Flood classification method based on graph attention neural network

Abstract

The invention discloses a flood classification method based on a graph attention neural network. The method comprises the following steps: constructing a flood event data set; constructing a flood map set; based on a graph attention neural network GAT, graph-level features of a flood map corresponding to each flood event are extracted; and carrying out flood classification by using a K-means clustering method based on the graph-level features of the flood graph set. According to the method, the graph neural network is introduced into the field of flood classification, deep learning modeling of drainage basin flood space correlation features is realized, multi-site synchronous monitoring flow data in a drainage basin is integrated by using the graph attention network, space-time evolution features of flood in a drainage basin river network are extracted, and the limitation of a traditional single-site analysis method is broken through. According to the method provided by the invention, the spatial heterogeneity of flood is relatively considered, and the classification result contains richer information.

Description

Technical Field

[0001] The present invention belongs to the technical field of flood classification methods, and particularly relates to a flood classification method based on a graph attention neural network. Background Art

[0002] The formation and development of flood disasters are the results of the non-linear coupling of multiple factors such as meteorological conditions, hydrological processes, topographic features, and human activities. Conducting systematic classification research on historical flood processes and dividing historical flood events into several types with similar characteristics is an important technical approach to deeply revealing the essence of flood disasters, and is helpful for achieving precise flood disaster prevention and control and sustainable utilization of water resources.

[0003] However, existing flood classification methods mainly rely on single-point measured flood characteristics (such as peak flow, flood volume, etc.) at the control stations at the outlets of river basins. This simplified processing method is difficult to effectively reflect the spatial heterogeneity characteristics of floods during the evolution process within the river basin. Against the background of the intensification of global climate change and the frequent occurrence of extreme precipitation events, the complexity and uncertainty of flood disasters have increased significantly. There is an urgent need to develop a flood intelligent classification method that integrates artificial intelligence technology, fully considering the impact of the spatial differences in the underlying surface conditions within the river basin on the flood formation process, so as to improve the prediction accuracy and response ability to flood disasters. [[ID=I2]] Summary of the Invention

[0004] The present invention aims to propose a flood classification method based on the river basin scale. A spatial association matrix is established according to the topological relationship of the water systems of multiple stations within the river basin, and the flood characteristics between different adjacent hydrological stations are aggregated to characterize the spatio-temporal distribution characteristics of the flood process within the river basin.

[0005] The technical solution of the present invention: A flood classification method based on a graph attention neural network, comprising the following steps: Step S1, collect the flow and precipitation data observed by hydrological stations and rain gauges in the study area, extract flood events, and construct a flood event dataset; Step S2, construct a flood map for each flood event, and then obtain a flood map set of multiple flood events; Step S3, based on the graph attention neural network GAT model, extract the graph-level features of the flood map corresponding to each flood event; Step S4, based on the graph-level features of the flood map set, use the elbow method to determine the optimal number of clusters, and classify the flood events by using a clustering method.

[0006] Further, step S1 includes: Step S1a, divide the study area river basin into multiple sub-basins according to hydrological stations, and the control section of each hydrological station or the control sections between two upstream and downstream hydrological stations are used as one sub-basin; Step S1b: Count the rainstorm events occurring in each sub-basin. The criterion for judging a rainstorm event is that the total observed precipitation of a single rain gauge in 24 hours is greater than 50 mm or the total observed precipitation in 12 hours is greater than 30 mm. Step S1c: Analyze the relationship between the rainstorm events in the sub-basin and the flow change process at the hydrological station at the outlet of the sub-basin, and screen out the flood events with obvious rising and falling patterns in the flow process or a peak flood flow greater than 100 m 3 / s to form a flood event dataset.

[0007] The division of sub-basins in Step S1a is based on DEM. After filling the depressions in the DEM, calculate the cumulative flow raster count flowing into each DEM pixel, and select the DEM pixel with the largest cumulative flow raster count near the hydrological station as the outlet of the sub-basin. The upstream cumulative flow raster range is the sub-basin range.

[0008] Furthermore, Step S2 includes: Step S2a: Construct a flood graph with hydrological stations as nodes and the water system topological relationship between hydrological stations as directed edges. Step S2b: The peak flood modulus, flood concentration degree, flood total volume, skewness coefficient, and coefficient of variation of the measured flood events at the hydrological station are the node features. Step S2c: Each flood event corresponds to a flood graph, and multiple flood graphs corresponding to multiple flood events constitute a flood graph set.

