A flash flood disaster zoning method integrating attributes and structures

Abstract

The present invention provides a method for flood disaster zoning that integrates attributes and structures, which relates to the technical field of flood disaster zoning. The method includes the following steps: (1) Preprocess rainfall and terrain factors and spatially aggregate them to the scale of each small watershed as the rainfall and terrain attributes of the small watershed; (2) According to the small watershed boundary data, use the center of the small watershed to represent the small watershed and serve as the nodes of the graph, and construct the edges between the nodes for adjacent small watersheds, so as to construct a graph structure to represent the spatial structure relationship of the small watershed; (3) Use the attributes and spatial structure relationship of the small watershed as inputs, and utilize three graph convolutional neural network modules to construct a graph autoencoder to simultaneously perform feature expression learning on the attributes and spatial structure of the small watershed; (4) Establish a graph decoder according to the output of each graph convolutional neural network module to simultaneously realize the reconstruction of the small watershed spatial structure and the clustering of small watersheds; (5) Merge small watersheds of the same category according to the best clustering result of the small watersheds to obtain the final flood disaster zoning scheme.

Description

Technical Field

[0001] The present invention relates to the technical field of mountain flood disaster zoning. Specifically, it is a mountain flood disaster zoning method that integrates attributes and structures. Background Art

[0002] Mountain floods are sudden surface runoff formed in small watersheds in hilly areas due to precipitation under special topographic conditions. Their prominent characteristics are high flow velocity, strong scouring force, and extremely strong destructive power (Arabameri et al., 2020; Zhang et al., 2022). Mountain flood disasters can cause huge losses to life and property, the ecological environment, and infrastructure, and often result in casualties (Bui et al., 2019; Zhai et al., 2021).

[0003] Mountain flood disaster zoning is to divide a region into several non-overlapping homogeneous regions by comprehensively analyzing the spatial differentiation law of mountain flood disaster influencing factors according to the zoning principles of within-region consistency and between-region difference (Alipour et al., 2020; Liu Changjun et al., 2021). In the same mountain flood zoning unit, mountain flood disasters usually have similar formation mechanisms, which can effectively reflect the differentiation law of mountain flood disasters and facilitate the transfer of parameters of mountain flood disaster forecasting and early warning models, providing a scientific basis for the prevention, control, and early warning of mountain flood disasters by "treating them separately" (Zhao Shipeng, 1996; Zhang Pingcang et al., 2006).

[0004] Currently, the existing mountain flood disaster zoning research in China is mainly the early national macro-scale zoning. However, due to the limitations of observation data, the divided zoning units are relatively simple and small in number, making it difficult to provide the spatial differentiation law of mountain flood disasters at a more refined scale for local government departments at the city and county levels. Summary of the Invention

[0005] The purpose of the present invention is to design a mountain flood disaster zoning method that integrates attributes and structures. By constructing a graph clustering neural network model, the structure of the small watershed and the attribute factors related to mountain flood disasters described by the small watershed are utilized simultaneously to obtain a mountain flood disaster zoning scheme, so as to solve the problem of the simplicity and small number of existing mountain flood disaster zoning units.

[0006] The present invention is realized through the following technical solutions:

[0007] A mountain flood disaster zoning method that integrates attributes and structures includes the following steps:

[0008] Step 1: Preprocess the rainfall and terrain factors related to mountain flood disasters in raster form, and spatially aggregate the extracted rainfall and terrain factors into each small watershed in polygon format as the small watershed attributes. The small watershed attributes include the rainfall and terrain attributes of the small watershed;

[0009] Step 2: Based on the small watershed boundary data in polygon format, represent the small watershed with the small watershed center as the node of the graph, and construct the edges between the nodes for adjacent small watersheds, thereby constructing a graph structure to represent the small watershed spatial structure;

[0010] Step 3: Using the small watershed attributes in Step 1 and the small watershed spatial structure in Step 2 as inputs, construct a graph autoencoder using three graph convolutional neural network modules to simultaneously perform feature expression learning on the small watershed attributes and the small watershed spatial structure;

[0011] Step 4: According to the output of each graph convolutional neural network module in Step 3, establish a graph decoder to simultaneously achieve the reconstruction of the small watershed spatial structure and small watershed clustering, obtain the optimal small watershed clustering result, and use it as the initial clustering result of the small watershed unit;

[0012] Step 5: According to the optimal small watershed clustering result in Step 4, merge the small watersheds of the same category to obtain the final mountain flood disaster zoning plan.

