is constructed according to the semantic constraints and angle constraints in the graph structure. The maximum length of the river section is obtained according to the length of the smooth chain stroke of the river section. The overall curvature of the river section is obtained based on the length of the connecting line between the first and last points of the smooth chain stroke of the river section and the area of ​​the shape enclosed, and the average length and average area of ​​the smooth chain stroke in the networked river system; and the proximity relationship between the river section and the road is obtained based on the minimum distance between the midpoint of the river section and the road.

4. The cross-scale network river system matching method based on matching pattern classification according to claim 1 is characterized in that: Before using the pre-trained matching pattern classification model to classify and identify the large-scale network river system data graph structure according to the graph node labels, it also includes: Acquire large-scale remote sensing image data and small-scale terrain data, pre-process the data and segment the river sections, and construct the large-scale remote sensing image data after pre-processing and segmentation into a graph structure, use manual and / or machine methods to identify the matching pattern of each river section in the large-scale remote sensing image data, use the network river system characteristic indicators of the large-scale remote sensing image data as graph node labels, and use the graph structure and graph node labels to construct a river section sample data set for graph neural network training; A graph neural network model is constructed based on GraphSAGE, and the graph neural network model is trained according to the model loss function and the river section sample dataset. The trained graph neural network model is used as a pre-trained matching pattern classification model.

5. The cross-scale network river system matching method based on matching pattern classification according to claim 4 is characterized in that: The graph neural network model includes: an input layer, a hidden layer and an output layer, wherein the input layer is used to receive graph structure and graph node label data, the hidden layer includes three GraphSAGE convolutional layers and one fully connected layer, and the output layer takes the hidden layer input as input and uses the softmax function to generate node classification probability distribution.

6. The cross-scale network river system matching method based on matching pattern classification according to claim 4 or 5, characterized in that: The graph neural network model is trained using the river section sample dataset, including: The graph structure and graph node labels are used as model inputs. Each node in the graph structure is taken as a core node. The graph structure is divided into k layers from the inside to the outside, and sampling nodes are selected from neighboring nodes according to the number of samples. The features of the sampling nodes in the kth layer are aggregated to the nodes of the corresponding k-1 layer, and the features of the sampling nodes in the k-1 layer are aggregated to the core nodes. The sampling node features of the core nodes in the k-1 layer are aggregated into the overall features of the core nodes. The classification function is used to map the aggregated feature information to the matching pattern category probability, output the network river system matching pattern classification type, compare the model output with the true matching pattern, and calculate the loss function. According to the calculated loss function, the chain rule is used to propagate the gradient from the model output layer to the model input layer to update the model parameters according to the gradient.

7. The cross-scale network river system matching method based on matching pattern classification according to claim 1 is characterized in that: According to the matching pattern of each river section of the large-scale network river system to be matched, the matching strategy corresponding to the matching pattern is executed based on spatial proximity and geometric similarity, including: When the matching mode is 1:1, obtain the candidate matching set of each target river section under the matching mode, measure the similarity between each target river section and the small-scale network river section in the corresponding candidate matching set, and select the small-scale network river section in the matching item of the corresponding target river section based on the similarity; If multiple target river sections match the same small-scale network river section, the target river section with the greatest similarity is identified as the 1:1 matching mode, and the matching modes of other target river sections in the multi-hop target river section are classified as m:

1. If there is a target river section that does not match the small-scale network river section, the matching mode of the target river section is classified as 1:

0. When the matching mode is m:1, determine whether the target river section of this category is adjacent to other m:1 river sections. If there is no adjacent situation, the matching mode of the target river section is classified as 1:1, and the corresponding matching strategy is adopted; if there is an adjacent situation, obtain the candidate matching set of each target river section under the m:1 matching mode, merge the other x target river sections connected to the target river section, and perform river section matching based on the combination of connected target river sections. The matching mode of the target river section that has not been successfully matched is classified as 1:0; When the matching mode is 1:0, the river sections misclassified by the matching mode are identified and their matching items are found. By obtaining the candidate matching set of each target river section under the 1:1 matching mode, the similarity of each target river section and the small-scale network river section in the corresponding candidate matching set is measured to obtain the target river section matching items based on the similarity.

8. The cross-scale network river system matching method based on matching pattern classification according to claim 7 is characterized in that: Get the candidate matching set of each target river section in the matching mode, including: Construct a river section buffer, and include at least one corresponding small-scale networked river system section data in each target river section buffer of this category, and use the small-scale networked river system sections falling into each target river section buffer as the candidate matching set of each target river section under the corresponding matching mode.

9. The cross-scale network river system matching method based on matching pattern classification according to claim 7, characterized in that: The river sections are matched based on the connected target river sections, including: When x>1 and the x target river sections are merged into a y river section line of the same level, the y river section line is measured for similarity with the small-scale network river section in the candidate matching set, and then the line formed by merging x-1 connected river sections is calculated and measured for similarity with the small-scale network river section in the candidate matching set, and the cycle is repeated z times until xz is equal to 1. The similarities are recorded and compared, and the target river section combination with the greatest similarity is selected as the target river section that successfully matches the small-scale network river section, and the corresponding small-scale network river section is used as the matching item of the target river section.

10. A cross-scale network river system matching system based on matching pattern classification, characterized in that: Contains: graph construction module and graph matching module, among which, A graph construction module is used to construct the large-scale network river system data to be matched into a graph structure, and extract graph node labels according to the network river system characteristic indicators, wherein the network river system characteristic indicators include river section grade, river section watershed density, river section connectivity, river section maximum length, river section overall curvature and river section-road proximity; The graph matching module is used to classify and identify the graph structure of large-scale networked river system data by using the pre-trained matching pattern classification model and according to the graph node labels, establish a mapping relationship between the graph structure features and the matching pattern, and execute the matching strategy corresponding to the matching pattern based on the matching pattern of each river section of the large-scale networked river system to be matched and based on spatial proximity and geometric similarity, so as to achieve matching with the small-scale networked river system to be matched.

Images