[0009] Furthermore, in Step S2b, the calculation formulas for each node feature are as follows: (1) Peak flood modulus M P :

[0010] In the formula, q t is the observed flow process at this node during this flood event, A is the basin area, and max( q t ) is the peak flood flow.

[0011] (2) Flood concentration degree D c :

[0012] In the formula: i' represents the number of time periods, i' = 1, 2,..., n ; n is the total number of time periods of this flood event, r i’ is the i'The total amount of flood in each period is divided equally among each period, taking the interval as a circle; R t is the total amount of flood in this event; D c The value range is [0,1]; the closer it is to 1, the more concentrated the flood is in certain periods; the closer it is to 0, the more uniform the distribution is or the more it diffuses in a ring shape.

[0013] (3) Total flood volume V total :

[0014] Where: t start is the starting time, t end is the termination time, the time interval Δt = t r+1 -t r ; q r It is r Traffic flow at a point in time; q r+1 It is r + Traffic flow at 1 time point; (4) Skewness coefficient C s :

[0015] Where: m is the flow sequence length, is the mean of the flow series, s is the standard deviation of the traffic series.

[0016] (5) Coefficient of variation C v :

[0017] Furthermore, step S3 includes: Step S3a: Collect the characteristic data of all flood events and all nodes, calculate the mean and standard deviation of each characteristic, and then perform Z-score standardization processing on each characteristic. The calculation formula is:

[0018] in, μ is the mean of the feature, is the standard deviation of the feature, x is the original data value, z is the standard score, indicating x Relative to the mean μ degree of deviation.

[0019] Step S3b: Use the Graph Attention Network (GAT) to model the spatial correlation relationship between hydrological stations during the same flood event. Through the attention mechanism, weight-aggregate the features of neighbor nodes to extract the multi-dimensional representation features of each hydrological station; input the features of all nodes into the graph-level attention pooling module to obtain the comprehensive feature representation of the entire flood event graph.

[0020] Step S3c: Process the flood event graph sets of multiple flood events to obtain the graph-level features of each flood graph.

[0021] Further, in Step S3b, in the GAT model, the feature representation of a node is similar to that in a general graph neural network. Both use the embedding method to vectorize the feature representation of the node, and its initial input is also the feature combination of each node. It is expressed by the formula:

[0022] where, h i is the feature of node i , k represents the number of nodes in the graph, F represents the number of feature representations of each node.

[0023] For node i and neighbor node , first calculate the attention score between nodes through the attention mechanism. The calculation formula is as follows:

[0024] where, h j is the feature of adjacent node j , W att is the weight of the attention between node i and adjacent node j , is the learnable attention vector, is the vector concatenation; e ij represents the unnormalized attention coefficient between node i and j ; LeakyReLU is the activation function; Then use the Softmax function to normalize the attention scores of neighbor nodes to obtain the attention coefficient of node j to node i : ; Use the attention coefficient to weight and aggregate the neighbor features to obtain the updated feature representation of the node i : ; Among them, f is the activation function; represents the normalized attention coefficient between nodes i and j; represents the node i updated feature vector; N ( i ) represents the neighbor set of node i , is the attention coefficient of node j to node i ; In order to aggregate the node features into a graph-level representation, use the attention pooling mechanism to obtain the original attention score of each node's updated feature through a linear transformation: ; In the formula, b is a learnable bias term, e i represents the original attention score of the i th node; W pool is the weight matrix of attention pooling; Use the softmax function to normalize the scores of all nodes to obtain the graph-level attention weights: ; Among them, represents the normalized score of the i th node; N represents the number of all nodes participating in the Softmax normalization; e j represents the original attention score of the j th node; The final graph-level feature is the weighted sum of all nodes' updated features: .

[0025] Furthermore, the step S4 includes: Step S4a, determine the optimal number of clusters by the elbow method; Step S4b, based on the optimal number of clusters determined by the elbow method, use the K-means method to classify flood events.

[0026] The principle of the elbow method is to observe the change trend of the clustering error SSE with the increase of the K value and find the "inflection point" (i.e., the "elbow").

[0027] The steps of the K-means clustering method are as follows: (1) Randomly initialize y points as cluster centroids; (2) Assign each point in the sample set to a cluster: Calculate the distance between each point and the centroid (commonly used Euclidean distance and cosine distance), and assign it to the cluster corresponding to the nearest centroid; (3) Update the centroid of the cluster: The centroid of each cluster is updated to the average value of all points in the cluster; (4) Repeatedly iterate steps (2)-(3). If the distance between the new centroid and the previous centroid is less than a certain threshold, it can be considered that the clustering result has reached the desired result, and the algorithm terminates. Otherwise, continue to iterate.

