Quantitative analysis method of the combined impact of flash flood spawning background data on flash flood zoning

Abstract

A quantitative analysis method for the combined impact of mountain flood gestation background data on mountain flood zoning includes: collecting mountain flood gestation background data of a target area, obtaining a mountain flood zoning result by applying a mountain flood zoning model based on unsupervised machine learning, classifying the mountain flood gestation background data according to its physical meaning, constructing mountain flood gestation background composite indicators respectively for each classification by applying principal component analysis, training a supervised machine learning model through the mountain flood zoning result and the composite indicators to obtain a supervised support model; inputting the supervised support model and all mountain flood gestation background composite indicators into an interpretation model, and quantifying the contribution of each mountain flood gestation background composite indicator in the mountain flood zoning process through the marginal contribution of each mountain flood gestation background composite indicator to the prediction result of the supervised support model, so as to obtain the impact of each mountain flood gestation background composite indicator on the zoning result. The present invention realizes the transparency and quantitative interpretation of the mountain flood zoning result.

Description

Technical Field

[0001] The present invention belongs to the fields of mountain intelligent disaster reduction, disaster risk management, soil and water conservation, ecological protection, geographic information remote sensing, artificial intelligence application, etc. Specifically, it relates to a quantitative analysis method for the combined influence of background data for mountain flood gestation on mountain flood zoning, providing efficient and intelligent technical support for mountain disaster prevention and reduction work. Background Art

[0002] Mountain floods are short-term extreme runoff processes in small mountainous watersheds triggered by extreme rainfall, dammed lake breaches, snow and ice melt, etc. They are characterized by strong suddenness and great destructiveness, which can cause significant casualties and threaten the high-quality development of mountainous social economy. The gestation background of mountain floods is complex, including factors such as topographical and geomorphic conditions, climate and meteorological conditions, underlying surface conditions, etc. However, the spatial differentiation of mountain flood gestation background factors is obvious and they interact with each other. The lack of a quantification method for the influence of mountain flood gestation background factors has hindered the understanding of the formation mechanism of mountain flood gestation and limited the research and development of scientific and efficient disaster prevention and mitigation measures.

[0003] As a key method for quantitatively analyzing the spatial heterogeneity of mountain flood gestation background, mountain flood zoning can reveal the characteristic similarities and differences of mountain flood gestation background factors in space by dividing homogeneous regions, thus having the potential to quantitatively analyze the influence of mountain flood gestation background factors. The background data for mountain flood gestation are generally geographical data with significant spatial correlation. Therefore, a mountain flood zoning method considering the spatial structure of input data can more comprehensively reflect the spatial interaction of input data, and the zoning results are more reliable. In recent years, Graph Neural Networks (GNNs) have shown unique advantages in processing spatially correlated data, being able to fully utilize the attribute features and spatial topological structure of input data, and being particularly suitable for the analysis of geographical data with complex spatial relationships. Graph neural networks can capture the complex interactions of mountain flood gestation background factors in space, such as the upstream and downstream associations between terrain undulations and river fluxes, the diffusion and local effects of meteorological factors, etc., thus generating more reliable zoning results. In addition, the efficiency and robustness of graph neural networks in processing large-scale high-dimensional data make them have great application potential in the field of mountain flood zoning.

[0004] However, the "black box" characteristic of graph neural networks poses challenges to the interpretability of the mountain flood zoning results. This limits the in-depth understanding of the complex non-linear relationship between the input breeding background data captured by advanced neural networks and mountain flood zoning, and is not conducive to quantitatively describing the spatial distribution characteristics and action mechanisms of the influence of various mountain flood breeding background factors on the zoning results. In addition, the amount of mountain flood breeding background data is large, and they influence each other, and their influence on mountain flood zoning also shows a complex non-linear coupling effect. Therefore, it is difficult to directly linearly add the single-factor influence of the mountain flood zoning results to quantify the combined influence. The combined influence of multiple breeding environment background factors with highly correlated physical processes has higher application value for mountain flood risk management decisions than single-factor influence, such as the combined influence of comprehensive underlying surface conditions such as vegetation cover, soil properties, and land use. However, there is currently no systematic quantitative analysis method for the combined influence of mountain flood breeding environment background factors, which severely restricts the scientific guiding role that the zoning results should play in mountain flood disaster risk management and prevention and control. Therefore, there is an urgent need to develop a quantitative analysis method for the combined influence of mountain flood breeding background data on mountain flood zoning. Summary of the Invention

[0005] The present invention aims to solve at least one of the technical problems existing in the prior art.

[0006] To this end, the present invention proposes a quantitative analysis method for the combined influence of mountain flood breeding background data on mountain flood zoning. Based on the mountain flood breeding background data such as the terrain, climate, hydrology, and underlying surface of the target area, a graph neural network is applied to carry out mountain flood zoning, and the SHAP (Shapley Additive Explanations) explanation model is applied to quantify the combined influence of mountain flood breeding background factors on the zoning results, so as to realize the transparency and interpretability of the mountain flood zoning results.

[0007] The present invention is realized through the following technical solutions:

[0008] A quantitative analysis method for the combined influence of mountain flood breeding background data on mountain flood zoning provided by the present invention includes the following steps:

[0009] Step S100: Collect the mountain flood breeding background data of the target area, preprocess it, and then use a mountain flood zoning model based on unsupervised machine learning to obtain the mountain flood zoning results of the target area. Classify the mountain flood breeding background data according to the physical meaning of the mountain flood breeding background data. For each classification, use principal component analysis to respectively construct mountain flood breeding background composite indicators. Train a supervised machine learning model through the mountain flood zoning results and the mountain flood breeding background composite indicators to learn the zoning results of the mountain flood zoning model, and obtain a supervised support model;

[0010] Step S200: Input the supervised support model and all the composite indexes of the background for flash flood breeding in the target area into the interpretation model. The interpretation model provides the contributions of each composite index of the background for flash flood breeding in the process of flash flood zoning based on the marginal contributions of each composite index of the background for flash flood breeding to the prediction results of the supervised support model, and obtains the influence of each composite index of the background for flash flood breeding on the flash flood zoning result.

[0011] In some embodiments, in step S100, the flash flood zoning result of the target area is obtained by using an unsupervised graph neural network. The specific steps include:

[0012] Step S110: Resample all the background data of flash flood breeding in the target area collected into grid data with a unified spatial resolution, and crop the grid data according to the scope of the target area to obtain the initial data R0.

[0013] Step S120: Conduct correlation analysis and multicollinearity analysis on each dimension of data in the initial data R0, identify and remove the data dimensions whose correlation and variance inflation factor do not meet the requirements, and obtain the input data R in grid form.

[0014] Step S130: For the input data R in grid form, regard each grid as a node, construct a node matrix V, determine the edges between nodes according to the distance between nodes, construct an edge set E and a normalized adjacency matrix A, and construct a node attribute matrix X based on the input data R, so as to construct the input data R in grid form into graph structure data G=(V, E, X).

[0015] Step S140: For the case where the number of partitions has been set in the target area, based on the set number of partitions, use the graph structure data G=(V, E, X) as the input, and perform clustering analysis through a graph neural network to obtain the zoning result at the grid scale; for the case where the number of partitions has not been set in the target area, for all the numbers of partitions, use the graph structure data G=(V, E, X) as the input, perform traversal calculation using a graph neural network, and through an optimization method, with the clustering validity index as the optimization target, search and determine the optimal number of partitions, so as to determine the zoning result at the grid scale.

[0016] In some embodiments, in step S130, calculate the distance between nodes according to the longitude and latitude attributes of the nodes in the input data R in grid form, and establish the edges between nodes according to the number of nearest neighbor nodes and the upper limit of the neighbor distance, following the eight-neighborhood or four-neighborhood rule, to obtain the edge set Used to represent a set of edges connecting different nodes, (i, j) represents an edge connecting node i and node j, where i≠j.

[0017] In some embodiments, in step S130, the construction steps of the normalized adjacency matrix A include:

[0018] Find the neighbor nodes closest to node i through the KD tree:

[0019] neighbors i = KDTree.query(C i , num_neighbors + 1, distance_upper_bound)

[0020] where neighbors i represents all neighbor nodes of node i; KDTree.query(·) represents the KD tree calculation function; each node i has two attributes, longitude log i and latitude lat i All nodes form a two-dimensional coordinate set C containing the longitude and latitude attributes of all nodes C i =(log i , lat i ) is the longitude and latitude attribute of node i, N is the number of grids contained in each dimension data in the input data R; num_neighbors is the number of neighbor nodes to be found in the neighborhood for each node; distance_upper_bound represents the distance upper limit for finding neighbor nodes;

[0021] Remove the self-node i from neighbors i to obtain the neighbor node set neighbors i ';

[0022] For node i and node j, if node i is adjacent to node j, that is, j ∈ neighbors i ', then a ij = 1, otherwise a ij = 0, thus constructing the adjacency matrix A0 = {a ij |(i, j) ∈ V 2} ∈ R N×N ;

[0023] Perform symmetric normalization on the adjacency matrix A0 to obtain the normalized adjacency matrix A:

[0024] A = D -1 / 2 A0D -1 / 2 ∈ R N×N

[0025] where D = {d' i |i ∈ V} is the degree matrix, which is a diagonal matrix, and its diagonal elements are the degrees d' of each nodei = ∑ j a ij , where \(i\in V\) and \(j\in V\).

