River network pattern recognition method combining second-order co-occurrence matrix and global features

Abstract

The application discloses a river network pattern recognition method combining a second-order co-occurrence matrix and global features, and comprises the following steps: obtaining vector data of a river basin, pre-processing and marking the vector data, and obtaining training samples and test samples; based on the vector data, constructing a graph structure of node-edge relationship according to the confluence condition of the river, and obtaining local adjacency relationships between river sections based on the graph structure; constructing a second-order co-occurrence matrix based on the local adjacency relationships and combining local attributes of the river sections, obtaining global attributes of the river network, and performing normalization processing on the global attributes of the river network to obtain normalized data; based on a convolution layer and a full connection layer, constructing a multi-input CNN network taking the second-order co-occurrence matrix and the normalized data as input features; training and verifying the multi-input CNN network through the training samples and the test samples, obtaining a recognition model, performing river network pattern recognition through the recognition model, and obtaining a recognition result. The application improves the accuracy of river type recognition.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of river network recognition, and particularly relates to a river network pattern recognition method combining a second-order co-occurrence matrix and global features. BACKGROUND

[0002] In geographical environment, a river is a continuous whole in spatial structure, biological composition and time scale, and has strong systematization, which is called a river system. The river system is an open system of economic society, natural resources and ecological environment, and due to the flowability of water, the system continuously exchanges matter and energy and transmits information with the outside world, and meanwhile, the system completes self-organization and self-coordination through the synergistic effect between components in the system. Different levels of trunk streams and tributaries are formed by river sources, lakes, wetlands and estuaries, and form numerous river network systems. From the river source to the downstream, the width, depth, flow rate and flow of the river system have the characteristics of continuous change. Distinguishing different river networks helps to identify different river network types and understand the natural evolution process. Meanwhile, the arrangement form of the water system often reflects certain geological structure and crustal movement properties, and through the analysis of the distribution pattern, the general situation of the geological structure and crustal movement can be inferred, so that the correct judgment of the type of the water system has an auxiliary role for judging the local geological structure. In addition, the physical, chemical and aquatic ecological system characteristics of different types of rivers are different, and the repair and protection strategies are also different, so the research on river classification has attracted the attention of many researchers.

[0003] The use of high-resolution remote sensing products can quickly obtain more indicators reflecting the characteristics of rivers and comprehensively and accurately reflect the characteristics of the aquatic ecological system. However, high-resolution remote sensing image data is expensive and labor-intensive, and is difficult to apply in large areas of the basin, and in the geographic information system database, information is usually stored as vector data, and the evaluation of spatial patterns must directly process individual geometric entities. In this regard, the drainage pattern semantic model, hierarchical and relationship model, attribute extraction and classification strategy and other methods are used to quantitatively describe and classify 8 types of rivers. It creates a precedent for automatic recognition of vector river network drainage patterns, sets thresholds based on geometric quantitative indicators of rivers to gradually divide different categories of water systems to extract geometric characteristics of river networks, and uses decision trees, analytic hierarchy process and fuzzy set theory to divide different river network patterns. According to the complex context relationships such as river section level and drainage density, the type of the river network is determined. According to the distribution characteristics of the confluence angle, the support vector machine method is used to classify some drainage systems. However, the above traditional methods generally need to rely on the experience of researchers to design multiple thresholds, lack objectivity and universality, and it is difficult to construct a stable and universal autonomous river network pattern recognition model. SUMMARY

[0004] To solve the above technical problems, the application provides a river network pattern recognition method combining a second-order co-occurrence matrix and global features to solve the problems of the prior art.

[0005] To achieve the above object, the application provides a river network pattern recognition method combining a second-order co-occurrence matrix and global features, comprising the following steps:

[0006] Vector data of a river basin is acquired, the vector data is preprocessed and labeled to obtain training samples and test samples;

[0007] Based on the vector data, a graph structure of node-edge relationship is constructed according to the confluence of the river, and local adjacency relationships between river sections are obtained based on the graph structure;

[0008] A second-order co-occurrence matrix is constructed based on the local adjacency relationships and combined with local attributes of the river sections, global attributes of the river network are obtained, and the global attributes of the river network are normalized to obtain normalized data;

[0009] A multi-input CNN network taking the second-order co-occurrence matrix and the normalized data as input features is constructed based on convolution layers and fully connected layers;

[0010] The multi-input CNN network is trained and verified through the training samples and the test samples to obtain a recognition model, the river network pattern is recognized through the recognition model, and a recognition result is obtained.