[0028] Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention introduces a graph attention neural network into the field of flood classification, realizes deep learning modeling of the spatial correlation characteristics of basin floods, uses the graph attention network to integrate the synchronous monitoring flow data of multiple stations in the basin, extracts the spatio-temporal evolution characteristics of floods in the basin river network, and breaks through the limitations of traditional single-station analysis methods. The method proposed by the present invention relatively considers the spatial heterogeneity of floods, and the classification results contain richer information. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is a general schematic diagram of the method flow of the present invention; Figure 2 It is a structural schematic diagram of constructing a flood map in the embodiment;

[0030] Figure 3 It is an optimal clustering result diagram of the elbow method of the present invention example. DETAILED DESCRIPTION OF THE INVENTION

[0031] The technical solutions of the present invention will be further described below in conjunction with the drawings and embodiments.

[0032] Refer to Figure 1 , a flood classification method based on a graph attention neural network, specifically including: Step S1. Collect the flow and precipitation data observed by 15 hydrological stations and 135 rain gauges in the Yiluo River Basin, extract flood events, and construct a flood event data set.

[0033] Specifically, it includes the following steps: Step S1a. Divide the Yiluo River Basin into 15 sub-basins according to hydrological stations, and each hydrological station control section or the control sections of two upstream and downstream hydrological stations is a sub-basin.

[0034] Implement sub - basin division based on DEM using Arcgis. First, use Fill in the Hydrology module under the Spatial Analyst Tools toolbox to fill depressions in the DEM. Then use Flow Direction to calculate the flow direction using the D8 method, and use Flow Accumulation to calculate the cumulative inflow raster count upstream of each raster. Determine the DEM pixel with the largest cumulative inflow raster count near the hydrological station and mark it as the tipping point. Then use Watershed to extract the sub - basins. The cumulative flow raster range upstream of the tipping point is the sub - basin range, completing the sub - basin division.

[0035] Step S1b: Statistically analyze the rainstorm events occurring in each sub - basin. The criterion for judging rainstorm events is that the total observed precipitation of a single rain gauge in 24 hours is greater than 50 mm or the total observed precipitation in 12 hours is greater than 30 mm. Step S1c: Analyze the relationship between the rainstorm events in the sub - basin and the flow variation process at the outlet hydrological station of the sub - basin. Screen out flood events where the flow process shows obvious rising and falling patterns or the peak flow is greater than 100 m 3 / s. Determine the starting point and recession point of this flood event and the flood flow process data, forming a flood event dataset, that is, a flood flow dataset.

[0036] In this embodiment, 16 rainstorm flood events are selected, and the time range is from 2020 to 2024.

[0037] Step S2: Construct the flood map data structure, and then obtain a flood map set of multiple flood events.

[0038] Specifically, it includes the following steps: Step S2a: Use hydrological stations as nodes and the water system topological relationship between hydrological stations as directed edges to form a flood map. Each flood event corresponds to a flood map, so 16 flood maps are constructed. The structure of each flood map is the same (node - edge relationship), and the dimension of the node features is also the same. There are five feature dimensions in the present invention. The difference between different flood maps lies in that the five - dimensional feature values of the node features are different. Taking the peak discharge modulus as an example, the peak discharge modulus of the same node in the flood maps corresponding to different flood events is different. The specific structure of the flood map is as Figure 2 shown, and the numbers in the figure represent node numbers.

[0039] Step S2b: Calculate the node characteristics of the hydrological station, including five dimensions: the peak discharge modulus, flood concentration degree, total flood volume, skewness coefficient, and coefficient of variation of the measured flow process at the node. In the same flood event, if a flood is not observed at an individual hydrological station, the five-dimensional characteristic values of its corresponding node are all taken as 0. The formula for calculating the node characteristics is the same for each node. However, since the observed flow processes at different nodes are different, the calculation results of their node characteristics are also different. As long as a flood is observed at any one hydrological station, it is considered that a flood event has occurred in this basin. Other hydrological stations may or may not observe the flood. For the hydrological stations that observe the flood, calculate the corresponding five characteristics as the node characteristics of the corresponding hydrological station; for those that do not observe the flood, all five node characteristics are taken as 0.

[0040] The calculation formula for the node characteristics is as follows: (1) Peak discharge modulus M P :

[0041] In the formula, t is time, q t is the flow process observed at this node during this flood event, A is the basin area, and max( q t ) is the peak discharge.

