Method and system for predicting disaster risk space, terminal and storage medium
Patent Information
- Application Number
- CN202511263397.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-05
- Publication Date
- 2026-01-20
Smart Images
Figure CN121365765A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of data processing, and particularly relates to a disaster risk space prediction method, system, terminal and computer readable storage medium. BACKGROUND
[0002] Disaster risk space prediction aims to depict the intensity distribution and regional differences of natural disasters in space, and can be used to support risk early warning plans, emergency decision support, and on-demand and timely rescue, etc. application scenarios, so as to improve the disaster emergency response and disaster prevention and reduction ability, and promote the development of fine disaster emergency management level and the resilience construction of smart city.
[0003] However, the existing disaster risk space prediction focuses on the modeling of internal indicators of the unit, ignoring the potential influence of the surrounding environment on the current unit, thereby resulting in low accuracy of disaster risk space prediction, which cannot meet the prediction requirements of disaster risk space.
[0004] Therefore, the prior art still needs to be improved and developed. SUMMARY
[0005] The main purpose of the present application is to provide a disaster risk space prediction method, system, terminal and computer readable storage medium, which aims to solve the problem that the existing disaster risk space prediction focuses on the modeling of internal indicators of the unit, ignoring the potential influence of the surrounding environment on the current unit, thereby resulting in low accuracy of disaster risk space prediction.
[0006] To achieve the above purpose, the present application provides a disaster risk space prediction method, which comprises the following steps:
[0007] An adjacency matrix of each geographic unit is constructed, and high-order neighborhood expansion processing is performed on the adjacency matrix to obtain an external environment of the geographic unit;
[0008] The external environment of the geographic unit is subjected to the same-order geographic unit weighting processing and neighborhood feature aggregation processing to obtain a geographic unit environment representation of each geographic unit;
[0009] The cosine similarity between each geographic unit environment representation is calculated, and a semantic space correlation graph is constructed according to the cosine similarity;
[0010] The node feature fusion processing and disaster risk space prediction processing are performed on the semantic space correlation graph and the geographic unit environment representation to obtain a disaster risk space prediction result.
[0011] Optionally, the method for predicting a disaster risk space, wherein the adjacency matrix of each geographical unit is constructed, and high-order neighborhood expansion processing is performed on the adjacency matrix to obtain a geographical unit external environment, and the method specifically comprises:
[0012] A disaster environment element in each geographical unit is obtained, and a geographical unit internal environment of each geographical unit is constructed according to the disaster environment element, wherein the disaster environment element comprises a surface topography, a climate element, a land cover, a human economy, and a disaster-causing factor;
[0013] A topological adjacency relationship between each geographical unit is obtained according to the geographical unit internal environment, and an adjacency matrix is constructed according to the topological adjacency relationship;
[0014] High-order neighborhood expansion processing is performed on the adjacency matrix to obtain a geographical unit external environment.
[0015] Optionally, the method for predicting a disaster risk space, wherein the same-order geographical unit weighting processing comprises attention coefficient calculation processing and weighting processing, and the neighborhood feature aggregation processing comprises feature updating processing, attention weight distribution, and weighted summation processing;
[0016] The same-order geographical unit weighting processing and the neighborhood feature aggregation processing are performed on the geographical unit external environment to obtain a geographical unit environment representation of each geographical unit, and the method specifically comprises:
[0017] A target neighborhood in the geographical unit external environment is determined, and an attention coefficient between a current geographical unit and other geographical units in the target neighborhood is calculated;
[0018] A same-order neighbor geographical unit corresponding to the current geographical unit is determined, and the same-order neighbor geographical unit is weighted according to the attention coefficient to obtain a weighted feature;
[0019] Feature updating processing is performed on the current geographical unit according to the weighted feature to obtain a target geographical unit feature;
[0020] A channel attention mechanism is adopted to perform attention weight distribution and weighted summation processing on the target geographical unit feature to obtain a geographical unit environment representation of each geographical unit.
[0021] Optionally, the method for predicting a disaster risk space, wherein the expression of the attention coefficient is:
[0022]
[0023] wherein i is a current geographical unit, j is another geographical unit, and l is a layer number at which the current geographical unit and the other geographical unit are located, is an attention coefficient, LeakyReLU is an activation function, W (l) is a weight matrix of a linear transformation, is an input feature of a current geographical unit at the l-th layer, and || is a vector concatenation operation, is a transposed weight vector, is an input feature of another geographical unit at the l-th layer.
[0024] Optionally, the method for predicting a disaster risk space, wherein the cosine similarity between the environment representations of each geographical unit is calculated, and a semantic space correlation graph is constructed according to the cosine similarity, specifically comprising:
[0025] The cosine similarity between the environment representations of each geographical unit is calculated to obtain a similarity result;
[0026] Each geographical unit is converted into a geographical unit node, and the environment representation corresponding to each geographical unit is converted into a geographical unit edge;
[0027] A preset number of target geographical unit nodes are extracted from the geographical unit nodes in descending order according to the similarity result, and a semantic space correlation graph is constructed according to the target geographical unit nodes, the similarity result, and the geographical unit edge.
[0028] Optionally, the method for predicting a disaster risk space, wherein the node feature fusion processing and the disaster risk space prediction processing are performed on the semantic space correlation graph and the environment representation of the geographical unit to obtain a disaster risk space prediction result, specifically comprising:
[0029] Local features of nodes in the semantic space correlation graph are extracted, and the local features of the nodes are standardized to obtain standardized local features of the nodes;
[0030] The standardized local features of the nodes and the environment representation of the geographical unit are fused to obtain a fusion feature vector;
[0031] An initial clustering processing is performed on the fusion feature vector by using a K-means clustering algorithm to obtain an initial label assignment;
[0032] An optimization processing is performed on the initial label assignment by using a label propagation algorithm to obtain a disaster risk space prediction result.
[0033] Optionally, the method for predicting a disaster risk space, wherein the expression of the disaster risk space prediction result is:
[0034]
[0035] wherein, is the label of i after the t+1th iteration, argmax is the maximum value solution, c is a preset label, N(i) is a neighbor node set of i, w ij is an edge weight between i and j, and δ is an indicator function, is the label of j after the tth iteration.
[0036] In addition, to achieve the above object, the application further provides a disaster risk space prediction system, wherein the disaster risk space prediction system comprises:
[0037] An external environment construction module is configured to construct an adjacency matrix of each geographical unit, perform high-order neighborhood expansion processing on the adjacency matrix, and obtain an external environment of the geographical unit.