[0011] Preferably, the global attributes of the river network include river network density, river network rectangularity, river network connected ring degree, node connection rate, water system connectivity and river channel frequency.

[0012] Preferably, the expression of the river network density is:

[0013] Rd = L / A

[0014] In the formula, L represents the total length of all river sections, and A is the total area of the basin.

[0015] Preferably, the local attributes of the river sections include length, connectivity, direction, rectangularity, relative position, angle, river section bending degree and branch and trunk angle difference.

[0016] Preferably, the local attributes of the river sections are further graded before the second-order co-occurrence matrix is constructed, and the grading method includes a natural breaking method and an equal interval division method.

[0017] Preferably, the expression of the value of each position in the second-order co-occurrence matrix is:

[0018]

[0019] In the formula, l, m, n respectively represent the level category of three continuous units, θ1 represents the direction of the first unit to the second unit in the three continuous units, θ2 represents the direction of the second unit to the third unit in the three continuous units, Indicates the number of level value combinations (l, m, n) when the first-order and second-order adjacent directions are θ1 and θ2 respectively, Indicates the total number of all three level size combinations corresponding to the direction angle combination.

[0020] Preferably, the identification model adopts a ReLu function as an activation function, model training adopts back propagation guided gradient descent, takes cross entropy as a loss function, adopts an Adam optimizer to optimize model parameters, and the output layer of the full connection layer and the fusion feature network of the second-order co-occurrence matrix refers to a dropout layer.

[0021] Preferably, the identification result includes dendritic shape, fan shape, parallel shape, grid shape and skeleton shape.

[0022] Compared with the prior art, the present application has the following advantages and technical effects:

[0023] The present application is easier to extract the hidden information of the elements by constructing a matrix, selects the local attributes of the river section, considers the upstream and downstream relationship between the river sections to construct a graph structure of node-edge relationship, utilizes the attributes between the continuously adjacent nodes, and counts the second-order adjacent situation, so as to construct a second-order co-occurrence matrix. The application of the second-order co-occurrence matrix expands the discussion of the neighborhood relationship between the river sections, which is not limited to the upstream and downstream relationship. In addition, the characteristics of the river vector data are quantified through the matrix, which also makes the abstract Gestalt principle be embodied.

[0024] The present application is aimed at river data, and the embodiment method describes different types of rivers from two aspects of local features and global features. The local features are obtained from the river section, such as the curve length, bending degree and relative position of the river section. The global features are obtained from the river, such as the river network density and water system connectivity. The combination of the local features and the global features of the river more reasonably and comprehensively describes the river, and improves the accuracy of the river type identification. BRIEF DESCRIPTION OF DRAWINGS

[0025] The drawings constituting a part of the present application are used to provide a further understanding of the present application, the illustrative embodiments of the present application and the description thereof are used to explain the present application, and do not constitute an improper limitation on the present application. In the drawings:

[0026] Figure 1 It is a structure diagram of the identification method of the embodiment of the present application.

[0027] Figure 2For the graph structure of the node-edge relationship of the embodiment of the application, (a) a river distribution map, the midpoint of the river section line is extracted as a node; (b) a node-edge relationship graph structure of the river is constructed according to the upstream and downstream relationship of the river section; (c) a degree matrix of the river section is obtained from the graph structure of the node-edge relationship;

[0028] Figure 3 For the river global feature description graph of the embodiment of the application, (a) a river section and its node distribution map; (b) a schematic diagram of constructing the minimum rectangular boundary of the river; (c) a schematic diagram of constructing the convex surface of the river;

[0029] Figure 4 For the river section local feature description of the embodiment of the application, (a) a river section, a river section starting point and a reference point distribution map; (b) a schematic diagram of constructing the minimum rectangular boundary of the river section and the main direction and the rectangular degree; (c) a schematic diagram of calculating the bending degree of the river; (d) a schematic diagram of representing the angle of the relative position of the river section;