[0042] The flow process is the sequence of flow changes over time, that is, the flow curve measured during the flood process.

[0043] (2) Flood concentration degree D c : ; In the formula: i' represents the number of time periods, i' = 1, 2,..., n ; n is the total number of time periods of this flood event, r i’ is the total flood volume in the i' th time period. The interval is regarded as a circle, and each time period is evenly distributed; R t is the total flood volume of this flood event; D c The value range is [0, 1]; the closer it is to 1, the more concentrated the flood is in certain time periods; the closer it is to 0, the more evenly distributed or ring-shaped diffusion it shows.

[0044] (3) Total flood volume V total :

[0045] Wherein: t start is the start time, t end is the end time, and the time interval Δt = t r+1 -t r ; q r is the flow rate at the r th time point; q r+1 is the flow rate at the r +1th time point; (4) Skewness coefficient C s :

[0046] Wherein: m is the length of the flow rate sequence, is the mean value of the flow rate sequence, s is the standard deviation of the flow rate sequence.

[0047] (5) Coefficient of variation C v : .

[0048] Step S2c: Each flood event corresponds to a flood map, and multiple flood maps corresponding to multiple flood events constitute a flood map set.

[0049] In this embodiment, taking the flood event with flood number 20200731 as an example, the calculation results of the five-dimensional characteristic values of its different nodes (15 hydrological stations, corresponding node numbers are 0 to 14) are shown in Table 1 below: Table 1 Calculation results of node characteristic values of flood event 20200731

[0050] Step S3: Based on the graph attention neural network GAT considering the attention mechanism, extract the graph-level features of the flood map corresponding to each flood event.

[0051] Specifically, it includes the following steps: Step S3a: Perform Z-score standardization processing for each feature dimension of the node respectively: extract the set of characteristic values of all flood events and all nodes in this dimension; secondly, independently calculate its mean value and standard deviation based on this set of characteristic values, and the calculation formula for Z-score standardization processing is: ; Among them, is the mean of the features, is the standard deviation of the features; x is the original data value, and z is the standard score, indicating x the degree of deviation relative to the mean μ . Each node has five-dimensional features, and each dimension corresponds to a flood feature index value. Then, the feature index value is standardized and input into the GAT model.

[0052] In this embodiment, the five-dimensional feature values of the standardized nodes are shown in Table 2 below: Table 2 Results of Standardized Node Feature Values for Flood Event 20200731

[0053] Step S3b: Use the Graph Attention Network (GAT) to model the spatial correlation relationship between hydrological stations under the same flood event. Through the attention mechanism, weight aggregation is performed on the features of neighbor nodes to extract the multi-dimensional representation features of each hydrological station; the features of all nodes are input into the graph-level attention pooling module to obtain the comprehensive feature representation of the entire flood event graph.

[0054] In this embodiment, the output dimension of the hidden layer is 64 dimensions, and other dimensions can also be set, such as 32 dimensions, etc. Then, the graph-level attention pooling module is used to process the 64-dimensional feature values of the hidden layer of all nodes to obtain the comprehensive feature of the entire flood event graph.

[0055] In the same flood event, one hydrological station corresponds to one node, and the measured flood process characteristics of the hydrological station are used as the features of the node. The GAT model uses attention weights to weight-aggregate neighbor features, outputs the hidden layer features of each node, and then uses the attention pooling method to process the 64-dimensional feature values of the hidden layer of all nodes to obtain the comprehensive feature of the entire flood event graph. This weighted process is the process of extracting the spatial relationship between each node and its adjacent nodes.

[0056] In the GAT model, the feature representation of a node is: ; Among them, h i is the feature of node i , k represents the number of nodes in the graph, F represents the number of feature representations of each node; For node i and neighbor node First, calculate the attention score between nodes through the attention mechanism. The calculation formula is as follows: ; Among them, h j is the adjacent node j feature, W att is the weight of the attention between node i and the adjacent node j ; is the learnable attention vector, is vector concatenation; e ij represents the unnormalized attention coefficient between node i and j ; LeakyReLU is the activation function; Then use the Softmax function to normalize the attention scores of neighbor nodes to obtain the attention coefficient of node j to node i : ; Use the attention coefficient to weight and aggregate the neighbor features to obtain the updated feature representation of node i : ; Among them, f is the activation function; represents the normalized attention coefficient between nodes i and j; represents the updated feature vector of node i ; N ( i ) represents the neighbor set of node i , is the attention coefficient of node j to node i . [[ID=...]]