[0038] An environment representation generation module is configured to perform same-order geographical unit weighting processing and neighborhood feature aggregation processing on the external environment of the geographical unit, and obtain a geographical unit environment representation of each geographical unit.
[0039] A semantic space correlation graph construction module is configured to calculate a cosine similarity between each geographical unit environment representation, and construct a semantic space correlation graph according to the cosine similarity.
[0040] A prediction result output module is configured to perform node feature fusion processing and disaster risk space prediction processing on the semantic space correlation graph and the geographical unit environment representation, and obtain a disaster risk space prediction result.
[0041] In the application, the adjacency matrix of each geographical unit is constructed, the high-order neighborhood expansion processing is performed on the adjacency matrix, and the external environment of the geographical unit is obtained. The same-order geographical unit weighting processing and the neighborhood feature aggregation processing are performed on the external environment of the geographical unit, and the geographical unit environment representation of each geographical unit is obtained. The cosine similarity between each geographical unit environment representation is calculated, and the semantic space correlation graph is constructed according to the cosine similarity. The node feature fusion processing and the disaster risk space prediction processing are performed on the semantic space correlation graph and the geographical unit environment representation, and the disaster risk space prediction result is obtained. By constructing the adjacency matrix of the geographical unit and performing the high-order neighborhood expansion processing, the internal and external environments of the geographical unit are obtained. Then, the same-order geographical unit weighting processing, the neighborhood feature aggregation processing and the similarity calculation are performed according to the internal and external environments, so that the geographical unit environment representation and the semantic space correlation graph are constructed, and the accuracy of the disaster risk space prediction result can be effectively improved. BRIEF DESCRIPTION OF DRAWINGS
[0042] Figure 1 is a flowchart of a preferred embodiment of the disaster risk space prediction method of the application;
[0043] Figure 2 is a schematic diagram of the overall structure of a preferred embodiment of the disaster risk space prediction method of the present application;
[0044] Figure 3 is a structural diagram of a preferred embodiment of the disaster risk space prediction system of the present application;
[0045] Figure 4 is a structural diagram of a preferred embodiment of the terminal of the present application. DETAILED DESCRIPTION
[0046] In order to make the objects, technical solutions and advantages of the present application clearer and more explicit, the present application will be further described in detail below with reference to the drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.
[0047] Disaster risk space prediction aims to depict the intensity distribution and regional differences of natural disasters in space, support risk early warning and contingency plans, emergency decision support, and on-demand and timely rescue, etc. application scenarios, improve disaster emergency response and disaster prevention and mitigation capabilities, and promote the development of fine disaster emergency management level and the resilience construction of smart city.
[0048] Geographical units refer to region entities with clear boundaries, such as grid cells, administrative divisions, and natural watersheds, etc., which are the basic carriers for spatial prediction analysis. The current urban digital infrastructure and data resources are unevenly distributed, and the disaster monitoring data within the geographical unit presents obvious spatial sparseness, especially in the urban fringe and community scale, which significantly increases the difficulty of accurate disaster risk prediction. The current research methods of disaster risk space prediction mainly fall into two directions:
[0049] 1. Risk estimation based on index system: By integrating meteorological data, hydrological data, topographic data and other multi-source data, a multi-dimensional index system is constructed for risk estimation, which can effectively avoid the limitations of sparse distribution of monitoring data. Researchers usually select climate, terrain, economy and other multi-dimensional factors as evaluation indexes, construct a disaster risk evaluation index system and calculate the risk index, and complete the geographical unit disaster risk zoning according to the risk index. Although the method based on index system has important application in risk assessment, its results are over-dependent on empirical knowledge, and the evaluation process has strong uncertainty and low credibility. In addition, disaster risk is formed by a variety of interdependent and complex variables and data, while these methods usually assume that the indexes are independent of each other, without fully considering the interaction between the indexes.
[0050] 2. Spatially dependent risk prediction: Considering the sparse spatial distribution of monitoring data, spatially dependent risk prediction data follows specific mathematical rules in the spatial dimension, and the units in the real world are abstracted as a connected network for prediction. Early studies usually use hydrodynamic models combined with topographic data to simulate the runoff and confluence process of each regional unit, and further coupled with the rainstorm model to realize urban flood prediction. Such models have a certain effect on simulating urban disaster processes, but the prediction accuracy in data-scarce areas is limited, and they lack universal applicability for different basins or flood events. With the development of deep learning technology, graph neural network models have gradually been applied to disaster risk spatial prediction due to their advantages in processing complex graph structure data. Researchers combined topography, river network structure, and historical meteorological data to extract spatial features through multi-layer convolution of graph convolution network, improving the accuracy and stability of disaster risk prediction.
[0051] The complex correlation between geographical units and the spatio-temporal sparse distribution of monitoring data make it difficult to predict disaster risks in a timely and effective manner. To meet the application needs of disaster risk assessment and management, existing research has developed from index selection and weight allocation to geographical spatial prediction. However, current disaster risk spatial prediction mainly focuses on internal index modeling, ignoring the potential influence of the surrounding environment on the current unit, making it difficult to achieve disaster risk spatial discovery in unmonitored areas. The spatial dependence of spatial units revealed by the first law of geography indicates that disaster risk propagation has significant nonlinear neighborhood correlation characteristics, which exacerbates the dynamic quantification complexity of multi-order geographical neighborhood units and limits the timeliness and accuracy of disaster risk spatial prediction.
[0052] Overall, current research has developed from the selection of different indicators and the calculation of indicator weights to the simulation of the real world, but focuses on internal index modeling, ignoring the potential influence of the surrounding environment on the current unit, resulting in low risk prediction accuracy and time lag, making it difficult to support disaster risk spatial prediction in unmonitored areas.