[0013] Further, in Step 2, the expression of the graph structure data obtained by constructing the graph structure is:

[0014] G=(V,E,X)

[0015] where, V={v1,...,v n} represents the set of n nodes in the graph (i.e., the set of small watershed centers); E={e ij} represents the set of edges between the nodes; the structure of G can be represented by the adjacency matrix A of the nodes; A ij =1 indicates that there is an edge between node i and node j, otherwise A ij =0; X={x1,...,x n} represents the attributes of the n nodes in the graph (i.e., the rainfall and terrain attributes of the small watershed).

[0016] Further, in Step 3, each graph convolutional neural network module includes a graph convolutional neural network layer, a batch normalization layer, and an activation layer.

[0017] Further, in Step 3, the autoencoder uses three graph convolutional neural network modules to implement feature learning expression for the input small watershed attributes and the small watershed spatial structure, as shown in the following formula:

[0018]

[0019] where, Z l+1 is the feature Z l at the l-th layer calculated through the graph convolutional neural network module to obtain the feature at the (l + 1)-th layer; is the sum of the adjacency matrix A and the identity matrix I; is the degree matrix, where W l are the parameters to be learned in the l-th layer of the graph neural network; σ is the activation function of the graph neural network; when l = 0, then Z 0 = X, which is the input small watershed attribute.

[0020] Furthermore, in step 4, the graph decoder uses the following formula to reconstruct the small watershed spatial structure by using feature inner product and activation function. The formula is:

[0021]

[0022] where Z is the output of the graph autoencoder; is the reconstructed adjacency matrix of the small watershed spatial structure, making it close to the original adjacency matrix A of the small watershed.

[0023] Furthermore, when the graph decoder performs graph clustering, it constructs a similarity index q i between the feature z u of node i and the clustering center μ iu , and calculates it through the graph clustering Q distribution function of the following formula. The formula is:

[0024]

[0025] For the similarity index q iu , a target value is constructed, that is, nodes of the same category are made closer, and nodes of different categories are made farther away to enhance the difference between different categories. Using the square of q iu , a P distribution function is constructed as the target distribution, making q iu as close as possible to the clustering center to achieve clustering-oriented graph feature supervised learning, as shown in the following formula:

[0026]

[0027] The graph clustering Q distribution function constructed by the graph decoder should be as close as possible to the P distribution function, and the KL function value is made as small as possible to achieve this, as shown in the following formula:

[0028]

[0029] where KL is a divergence function used to describe the difference between the P and Q distribution functions.

[0030] The present invention has the following advantages and beneficial effects:

[0031] In the present invention, attributes and structures are integrated to convert small watershed data into graph-structured data, so as to uniformly express the small watershed structure and the rainfall and terrain attributes related to flash flood disasters described by the small watershed. Through the graph convolutional neural network module, the integration and expression learning of the attributes and spatial structure features of the small watershed are realized. Through the joint optimization of the reconstruction of the small watershed spatial structure and the clustering of the small watershed, a better flash flood disaster clustering result is achieved, and a spatial zoning map of flash flood disasters is obtained, thereby describing the spatial differentiation law of flash flood disasters, solving the problem that the current flash flood disaster zoning units are brief and small in number, and providing a more scientific basis for formulating local prevention and control strategies for flash flood disasters. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0033] Figure 1 is the main flowchart of the flash flood disaster zoning method of the present invention;

[0034] Figure 2 is the division result of the flash flood disaster zoning method of the present invention. Among them, (a) represents the initial clustering result of the small watershed unit, and (b) represents the final flash flood disaster zoning plan. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0035] To make the objectives, technical solutions and advantages of the present invention clearer, the technical solutions of the present invention will be described in detail below. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other implementation manners obtained by those of ordinary skill in the art without creative efforts based on the embodiments of the present invention belong to the scope protected by the present invention.

[0036] Embodiment 1:

[0037] A flash flood disaster zoning method that integrates attributes and structures. By constructing a graph clustering neural network model, the structure of the small watershed and the attribute factors related to flash flood disasters described by the small watershed are utilized simultaneously to obtain a flash flood disaster zoning plan, so as to solve the problem that the current flash flood disaster zoning units are brief and small in number. The main flowchart is as Figure 1 shown, and it includes the following steps:

[0038] Step 1: Preprocess the rainfall and terrain factors related to mountain flood disasters, and spatially aggregate the rainfall and terrain factors extracted during the preprocessing to the scale of each small watershed as the small watershed attributes. Among them, the small watershed attributes (map attributes) include the rainfall and terrain attributes of the small watershed:

[0039] Since the original rainfall and terrain factors of the small watershed are in raster form, they need to be aggregated to the polygon-format small watershed as the attribute factors describing mountain flood disasters in the small watershed.