[0057] To aggregate node features into a graph-level representation, use the attention pooling mechanism to obtain the original attention score of each node's updated feature through a linear transformation: ; In the formula, b is a learnable bias term, e i represents the original attention score of the i th node; W pool is the weight matrix of attention pooling.

[0058] Use the softmax function to normalize the scores of all nodes to obtain the graph-level attention weight: ; Among them, represents the score after normalization of the i -th node; N represents the number of all nodes participating in Softmax normalization; e j represents the original attention score of the j -th node.

[0059] The final graph-level feature is the weighted sum of the updated features of all nodes: .

[0060] Step S 3c: Process the flood atlas of multiple flood events to obtain the graph-level features of each flood map. In this embodiment, 16 flood map-level features are obtained.

[0061] The above calculations are implemented using the PyTorch Geometric module in Python. This module provides efficient graph data structures and GNN model implementations (such as GCN, GAT, GraphSAGE), and supports sparse matrix operations, heterogeneous graphs, dynamic graphs, etc. By calling the relevant functions of the PyTorch Geometric library package, the main parameters modified in the program are: (1) the input settings of the GAT model, the number of input dimensions (5 dimensions in this embodiment); (2) the structure settings of the GAT model: the number of graph convolutional layers (2 in this embodiment), the number of attention heads in each graph convolutional layer (4 in this embodiment); (3) the output settings of the GAT model: the output dimension of the hidden layer (64 dimensions in this embodiment).

[0062] Step S 4: Based on the graph-level features of the flood atlas, use the elbow method to determine the optimal number of clusters, and use the K-means clustering method to classify flood events.

[0063] Specifically, it is decomposed into the following steps: Step S 4a: Determine the optimal number of clusters through the elbow method; Step S 4b: Based on the optimal number of clusters determined by the elbow method, use the K-means method to classify flood events.

[0064] The principle of the elbow method in Step S 4a is to observe the change trend of the clustering error SSE with the increase of the K value and find the "inflection point" (i.e., the "elbow"); both the elbow method in Step S 4a and the K-means clustering method in Step S 4b belong to the prior art, and the principles and calculation processes will not be explained in detail here.

[0065] The above calculations use the matplotlib module in Python to visualize the clustering errors corresponding to different K values, determine the optimal number of classes for flood classification, and then use the KMeans function in the scikit-learn module to output the flood classification results.

[0066] The optimal number of flood clusters is determined to be 4 through the elbow method, as shown in Figure 3 . The classification results of 16 flood events are shown in Table 3.

[0067] Table 3 Flood classification results based on the graph attention neural network

[0068] So far, the 16 flood events have been classified into 4 categories (0, 1, 2, 3). The flood classification itself does not directly assign clear physical meanings to each type. However, by analyzing the flood characteristics corresponding to each category (such as spatial distribution, evolution speed, peak time, etc.), the formation mechanisms of different flood types can be summarized, thus deepening the understanding of the essence of basin flood disasters. The present invention can assist relevant departments in formulating more targeted flood control measures and water resource management strategies, improving flood warning and emergency response capabilities, and reducing the impact of flood disasters.

Claims

1. A flood classification method based on graph attention neural network, characterized in that, The following steps are involved: Step S1: Collect the flow and precipitation data observed by the hydrological stations and rain gauges in the study area, extract flood events, and construct a flood event dataset; Step S2: constructing a flood map for each flood event, thereby obtaining a flood atlas for multiple flood events; Step S3: Extracting graph-level features of the flood map corresponding to each flood event based on the graph attention neural network (GAT) model; Step S4: Based on the graph-level features of the flood atlas, the optimal number of clusters is determined using the elbow rule, and the flood events are classified using the clustering method.

2. The flood classification method based on graph attention neural network according to claim 1, wherein Step S1 includes: Step S1a, dividing the study area basin into multiple sub-basins according to the hydrological station, with each hydrological station control interval or the control intervals of two upstream and downstream hydrological stations as one sub-basin; Step S1b: Count the rainstorm events in each sub-basin. The criteria for judging a rainstorm event are that the total precipitation observed in 24 hours at a single rain gauge station is greater than 50 mm or the total precipitation observed in 12 hours is greater than 30 mm. Step S1c: Analyze the relationship between the rainstorm events in the sub - basin and the flow variation process at the hydrological station at the sub - basin outlet, and screen out the flood events where the flow process shows obvious rising and falling patterns or the peak flood flow is greater than 100 m 3 / s to form a flood event data set.