[0053] To solve the above problems, the present application takes into account the correlation of the environment inside and outside the geographic unit and the fine management demand of disaster, and in view of the difficulty of risk space prediction in the area not covered by monitoring data and the problem of not being timely and effective, a disaster risk space prediction method fusing the environment inside and outside the geographic unit is provided. According to the characteristics of the environment inside and outside the geographic unit and the high-order neighborhood relationship, the environment representation of the geographic unit and the disaster risk prediction based on semantic similarity are realized, and the process mainly includes: 1, on the basis of defining the internal environment of the geographic unit, the high-order neighborhood is introduced to construct the external environment of the geographic unit; 2, in view of the characteristics that the correlation between geographic units is affected by the difference of distance, the multi-order neighbor is realized by the graph attention network and the channel attention adaptive weighted aggregation to realize the environment representation of the geographic unit; 3, the similarity between different geographic unit environment representations is calculated to construct the semantic space correlation graph, and the multi-stage label propagation is realized by fusing global and local features to realize disaster risk space prediction. The present application fully considers the influence difference of geographic units at different distances on the current unit and the adjacency relationship between units, and can realize fine scale disaster risk space prediction based on the environment inside and outside the geographic unit, thereby further supporting the application in the fields of disaster prevention and reduction, emergency management, disaster rescue and the like.
[0054] The prediction method of disaster risk space provided by the preferred embodiment of the present application, as shown in Figure 1 The prediction method of disaster risk space includes the following steps:
[0055] Step S10, construct the adjacency matrix of each geographic unit, and perform high-order neighborhood expansion processing on the adjacency matrix to obtain the external environment of the geographic unit.
[0056] The present application takes into account the correlation of the environment inside and outside the geographic unit and the fine management demand of disaster, and in view of the difficulty of risk space prediction in the area not covered by monitoring data and the problem of not being timely and effective, a disaster risk space prediction method fusing the environment inside and outside the geographic unit is provided. According to the characteristics of the environment inside and outside the geographic unit and the high-order neighborhood relationship, the environment representation of the geographic unit and the disaster risk prediction based on semantic similarity are realized, and the process mainly includes: 1, on the basis of defining the internal environment of the geographic unit, the high-order neighborhood is introduced to construct the external environment of the geographic unit; 2, in view of the characteristics that the correlation between geographic units is affected by the difference of distance, the multi-order neighbor is realized by the graph attention network and the channel attention adaptive weighted aggregation to realize the environment representation of the geographic unit; 3, the similarity between different geographic unit environment representations is calculated to construct the semantic space correlation graph, and the multi-stage label propagation is realized by fusing global and local features to realize disaster risk space prediction. The present application fully considers the influence difference of geographic units at different distances on the current unit and the adjacency relationship between units, and can realize fine scale disaster risk space prediction based on the environment inside and outside the geographic unit, thereby further supporting the application in the fields of disaster prevention and reduction, emergency management, disaster rescue and the like.
[0057] Specifically, the disaster environment elements in each geographic unit are obtained, and the internal environment of each geographic unit is constructed according to the disaster environment elements, wherein the disaster environment elements include surface topography, climate elements, land cover, human economy and disaster-causing factors.
[0058] As shown in Figure 2 The internal environment of the geographic unit is defined as follows:
[0059] Disaster risk represents the risk level of disaster activities, which is closely related to the disaster environment and disaster-causing factors; the internal environment of a geographical unit refers to a set of core environmental elements that affect the formation and evolution process of disasters in the unit, and comprehensively considers the natural and human driving factors involved in the disaster system. The internal environment of a geographical unit is defined as follows:
[0060] InEnv i ={Terrain, Climate, Landcover, Humanity, Factor};
[0061] wherein, InEnv i represents the internal environment of geographical unit i, Terrain, Climate, Landcover, Humanity and Factor represent terrain, climate elements, land cover, humanity and main disaster-causing factors, respectively.
[0062] According to the internal environment of the geographical unit, the topological adjacency relationship between each geographical unit is obtained, and an adjacency matrix is constructed according to the topological adjacency relationship;
[0063] As shown in Figure 2 , the establishment process of the geographical unit adjacency matrix is as follows (based on the topological adjacency relationship of geographical units in the real space, an adjacency matrix is constructed):
[0064] Any element in the adjacency matrix represents the connection relationship and weight between geographical units (the weight represents the coefficient on the edge between two adjacent geographical units. In the present application, only the adjacent relationship is represented, so it is directly set to 1 after modification), which is as follows:
[0065]
[0066] wherein, E represents a set of adjacent relationships between geographical units, & represents whether adjacent or not, if geographical unit i and geographical unit j are adjacent, then A i,j is set to 1; if not adjacent, then A i,j is set to 0.
[0067] The adjacency matrix is subjected to high-order neighborhood expansion processing to obtain the external environment of the geographical unit.
[0068] As shown in Figure 2 , the construction process of the external environment of the geographical unit is as follows:
[0069] The k-order neighbor of geographical unit i refers to a set of points with a shortest path distance of k edges from it, and the k-order neighborhood of geographical unit i refers to a set of all units that can be reached from i through at most k edges, including all neighbor units from the first order to the k order.
[0070] The multi-order neighbors of the geographical unit not only include directly adjacent units, but also cover indirectly connected units through intermediate units. The geographical unit information in a larger spatial range is aggregated by establishing a high-order neighborhood (the meaning of the high-order neighborhood is a multi-order geographical unit, where the order refers to the step distance between units) to represent the external environment, and the process is as follows:
[0071]
[0072] wherein, represents the adjacency matrix after the high-order neighborhood is expanded, that is, the external environment in the present application, K is the order of the high-order neighborhood, A k is the adjacency matrix of the kth-order neighborhood, k∈{1,2,3,…,K}, which is obtained by multiplying the geographical unit adjacency matrix A by itself k-1 times.
[0073] Step S20, the same-order geographical unit weighting processing and neighborhood feature aggregation processing are performed on the external environment of the geographical unit, to obtain the geographical unit environment representation of each geographical unit.
[0074] Considering the influence difference of different geographical units in the neighborhood on the current unit, the present application adopts a graph attention network to distinguish the influence weight of different geographical units in the same-order neighborhood, and performs weighted aggregation on the same-order unit features according to the attention weight difference between geographical units. A channel attention mechanism is introduced to quantify the contribution degree of each order neighborhood to the current unit, to obtain the internal and external environment representation of the geographical unit aggregated by the high-order neighborhood.
[0075] Specifically, a target neighborhood in the external environment of the geographical unit is determined, and an attention coefficient between the current geographical unit and other geographical units in the target neighborhood is calculated.