[0026] In some embodiments, in step S140, for the case where the number of partitions has been set for the target area, the number of partitions is set to \(K\). Based on this number of partitions, taking the graph structure data \(G=(V, E, X)\) as the input, clustering analysis is performed through a graph neural network to obtain a partition result at the grid scale. The specific steps include:

[0027] Step S1411: For the original graph structure data \(G = \{X, A\}\) represented by the node attribute matrix \(X\in R N×d and the normalized adjacency matrix \(A\in R N×N , where \(N\) is the number of grids contained in each dimension data of the input data \(R\), enhanced data is generated through data augmentation operations Generate enhanced data

[0028]

[0029] Step S1412: Use the graph neural network to calculate the node representations for the original graph structure data \(G\) and the enhanced data respectively. The graph neural network updates the node representations through multiple layers of graph convolutions:

[0030] \(H (l+1) =\sigma(AH (l) W (l) )

[0031] where \(H (l) is the node representation of the \(l\)-th layer in the graph neural network; \(W (l) is the weight matrix of the \(l\)-th layer in the graph neural network; \(\sigma(\cdot)\) is the activation function;

[0032] After multiple layers of graph convolutions, the node representations \(H\) of the original graph structure data \(G\) and the enhanced data are obtained.

[0033]

[0034] where is the graph neural network function;

[0035] is the contrastive learning between the attribute vector \(x i of the constrained node \(i\) and its enhanced attribute vector . The contrastive loss function adopted is:

[0036]

[0037]

[0038] Among them, h i and are respectively the node representations of the original graph structure data and the enhanced data of node i; Sim(·, ·) represents the similarity between nodes;

[0039] Step S1413: For the node representations H = {h1, h2,..., h N} learned by the graph neural network, use the K-means method to initialize the set of clustering centers C, and take the clustering centers as the learnable parameters of the graph neural network. By calculating the distance metric assign each node representation h i to the nearest clustering center c k ; optimize the total clustering loss of the graph neural network through backpropagation and gradient descent In each iteration, update the set of clustering centers C and the classification of nodes to obtain the final set of clustering centers and the clustering label representations of each node;

[0040] The total clustering loss has the following expression:

[0041]

[0042]

[0043]

[0044] Among them, is the clustering expansion loss, is the clustering contraction loss, and τ is a hyperparameter used to control the balance between expansion and contraction; c A ∈ C and c B ∈ C are respectively the clustering centers of cluster A and cluster B, are all the nodes in cluster A;

[0045] In each iteration, for node update the clustering center c A according to the following formula:

[0046]

[0047] Among them, |C A | is the number of nodes in cluster A;

[0048] Step S1414: Visualize the clustering labels of each node in space according to the longitude and latitude of the node to obtain the flash flood zoning result at the grid scale.

[0049] In some embodiments, in step S140, in the case where the number of partitions is not set for the target area, for all numbers of partitions, using the graph structure data G = (V, E, X) as the input, traverse and calculate using a graph neural network. Through an optimization method, with the clustering validity index as the optimization objective, search and determine the optimal number of partitions, thereby determining the partition result at the grid scale. The specific steps include:

[0050] Step S1421, set as the initial number of clusters, where δ and ε are proportionality coefficients; based on the set number of clusters perform clustering analysis according to steps S1411 to S1414, calculate the corresponding clustering validity index based on the clustering result, and set the objective function as the weighted sum of the clustering validity indices:

[0051]

[0052] where, is the current number of clusters, that is, the current solution; α and β are weight coefficients; and are the validity indices calculated based on the clustering result of the graph neural network, is the sum of the average coefficient of variation of the selected background data for flash flood gestation, used to quantify the dispersion degree of the selected background data for flash flood gestation, evaluate the tightness and dispersion between clustering results through the distance between cluster centers, and the lower the values of and, the better the clustering effect. The calculation formulas are as follows:

[0053]

[0054]

[0055]

[0056] where, |C w | is the number of nodes in the clustering category w, d is the dimension of the background data for flash flood gestation, is the coefficient of variation of the r-th dimension of the gestation background data in the clustering category w; is the standard deviation of the r-th dimension of the gestation background data in the clustering category w, is the mean of the r-th dimension of the gestation background data in the clustering category w; is the average distance from all nodes in the clustering category w to the cluster center, is the average distance from all nodes in the clustering category t to the cluster center, d wtis the distance between the centroid of clustering category w and the centroid of clustering category t;

[0057] Step S1422, starting from the current solution Generate a new neighborhood solution

[0058]

[0059] where γ takes -1 or 1;

[0060] Step S1423, calculate the current solution and the neighborhood solution of the objective function value and and calculate the difference ΔE between the two:

[0061]

[0062] Decide whether to accept the neighborhood solution according to the following criterion

[0063] If ΔE ≤ 0, accept the neighborhood solution as the new current solution;

[0064] If ΔE > 0, accept the neighborhood solution according to the probability P(ΔE) as the new current solution, and the probability is determined by the parameter T, and the formula is as follows:

[0065]

[0066] where the parameter T gradually decreases after each iteration as the algorithm progresses;

[0067] Step S1424, update the parameter T in an exponentially decaying manner:

[0068] T = α T T

[0069] where α T is a constant, and α T ∈(0, 1);

[0070] Step S1425, continuously repeat steps S1421 to S1424 until the iteration meets the termination condition, the iteration ends, and the optimal number of clusters and the corresponding flash flood zoning results of the grid scale are obtained.

[0071] In some embodiments, after obtaining the flash flood zoning results, it further includes:

[0072] Step S150: Obtain the historical mountain flood event data of the target area, and use this historical mountain flood event data to evaluate the obtained zoning result. Specifically, overlay the historical mountain flood event data of the target area in the zoning result, calculate the number and density of historical mountain flood disaster events in each zoning unit, apply the hot spot analysis method to calculate the z-value of the spatial distribution of historical mountain flood events in each zoning unit, and determine the spatial distribution structure of the significance of historical mountain flood events. If there are significant differences in the number and density of historical mountain flood events in each zoning unit in the zoning result, and there is a significant spatial distribution structure of historical mountain flood events in some zoning units, it indicates that the zoning result corresponds well with the spatial distribution of historical mountain flood events, indicating that the zoning result is accurate. If there are no significant differences in the number and density of historical mountain flood events in each zoning unit in the zoning result, and there are no significantly higher positive values or lower negative values of the z-value in each zoning unit, re-perform clustering analysis considering aspects such as the selection of mountain flood breeding background data, the determination of graph neural network model parameters, and / or the optimization of the number of zoning to obtain the mountain flood zoning result.

[0073] In some embodiments, the supervised support model is obtained according to the following steps:

[0074] Step S160: Construction of the mountain flood breeding background composite index:

[0075] According to the physical meaning of the mountain flood breeding background input data R = {r1, r2,..., r d}, divide the mountain flood breeding background input data R into Q major categories, where 1 < Q < d, and d is the dimension of the mountain flood breeding background data in R. Let be the set of all mountain flood breeding background data in the major category, and its expression is as follows:

[0076]

[0077] where, is the dimension of the mountain flood breeding background data included in the major category ;

[0078] For apply principal component analysis to the mountain flood breeding background data it contains, and construct the mountain flood breeding background composite index of the major category according to the following formula

[0079]

[0080] where, is the th principal component score of ; is the th of The weight of a principal component among all principal components is calculated by the proportion of the percentage of variance of this principal component in the cumulative percentage of variance of the principal components;

[0081] Construct a composite index matrix by taking the composite index of the background of mountain flood breeding for each major category as a column Convert this composite index matrix into a form expressed by rows, that is where N is the number of nodes, Q is the number of composite indexes, and Z′ i ∈R Q×1 is the vector of the composite index of the background of mountain flood breeding for node i, which is composed of the elements in the i-th row of the composite index matrix;

[0082] Step S170, training of a supervised machine learning model:

[0083] Take the clustering label y = {y1, y2,..., y N}∈R N×1 and the corresponding vector of the composite index of the background of mountain flood breeding Z′ i as the target variables, set K as the number of clusters, and train the supervised machine learning model ML():

[0084] y i = ML(Z′ i , θ)

[0085] where y i is the clustering label for node i in the graph structure data G, and this clustering label is used as the supervision signal; θ is the parameter of the supervised machine learning model.

[0086] In some embodiments, in step S200, for each node i, let the contribution value of the interpretation model to the clustering label y i of node i predicted by the -th composite index of the background of mountain flood breeding be

[0087]

[0088] where, is the value of the -th composite index of the background of mountain flood breeding on node i; S is the feature subset, indicating not including the a subset of other composite indicators of the background for flash flood generation; g(S) is the result of predicting node i by the supervised support model trained on the feature subset S; is the case where the feature subset S is added with the prediction result of the supervised support model for node i; represents the expectation for all possible feature subsets S;

[0089] By calculating the contribution values of all nodes, the global influence and local influence of each composite indicator of the background for flash flood generation are obtained, where:

[0090] Let the global influence of the th composite indicator of the background for flash flood generation be representing the average contribution size of the th composite indicator of the background for flash flood generation in the entire target area, and the calculation formula is as follows:

[0091]

[0092] Then the global influence of all composite indicators d is the dimension of the background data for flash flood generation;

[0093] For the local influence, take the contribution value as the local contribution value representing the specific contribution of the th composite indicator of the background for flash flood generation to the zoning result of node i.

[0094] In some embodiments, the quantitative analysis method further includes:

[0095] Step S300, for the global influence of all composite indicators obtained by the interpretation model determine the ranking of the influence of all composite indicators on the zoning result, and determine the influence difference of each composite indicator of the background for flash flood generation; for the local influence, visualize the influence of each node and each composite indicator of the background for flash flood generation in space according to the longitude and latitude of the node, and analyze the spatial variability of the influence of each composite indicator of the background for flash flood generation.