[0030] Figure 5 For the division and judgment schematic diagram of the direction angle of the embodiment of the application, (a) a schematic diagram of 8 regions of the angle value range; (b) an original river data graph; (c) a graph structure of the node-edge relationship of the river; (d) an example schematic diagram of determining the angle of the connection relationship;

[0031] Figure 6 For the second-order co-occurrence matrix calculation schematic diagram of the river size of the embodiment of the application;

[0032] Figure 7 For the river sample type graph of the embodiment of the application;

[0033] Figure 8 For the CNN double-channel network structure graph of the embodiment of the application;

[0034] Figure 9 For the network training result schematic diagram of the embodiment of the application, (a) a curve graph of the training accuracy and the test accuracy, (b) a curve graph of the training process loss value and the test process loss value. DETAILED DESCRIPTION

[0035] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0036] It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a group of computer executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that here.

[0037] Embodiment one

[0038] As Figure 1 shown in the embodiment, a river network pattern recognition method combining a second-order co-occurrence matrix and global features is provided, comprising the following steps:

[0039] S1, respectively from OSM and USGS, obtain the vector data of the river basin, pre-process and mark the vector data, and obtain training samples and test samples;

[0040] S2, based on the vector data, a graph structure of node-edge relationship is constructed according to the confluence of the river, and the local adjacency relationship between the river reaches is obtained based on the graph structure;

[0041] Specifically, the river data is subjected to topological inspection and editing, and is composed into a network structure connected to each other. The "graph" structure of the river network is constructed, wherein the river reaches are nodes of the "graph" structure, and if two river reaches are connected, the corresponding "nodes" are connected to form edges of the "graph" structure. Figure 2 The construction process of the "graph" structure of the river network is shown, Figure 2 the 36 river reaches in (a) are respectively marked with nodes N i (i∈{0,1,2…,35}); Figure 2 The black lines in (b) are the edges of the "graph" structure, indicating that the river reaches corresponding to the two "nodes" at the ends thereof have a connection relationship. In the graph, the edge ei j represents that the river reaches corresponding to the nodes Ni,N j have a connection relationship; Figure 2 (c) is the adjacency matrix corresponding to the graph structure, expressing the connection relationship of all river reaches in the river network.

[0042] S3, based on the local adjacency relationship, a second-order co-occurrence matrix is constructed combining the local attributes of the river reaches, the global attributes of the river network are obtained, and the global attributes of the river network are normalized to obtain normalized data;

[0043] The embodiment extracts the features of the river network from the local and global perspectives. The local features include the length, direction and curvature of a single river reach, while the global features are the area of the river basin range, the river network density and other features. The global features of the river network and the local features of the river reaches are described below.

[0044] Global features of the river:

[0045] The global features selected in the embodiment include river network density, river network rectangularity, river network connectedness, node connection rate, water system connectedness and river channel frequency.

[0046] (1) River network density: the ratio of the total length of the river to the total area of the basin, reflecting the density of the river. The calculation formula of the river network density is as shown in formula (1).

[0047] Rd = L / A (1)

[0048] In formula (1), L is the total length of all river reaches in the river, and A is the total area of the river basin. In specific applications, the area of the convex hull of the river network is used, such as Figure 3 (c) identifies the convex hull corresponding to the given river network.

[0049] (2) River network rectangularity: reflects the ratio of the longitudinal range and the transverse range of the river, which is represented by the aspect ratio of the minimum circumscribed rectangle of the river network. The calculation formula is shown in formula (2). Figure 3 (b) gives the minimum rectangle of the river network.

[0050] E = Ld / Ls (2)

[0051] In the formula, L d , L s correspond to the length and width of the minimum rectangle, respectively.

[0052] (3) River network connectedness: reflects the level of material and energy exchange between river reaches.

[0053]

[0054] (4) Node connectivity: reflects the ease of connection between river reaches in the water system.

[0055] βl / v (4)

[0056] (5) Water system connectedness: indicates the degree of connection of the river network.

[0057] γ = l / 3(v-2) (5)

[0058] (6) River channel frequency: indicates the number of river channels per unit area. The more the number of channels, the better the structural connectivity between lakes.