[0076] As Figure 2 shown, considering the influence difference of different geographical units in the neighborhood on the current unit (the neighbor geographical unit is defined by the order, that is, the step distance instead of the spatial distance in reality. The step distance of the first-order unit is 1, that is, the unit directly adjacent to the current unit. The step distance of the second-order unit is 2), the influence weight of each geographical unit in the neighborhood range on the current unit is reflected by calculating the attention coefficient, and the expression of the attention coefficient is:
[0077]
[0078] wherein, i is the current geographical unit, j is the other geographical unit, and l is the layer number where the current geographical unit and the other geographical unit are located, is the attention coefficient, and LeakyReLU is an activation function used to calculate the attention coefficient W (l) is a weight matrix of linear transformation, is the input feature of the current geographical unit at the l-th layer, is the vector concatenation operation, is the transpose weight vector, is the input feature of other geographical units at the l-th layer.
[0079] The same-order neighbor geographical units corresponding to the current geographical unit are determined, and the same-order neighbor geographical units are weighted according to the attention coefficient to obtain a weighted feature; the current geographical unit is updated according to the weighted feature to obtain a target geographical unit feature.
[0080] As shown in Figure 2 , the weighting process of the same-order unit is as follows:
[0081] According to the difference of the attention coefficient, the features of the same-order neighbor geographical units of the current geographical unit are weighted and accumulated to update the features of the current unit, as shown below:
[0082]
[0083] wherein, is the normalized attention coefficient, N(i) represents a neighbor unit set of geographical unit i, represents a feature vector of geographical unit i at the l+1-th layer, is the feature of geographical unit i and its neighbor units at the l-th layer is calculated by the weight matrix W (l) After weighted summation, it is calculated again through the activation function σ.
[0084] The channel attention mechanism is adopted to perform attention weight distribution and weighted summation processing on the target geographical unit feature to obtain a geographical unit environment representation of each geographical unit.
[0085] As shown in Figure 2 , the aggregation process of the neighborhood feature is as follows:
[0086] To dynamically balance the contribution of features of each layer (i.e. geographical units at different distances), a channel attention mechanism is introduced to assign an attention weight to the output feature of each layer, and the features of different layers are aggregated by weighted summation, as shown below:
[0087]
[0088] wherein, H i represents an environment representation of aggregating geographical unit itself attributes and multi-order neighbor features, L is the order of high-order neighborhood, is the l-th layer feature, W c is the weight matrix of the channel attention, which is normalized by the sigmoid function to obtain the weight β (l) of the l-th layer feature.
[0089] Step S30, calculate the cosine similarity between each of the geographical unit environment representations, and construct a semantic space correlation graph according to the cosine similarity.
[0090] The present application calculates the semantic similarity between different unit environment representations by cosine distance, and constructs a semantic correlation graph between unit environments based on the K-nearest neighbor algorithm, thereby realizing the conversion of geographical space multi-dimensional attributes to semantic space environment representations.
[0091] Specifically, the cosine similarity between each of the geographical unit environment representations is calculated to obtain a similarity result; each geographical unit is converted into a geographical unit node, and the geographical unit environment representation corresponding to each geographical unit is converted into a geographical unit edge; a predetermined number of target geographical unit nodes are extracted from the geographical unit nodes in descending order according to the similarity result, and a semantic space correlation graph is constructed according to the target geographical unit nodes, the similarity result, and the geographical unit edge.
[0092] As shown in Figure 2 , the construction process of the semantic space correlation graph is as follows:
[0093] The cosine similarity is used to measure the similarity between each geographical unit environment representation, and the similarity calculation process of the environment representations of any two geographical units is as follows:
[0094]
[0095] wherein, H i and H j represent the feature vectors of the internal and external environments of the i th and j th geographical units, H i · H j represents the dot product of the two vectors, ||H i || and ||H j || represent the Euclidean norm of vectors H i and H j .
[0096] Further, the present application takes each geographical unit as a node in the graph, and its internal and external environment representations as node vectors. For node i in the graph, sort all other nodes in descending order according to the environment similarity, and select the top K nodes with the highest similarity, establish a connection edge between node i and the K nodes, and take the environment similarity as the weight of the edge, that is, the construction of the semantic space correlation graph is completed.
[0097] Step S40, node feature fusion processing and disaster risk space prediction processing are performed on the semantic space correlation graph and the geographical unit environment representation, to obtain a disaster risk space prediction result.
[0098] On this basis, the geographical unit environment representation and the local features in the semantic correlation graph are fused for clustering, and label propagation is performed on the unmarked nodes to realize disaster risk space prediction.
[0099] Specifically, local features of nodes in the semantic spatial correlation graph are extracted, and the local features of the nodes are standardized to obtain standardized local features of the nodes; the standardized local features of the nodes and the geographical unit environment representation are fused to obtain a fusion feature vector.
[0100] As shown in Figure 2 , the fusion process of node features is as follows:
[0101] The local features of each node are extracted from the constructed semantic spatial correlation graph, including the degree and clustering coefficient of the node, and the process is as follows:
[0102]
[0103] Among them, the degree d i of the node reflects the number of direct neighbors of the node i, and the clustering coefficient c i measures the connection tightness between the neighbors of the node i, A ij represents whether there is a connection edge between the node i and the node j (1 if connected, otherwise 0), and e i represents the number of connection edges between the neighbor nodes of the node i.
[0104] On this basis, the local features of the nodes are standardized to eliminate the differences between different feature quantities, and are fused with the environment representation of the nodes, as follows:
[0105]
[0106] Among them, f i is the fusion feature vector, H i is the environment representation of the geographical unit i, is the standardized degree of the node i, is the standardized clustering coefficient of the node i.
[0107] The K-means clustering algorithm is used to perform initial clustering processing on the fusion feature vector to obtain an initial label assignment; the label propagation algorithm is used to optimize the initial label assignment to obtain a disaster risk space prediction result.
[0108] Among them, the expression of the disaster risk space prediction result is:
[0109]
[0110] Among them, is the label of node i after the t+1th iteration, argmax is the maximum value solution, c is the preset label, N(i) is the neighbor node set of i, w ij is the edge weight between i and j, and δ is an indicator function, is the label of node j after the tth iteration, is an indicator function, when the label of node j in the tth iteration is c, the value of δ is 1, otherwise 0. After multiple iterations, the label of the node will gradually tend to be stable, and finally the optimized label assignment is obtained, that is, the result of disaster risk space prediction.