3. A flood classification method based on a graph attention neural network according to claim 1, characterized in that In step S2, the flood map construction process includes: Step S2a: construct a flood graph with hydrological stations as nodes and the water system topology relationships between hydrological stations as directed edges; Step S2b: The total flood volume, flood peak modulus, flood concentration, skewness coefficient and coefficient of variation of the flood events measured at the hydrological station are used as node features; Step S2c: Each flood event corresponds to a flood map, and multiple flood maps corresponding to multiple flood events constitute a flood map set.

4. A flood classification method based on a graph attention neural network according to claim 3, characterized in that In step S2b, the calculation formula for each node feature is as follows: (1)Peak runoff modulus M P : ； In the formula, q t is the flow process observed at this node during this flood event, A is the basin area, and max( q t ) is the peak flow rate; (2)Flood concentration D c : ； In the formula: i' represents the number of time periods, i' = 1, 2,..., n ; n is the total number of time periods of the flood event, r i’ is the total flood volume of the i' th time period. The interval is regarded as a circle and evenly distributed for each time period; R t is the total flood volume of the flood event; D c The value range is [0, 1]; The closer it is to 1, the more concentrated the floods are in certain periods of time; The closer it is to 0, the more uniform the distribution is or the more ring-like diffusion it exhibits; (3) Total flood volume V total : ； Wherein: t start is the start time, t end is the end time, and the time interval Δt = t i+1 - t i ; q r is the flow rate at the r th time point; q r+1 is the flow rate at the r +1 th time point; (4) Skewness coefficient C s : ； Wherein: m is the length of the flow rate sequence, is the mean value of the flow rate sequence, s is the standard deviation of the flow rate sequence; (5) Coefficient of variation C v : 。 5. The flood classification method based on graph attention neural network according to claim 3, characterized in that Step S3 includes: Step S3a: Collect the characteristic data of all flood events and all nodes, calculate the mean and standard deviation of each characteristic, and then perform Z-score standardization processing on each characteristic. The calculation formula is: ； wherein, is the mean of the features, is the standard deviation of the features; x is the original data value, and z is the standard score; Step S3b: Use the graph attention neural network (GAT) to model the spatial correlation between hydrological stations under the same flood event. Use the attention mechanism to perform weighted aggregation on neighbor node features to extract multi-dimensional representation features of each hydrological station. Input the features of all nodes into the graph-level attention pooling module to obtain a comprehensive feature representation of the entire graph of the flood event. Step S3c: Process the flood atlas of multiple flood events to obtain the graph-level features of each flood graph.

6. The flood classification method based on graph attention neural network according to claim 5, characterized in that Step S3b is specifically as follows: In the GAT model, the node features are represented as: ； Among them, h i is a feature of the node i and represents k the number of nodes in the graph, F and represents the number of feature representations of each node; For nodes i and neighbor nodes First, calculate the attention scores between nodes through the attention mechanism. The calculation formula is as follows: ; Among them, h j is an adjacent node j feature, W att is the weight of the attention between the node i and the adjacent node j ; is a learnable attention vector, is vector concatenation; e ij represents the unnormalized attention coefficient between the nodes i and j ; LeakyReLU is an activation function; Then, the Softmax function is used to normalize the attention scores of neighbor nodes to obtain the attention coefficients of node j for node i : ； Using the attention coefficient to weight and aggregate the neighbor features to obtain the updated feature representation of the node i : ； Among them, f is the activation function; represents the normalized attention coefficient between nodes i and j; represents the node i updated feature vector; N ( i ) represents the neighbor set of node i , is the attention coefficient of node j to node i ; In order to aggregate node features into graph-level representations, an attention pooling mechanism is used to obtain the original attention score of each node’s updated features through a linear transformation: ； where b is a learnable bias term, e i denotes the original attention score of the i -th node; W pool is the weight matrix of attention pooling; Use the softmax function to normalize the scores of all nodes to obtain the graph-level attention weights: ； Among them, represents the score after normalization of the i -th node; N represents the number of all nodes participating in Softmax normalization; e j represents the original attention score of the j -th node; The final graph-level feature is the weighted sum of all node update features: 。 7. A flood classification method based on a graph attention neural network according to claim 1, characterized in that, The step S4 comprises: Step S4a: Based on the graph-level features of the flood atlas, the optimal number of clusters is determined by the elbow rule; Step S4b: Based on the optimal number of clusters determined by the elbow rule, the K-means method is used to classify flood events.

Images