[0096] The beneficial effects of the present invention are as follows:

[0097] 1. The flash flood zoning method based on the fusion of graph neural network and SHAP interpretation model of the present invention uses the macroscopic background data for flash flood generation as input data to quantify the composite influence of the background data for flash flood generation on the zoning result and its spatial variability. This zoning method combines the advantages of the graph neural network model in fully utilizing the spatial structure characteristics of data and the characteristics of the interpretation model in quantitatively interpreting deep network models, and systematically improves the accuracy and interpretability of the flash flood zoning result.

[0098] 2. The graph neural network method applied in the present invention takes into account both the attribute characteristics and spatial structure of the background data for mountain flood gestation during the zoning process. The zoning results fully reflect the complex spatial interactions of the background data for mountain flood gestation, and accurately capture the spatial distribution characteristics of the background data for mountain flood gestation. By using an optimization algorithm to automatically determine the optimal number of zones, it avoids the interference of subjective human factors in traditional zoning methods, ensuring the scientific nature and objectivity of the zoning process.

[0099] 3. The present invention applies the principal component analysis method to construct a composite index for the background of mountain flood gestation, which reflects the non-linear relationships of these indexes while retaining as much information as possible of the multi-dimensional single-factor indexes corresponding to the composite index. Combining with the SHAP interpretation model, it analyzes the zoning results obtained by the graph neural network, quantifies the composite influence and its spatial variability of the background data for mountain flood gestation on the mountain flood zoning results, and proposes a scientific and reasonable method for constructing the composite index and quantifying the composite influence.

[0100] 4. The quantitative analysis method proposed in the present invention is based on an unsupervised method, with high automation and integration levels, strong result interpretability, high calculation efficiency, strong method generalization ability, and is easy to be popularized and used in mountainous areas. BRIEF DESCRIPTION OF THE DRAWINGS

[0101] Figure 1 is the overall flowchart of a quantitative evaluation method for the composite influence of the background data for mountain flood gestation on mountain flood zoning provided by the embodiment of the first aspect of the present invention.

[0102] Figure 2 is a schematic diagram of the method for obtaining the mountain flood zoning results in the quantitative evaluation method provided by the embodiment of the first aspect of the present invention.

[0103] Figure 3 is an example in the present invention taking the Hengduan Mountains as an example, applying Figure 2 the method shown, the clustering validity indexes at the clustering numbers from 2 to 20, where (a) is the DBI index and (b) is the CQI index.

[0104] Figure 4 is an example in the present invention taking the Hengduan Mountains as an example, based on Figure 2 the method shown, (a) the mountain flood zoning result map and the spatial distribution of historical mountain flood events; (b) the number and density of historical mountain flood events in each zoning unit; (c) the spatial statistical z value of historical mountain flood events in each zoning unit.

[0105] Figure 5 is an example in the present invention taking the Hengduan Mountains as an example, the global composite influence of the background data for mountain flood gestation on the mountain flood zoning results obtained based on the SHAP interpretation model.

[0106] Figure 6-1 and Figure 6-2In the present invention, taking the Hengduan Mountains as an example, it shows the spatial variation of the combined influence of the background data of mountain flood gestation on the mountain flood zoning results based on the SHAP interpretation model, (a) meteorology; (b) climate; (c) terrain; (d) vegetation; (e) soil. Detailed implementation manners

[0107] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0108] On the contrary, the present application covers any alternatives, modifications, equivalent methods and solutions made within the essence and scope of the present application as defined by the claims. Further, in order to enable the public to better understand the present application, in the following detailed description of the present application, some specific details are described in detail. Those skilled in the art can fully understand the present application without the description of these details.

[0109] See Figure 1 , a quantitative analysis method for the combined influence of the background data of mountain flood gestation on mountain flood zoning proposed in the first aspect embodiment of the present invention includes the following steps:

[0110] Step S100: Collect the background data of mountain flood gestation in the target area, preprocess it, and then use an unsupervised machine learning mountain flood zoning model to obtain the mountain flood zoning results of the target area. Classify the background data of mountain flood gestation according to its physical meaning. For each classification, use principal component analysis to construct a composite index of mountain flood gestation background respectively. For the mountain flood zoning model, train a supervised machine learning model through the mountain flood zoning results and the composite index of mountain flood gestation background to learn the zoning results of the mountain flood zoning model, and obtain a supervised support model to replace the mountain flood zoning model;

[0111] Step S200: Input the supervised support model and all the composite indexes of the background data of mountain flood gestation in the target area into the interpretation model. The interpretation model provides the contribution (importance) of each composite index of the background data of mountain flood gestation in the mountain flood zoning process through the marginal contribution of each composite index of the background data of mountain flood gestation to the prediction result of the supervised support model, and obtains the influence of each composite index of the background data of mountain flood gestation on the mountain flood zoning result.

[0112] In some embodiments, see Figure 2 , in step S100, the mountain flood zoning model uses an unsupervised graph neural network. The steps of using the mountain flood zoning model to obtain the mountain flood zoning results of the target area include:

[0113] Step S110: Resample all the background data of mountain flood gestation in the target area collected into raster data with a unified spatial resolution, and crop the raster data according to the scope of the target area to obtain the initial data R0;

[0114] Step S120: Conduct correlation analysis and multicollinearity analysis on the data of each dimension in the initial data R0, identify and eliminate the data dimensions whose correlation and variance inflation factor do not meet the requirements, and obtain the input data R in raster form;

[0115] Step S130: For the input data R in raster form, regard each raster therein as a node, construct a node matrix V, determine the edges between nodes according to the distances between nodes, construct an edge set E and a normalized adjacency matrix A, and construct a node attribute matrix X based on the input data R, so as to construct the raster-form input data R into a graph structure data G=(V, E, X);

[0116] Step S140: In the case where the number of partitions has been set for the target area, based on the set number of partitions, using the graph structure data G=(V, E, X) as the input, conduct clustering analysis through a graph neural network to obtain the partition result at the raster scale; in the case where the number of partitions has not been set for the target area, for all numbers of partitions, using the graph structure data G=(V, E, X) as the input, perform traversal calculations using a graph neural network, and through an optimization algorithm, with the clustering validity index as the optimization objective, search and determine the optimal number of partitions, so as to determine the partition result at the raster scale.

[0117] In some embodiments, in step S110, the background data of mountain flood gestation in the target area collected includes data in aspects such as terrain, climate, meteorology, hydrology, and underlying surface. When conditions permit, raster data with a relatively high spatial resolution is mainly used. When raster data is not easily obtained, it is also allowed that some data uses vector data. Subsequently, all the collected data is resampled into raster data with a unified spatial resolution and cropped according to the scope of the target area to obtain the initial data R0={r1, r2,...r m}, where m is the number of data dimensions of R0.

[0118] In some embodiments, in step S120, when conducting correlation analysis and multicollinearity analysis on the data of each dimension in the initial data R0, identify and eliminate the data dimensions with a correlation>0.7 and a variance inflation factor>10, so as to obtain the input data R in raster form={r1, r2,...r d}, where d is the dimension of the background data of mountain flood gestation in R.

[0119] In some embodiments, step S130 specifically includes the following steps:

[0120] Step S131: Assume that each dimension of the input data R has N grid points. Each grid point is regarded as a node in the graph structure, and a node matrix V = {v1,..., v i ,..., v N} is constructed. Each node vector v i corresponds to the d-dimensional data of node i in the input data R;

[0121] Step S132: Construct a KD-tree to find neighbor nodes: To effectively find the neighbor nodes of each node, a KD-tree (K-dimensional tree e ) is used to construct a spatial index structure between nodes. Calculate the distance between nodes according to the longitude and latitude attributes of the nodes. According to the number of nearest neighbor nodes and the upper limit of the neighbor distance, establish edges between nodes according to the rule of octal neighborhood (or quadrilateral neighborhood) to obtain an edge set used to represent a group of edges connecting different nodes. (i, j) represents an edge connecting node i and node j, where i ≠ j. Find the neighbor nodes closest to node i through the KD-tree:

[0122] neighbors i = KDTree.query(C i , num_neighbors + 1, distance_upper_bound)

[0123] where, neighbors i represents all the neighbor nodes of node i; KDTree.query(·) represents the KD-tree calculation function; each node i has two attributes, longitude log i and latitude lat i . All nodes form a two-dimensional coordinate set C containing the longitude and latitude attributes of all nodes C i = (log i , lat i ) is the longitude and latitude attribute of node i; num_neighbors is the number of neighbor nodes to be found in the neighborhood for each node. One more neighbor node is taken to exclude the node itself; distance_upper_bound represents the upper limit of the distance to find neighbor nodes, and only nodes within this distance range are searched. Then, remove the self-node i to obtain the neighbor node set neighbors i ′.

[0124] Step S133: Construction and normalization of the adjacency matrix: neighbors i`neighbors`' is the set of neighbor nodes of node i. For node i and node j, if node i is adjacent to node j, that is, j ∈ `neighbors` i `, then a ij = 1, otherwise a ij = 0. Thus, the adjacency matrix A0 = {a ij | (i, j) ∈ V 2} ∈ R N×N . Normalize the adjacency matrix A0 to enhance the stability of the graph neural network. Usually, the symmetric normalization method is adopted, and the formula is:

[0125] A = D -1 / 2 A0D -1 / 2 ∈ R N×N

[0126] where A is the normalized adjacency matrix; D = {d' i | i ∈ V} is the degree matrix, that is, the degree of each node (i.e., the number of edges connected to the node). The degree matrix is a diagonal matrix, and its diagonal elements are the degrees d' i = ∑ j a ij , i ∈ V, j ∈ V. Through the normalization process, the adjacency matrix is made more suitable for the processing of graph neural networks.