[0111] As Figure 2 shown, the prediction process of the disaster risk space is as follows:
[0112] Based on the K-means clustering algorithm, the initial clustering processing is performed on the fused feature vector of the node to obtain the initial label assignment, and the target of the clustering process is to minimize the distance between the node and the cluster center, as follows:
[0113]
[0114] Where, f i is the vector representation of node i fused with local features and global features, K is the number of clusters for clustering, C k represents the node set of cluster k, μ k is the center of cluster k, and represents the mean of all node feature vectors in the cluster.
[0115] The present application assigns a cluster label to each node through clustering, and the clustering result can be represented as a label set of nodes {y i}, wherein y i represents the cluster label of node i.
[0116] Further, based on the topological structure of the semantic space correlation graph, a label propagation algorithm is introduced to transfer the class information of the labeled nodes to the unlabeled nodes or noise nodes, so as to realize the completion and correction of the disaster risk label.
[0117] The present application has the beneficial effects that: the present application provides a disaster risk space prediction method fusing the internal and external environment of geographic units, models the internal and external environment by integrating the properties of the geographic units themselves and the multi-order neighbor features, and obtains the internal and external environment representation of the units based on the graph attention network and the channel attention adaptive aggregation of multi-order neighbors, so as to fuse the global and local features to predict the disaster risk space distribution through multi-stage label propagation. The present application not only considers the multi-dimensional properties inside the geographic units and the external correlation features, but also considers the influence difference of geographic units at different distances on the current unit, and good effects are achieved in the representation of the geographic unit environment and the prediction of the disaster risk space.
[0118] In summary, the present application takes into account the correlation between the internal and external environment of the geographical unit and the fine management needs of disasters, defines the internal environment of the geographical unit and constructs the external environment, realizes the geographical unit environment representation based on the graph attention network and the channel attention adaptive weighted aggregation of multi-order neighbors, and realizes the multi-stage label propagation for disaster risk space prediction by fusing global and local features. The present application is helpful for discovering the disaster situation of urban grassroots communities or marginal areas, thereby further supporting the application in the fields of disaster prevention and reduction, emergency management, disaster rescue, etc.
[0119] Example introduction: The present application takes the collected 2021 flood disaster related data of a certain city as a specific implementation example implementation scene, and carries out the embodiment method steps of geographical unit internal and external environment modeling, aggregation of high-order neighborhood geographical unit environment representation, disaster risk prediction based on semantic space similarity, etc. The specific steps are as follows:
[0120] Step 1, geographical unit internal and external environment modeling:
[0121] 1. Definition of internal environment of geographical unit: Considering that disaster is a complex earth surface hetero-system composed of disaster environment, disaster-causing factors, disaster-bearing body and disaster situation, and that the risk of flood disaster is closely related to disaster environment and disaster-causing factors, the internal environment of geographical unit i is defined as follows:
[0122] InEnv i ={Terrain, Climate, Landcover, Humanity, Factor};
[0123] InEnv i represents the internal environment of geographical unit i, Terrain, Climate, Landcover, Humanity and Factor represent terrain, climate elements, land cover, humanity and main disaster-causing factors, respectively. Terrain describes the terrain characteristics of the geographical unit, including average elevation, elevation standard deviation, average slope and average slope direction; climate characteristics reveal the climate conditions in the region, including average annual evaporation, average annual precipitation, average annual air pressure and average annual humidity; land cover reflects the distribution of surface cover types, including impervious surface area and water area; humanity describes human activities and population distribution, including population density and average annual nighttime light index. Precipitation is the main disaster-causing factor of flood disaster. The specific source data of various unit attributes is shown in Table 1.
[0124] Table 1: Source data introduction of unit attributes
[0125]
[0126] The study area is divided into a 500m x 500m grid, resulting in 31,021 geographic units in the city. Based on the above multi-source data, spatial statistical methods are used to calculate the surface topography, climate factors, land cover, human economy and average precipitation in each geographic unit as the internal environment of the geographic unit.
[0127] 2. Establishment of geographic unit adjacency matrix:
[0128] Based on the topological adjacency relationship of the 500m x 500m grid obtained by division in the real space, an initial adjacency matrix is constructed to represent the spatial connection between geographic units, as follows:
[0129]
[0130] where E represents the set of adjacent relationships between geographic units, and if geographic unit i and geographic unit j are adjacent, A i,j represents the weight w i,j of the connection edge between geographic units; if not adjacent, A i,j is set to 0. Since the adjacency relationship between geographic units is mutual, the adjacency matrix A of the undirected graph formed is symmetric, i.e. A i,j =A j,i .
[0131] 3. Construction of geographic unit external environment:
[0132] On the basis of the initial adjacency matrix, further iteration and traversal methods are used to identify and incorporate the multi-order neighbors of all geographic units, i.e. not only including the "neighbors" directly connected, but also considering the "neighbors of neighbors" connected through multiple intermediate units. In this process, the k-order neighbors of geographic unit i refer to the set of points with a shortest path distance of k edges from it, while the k-order neighborhood of geographic unit i refers to the set of all units that can be reached from i through at most k edges, which includes all neighbor units from the 1st order to the kth order, as follows:
[0133]
[0134] where, represents the adjacency matrix after high-order expansion, K is the order of the high-order neighborhood, A k is the adjacency matrix of the kth-order neighborhood, k ∈ {1, 2, 3, …, K}. By calculating the kth power of the initial adjacency matrix, it is essentially multiplying the initial adjacency matrix by itself k-1 times, i.e. A 1 =A, A 2 =A 1 *A, A 3 =A 2 *A, …, Ak = A k-1 * A.
[0135] Step 2, aggregation of high-order neighborhood geographical unit environment representation:
[0136] 1. Attention coefficient calculation:
[0137] The self-attention mechanism based on the graph attention network assigns different weights to each neighbor of a geographical unit, thereby distinguishing the importance of different neighbor units. The steps of attention coefficient calculation are as follows:
[0138] Neighbor unit weight calculation:
[0139] For the spatial heterogeneity correlation characteristics of flood risk, the influence weight of each neighbor unit on the current unit is reflected by calculating the attention coefficient, as shown in the formula below:
[0140]
[0141] where, and are the input features of geographical units i and j at the l-th layer, W l) is the weight matrix of linear transformation, || represents the vector concatenation operation, is the transpose weight vector for dot product operation with the concatenated feature vector, LeakyReLU is the activation function, and is used to calculate the attention coefficient
[0142] The initial feature vector of each geographical unit is composed of its own attribute data. For example, the initial features of geographical unit i are [168.35, 285.52, 20.13, 11.15, 0.190, 60, 1.21, 1237.45, 1295.5, 983.85, 60.21, 93.61], indicating that the average slope direction in this geographical unit is 168.35 degrees, the average elevation is 285.52 meters, the elevation standard deviation is 20.13 meters, the average slope is 11.15 degrees; the impervious surface area is 0.19 square kilometers, the water area is 0 square kilometers; the average population density is 60 people per square kilometer, the average nighttime light index is 1.21; the annual evaporation is 1237.45 millimeters, the annual precipitation is 1295.5 millimeters, the annual average pressure is 983.85 hundred pascals, the annual average relative humidity is 60.21%, and the daily precipitation is 93.61 millimeters.