[0059]

[0060] In formula (5), l is the number of river reaches; v is the total number of nodes (such as Figure 3 (a) river nodes), and in formula (6), N represents the number of nodes connecting between river reaches (such as Figure 3 (c)).

[0061] River reach characteristics:

[0062] Local characteristics are geometric parameters of a single river reach, including length, connectivity, direction, rectangularity, relative position, angle, difference between tributary and trunk stream angle, and river reach curvature.

[0063] (1) Curve length: the length of the river reach curve.

[0064] (2) Degree: the number of river segments connected to each river segment, calculated from the degree matrix of (c). Figure 2

[0065] (3) Direction: the direction of a river segment is represented by the long side direction of the minimum rectangle formed by the river segment curve, as shown in (b). Figure 4

[0066] (4) Rectangularity: the aspect ratio of the minimum rectangle formed by the river segment curve, as shown in (a-b). Figure 4

[0067] e = ld / is (7)

[0068] In equation (7), l d is the length of the minimum rectangle of a river segment, and l s is the width of the minimum rectangle of a river segment.

[0069] (5) Distance relative position: the distance difference between the head and tail points of a river segment and the outlet.

[0070] Al = loB - loA (8)

[0071] (6) River segment relative angle: the angle between the head and tail points of a river segment and the outlet, as shown in equation (9), and the schematic diagram is given in (d).

[0072]

[0073] In equations (8-9), A and B are the end points of a river segment, and O is the outlet of the river. l OA and l OB represent the distances from the two end points of a river segment to the outlet, and L is the distance between points A and B.

[0074] (7) Angle difference between tributary and main stream: the angle between the main direction of the river network and the main direction of each river segment, reflecting the consistency of the river segment with the main direction of the river network.

[0075] (8) River segment bending degree: the ratio of the curve length of a river segment to the straight-line distance between the end points of the river segment (equation (10)), as shown in (c). Figure 4

[0076]

[0077] In the equation, l is the curve length of a river segment, and d is the distance between the two ends of the river segment.

[0078] Construction of the second-order co-occurrence matrix

[0079] ​​​​For the 8 local attributes of the river section, the second-order co-occurrence matrix is constructed as the statistical characteristics of pattern recognition. In order to ensure that the co-occurrence matrix formed by different attribute values has the same dimension size, the 8 attributes are reclassified according to the same level number, and the attribute values of each level are re-assigned to the corresponding numerical value.

[0080] The selected local attributes of the river section in this embodiment involve size, shape, direction and adjacency relationship, as shown in Table 1:

[0081] Table 1

[0082]

[0083] Before constructing the matrix, level division is needed, and the 8 attribute values representing local features are converted into corresponding levels. For example, for size, the natural breaking method is used to divide the curve length into 0-K classes, represented as {0, 1, 2…, K-1} levels; similarly, the rectangularity, relative position, and bending degree attributes are converted into the same level; for direction, the equal interval division method is used, and each level has an angle range: the river section direction, river section angle, and angle difference attribute values are divided into {0, 1, 2…, K-1}.

[0084] The second-order discussed in this embodiment needs to consider the combination of three units, and the combination of three unit directions can form different second-order co-occurrence matrices. The second-order co-occurrence matrix can be generated by constructing the graph structure of the river section and its relationship. Let p0, p1, p2 be three consecutive units, with p0 as the center unit, the direction of the line connecting p0 to p1 is the first-order direction, and the direction is θ1, the direction of the line connecting p1 to p2 is the second-order direction, and the direction is θ2, θ1, θ2 can be in the 8 quadrants, which are represented by the center direction, i.e. 0°, 45°, 90°, 135°, 180°, 225°, 270°, 315°, a total of 8 directions. As shown in (a). Figure 5 (b)-(d)) are shown in (b)-(d). Figure 5 (b)-(d)) are shown in (b)-(d). Figure 5 (d) are 4 different directions. θ1 and θ2 have 8 different value conditions, i.e. θ1, θ2 ∈ 0°, 45°, 90°, 135°, 180°, 225°, 270°, 315°, and there are 8×8 (=64) combination modes, each combination corresponds to a second-order co-occurrence matrix The values of each position in the matrix are The calculation is shown in formula (11):

[0085]

[0086] where l, m, n represent the rank category of three consecutive units respectively, which also correspond to the coordinates in the three-dimensional co-occurrence matrix, l, m, n ∈ {0, 1, 2…, K-1}, denotes the number of rank value combinations (l, m, n) when the first-order and second-order adjacent directions are θ1, θ2 respectively, denotes the total number of all three rank size combinations under the corresponding direction angle combination, obviously is a three-dimensional matrix of K × K × K.