[0040] Step 2: According to the boundary data of the small watershed in polygon format, use the center of the small watershed to represent the small watershed and serve as the nodes of the graph, and construct the edges between the nodes for adjacent small watersheds. In this way, construct a graph structure to represent the spatial structure of the small watershed and obtain the small watershed graph structure data:

[0041] G=(V,E,X)

[0042] Among them, V={v1,...,v n} represents the set of n nodes in the graph (i.e., the set of small watershed centers); E={e ij} represents the set of edges between the nodes; the structure of G can be represented by the adjacency matrix A of the nodes; A ij =1 indicates that there is an edge between node i and node j, otherwise A ij =0; X={x1,...,x n} represents the attributes of the n nodes in the graph (i.e., the rainfall and terrain attributes of the small watershed).

[0043] Common clustering methods mainly use attribute information for clustering, but ignore spatial structure information such as the spatial neighborhood relationship of clustering units, and cannot well express the spatial autocorrelation characteristics of surface elements. In this step 2, by abstracting the small watershed as a node and considering the adjacent relationship of the small watersheds, a graph structure is constructed to express the spatial structure relationship of the small watersheds and integrate it into the mountain flood zoning process.

[0044] Step 3: Using the small watershed attributes (the rainfall and terrain attributes of the small watershed) in Step 1 and the small watershed spatial structure in Step 2 as inputs, use three graph convolutional neural network modules to construct a graph autoencoder to simultaneously perform feature expression learning on the small watershed attributes and the small watershed spatial structure.

[0045] In Step 3, each graph convolutional neural network module includes a graph convolutional neural network layer, a batch normalization layer, and an activation layer.

[0046] The autoencoder uses three graph convolutional neural network modules to realize the feature learning expression of the input small watershed attributes and the small watershed spatial structure, as shown in formula (1):

[0047]

[0048] Among them, Z l+1 is the feature Z of the l-th layer l calculated by the graph convolutional neural network module and is the feature of the (l + 1)-th layer; is the sum of the adjacency matrix A and the identity matrix I; is the degree matrix, where W l is the parameter to be learned in the l-th layer of the graph neural network; σ is the activation function of the graph neural network; when l = 0, then Z 0 = X, which is the input small watershed attribute.

[0049] It should be noted that, as can be seen from formula (1), the model measures the neighborhood relationship between nodes (i.e., the structural relationship between adjacent small watersheds) through and Z l W l considers the attribute features between adjacent nodes, and multiplies the two and outputs with the activation function to achieve the purpose of feature expression considering both attributes and structure.

[0050] The graph autoencoder proposed in step 3 uses three graph convolutional neural network modules to learn and express different dimensional features of the attributes and structures of the small watershed. Each graph convolutional neural network module includes a graph convolutional neural network layer, a batch normalization layer, and an activation layer to efficiently and accurately learn and express features. At the same time, adjacent modules are connected using a residual structure to accelerate feature learning and model convergence.

[0051] Step 4: According to the output of each graph convolutional neural network module in step 3, build a graph decoder to simultaneously realize the reconstruction of the small watershed spatial structure and the clustering of small watersheds, obtain the best clustering result of the small watersheds, and use it as the initial clustering result of the small watershed units.

[0052] The graph autoencoder realizes the hidden expression of the graph structure and graph attributes and outputs features. This feature needs to be supervised to achieve the model goal. The goal of the model is to reconstruct the graph structure and graph clustering. For this purpose, a graph decoder is constructed. The graph decoder uses inner multiplication of features and an activation function to reconstruct the small watershed spatial structure, as shown in formula (2):

[0053]

[0054] Among them, Z is the output of the graph autoencoder; is the reconstructed adjacency matrix of the small watershed spatial structure, making it close to the original adjacency matrix A of the small watershed.

[0055] When the graph decoder performs graph clustering, it constructs the similarity index q between the feature z i of node i and the clustering center μ u ​iu , calculated by the graph clustering Q distribution function of formula (3):

[0056]

[0057] To achieve graph feature supervised learning with clustering guidance, it is necessary to use the similarity index q iu Construct a target value, that is, bring nodes of the same category closer and keep nodes of different categories farther away to enhance the distinction between different categories; use q iu The square of to construct a P distribution function as the target distribution, so that q iu is as close as possible to the clustering center, as shown in formula (4):

[0058]

[0059] The graph clustering Q distribution function constructed by the graph decoder should be as close as possible to the P distribution function, and the KL function value should be as small as possible to achieve this, as shown in formula (5):

[0060]

[0061] Among them, KL is a divergence function used to describe the difference between the P and Q distribution functions.