[0127] Step S134, construct the node attribute matrix X: Use the input data R as the attribute vector x i of each node. The attribute vector of a single node is represented as a d-dimensional vector:

[0128] x i = {x i，1 , x i，2 ,..., x i，d}

[0129] where d is the dimension of the attribute, that is, the number of attributes, and x i，1 , x i，2 ,..., x i，d are the elements of the 1st to the dth dimensions in the attribute vector of node i respectively. The node attribute matrix X = {x1, x2,..., x N} ∈ R N×d , is an N×d matrix, and the ith row in it represents the d-dimensional feature vector of node i.

[0130] In some embodiments, in step S140, for the case where the number of partitions has been set for the target area, based on the set number of partitions, using the graph structure data G = (V, E, X) as the input, perform clustering analysis through the graph neural network to obtain the partition result at the grid scale. The specific steps include:

[0131] Step S1411, Data Augmentation: Generate augmented samples through data augmentation methods. Common data augmentation methods include node feature noise, node dropout, adjacency matrix perturbation, etc. For the original graph structure data G = {X, A} represented by the node attribute matrix X ∈ R N×d and the normalized adjacency matrix A ∈ R N×N , generate augmented data through data augmentation which is expressed as augmenting the graph structure data using data augmentation operations:

[0132]

[0133] Step S1412, Contrastive Learning: Use a graph neural network (GNN) to calculate node representations for the original graph structure data G and the augmented data respectively. The graph neural network updates node representations through multiple layers of graph convolutions:

[0134] H (l+1 ) = σ(AH (l) W (l) )

[0135] where H (l) is the node representation of the l-th layer in the graph neural network; A is the normalized adjacency matrix; W (l) is the weight matrix of the l-th layer in the graph neural network; σ(·) is the activation function (such as ReLU).

[0136] After multiple layers of graph convolutions, obtain the node representations H, of the original graph structure data G and the augmented data for subsequent clustering operations, which is expressed as:

[0137]

[0138] where H and are the node representations of the original graph structure data G and the augmented data respectively; is the graph neural network function.

[0139] Setting of the contrastive learning loss function: Contrastive learning trains the graph neural network function by maximizing the similarity of the corresponding node representations between the original graph structure data G and the augmented data , while minimizing the similarity between different nodes. For each node i, the node representation h i of the original graph structure data and the node representation of the augmented data can be expressed as:

[0140]

[0141] where x i is the attribute vector of node i; is the enhanced data attribute vector of node i; A is the normalized adjacency matrix; is the normalized adjacency matrix of the enhanced data.

[0142] For node i, the contrastive learning between its original attribute vector x i and the enhanced attribute vector aims at the contrastive loss function (such as contrastive loss or information maximization loss):

[0143]

[0144] where h i and are the node representations of the original graph structure data and the enhanced data of node i respectively; Sim(·, ·) represents the similarity between nodes (such as cosine similarity); the numerator term of the contrastive loss function encourages higher similarity between the original representation and the enhanced representation of node i, and the denominator term suppresses the similarity between node i and other nodes.

[0145] Step S1413, clustering analysis, the specific process is as follows:

[0146] 1) K-means clustering initialization: For the node representations H = {h1, h2,..., h N} learned by the graph neural network (GNN), apply the K-means method to initialize the set of clustering centers C and assign the clustering centers as learnable neural network parameters. For the target area, set the number of clusters to K, that is, the number of partitions already set, and the initialization includes the following two steps:

[0147] First, randomly select node representations to form the initial set of clustering centers Each randomly selected node representation is a d-dimensional vector;

[0148] After that, by calculating the distance metric (such as Euclidean distance), assign each node representation h i to the nearest clustering center

[0149]

[0150] Take the clustering center C as the learnable parameter of the graph neural network, which means that the clustering center C will be updated as the graph neural network is trained to better fit the distribution of the nodes.

[0151] 2) Graph Neural Network Clustering Loss Setting

[0152] Apply the graph neural network clustering module to optimize the clustering distribution by minimizing the clustering expansion loss and the clustering contraction loss. Specifically, in an adversarial manner, the clustering expansion loss attempts to push different cluster centers apart to disperse the clustering distribution, while the clustering contraction loss aims to compress the clustering distribution by pulling samples back to the cluster centers.

[0153] Total Clustering Loss is the clustering expansion loss and the clustering contraction loss weighted sum, which can be expressed as:

[0154]

[0155] where τ is a hyperparameter used to control the balance between expansion and contraction.

[0156] Clustering Expansion Loss Attempts to push different cluster centers apart so that the nodes within each cluster are as similar as possible, while the nodes between different clusters are as different as possible. Expand the clustering distribution by maximizing the distance between cluster centers. The clustering expansion loss is defined as:

[0157]

[0158] where c A ∈C and c B ∈C are the cluster centers of cluster A and cluster B respectively; ||·|| represents the Euclidean distance.

[0159] Clustering Contraction Loss Aims to make the distance between the nodes within the same cluster as close as possible to the cluster center. Compress the node distribution within the cluster by pulling the samples back to the cluster center. The clustering contraction loss can be defined as:

[0160]

[0161] where are all the nodes in cluster A; node i belongs to cluster A, that is hi is the node representation of node i; c A ∈C is the cluster center of cluster A.

[0162] 3) Iterative Optimization of Clustering Results: Optimize the total clustering loss through backpropagation and gradient descent Update the cluster center C and the classification of nodes in each iteration. Specifically:

[0163] In each round of iteration, the distance from each node to each cluster center is first calculated, and the node is assigned to the cluster center with the closest distance;

[0164] For node i, Update the cluster center c according to the following formula A :

[0165]

[0166] Among them, |C A | is the number of nodes in cluster A; are all nodes in cluster A.

[0167] After multiple rounds of iterations, the cluster center C will be gradually optimized, and finally the most suitable clustering result will be obtained.

[0168] Step S1414: Final clustering results and node classification

[0169] After multiple rounds of iterative optimization, the clustering results finally converged and the final cluster center set was obtained. Where K is the number of clusters, that is, the number of partitions that have been set, and the cluster label of each node is expressed as y = {y1, y2, ..., y N}∈R N×1 , where N is the number of nodes. The cluster label of each node is visualized spatially according to the longitude and latitude of the node to obtain the flash flood zoning result at the grid scale.

[0170] In some embodiments, in step S140, for the case where the number of partitions is not set for the target area, the optimal number of partitions is determined by the following optimization algorithm. The core idea of ​​the optimization algorithm is to start from an initial solution and randomly search for the global optimal solution of the objective function in the solution space in combination with a certain probability jump characteristic, that is, the local optimal solution can be probabilistically jumped out and eventually tend to the global optimal solution. By giving the search process a time-varying and eventually zero probabilistic jump, a poor solution is probabilistically accepted to avoid falling into the local optimum and eventually find the global optimal solution. The specific steps are as follows:

[0171] Step S1421, objective function setting: setting is the initial number of clusters, in δ∈[0, 0.1], ε∈[2, 10], δ and ε are proportional coefficients. Based on the set number of clusters Perform cluster analysis according to steps S1411 to S1414, calculate the corresponding cluster effectiveness index (CQI and DBI) based on the clustering results, and set the objective function is the weighted sum of clustering effectiveness indicators:

[0172]

[0173] wherein, is the current number of clusters (i.e., the current solution); α and β are weight coefficients that control the contributions of the two cluster validity indicators to the objective function.

[0174] and are validity indicators calculated based on the clustering results of the graph neural network. is the sum of the average coefficients of variation of the selected background data for flash flood gestation in the clustering results, quantifying the dispersion degree of these gestation background data. The compactness and dispersion between clustering results are evaluated by the distances between cluster centers. and The lower the values of and , the better the clustering effect.

[0175]

[0176] wherein, N is the total number of nodes, is the currently set number of clusters, |C w | is the number of nodes in cluster category w, d is the dimension of the background data for flash flood gestation, is the coefficient of variation of the r-th dimension of the gestation background data in cluster category w, and the calculation formula is as follows:

[0177]

[0178] wherein, is the standard deviation of the r-th dimension of the gestation background data in cluster category w, is the mean of the r-th dimension of the gestation background data in cluster category w.

[0179]

[0180] wherein, is the currently set number of clusters, is the average distance from all nodes in cluster category w to the cluster center, is the average distance from all nodes in cluster category t to the cluster center, d wt is the distance between the centroid of cluster category w and the centroid of cluster category t.

[0181] Step S1422, neighborhood solution selection: Generate a new neighborhood solution from the current solution

[0182] ​

[0183] Among them, γ is a constant, taking -1 or 1.

[0184] Step S1423, acceptance criterion setting

[0185] Calculate the change of the objective function: Calculate the objective function values of the current solution and the neighborhood solution and and calculate the difference ΔE between them:

[0186]

[0187] Decide whether to accept the neighborhood solution according to the following criterion

[0188] If ΔE ≤ 0, accept the neighborhood solution That is, select as the new current solution;

[0189] If ΔE > 0, accept the neighborhood solution according to the probability P(ΔE) as the new current solution: Generate a random number r ∈ (0, 1). When P(ΔE) ≥ r, accept the neighborhood solution as the new current solution. When P(ΔE) < r, do not accept the neighborhood solution Keep the current solution unchanged. r will be updated each iteration, and the probability is determined by the parameter T. The formula is as follows:

[0190]

[0191] Among them, the parameter T gradually decreases after each iteration as the algorithm progresses.