[0087] Take the river structure in Figure 6 as an example, Figure 6 (b) is its graph structure, the curve length of the river section is extracted, so that the size rank of each node is given, and the size rank of each node is shown in Figure 6 (c), the size rank of the river is 4, and the value range is {0, 1, 2, 3}, and a second-order co-occurrence matrix is constructed. The relationship between two consecutive adjacent nodes constitutes a second-order unit, and eight quadrants are divided, and the quadrant to which the adjacent node direction belongs is determined, and the direction combination between the two consecutive adjacent relationships will appear 8 × 8 (=64) cases. As follows Figure 6 ((d)-(f)) are the calculation processes of the second-order co-occurrence matrix of the river size of two combinations of three direction angles (θ1=0°, θ2=0°; θ1=0°, θ2=45°; θ1=0°, θ2=90°) respectively.

[0088] As shown in Figure 6 (d), in the combination of θ1=0°, θ2=0° direction angle, the length rank combination of the river Figure 6 (a) is [3, 3, 1], [3, 3, 0], [3, 1, 0], [1, 0, 0], and the value of the second-order co-occurrence matrix is calculated according to formula (11) , that is, the result value corresponding to the position of the lth row, the mth column and the nth dimension in the three-dimensional matrix. For example Then the value of the third row, the third column and the first latitude in the three-dimensional matrix is 0.25. Thus, the rank value combination of the second-order unit of the entire river is traversed to construct 64 second-order co-occurrence matrices, Figure 6 ((d)-(f)) shows the results of three second-order co-occurrence matrices.

[0089] S4, based on the convolution layer and the full connection layer, a multi-input CNN network taking the second-order co-occurrence matrix and the normalized data as the input features is constructed;

[0090] The input features of the network are composed of two parts, the second-order co-occurrence matrix calculated by 8 local attributes of the river section, and the global features normalized by 6 global attributes of the river. The local features are divided into 6 levels, i.e. the feature size of each sample is 8x6x6x6x64, where 8 represents 8 local attributes, 6x6x6 represents the level size combination of three node units, and 64 represents the combination of two direction angles in the second-order unit. The second-order co-occurrence matrix data is transformed into a 3072x6x6 matrix through matrix transformation, and is passed through the convolutional layer and the pooling layer of the CNN. For the 6 global features, after normalization, they are passed through the fully connected layer, and finally fused with the second-order co-occurrence matrix data through the softmax network layer to output the class of the pattern. The model uses the ReLu function as the activation function, the back propagation guided gradient descent is used for model training, the cross entropy is used as the loss function, the Adam optimizer is used to optimize the model parameters, and the Dropout is used in the output layer of the fully connected layer of the second-order co-occurrence matrix and the fusion feature network respectively to prevent overfitting, and the Dropout rates are 40% and 20% respectively. The other parameters of the model, such as learning rate and bath_size, are 0.001 and 64 respectively.

[0091] S5, training and verifying the multi-input CNN network through the training samples and the test samples to obtain a recognition model, performing river network pattern recognition through the recognition model to obtain a recognition result. The recognition result includes dendritic shape, fan shape, parallel shape, grid shape and skeleton shape.

[0092] The above method is verified as follows in this embodiment.

[0093] 1, sample data construction

[0094] Considering the differences in water source, formation method, flow, flow direction and hydrological characteristics in the river system, the river system is divided into 5 categories, namely dendritic shape, fan shape, parallel shape, grid shape and skeleton shape. The characteristics are as follows: the dendritic water system has more branches, the main stream, branch and branch intersect at acute angles, and the arrangement is like a dendritic branch; the fan-shaped water system has a fan-shaped river basin profile formed by the combination of the main stream and the branch, such as the river and sea water system, and each branch flows into the sea; the parallel water system has almost parallel branches; the grid water system has almost right-angled intersections between the main stream and the branch; the skeleton water system has small river bends, and the branches are evenly distributed on both sides of the main stream, as shown in Figure 7 .