[0143] Attention coefficient normalization:
[0144] To facilitate the comparison and aggregation of weights between different neighbor units, the attention coefficients are normalized by the softmax function, so that the sum of the attention weights of all neighbor units of a geographic unit is 1, and the process is as follows:
[0145]
[0146] wherein, is the normalized attention coefficient, represents the contribution weight of neighbor unit j to the feature update of the current unit i, and N(i) represents the neighbor unit set of unit i.
[0147] 2. Weighting of geographic units of the same order:
[0148] According to the difference of the attention coefficients, the features of each neighbor geographic unit are weighted and accumulated, and the information of the neighbor geographic units is aggregated into the current unit, so that the environmental representation of the geographic unit is fused with the neighbor features while retaining the properties of the unit itself, and the process is as follows:
[0149]
[0150] wherein, represents the feature vector of node i at the lth layer, which is calculated by weighting and summing the features of node i and its neighbor nodes at the lth layer through the attention mechanism, and then passing through the activation function σ.
[0151] 3. Neighborhood feature aggregation:
[0152] The importance of the neighborhood information captured by different attention layers to the prediction result is different. In order to dynamically balance the contribution of each layer feature (i.e. geographic units of different distances), a channel attention mechanism is introduced to assign an attention weight to the output feature of each layer, and the features of different layers are aggregated by weighted summation, and the process is as follows:
[0153]
[0154] wherein, L is the order of high-order neighborhood, is the lth layer feature, β (l) is the weight learned by the channel attention mechanism, which represents the importance of the lth layer feature, W c is the weight matrix of the channel attention, which is normalized by the sigmoid function. Through the above calculation steps, the representation H i combining the environmental feature weights of spatial units of different distance ranges can be obtained.
[0155] Furthermore, to address the problem of scarce labeled samples in flood disaster scenarios, this embodiment adopts a training strategy based on semi-supervised learning, which uses limited labeled data and a large amount of unlabeled data to jointly train the model, thereby achieving effective representation of unlabeled nodes under limited label conditions.
[0156] Step 3: Disaster risk prediction based on semantic space similarity:
[0157] 1. Semantic space association graph construction:
[0158] After obtaining the environmental characterization H of each geographical unit i Building upon this foundation, geographic units are treated as nodes in a graph structure. Based on the K-nearest neighbor concept, the K closest nodes are selected and connected by edges to construct a semantic association graph of geographic units based on environmental similarity. This achieves the mapping from geographic units in natural space to nodes in the semantic space. The steps for constructing the semantic space association graph are as follows:
[0159] Environmental similarity calculation:
[0160] After obtaining the environmental representations of each geographic unit, cosine similarity is used to measure the similarity between the environmental representations of each geographic unit. The calculation of the environmental representation similarity between any two geographic units is as follows:
[0161]
[0162] Among them, H i and H j H represents the feature vectors of the internal and external environments of the i-th and j-th geographic units, respectively. i ·H j H represents the dot product of two vectors, ||H i ||and||H j || represents vector H i and H j The Euclidean norm of a vector is the square root of the sum of the squares of its components. Environmental similarity values range from -1 to 1, with values closer to 1 indicating greater similarity. Pairwise comparisons are performed on the environmental representations of all geographic units, and their cosine similarity is calculated sequentially.
[0163] Graph structure establishment:
[0164] Each geographical unit within the study area is treated as a node in the graph, and its internal and external environmental representation vectors are used as node features. For node i in the graph, the environmental similarity (EnvSim(H)) is first calculated. i H j Sort all other nodes in descending order, then select the K nodes N with the highest similarity. K (i). At node i and N KAdd a connection edge between each node j in (i) ("undirected edge" refers to the edge without direction, relative to the directed edge), and set its weight to the corresponding similarity value, to establish the environment representation similarity driven semantic space correlation graph.
[0165] To evaluate the impact of the number of nearest neighbors K on the prediction of disaster risk space, and avoid the graph structure being too dense or sparse, 1 to 10 nearest neighbor nodes are selected as the value of K. For each K value, the corresponding semantic space correlation graph is first constructed, then clustering analysis is performed, and the accuracy, average silhouette coefficient and integrity three indicators are used for evaluation (the higher the value of each indicator, the better the representation effect). The value of each indicator is higher, the higher the value of each indicator, the better the representation effect, to determine the most suitable number of nearest neighbor nodes K.
[0166] 2, node feature fusion:
[0167] On the basis of the semantic space correlation graph, the local features of the nodes in the graph are extracted and fused with the node environment representation to obtain a feature representation that combines the local information and global information of the nodes.
[0168] The local features of each node are extracted from the constructed semantic space correlation graph, including the degree and clustering coefficient of the node, as shown in the following process:
[0169]
[0170] Among them, the degree of node d i reflects the number of direct neighbors of node i, the clustering coefficient c i measures the connection tightness between the neighbors of node i, A ij represents whether there is a connection edge between node i and node j (1 if connected, otherwise 0), e i represents the number of connection edges between the neighbor nodes of node i, in order to avoid the denominator being zero, when d i <2, c i = 0.
[0171] After extracting the local features of the nodes, standardization processing is performed to eliminate the differences between different feature dimensions, and the node environment representation is fused, as shown below:
[0172]
[0173] Among them, f i is the fused feature vector, H i is the environment representation of geographic unit i, is the standardized degree of node i, is the standardized clustering coefficient of node i. By combining the global features (environment representation) of the nodes with the local features (degree and clustering coefficient), the final feature vector fi as input for the subsequent clustering analysis.