[0192] Step S1424, parameter update

[0193] T gradually decreases as the algorithm progresses, usually in an exponential decay manner:

[0194] T = α T T

[0195] Among them, α T is a constant less than 1, usually set as α T ∈ (0, 1). For example, α T = 0.95.

[0196] Step S1425, continuously repeat steps S1421 to S1424 until the iteration meets the termination condition. The iteration ends, and the optimal number of clusters is obtained and the corresponding flash flood zoning results at the grid scale, where the termination conditions are as follows:

[0197] Reaching the maximum number of iterations N iter ;

[0198] Or the parameter T is reduced to a preset minimum value T min .

[0199] Thus, the optimization algorithm optimizes the selection of the number of clusters by minimizing the objective function composed of the clustering validity index to optimize the choice of the number of clusters, and the optimal number of clusters corresponding objective function value is the smallest. Based on the obtained optimal number of clusters i.e., the optimal number of partitions, the flash flood zoning results at the grid scale can be obtained.

[0200] In some embodiments, after obtaining the flash flood zoning results at the grid scale of the target area based on step S140, the following steps are further included:

[0201] Step S150, if historical flash flood event data is available for the target area, apply the historical flash flood event data to verify the accuracy of the obtained zoning results. The specific steps are as follows:

[0202] Overlay and compare the historical flash flood event data with the flash flood zoning results obtained in step S140, and analyze the correspondence between each partition unit in the zoning results and the spatial distribution pattern of historical flash flood events, that is, the zoning results can reflect the spatial distribution pattern of historical flash flood events, so as to evaluate the zoning results. Specifically:

[0203] First, calculate the number and density of historical flash flood disaster events in each partition unit.

[0204] After that, apply the hotspot analysis method to calculate the z-value of the spatial distribution of historical flash flood events in each partition unit, and determine the spatial distribution structure of the significance of historical flash flood events: for partition units with positive z-values, the higher the z-value, the more significant the spatial clustering characteristics of flash flood events; for partition units with negative z-values, the lower the z-value, the sparser the spatial distribution of flash flood events. The GIS software hotspot analysis tool calculates the Getis-Ord Gi* statistic of the target area based on historical flash flood events, and through the obtained z-score, the location where the elements cluster in space can be analyzed.

[0205] If the number and density of historical mountain flood events in each partition unit of the partition result are significantly different, and there is a significant spatial distribution structure of historical mountain flood events in some partition units, it indicates that the partition result corresponds well to the spatial distribution of historical mountain flood events. If there is no significant difference in the number and density of historical mountain flood events in each partition unit of the partition result, and there are no significantly higher positive or lower negative z-values in each partition unit, that is, the partition result cannot better reflect the spatial distribution pattern of historical mountain flood events, then consider from the aspects of the selection of mountain flood breeding background data, the determination of graph neural network model parameters, the optimization of the number of partitions, etc., and re-conduct clustering analysis to obtain the mountain flood partition result.

[0206] It can be understood that when performing mountain flood zoning through steps S110 to S150, the attribute characteristics and spatial structure of the mountain flood breeding background data are considered simultaneously. The partition result fully reflects the complex interactions in space of the mountain flood breeding background data, not only improving the zoning accuracy, but also realizing the accurate capture of the spatial distribution characteristics of the mountain flood breeding background data, providing data support for subsequent analysis of the combined effects and spatial variability of the mountain flood breeding background data.

[0207] In some embodiments, in step S100, constructing the mountain flood breeding background composite index by applying principal component analysis is obtained according to the following steps:

[0208] Step S160. Classify the mountain flood breeding background data according to its physical meaning, and apply principal component analysis to construct the mountain flood breeding background composite index.

[0209] For the mountain flood breeding background input data R = {r1, r2,..., r d}, where d is the dimension of the mountain flood breeding background data in R, divide it into Q major categories (1 < Q < d) according to the data physical meaning (such as terrain, climate, hydrology, underlying surface, etc.), and each major category is expressed as is the major category, and it is the set of all mountain flood breeding background data in the major category, and its expression is as follows:

[0210]

[0211] Among them, is the dimension of the mountain flood breeding background data included in the major category .

[0212] For each major category, principal component analysis (PCA) is applied separately to obtain the set of principal component scores for that major category. Principal component analysis finds the principal components that best represent the data variability by performing eigenvalue decomposition on the covariance matrix of the factors. The principal component scores are obtained by multiplying each principal component coefficient by the corresponding standardized original background data of mountain flood gestation. The comprehensive score is calculated by summing the product of each principal component score and the principal component weight. The comprehensive score is the composite index of the background of mountain flood gestation for each major category.

[0213] For the major category Apply principal component analysis to the background data of mountain flood gestation it contains to obtain the set of all principal component scores for that major category Denoted as:

[0214]

[0215] Where, is the number of principal components obtained by dimensionality reduction from the original background data of mountain flood gestation in the major category ; is the th principal component score of the major category .

[0216] For the major category The composite index of the background of mountain flood gestation for that major category is calculated by summing the product of each principal component score in that major category and the weight of each principal component Denoted as:

[0217]

[0218] Where, is the th principal component score of the major category ; is the weight of the th principal component of the major category in all principal components, calculated by the proportion of the variance percentage of this principal component in the cumulative variance percentage of all principal components.

[0219] Based on the above, the composite index of the background of mountain flood gestation for the major category is determined and denoted as:

[0220]

[0221] For major categories containing multiple background data of mountain flood gestation, the above operations need to be repeated to construct the composite index of the background of mountain flood gestation for each major category through principal component analysis Construct a composite index matrix with them as separate columns For this composite index matrix Convert it into a row-based form, i.e., where N is the number of nodes (grid points), Q is the number of composite indices, and Z′ i ∈R Q×1 is the composite index vector of the flash flood gestation background for node i, which is composed of the elements in the i-th row of the composite index matrix . The composite index of the flash flood gestation background obtained based on principal component analysis can, to a certain extent, summarize the attribute characteristics contained in various multi-dimensional flash flood gestation background data.

[0222] In some embodiments, since the existing interpretation model can only be applied to supervised machine learning models, a supervised support model needs to be used to replace the unsupervised graph neural network. The supervised support model is obtained according to the following steps:

[0223] Step S170: Use the clustering results obtained by the unsupervised graph neural network and all the composite index vectors of the flash flood gestation background to train a supervised machine learning model. Specifically:

[0224] Preferably, the supervised machine learning model uses a random forest model. Take the clustering label y = {y1, y2,..., y N}∈R N×1 of each node i and the corresponding composite index vector Z′ i ∈R Q×1 as the target variables. Set K as the number of clusters, and take or Divide the training set and the test set in a ratio of 8:2, and train the supervised machine learning model ML():

[0225] yi = ML(Z′ i , θ)

[0226] where yi is the clustering label for node i in the graph structure data G, and this clustering label is used as the supervision signal; Z′i is the composite index vector of the flash flood gestation background for node i; and θ is the parameter of the supervised machine learning model.

[0227] By training the supervised machine learning model with the clustering results based on GNN and all the composite index vectors of the flash flood gestation background, a supervised support model is obtained. To a certain extent, this supervised support model can equivalently replace the unsupervised graph neural network model in the flash flood zoning process, so that the interpretation model can be applied to this supervised support model to obtain the interpretation result.

[0228] In some embodiments, step S200 specifically includes:

[0229] For the supervised support model applying the SHAP interpretation model, the SHAP interpretation model provides an explanation through the marginal contribution of each mountain flood gestation background composite index to the prediction result of the supervised support model. For each node i (i.e., each grid point), represents the th mountain flood gestation background composite index's contribution to the clustering label y i prediction of this node, and the calculation is as follows:

[0230]

[0231] where, is the value of the th mountain flood gestation background composite index at node i; is the contribution value of the th mountain flood gestation background composite index to the prediction of the clustering label yi of node i; S is a subset of features, representing a subset of other mountain flood gestation background composite indexes that do not include the th mountain flood gestation background composite index, that is, other mountain flood gestation background composite indexes except the th mountain flood gestation background composite index; g(S) is the result of the supervised support model trained on the feature subset S to predict node i; is the prediction result of the supervised support model for node i when the feature subset S is added with , that is, the change in the output after the supervised support model adds ; represents the expectation for all possible feature subsets S, and all possible permutations and combinations of composite indexes need to be considered, and the marginal contribution of the composite index in all combinations is calculated through the expected value.

[0232] By calculating the contribution values of all nodes (grid points), the global impact (the average impact of the mountain flood gestation background composite index on the entire target area partition result) and local impact (the impact of the mountain flood gestation background composite index on each grid point) of each mountain flood gestation background composite index can be obtained.

[0233] Global impact: By averaging the contribution values of all nodes, the global impact of the th mountain flood gestation background composite index can be obtained represents the average contribution size of the th mountain flood gestation background composite index in the entire target area:

[0234]

[0235] Repeat the above calculations for all composite indicators of the background of mountain flood breeding, and the global impact of all composite indicators can be obtained. Q is the number of composite indicators of the background of mountain flood breeding.

[0236] Local impact: For each individual node i, its local contribution value can be calculated. This represents the specific contribution of the th composite indicator of the background of mountain flood breeding to the zoning result of node i.

[0237] In some embodiments, referring to Figure 1 the quantitative analysis method provided by the embodiments of the present invention further includes:

[0238] Step S300: For the global impact of all composite indicators of the background of mountain flood breeding obtained by the interpretation model Determine the ranking of the impact of the composite indicators of the background of mountain flood breeding on the zoning result, and determine the impact difference of each composite indicator of the background of mountain flood breeding; for the local impact, in the GIS software, visualize each node and the impact of each composite indicator of the background of mountain flood breeding in space according to the longitude and latitude of the node, and analyze the spatial variability of the impact of each dimension of the composite indicator of the background of mountain flood breeding.