[0095] The training data of the embodiment is from OSM (https: / / www.openstreetmap.org), a total of 1500 data are collected for training of the network model, and each type of river has 300 data. The test data is from USGS (https: / / apps.nationalmap.gov), mainly collected from the Boise area of the capital of Idaho, USA, a total of 510 data are collected, including 106 dendritic rivers, 71 fan-shaped rivers, 107 parallel rivers, 100 grid-shaped rivers and 126 skeleton-shaped rivers.

[0096] 2. Combination of features and CNN network construction

[0097] For the training data and the test data, the local features of the river section (curve length, degree, river section direction, river section rectangularity, river section relative position, angle, angle difference, and river section bending degree) and the global features of the river (total length of the river, length-width ratio of the river basin range, river connection ring degree, node connection rate, river system connectivity, and river frequency) are calculated respectively.

[0098] For the local feature data, it is necessary to perform grade division and construct the corresponding second-order co-occurrence matrix respectively. In this embodiment, all attributes are divided into 6 grades, and each node in the graph structure is assigned a grade size of {0, 1, 2, 3, 4, 5}. According to the constructed second-order co-occurrence matrix of the river data, the global features of the river are combined and input into the deep learning network for recognition of the river pattern.

[0099] CNN (Convolutional Neural Networks), i.e., convolutional neural network, is a deep learning model or a multi-layer perceptron similar to artificial neural network, which is often used to analyze visual images and has a wide range of applications in image segmentation and other fields. In this embodiment, a CNN network with multiple input data is built based on the Pytorch deep learning environment for recognition of the river pattern, and the basic framework of the network is as shown in Figure 8 .

[0100] The input features of the network are composed of two parts, the second-order co-occurrence matrix calculated by the 8 local attributes of the river section, and the global features normalized by the 6 global attributes of the river. The local features are divided into 6 levels, i.e. the feature size of each sample is 8x6x6x6x64, where 8 represents 8 local attributes, 6x6x6 represents the level size combination of three node units, and 64 represents the combination of two direction angles in the second-order unit. The second-order co-occurrence matrix data is transformed into a 3072x6x6 matrix through matrix transformation, and is passed through the convolutional and pooling layers of the CNN. For the 6 global features, after normalization, they are passed through the fully connected layer, and finally fused with the second-order co-occurrence matrix data through the softmax network layer to output the class of the pattern. The model uses ReLu function as the activation function, the model training uses back propagation guided gradient descent, and the cross entropy is used as the loss function. The Adam optimizer is used to optimize the model parameters. In order to prevent overfitting, Dropout is used in the fully connected layer of the second-order co-occurrence matrix and the output layer of the fusion feature network, and the Dropout rates are 40% and 20% respectively. The other parameters of the model, such as learning rate and bath_size, are 0.001 and 64 respectively.

[0101] Figure 9 The changes of accuracy and loss during the model training process are shown respectively. With the increase of training iteration times, the training accuracy and test accuracy are gradually improved. After about 200 epochs, the accuracy gradually stabilizes, and the loss value also does not decrease, considering that the model has been fully trained.

[0102] 3. Recognition of river pattern

[0103] In order to further verify the feasibility of the model, the model is applied to specific river pattern recognition, and the vector river network elements collected from USGS are used as verification data to test the generalization of the model.

[0104] Table 2 is the confusion matrix of the classification results of the model in this embodiment on the test set. The classification accuracy of the model on the test set reaches 86.86%, among which the classification performance of the skeleton river is the highest, and the recall rate can reach 92.85%, and the classification accuracy of other classes can all reach more than 80%. The main error classification occurs between the dendritic and grid water systems.

[0105] Table 2

[0106]

[0107] The test sample data is mainly from Boise, the capital of Idaho, USA. The classification results of the model can identify most of the river types in the local area.