[0062] In step 4, it is proposed to establish a graph decoder to simultaneously realize the reconstruction of the small watershed spatial structure and the clustering of small watersheds, to use the inner product between node features and activation functions to reconstruct the structural relationship between nodes, and to propose self-supervised clustering to cluster the graph, and to jointly optimize the model with the graph structure reconstruction error and the graph clustering error, and finally obtain the clustering result of the small watershed.

[0063] Step 5. According to the best clustering result of the small watershed in step 4, merge small watersheds of the same category to obtain the final mountain flood disaster zoning plan.

[0064] The results of mountain flood disaster zoning using the method of the present invention are as Figure 2 shown: Figure 2 (a) is the clustering result obtained by fusing attribute and structural features with small watersheds as the basic clustering units; Figure 2 (b) is the final mountain flood disaster zoning plan obtained by merging small watersheds of the same type.

[0065] The above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, and all should be covered within the protection scope of the present invention.

Claims

1. A method for flash flood disaster zoning that integrates attributes and structures, characterized in that It includes the following steps: Step 1: Preprocess the rainfall and terrain factors related to mountain flood disasters in raster form, and spatially aggregate the extracted rainfall and terrain factors into each small watershed in polygon format as small watershed attributes, where the small watershed attributes include the rainfall and terrain attributes of the small watershed; Step 2: According to the boundary data of the small watershed in polygon format, use the center of the small watershed to represent the small watershed and serve as the node of the graph, and construct the edges between the nodes for adjacent small watersheds, thereby constructing a graph structure to represent the small watershed spatial structure; Step 3: Using the small watershed attributes in Step 1 and the small watershed spatial structure in Step 2 as inputs, utilize three graph convolutional neural network modules to construct a graph autoencoder to simultaneously perform feature expression learning on the small watershed attributes and the small watershed spatial structure. Each graph convolutional neural network module includes a graph convolutional neural network layer, a batch normalization layer, and an activation layer; Step 4: According to the output of each graph convolutional neural network module in Step 3, establish a graph decoder to simultaneously achieve the reconstruction of the small watershed spatial structure and the clustering of small watersheds, obtain the optimal clustering result of the small watershed as the initial clustering result of the small watershed unit; Step 5: According to the optimal clustering result of the small watershed in Step 4, merge the small watersheds of the same category to obtain the final mountain flood disaster zoning plan.

2. The method for mountain flood disaster zoning integrating attributes and structures according to claim 1, characterized in that: In Step 2, the expression of the graph structure data obtained by constructing the graph structure is: G=(V,E,X) Among them, V = {v1,..., v n} represents the set of n nodes in the graph, that is, the central set of the small watershed; E = {e ij} represents the set of edges between nodes; the structure of G is represented by the adjacency matrix A of the nodes; A ij = 1 indicates that there is an edge between node i and node j, otherwise A ij = 0; X = {x1,..., x n} represents the attributes of n nodes in the graph, that is, the rainfall and terrain attributes of the small watershed.

3. A method for flood disaster zoning integrating attributes and structures according to claim 2, characterized in that: In Step 3, the autoencoder uses three graph convolutional neural network modules to implement feature learning and expression for the input small watershed attributes and the small watershed spatial structure, as shown in the following formula: Among them, Z l+1 is the feature Z of the l-th layer l obtained by calculating through the graph convolutional neural network module as the feature of the (l + 1)-th layer; is the sum of the adjacency matrix A and the identity matrix I; is the degree matrix, where W l is the parameter to be learned in the l-th layer of the graph neural network; σ is the activation function of the graph neural network; when l = 0, then Z 0 = X, which is the input small watershed attribute.

4. A method for flood disaster zoning integrating attributes and structures according to claim 2, characterized in that: In Step 4, the graph decoder uses the following formula to reconstruct the small watershed spatial structure by using feature inner multiplication and activation function, and the formula is: where Z is the output of the graph autoencoder; is the reconstructed adjacency matrix of the small watershed spatial structure to approximate the original adjacency matrix A of the small watershed.

5. A method for flood disaster zoning integrating attributes and structures according to claim 4, characterized in that: When the graph decoder performs graph clustering, it constructs the feature z of node i i and the clustering center μ u The similarity index q iu between them is calculated through the graph clustering Q distribution function of the following formula, and the formula is: For the similar index q iu Construct a target value, that is, bring the nodes of the same category closer and keep the nodes of different categories farther away to enhance the difference between different categories. Use q iu The square of to construct a P distribution function as the target distribution, making q iu As close as possible to the cluster center to achieve cluster-oriented graph feature supervised learning, as shown in the following formula: The graph clustering Q distribution function constructed by the graph decoder should be as close as possible to the P distribution function, and the KL function value should be made as small as possible to achieve this, as shown in the following formula: Among them, KL is a divergence function used to describe the difference between the P and Q distribution functions.

Images