[0239] The following is a specific embodiment of the quantitative evaluation method of the present invention:

[0240] This embodiment is used to quantitatively evaluate the impact of the background data of mountain flood breeding on mountain flood zoning based on the mountain flood zoning result in the Hengduan Mountains. The specific steps of this embodiment are as follows:

[0241] Step S100: Collect the background data of mountain flood breeding in the Hengduan Mountains, and apply an unsupervised graph neural network to obtain the mountain flood zoning result in the Hengduan Mountains. Among them, the zoning result is the clustering label y = {y1, y2,..., y 108503} ∈ R 108503×1 . Classify the background data of mountain flood breeding according to physical meaning. For all the background data of mountain flood breeding within each major category, apply principal component analysis to construct the composite indicators of the background of mountain flood breeding for each major category. Train a random forest model based on the obtained zoning result and all the composite indicators of the background of mountain flood breeding, so as to train a supervised support model that can replace the original graph neural network. This step specifically includes:

[0242] Step S110: Extensively collect the background data of mountain flood breeding in aspects such as terrain, climate, meteorology, hydrology, and underlying surface in the Hengduan Mountains. When conditions permit, mainly use grid data with a relatively high spatial resolution, and it is also allowed that some input data are vector data. A total of 61-dimensional data are collected, including extreme rainfall data under 30 multi-statistical frequencies and 31 other breeding background data. Resample all the data to 2×2 km2 The raster data of spatial resolution is cropped according to the scope of the Hengduan Mountains to obtain the initial data R0 = {r1, r2,... r 61}, and the data dimension of R0 is 61.

[0243] Step S120: Perform correlation analysis and multicollinearity analysis on the data of each dimension in the initial data R0, identify and remove the data with correlation > 0.7 and variance inflation factor > 10. Specifically: Based on literature reading and research on the Hengduan Mountains, 12 typical extreme rainfall data are selected from 30 extreme rainfall data, which are the maximum 3-hour / 6-hour / 12-hour / 24-hour rainfall of once-in-a-century and once-in-two-century, and the maximum 3-hour / 6-hour / 12-hour / 24-hour average annual rainfall; at the same time, 6 soil attribute data and stream power index with high correlation are removed. The input data R in raster form is obtained as R = {r1, r2,..., r i ,... r 36}, and 36 is the number of data dimensions of the input data R.

[0244] Step S130: Construct the graph structure data G = (V, E, X) from the raster-form input data R. The specific steps are as follows:

[0245] Step S131: Construct the node matrix V: For the raster-form input data R, there are 108,503 raster points in the 2×2 km raster data of the Hengduan Mountains. All raster data are converted into the graph structure data G = (V, E, X). Among them, each raster point is regarded as a node in the graph structure, and the node matrix V = {v1,..., v 2} is constructed, and each node is associated with the corresponding 36-dimensional attributes. 108503}

[0246] Step S132: Construct a KD tree to find neighbor nodes: Calculate the distance between nodes according to the longitude and latitude of the nodes, and construct the edge set between nodes according to the number of nearest neighbor nodes and the upper limit of the neighbor distance Find the nearest neighbor nodes of each node through the KD tree to obtain the neighbor node set of each node.

[0247] Step S133: Construction and normalization of the adjacency matrix: For each node i, if there is an adjacency relationship between node i and node j, then a ij = 1, otherwise a ij = 0, so as to construct the adjacency matrix A0 ∈ {0, 1} 108503×108503 . Subsequently, perform symmetric normalization processing on the adjacency matrix A0 to obtain the normalized adjacency matrix A ∈ R 108503×108503 .

[0248] Step S134: Construct the node attribute matrix X: Take the 36 input mountain flood gestation background data screened on each node as the attribute vector x of each respective node i , thereby constructing the node attribute matrix X = {x1, x2,..., x 108503} ∈ R 108503×36 , where the i-th row represents the 36-dimensional feature vector of node i.

[0249] Through the above process, the conversion of the input data R in raster form to the graph structure data G is achieved.

[0250] Step S140: For the case where the number of partitions in the Hengduan Mountains is not set, for all numbers of partitions (2 - 20), using the graph structure data G = (V, E, X) as the input, traverse and calculate using a graph neural network. Through an optimization algorithm, with the clustering validity index as the optimization objective, search and determine the optimal number of partitions, thereby determining the partition result at the raster scale. The specific steps refer to the aforementioned steps S1421 - S1425 and will not be elaborated here. Among them, the parameters of the graph neural network in this embodiment are set as follows: Specify the training device device = cuda (using GPU acceleration), the trade-off parameter tradeoff = 1e-10, the activation function activate = relu, the number of units in the model hidden layer hid-units = 512, the number of layers in the encoder part encoder_layer = 1, the number of layers in the projector part projjector_layer = 1, the learning rate lr = 1e-2, the total number of training epochs epochs = 100, the number of neighbors of each node in the graph num_neighbors = 8, and the upper limit of the distance during neighbor search distance_upper_bound = 100.0. Determine the optimal clustering number that is, the optimal number of partitions, and obtain the mountain flood partition result at the raster scale corresponding to this optimal number of partitions. See , Figure 3 , Figure 3 where (a) in shows the values of the clustering validity index DBI under the number of partitions 2 - 20, Figure 3 and (b) in shows the values of the clustering validity index CQI under the number of clusters 2 - 20. See Figure 4 (a) in, which is the mountain flood partition result of the Hengduan Mountains obtained based on the graph neural network model.

[0251] Step S150: When historical mountain flood event data is available in the target area, apply the historical mountain flood events to further verify the accuracy of the partition result. See Figure 4 , Figure 4In (a), it is the result of the mountain torrent zoning in the Hengduan Mountains obtained based on the graph neural network model. Among them, there are a total of 12 zoning units, and each zoning unit is respectively labeled as SW-1, W-2, NW-3, NW-4, M-5, M-6, SW-7, SE-8, NE-9, NE-10, E-11, M-12. The points in the figure are historical mountain torrent events. Figure 4 In (b), it is the quantity and density of historical mountain torrent events within each zoning unit. Figure 4 In (c), it is the spatial statistical z-value of historical mountain torrent events within each sub-unit. The specific steps are as follows:

[0252] Overlay the historical mountain torrent event data with the mountain torrent zoning result in the Hengduan Mountains, and analyze the correspondence between each zoning unit in the zoning result and the spatial distribution pattern of historical mountain torrent events, that is, the zoning result can reflect the spatial distribution pattern of historical mountain torrent events. First, calculate the quantity and density of historical mountain torrent events within each zoning unit. It can be seen that the number of historical mountain torrent events in the zoning units SW-1, SW-7, and SE-8 is large (364, 187, and 433 respectively), and the density is relatively high (0.00896, 0.00627, and 0.00947 events per square kilometer respectively). The number of historical mountain torrent events in the zoning units NW-3, NW-4, M-6, and M-12 is small (0, 110, 35, and 5 respectively), and the distribution density is relatively low (0, 0.00153, 0.00140, and 0.00018 events per square kilometer respectively), indicating that there are significant differences in the quantity and density of historical mountain torrent events within each zoning unit. Then, apply GIS software for hotspot analysis, calculate the spatial statistical z-value of the distribution of historical mountain torrent events within each zoning unit, and determine the spatial distribution structure of the significance of historical mountain torrent events. For the zoning units with positive z-values, the higher the z-value, the more significant the spatial clustering characteristics of mountain torrent events; for the zoning units with negative z-values, the lower the z-value, the sparser the spatial distribution of mountain torrent events. It can be seen that the z-values in the zoning units SW-1, SW-7, and SE-8 are 4.336, 3.451, and 3.950 respectively, belonging to the spatial aggregation areas of historical mountain torrent events. The z-values in the zoning units NW-3, NW-4, M-6, and M-12 are -3.000, -2.331, -1.942, and -1.993 respectively, belonging to the spatial sparse distribution areas of historical mountain torrent events. The z-values in these several zoning units have significantly large positive or small negative values. This shows that the mountain torrent zoning can reflect the spatial distribution characteristics of historical mountain torrent events (such as areas of dense or sparse distribution), indicating the reliability and rationality of the mountain torrent zoning result.

[0253] Step S160: For the selected 36 background data of mountain flood gestation, divide them into five major categories according to physical meaning: topography (including seven data: altitude, slope, aspect, altitude difference, terrain wetness index, geomorphic type, and runoff), climate (including seven data: annual average rainfall, annual average temperature, annual average potential evapotranspiration, annual average actual evapotranspiration, vegetation transpiration, canopy interception evaporation, and soil evaporation), meteorology (including 12 extreme rainfall data), soil (including seven data: soil type, soil saturated hydraulic conductivity, soil surface gravel content, soil depth, soil texture, soil water content, and soil erosion modulus), and vegetation (including three data: vegetation type, normalized difference vegetation index, and leaf area index). For each major category, apply principal component analysis respectively. Multiply each principal component coefficient by the corresponding standardized original background data of mountain flood gestation to obtain all principal component scores. Sum the products of all principal component scores and the weights of each principal component to calculate the comprehensive score of each major category. The comprehensive score is the composite index of the background of mountain flood gestation for each major category.

[0254] For example, the calculation formula for the climate composite index is as follows:

[0255] Fcl = (51.683588 / 88.571910) * Fcl_1 + (19.941340 / 88.571910) * Fcl_2 + (9.900932 / 88.571910) * Fcl_3 + (7.046050 / 88.571910) * Fcl_4. Among them, Fcl is the climate composite index, and Fcl_1, Fcl_2, Fcl_3, and Fcl_4 are the four principal components obtained by performing principal component analysis on all background data of mountain flood gestation within the climate major category.