[0108] 4. Comparative experiment

[0109] The model quantifies the local features of the river section by the second-order co-occurrence matrix, and identifies the river pattern by combining the global features of the river. From the experimental results of pattern recognition, it can be seen that the model has good recognition effect. In order to further verify the high efficiency of the model in pattern recognition, the embodiment adopts GCN network for comparative experiment of pattern recognition. Based on the GCN network, the vector data of the river is directly used to construct the graph structure of the river. In order to ensure the fairness of the comparative experiment, the node attributes in the graph structure are the same as the 8 attribute values of the local features in the embodiment. The graph structure of the river is subjected to convolution, pooling and other operations, and the type of the river is determined.

[0110] The test data is consistent with the above data. The accuracy and recall rate of the two methods for river data recognition are compared. Table 3 is the test results of the GCN network model. The recall rate of the skeleton type is the highest, which is 80.95%, but it is still lower than the method of the embodiment, and the total accuracy is only 71.37%.

[0111] Table 3

[0112]

[0113] From Table 3, it can be seen that the recall rate of the dendritic type in the comparative experiment is only 64.15%.

[0114] Although the GCN network can directly extract the relevant attribute values of the vector data as features for pattern recognition, the input features are too limited and cannot accurately classify. On the contrary, the embodiment combines local features and global features, and introduces the second-order co-occurrence matrix to quantify the local attribute values, so the recognition effect is better.

[0115] The above is only the preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto. Any changes or replacements within the technical scope disclosed in the present application can be easily thought by those skilled in the art, which should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A river network pattern recognition method combining second-order co-occurrence matrix and global features, characterized by: The following steps are involved: Acquire vector data of a river basin, preprocess and label the vector data, and obtain training samples and test samples; Based on the vector data, constructing a graph structure of node-edge relationships according to the confluence of rivers, and obtaining local adjacency relationships between river sections based on the graph structure; Based on the local adjacency relationship and the local attributes of the river section, a second-order co-occurrence matrix is ​​constructed. The second-order co-occurrence matrix constructs three-dimensional statistical features through 64 direction combinations (θ1, θ2), where θ1 and θ2 represent the first-order and second-order flow direction angles of continuous river sections, respectively, to obtain the global attributes of the river network, and the global attributes of the river network are normalized to obtain normalized data; The local attributes of the river section include the length, connectivity, direction, rectangularity, relative position, angle, curvature of the river section and the angle difference between the tributary and the main stream; The global attributes of the river network include river network density, river network rectangularity, river network connectivity, node connectivity, water system connectivity and river channel frequency; Before constructing the second-order co-occurrence matrix, it is necessary to classify the local attributes of the river section. The classification methods include the natural fracture method and the equal interval division method. Based on the convolutional layer and the fully connected layer, a multi-input CNN network is constructed with the second-order co-occurrence matrix and normalized data as input features; The multi-input CNN network is trained and verified through training samples and test samples to obtain a recognition model, and river network pattern recognition is performed through the recognition model to obtain a recognition result.

2. The river network pattern recognition method combining the second-order co-occurrence matrix and global features according to claim 1 is characterized in that: The expression of the river network density is: R d =L / A Where L is the total length of all river sections, and A is the total area of ​​the river basin.

3. The river network pattern recognition method combining second-order co-occurrence matrix and global features according to claim 1 is characterized in that: The expression of the value of each position in the second-order co-occurrence matrix is: Where l, m, and n represent the level categories of the three consecutive units, θ1 represents the direction from the first unit to the second unit in the three consecutive units, and θ2 represents the direction from the second unit to the third unit in the three consecutive units. It represents the number of level value combinations (l, m, n) when the first-order and second-order adjacent directions are θ1 and θ2 respectively. It represents the total number of all three level size combinations under the corresponding direction angle combination.

4. The river network pattern recognition method combining second-order co-occurrence matrix and global features according to claim 1 is characterized in that: The recognition model uses the ReLu function as the activation function, the model training uses back-propagation guided gradient descent, cross entropy is used as the loss function, the Adam optimizer is used to optimize the model parameters, and a dropout layer is referenced in the fully connected layer through which the second-order co-occurrence matrix passes and the output layer of the fusion feature network.

5. The river network pattern recognition method combining second-order co-occurrence matrix and global features according to claim 1 is characterized in that: The recognition results include dendritic, fan-shaped, parallel, grid and skeleton shapes.

Images