[0174] 3. Disaster risk spatial prediction:
[0175] The disaster risk spatial prediction of geographical units is converted into a node class determination task in the semantic space, and a multi-stage label propagation method is designed to obtain the label information of all nodes, i.e., the disaster risk of geographical units. The steps of the multi-stage label propagation are as follows:
[0176] Geographical unit disaster risk classification:
[0177] According to the proportion of water accumulation area in the geographical unit, the flood disaster risk level is divided into three categories: extremely low risk, light risk and high risk, as shown in Table 2 below.
[0178] Table 2. Disaster risk classification of geographical units
[0179] Proportion of affected area Category Explanation <0.5％ K1 Very low risk 0.5％～5％ K2 Light to moderate risk 5% or more K3 Medium to high risk
[0180] Node clustering:
[0181] Based on the K-means clustering algorithm, the feature vector f i is clustered, and the nodes are assigned to different clusters according to their comprehensive features. The goal of the clustering process is to minimize the distance between the nodes and the cluster centers. The specific optimization objective function is as follows:
[0182]
[0183] where f i is the vector representation of node i after feature fusion, K is the number of clusters for clustering, C k represents the node set of cluster k, i.e., all nodes assigned to cluster k, μ k is the center of cluster k, representing the mean of all node feature vectors in the cluster. Through clustering, each node is assigned a cluster label, and the clustering result is represented as the node label set {y i}, where y i represents the cluster label of node i.
[0184] Label propagation:
[0185] Based on the obtained clustering labels, a label propagation algorithm is introduced to propagate the node information of existing labels to unmarked nodes or noise nodes through the structure information of the graph. In the label propagation process, nodes constantly update their own labels according to the label information of their neighbors. The update rule of label propagation is as follows:
[0186]
[0187] where, represents the label of node i after the t+1th iteration, N(i) represents the neighbor node set of node i, w ij represents the edge weight between node i and node j, is an indicator function. When the label of node j in the tth iteration is c, the value of δ is 1, otherwise 0. After multiple iterations, the label of the node will gradually tend to be stable and remain unchanged, that is, the optimized label assignment of the whole graph is obtained.
[0188] In this embodiment, 10%, 30% and 50% of the geographical units are selected as training data, and the remaining geographical units are used for disaster risk prediction. The weighted accuracy of the trained model is 0.6550, 0.7447 and 0.8142 respectively.
[0189] The present application takes into account the correlation between the internal and external environment of the geographical unit and the fine management requirements of the disaster, and defines the internal environment of the geographical unit and constructs the external environment. In view of the characteristic that the correlation strength between geographical units decreases with distance, the geographical unit environment representation is realized based on graph attention network and channel attention adaptive weighted aggregation of multi-order neighbors; and then the similarity between different geographical unit environment representations is calculated to construct a semantic space correlation graph, and multi-stage label propagation is realized by fusing global and local features to realize disaster risk space prediction. The present application is helpful for discovering the disaster situation of urban grassroots communities or marginal areas, and thus further supports the application in the fields of disaster prevention and mitigation, emergency management, disaster rescue, etc.
[0190] Further, as shown in Figure 3 based on the above-mentioned disaster risk space prediction method, the present application also correspondingly provides a disaster risk space prediction system, wherein the disaster risk space prediction system comprises:
[0191] An external environment construction module 51 is configured to construct an adjacency matrix of each geographical unit, and perform high-order neighborhood expansion processing on the adjacency matrix to obtain an external environment of the geographical unit.
[0192] An environment representation generation module 52 is configured to perform same-order geographical unit weighting processing and neighborhood feature aggregation processing on the external environment of the geographical unit to obtain a geographical unit environment representation of each geographical unit.
[0193] A semantic space correlation graph construction module 53 is configured to calculate the cosine similarity between each geographical unit environment representation, and construct a semantic space correlation graph according to the cosine similarity.
[0194] The prediction result output module 54 is configured to perform node feature fusion processing on the semantic space correlation graph and the geographical unit environment representation and disaster risk space prediction processing, to obtain a disaster risk space prediction result.
[0195] Further, as shown in the disaster risk space prediction method and system, the present application also provides a terminal, which comprises a processor 10, a memory 20 and a display 30. Figure 4 Figure 4 Only part of the components of the terminal are shown, but it should be understood that all the shown components are not required, and more or less components can be implemented instead.
[0196] The memory 20 can be an internal storage unit of the terminal in some embodiments, such as a hard disk or a memory of the terminal. The memory 20 can also be an external storage device of the terminal in other embodiments, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. Further, the memory 20 can include both an internal storage unit and an external storage device of the terminal. The memory 20 is configured to store application software and various data installed on the terminal, such as program codes of the terminal, etc. The memory 20 can also be configured to temporarily store data that has been output or is to be output. In an embodiment, the memory 20 stores a disaster risk space prediction program 40, which can be executed by the processor 10, so as to implement the disaster risk space prediction method.
[0197] The processor 10 can be a central processing unit (CPU), a microprocessor or other data processing chip in some embodiments, configured to run program codes or process data stored in the memory 20, such as to execute the disaster risk space prediction method, etc.
[0198] The display 30 can be an LED display, a liquid crystal display, a touch liquid crystal display, an OLED (Organic Light-Emitting Diode) touch, etc. in some embodiments. The display 30 is configured to display information of the terminal and a visualized user interface.
[0199] In an embodiment, the processor 10 implements the steps of the disaster risk space prediction method when executing the disaster risk space prediction program 40 in the memory 20.
[0200] In summary, the present application provides a disaster risk space prediction method, system and terminal, the method comprising: constructing an adjacency matrix of each geographic unit, and performing high-order neighborhood expansion processing on the adjacency matrix to obtain an external environment of the geographic unit; performing the same-order geographic unit weighting processing and neighborhood feature aggregation processing on the external environment of the geographic unit to obtain a geographic unit environment representation of each geographic unit; calculating the cosine similarity between each geographic unit environment representation, and constructing a semantic space correlation graph according to the cosine similarity; performing node feature fusion processing and disaster risk space prediction processing on the semantic space correlation graph and the geographic unit environment representation to obtain a disaster risk space prediction result. The present application constructs an adjacency matrix of the geographic unit and performs high-order neighborhood expansion processing to obtain the internal and external environments of the geographic unit, and then performs the same-order geographic unit weighting processing, neighborhood feature aggregation processing and similarity calculation according to the internal and external environments to construct the geographic unit environment representation and the semantic space correlation graph, thereby effectively improving the accuracy of the disaster risk space prediction result.