[0256] Step S170: Existing interpretation models can only be applied to the results of supervised machine learning models. Therefore, based on the clustering results obtained from the graph neural network and all the composite indices of the background of mountain flood gestation, train a random forest model. Let the clustering label of each node y = {y1, y2,..., y 108503} ∈ R 108503×1 and the corresponding composite indices of the background of mountain flood gestation {Z′1, Z′2,..., Z′ 108503} ∈ R 108503×5 be the target variables. Set the number of clusters to 12, divide the training set and the test set according to the ratio of 8:2, and train a random forest model with the number of decision trees in the random forest model n_estimators = 100. The accuracy of the obtained random forest model is accuracy = 0.84, indicating that the random forest model trained based on the clustering results of the graph neural network can equivalently replace the original graph neural network, so as to apply the interpretation model to this random forest model to obtain the interpretation result.

[0257] Step S200: Apply the SHAP interpretation model to the trained random forest model and the composite index of the background of mountain flood gestation. The interpretation model provides an explanation by the marginal contribution of each dimension of the composite index of the background of mountain flood gestation to the prediction result of the random forest model. By calculating the contribution values of all nodes (grid points), the global impact (the impact of the composite index of the background of mountain flood gestation on the mountain flood zoning result of the entire Hengduan Mountains) and the local impact (the impact of the composite index of the background of mountain flood gestation on each grid point) of each dimension of the composite index of the background of mountain flood gestation can be obtained. See Figure 5 , which is the global composite impact of the mountain flood gestation background data on the mountain flood zoning result based on the SHAP interpretation model, taking the Hengduan Mountains as an example.

[0258] Step S300: Based on the global impact, determine the influence ranking of the composite index on the zoning result, and analyze the influence differences of each composite index. See Figure 5 , which is the global composite impact of the mountain flood gestation background data on the mountain flood zoning result based on the SHAP interpretation model, taking the Hengduan Mountains as an example. The influence ranking of the five composite indexes on the mountain flood zoning of the Hengduan Mountains is meteorology (the proportion of the absolute value of SHAP of this factor in the sum of the absolute values of SHAP of the 5 composite factors: 40.12%), climate (26.84%), soil (13.94%), vegetation (10.03%) and terrain (9.08%). The meteorological composite index is obtained by principal component analysis of 12 extreme rainfall factors, representing the composite impact of multiple extreme rainfall factors on the occurrence of mountain floods, highlighting the important composite impact of extreme precipitation in driving the occurrence of mountain floods in the Hengduan Mountains. For the local impact, in the GIS software, the impacts of the composite indexes of the background of mountain flood gestation in all dimensions of each node are visualized spatially according to the longitude and latitude of the node, and the spatial variability of the impacts of the composite indexes of the background of mountain flood gestation in each dimension is analyzed. See Figure 6-1 and Figure 6-2 , which is the spatial variation of the composite impact of the mountain flood gestation background data on the mountain flood zoning result based on the SHAP interpretation model, taking the Hengduan Mountains as an example. The meteorological composite index shows a significant positive impact on both the east and west sides of the Hengduan Mountains. See Figure 6-1 (a) in Figure 6-1 ; the climate composite index has a significant positive impact in some areas in the northwest and southwest of the Hengduan Mountains. See Figure 6-2 (b) in Figure 6-2 ; the terrain composite index shows a prominent positive impact along the Yalong River, Dadu River and Minjiang River in the central and northwest of the Hengduan Mountains. See Figure 6-2In (e). This shows that the SHAP interpretation model can quantify the combined effects of multiple flash flood influencing factors and analyze the spatial variability of each combined effect, demonstrating the effectiveness and interpretability of the construction of composite indicators and the application of the interpretation model. This is of great significance for the study of flash flood formation mechanisms.

[0259] In summary, the embodiments of the present invention first obtain the flash flood zoning results of the target mountainous area based on unsupervised graph neural networks. During the zoning process, a graph neural network model that can simultaneously consider the attribute characteristics and topological structure of the input data is applied for clustering analysis. The obtained flash flood zoning results in the Hengduan Mountains can reflect the spatial distribution patterns of historical flash flood events and flash flood formation backgrounds. An optimization algorithm is used to automatically determine the most suitable number of zones for the Hengduan Mountains, making the zoning process more intelligent and objective, and less affected by human factors. Therefore, the automated flash flood zoning method adopted in the embodiments of the present invention can provide technical support for flash flood risk assessment in mountainous areas and the layout of refined flood control and disaster reduction measures. In addition, the embodiments of the present invention apply principal component analysis to construct composite indicators, and introducing the SHAP interpretation model can provide a quantitative interpretation of the results of the graph neural network. SHAP accurately evaluates the influence of each composite indicator of the formation background on the zoning results by simulating and quantifying the perturbation effects of input factors, thereby revealing the combined effects of the coupling of multiple flash flood formation background data on the zoning results. Combining the SHAP interpretation model can achieve a quantitative evaluation of the combined effects of flash flood formation background factors, provide a scientific basis for the zoning results, clarify the mechanism of the combined action of multiple factors in flash flood formation, and assist in the precise management of flash flood disasters and the optimization of defense measures.

[0260] In the description of this specification, the descriptions referring to terms such as "one embodiment", "some embodiments", "illustrative embodiments", "examples", "specific examples", or "some examples" etc. mean that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in a suitable manner in any one or more embodiments or examples.

[0261] Although the embodiments of the present application have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present application. The scope of the present application is defined by the claims and their equivalents.

Claims

1. A quantitative analysis method for the combined influence of background data of mountain flood gestation on mountain flood zoning, characterized in that It includes the following steps: Step S100: Collect the background data of mountain flood gestation in the target area, preprocess it, and then apply the mountain flood zoning model based on unsupervised machine learning to obtain the mountain flood zoning result of the target area. Classify the background data of mountain flood gestation according to its physical meaning, and construct the composite index of mountain flood gestation background for each classification by applying principal component analysis. Train the supervised machine learning model through the mountain flood zoning result and the composite index of mountain flood gestation background to learn the zoning result of the mountain flood zoning model and obtain the supervised support model; Step S200: Input the supervised support model and all the composite indexes of mountain flood gestation background in the target area into the interpretation model. The interpretation model provides the contribution of each composite index of mountain flood gestation background in the process of mountain flood zoning through the marginal contribution of each composite index of mountain flood gestation background to the prediction result of the supervised support model, and obtains the influence of each composite index of mountain flood gestation background on the mountain flood zoning result.

2. The quantitative analysis method according to claim 1, characterized in that, In step S100, the mountain flood zoning result of the target area is obtained by using an unsupervised graph neural network. The specific steps include: Step S110: Resample all the background data of mountain flood gestation in the collected target area into grid data with a unified spatial resolution, and crop the grid data according to the scope of the target area to obtain the initial data R0; Step S120: Conduct correlation analysis and multicollinearity analysis on each dimension data in the initial data R0, identify and remove the data dimensions whose correlation and variance inflation factor do not meet the requirements, and obtain the input data R in grid form; Step S130: For the input data R in grid form, regard each grid as a node, construct a node matrix V, determine the edges between nodes according to the distance between nodes, construct an edge set E and a normalized adjacency matrix A, and construct a node attribute matrix X based on the input data R, so as to construct the grid form input data R into graph structure data G=(V, E, X); Step S140: For the case where the number of partitions in the target area has been set, based on the set number of partitions, use the graph structure data G=(V, E, X) as the input, and conduct clustering analysis through the graph neural network to obtain the zoning result at the grid scale; for the case where the number of partitions in the target area has not been set, for all the number of partitions, use the graph structure data G=(V, E, X) as the input, conduct traversal calculation using the graph neural network, and through the optimization method, take the clustering validity index as the optimization target, search and determine the optimal number of partitions, so as to determine the zoning result at the grid scale.

3. The quantitative analysis method according to claim 2, wherein In step S130, the distances between nodes are calculated according to the longitude and latitude attributes of the nodes in the grid-form input data R. Based on the number of nearest nodes and the upper limit of the adjacent distance, edges between nodes are established according to the octal neighborhood or quadrilateral neighborhood rule, and an edge set is obtained. It is used to represent a set of edges connecting different nodes. (i, j) represents an edge connecting node i and node j, where i ≠ j.