[0201] It should be noted that in this document, the terms "comprising", "including", or any other variant thereof are intended to cover non-exclusive inclusions, so that processes, methods, articles or terminals including a series of elements not only include those elements, but also include other elements not explicitly listed, or include elements inherent to such processes, methods, articles or terminals. Without more limitations, the element defined by the statement "including a" does not exclude the presence of another identical element in the process, method, article or terminal including the element.
[0202] Of course, those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by a computer program instructing relevant hardware (such as a processor, a controller, etc.) to complete, and the program can be stored in a computer-readable computer-readable storage medium, and the program can include the processes of the above-mentioned method embodiments when executed. The computer-readable storage medium can be a memory, a magnetic disc, an optical disc, etc.
[0203] It should be understood that the application of the present application is not limited to the above examples, and those skilled in the art can improve or modify the above description, all of which should be within the scope of protection of the appended claims of the present application.
Claims
1. A method of predicting a disaster risk space, characterized by, The method for predicting the disaster risk space comprises the following steps: An adjacency matrix of each geographical unit is constructed, and high-order neighborhood expansion processing is performed on the adjacency matrix to obtain an external environment of the geographical unit; The external environment of the geographical unit is subjected to same-order geographical unit weighting processing and neighborhood feature aggregation processing to obtain a geographical unit environment representation of each geographical unit; Cosine similarity between each geographical unit environment representation is calculated, and a semantic space correlation graph is constructed according to the cosine similarity; Node feature fusion processing and disaster risk space prediction processing are performed on the semantic space correlation graph and the geographical unit environment representation to obtain a disaster risk space prediction result.
2. The method of claim 1, wherein, The construction of the adjacency matrix of each geographical unit and the high-order neighborhood expansion processing of the adjacency matrix to obtain the external environment of the geographical unit specifically comprises the following steps: Disaster environment elements in each geographical unit are obtained, and an internal environment of each geographical unit is constructed according to the disaster environment elements, wherein the disaster environment elements include surface topography, climate elements, land cover, human economy and disaster-causing factors; A topological adjacency relationship between each geographical unit is obtained according to the internal environment of the geographical unit, and an adjacency matrix is constructed according to the topological adjacency relationship; High-order neighborhood expansion processing is performed on the adjacency matrix to obtain the external environment of the geographical unit.
3. The method of claim 1, wherein, The same-order geographical unit weighting processing comprises attention coefficient calculation processing and weighting processing; and the neighborhood feature aggregation processing comprises feature updating processing, attention weight distribution and weighted summation processing. The same-order geographical unit weighting processing and the neighborhood feature aggregation processing of the external environment of the geographical unit to obtain the geographical unit environment representation of each geographical unit specifically comprise the following steps: A target neighborhood in the external environment of the geographical unit is determined, and an attention coefficient between a current geographical unit and other geographical units in the target neighborhood is calculated; A same-order neighbor geographical unit corresponding to the current geographical unit is determined, and the same-order neighbor geographical unit is subjected to weighting processing according to the attention coefficient to obtain a weighted feature; Feature updating processing is performed on the current geographical unit according to the weighted feature to obtain a target geographical unit feature; Channel attention mechanism is adopted to perform attention weight distribution and weighted summation processing on the target geographical unit feature to obtain the geographical unit environment representation of each geographical unit.
4. The method of claim 3, wherein, The expression of the attention coefficient is as follows: where i is the current geographical unit, j is the other geographical unit, and l is the layer number of the current geographical unit and the other geographical unit, is an attention coefficient, LeakyReLU is an activation function, and W (l) is a weight matrix of a linear transformation, is an input feature of the current geographical unit at the lth layer, and || is a vector concatenation operation, is a transposed weight vector, is an input feature of the other geographical unit at the lth layer.
5. The method of claim 1, wherein, The calculation of the cosine similarity between each geographical unit environment representation and the construction of the semantic space correlation graph according to the cosine similarity specifically comprise the following steps: The cosine similarity between each geographical unit environment representation is calculated to obtain a similarity result; Each geographical unit is converted into a geographical unit node, and each geographical unit environment representation corresponding to each geographical unit is converted into a geographical unit edge; A preset number of target geographical unit nodes are extracted from the geographical unit nodes in descending order according to the similarity result, and a semantic space correlation graph is constructed according to the target geographical unit nodes, the similarity result and the geographical unit edge.
6. The method of claim 4, wherein, The node feature fusion processing and disaster risk space prediction processing are performed on the semantic space correlation graph and the geographical unit environment representation, and a disaster risk space prediction result is obtained, and the method specifically comprises the following steps: Local features of nodes in the semantic space correlation graph are extracted, and the local features of the nodes are standardized to obtain standardized local features of the nodes; The standardized local features of the nodes and the geographical unit environment representation are fused to obtain a fusion feature vector; An initial clustering processing is performed on the fusion feature vector by using a K-means clustering algorithm to obtain an initial label assignment; The initial label assignment is optimized by using a label propagation algorithm to obtain a disaster risk space prediction result.
7. The method of claim 6, wherein, An expression of the disaster risk space prediction result is as follows: wherein, is the label of i after the t+1th iteration, argmax is the solution of maximum value, c is the preset label, N(i) is the neighbor node set of i, w ij is the edge weight between i and j, δ is an indicator function, is the label of j after the tth iteration.
8. A disaster risk spatial prediction system characterized by, The disaster risk space prediction system comprises: An external environment construction module configured to construct an adjacency matrix of each geographical unit and perform high-order neighborhood expansion processing on the adjacency matrix to obtain a geographical unit external environment; An environment representation generation module configured to perform same-order geographical unit weighting processing and neighborhood feature aggregation processing on the geographical unit external environment to obtain a geographical unit environment representation of each geographical unit; A semantic space correlation graph construction module configured to calculate a cosine similarity between each geographical unit environment representation and construct a semantic space correlation graph according to the cosine similarity; A prediction result output module configured to perform node feature fusion processing and disaster risk space prediction processing on the semantic space correlation graph and the geographical unit environment representation to obtain a disaster risk space prediction result.
9. A terminal, characterized by comprising: The terminal comprises a memory, a processor, and a disaster risk space prediction program stored on the memory and executable on the processor, and the disaster risk space prediction program is executed by the processor to implement the steps of the disaster risk space prediction method according to any one of claims 1-7.
10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a disaster risk space prediction program, and the disaster risk space prediction program is executed by the processor to implement the steps of the disaster risk space prediction method according to any one of claims 1-7.