4. The quantitative analysis method according to claim 3, characterized in that, In step S130, the construction steps of the normalized adjacency matrix A include: Find the nearest neighbor node of node i through the KD tree: neighbors i = KDTree.query(C i , num_neighbors + 1, distance_upper_bound) Among them, neighbors i represents all neighbor nodes of node i; KDTree.query(·) represents the KD-tree calculation function; each node i has longitude log i and latitude lat i Two attributes, all nodes form a two-dimensional coordinate set C, which contains the longitude and latitude attributes of all nodes C i =(log i , lat i ) is the longitude and latitude attribute of node i, N is the number of grids contained in each dimension data in the input data R; num_neighbors is the number of neighbor nodes found for each node in the neighborhood; distance_upper_bound represents the upper bound of the distance for finding neighbor nodes; Remove the own node i from neighbors i to obtain the neighbor node set neighbors i ' of node i; For nodes i and j, if node i is adjacent to node j, i.e., j ∈ neighbors i ′, then a ij = 1, otherwise a ij = 0, thus constructing the adjacency matrix A0 = {a ij |(i, j) ∈ V 2} ∈ R N×N ; Perform symmetric normalization processing on the adjacency matrix A0 to obtain the normalized adjacency matrix A: A = D -1 / 2 A0D -1 / 2 ∈R N×N where \(D = \{d' i | i\in V\}\) is the degree matrix, which is a diagonal matrix with diagonal elements being the degrees \(d'\) of each node i =\sum j a ij , i\in V, j\in V 5. The quantitative analysis method according to claim 2, characterized in that In step S140, when the number of partitions has been set for the target area, the number of partitions is set to K. Based on this number of partitions, using the graph structure data G=(V, E, X) as the input, clustering analysis is performed through a graph neural network to obtain the partition result at the grid scale. The specific steps include: Step S1411. For the original graph structure data G = {X, A} represented by the node attribute matrix X ∈ R N×d and the normalized adjacency matrix A ∈ R N×N , where N is the number of grids contained in each dimension data in the input data R, enhanced data is generated through the data augmentation operation to generate enhanced data Step S1412: Use a graph neural network to calculate node representations for the original graph structure data G and the augmented data respectively. The graph neural network updates the node representations through multiple layers of graph convolutions: H (l+1) = σ(AH (l) W (l) ) where, H (l) is the node representation of the l-th layer in the graph neural network; W (l) is the weight matrix of the l-th layer in the graph neural network; σ(·) is the activation function; After multiple layers of graph convolution, the original graph structure data G and the enhanced data are obtained. The node representation H of Among them, is a graph neural network function; To constrain the attribute vector x of node i i and its enhanced attribute vector for contrastive learning, the contrastive loss function used is as follows: where h i and are the node representations of the original graph structure data and the enhanced data of node i, respectively; Sim(·, ·) represents the similarity between nodes; Step S1413. For the node representations H = {h1, h2,..., h N} learned by the graph neural network, apply the K-means method to initialize the set C of clustering centers, and use the clustering centers as learnable parameters of the graph neural network. By calculating the distance metric Assign each node representation h i to the nearest clustering center c k ; Optimize the total clustering loss of the graph neural network through backpropagation and gradient descent Update the set C of clustering centers and the classification of nodes in each iteration to obtain the final set of clustering centers and the clustering label representations of each node; The total clustering loss is expressed as: Among them, is the clustering expansion loss, is the clustering contraction loss, and τ is a hyperparameter used to control the balance between expansion and contraction; c A ∈ C and c B ∈ C are the cluster centers of cluster A and cluster B respectively, are all the nodes in cluster A; In each iteration, for node update the cluster center c according to the following formula A :[[]]END]] where, |C A | is the number of nodes in cluster A; Step S1414: Visualize the clustering labels of each node spatially according to the longitude and latitude of the node to obtain the flash flood partition result at the grid scale.

6. The quantitative analysis method according to claim 5, characterized in that, In step S140, when the number of partitions has not been set for the target area, for all numbers of partitions, using the graph structure data G=(V, E, X) as the input, traversal calculations are performed using a graph neural network. Through an optimization method, with the clustering validity index as the optimization target, the optimal number of partitions is searched and determined, thereby determining the partition result at the grid scale. The specific steps include: Step S1421: Set as the initial number of clusters, where δ and ε are proportionality coefficients; based on the set number of clusters perform clustering analysis according to steps S1411 to S1414, calculate the corresponding clustering validity index based on the clustering result, and set the objective function as the weighted sum of the clustering validity indices: Among them, is the current number of clusters, that is, the current solution; α and β are weight coefficients; and are validity indicators calculated based on the clustering results of the graph neural network, is the sum of the average coefficient of variation of the selected background data for flash flood breeding in the clustering results, used to quantify the dispersion degree of the selected background data for flash flood breeding, evaluates the tightness and dispersion between clustering results through the distance between cluster centers, and The lower the value of, the better the clustering effect, and the calculation formulas are as follows: where, |C w is the number of nodes in the clustering category w, d is the dimension of the background data for mountain flood gestation, is the coefficient of variation of the r-th dimension gestation background data in the clustering category w; is the standard deviation of the r-th dimension gestation background data in the clustering category w, is the mean of the r-th dimension gestation background data in the clustering category w; is the average distance from all nodes in the clustering category w to the clustering center, is the average distance from all nodes in the clustering category t to the clustering center, d wt is the distance between the centroid of the clustering category w and the centroid of the clustering category t; Step S1422, from the current solution Generate a new neighborhood solution Among them, γ takes -1 or 1; Step S1423, calculate the current solution and the objective function values of the neighborhood solutions Calculate the difference ΔE between the two: and Calculate the difference ΔE between the two: Determine whether to accept the neighborhood solution by the following criteria If ΔE ≤ 0, accept the neighborhood solution as the new current solution; If ΔE > 0, accept the neighborhood solution according to the probability P(ΔE). As the new current solution, the probability is determined by the parameter T, and the formula is as follows: Among them, the parameter T gradually decreases after each iteration as the algorithm progresses; Step S1424: Update the parameter T in an exponentially decaying manner: T = α T T where α T is a constant, and α T ∈(0, 1); Step S1425: Continuously repeat steps S1421 to S1424 until the iteration meets the termination condition. The iteration ends, and the optimal number of clusters is obtained. And the corresponding flash flood zoning results at the grid scale.

7. The quantitative analysis method according to any one of claims 2 to 6, characterized in that After obtaining the flash flood partition result, it further includes: Step S150: Obtain the historical flash flood event data of the target area, and use this historical flash flood event data to evaluate the obtained partition result. Specifically, overlay the historical flash flood event data of the target area on the partition result, calculate the number and density of historical flash flood disaster events in each partition unit, apply the hot spot analysis method to calculate the z-value of the spatial distribution of historical flash flood events in each partition unit, and determine the spatial distribution structure of the significance of historical flash flood events; if the number and density of historical flash flood events in each partition unit in the partition result are significantly different, and there is a significant spatial distribution structure of historical flash flood events in some partition units, it indicates that the partition result corresponds well to the spatial distribution of historical flash flood events, indicating that the partition result is accurate; if there is no significant difference in the number and density of historical flash flood events in each partition unit in the partition result, and there are no significantly higher positive values or lower negative values of the z-value in each partition unit, then consider aspects such as the selection of flash flood gestation background data, the determination of graph neural network model parameters, and / or the optimization of the number of partitions, and perform clustering analysis again to obtain the flash flood partition result.

8. The quantitative analysis method according to claim 7, wherein The supervised support model is obtained according to the following steps: Step S160: Construction of the flash flood gestation background composite index: According to the physical meaning of the input data \(R = \{r_1, r_2, \ldots, r\) d \(\}\), the input data \(R\) of the mountain flood gestation background is divided into \(Q\) categories, where \(1 < Q < d\), and \(d\) is the dimension of the mountain flood gestation background data in \(R\). Let be the set of all mountain flood gestation background data in the \(Q\)th category, and its expression is as follows: Among them, is the dimension of the background data for flash flood breeding included in the large category ; the dimension of the background data for flash flood breeding included in the large category For Apply principal component analysis to the background data of mountain flood contained therein, and construct the composite index of the background of mountain flood of the th major category according to the following formula ​ Among them, is the -th principal component score; is the -th principal component's weight among all principal components, calculated by the proportion of the variance percentage of this principal component in the cumulative variance percentage of principal components; Construct a composite index matrix by taking the composite index of the background of mountain flood breeding in each major category as a column respectively Convert this composite index matrix into a form expressed by rows, that is where N is the number of nodes, Q is the number of composite indices, and Z′ i ∈R Q×1 is the composite index vector of the background of mountain flood breeding of node i, which is composed of the elements in the i-th row of the composite index matrix ; Step S170: Training of the supervised machine learning model: Let the clustering label y = {y1, y2,..., y N} ∈ R N×1 and the corresponding composite index vector Z′ of the background for flash flood gestation i be used as the target variables, set K as the number of clusters, and train the supervised machine learning model ML(): y i = ML(Z' i ,θ) where y i is the clustering label for node i in the graph structure data G, and this clustering label is used as the supervision signal; θ is the parameter of the supervised machine learning model.

9. The quantitative analysis method according to claim 7, characterized in that In step S200, for each node i, let the contribution value predicted by the interpretation model for the clustering label y of the node i with respect to the th mountain flood gestation background composite index be i The calculation formula is as follows: as follows: Among them, is the value of the th mountain flood breeding background composite index at node i; S is a feature subset, representing a subset of other mountain flood breeding background composite indices that do not include the th mountain flood breeding background composite index; g(S) is the result of predicting node i by the supervised support model trained on the feature subset S; is the prediction result of the supervised support model for node i when the feature subset S is added with ; represents the expectation for all possible feature subsets S; By calculating the contribution values of all nodes, the global influence and local influence of each flash flood gestation background composite index are obtained, where: Set up The global impact of the composite index of the flash flood breeding background is Indicates The average contribution of the composite index of flash flood breeding background in the entire target area is calculated as follows: The global impact of all composite indicators is d is the dimension of the background data of flash floods; The local impact is taken as the contribution value as the local contribution value indicating the specific contribution of the th mountain flood breeding background composite index to the zoning result of node i.

10. The quantitative analysis method according to claim 9, characterized in that, The quantitative analysis method further includes: Step S300, for the global impact of all composite indicators obtained by the interpretation model Determine the ranking of the impacts of all composite indicators on the zoning results, and determine the impact differences of each composite indicator of the mountain flood breeding background; for the local impact, visualize the impacts of each node and each composite indicator of the mountain flood breeding background in space according to the longitude and latitude of the node, and analyze the spatial variability of the impacts of each composite indicator of the mountain flood breeding background.

Images