A geological disaster risk prediction and evaluation method

Abstract

The application discloses a geological disaster risk prediction and evaluation method, which comprises the following steps: collecting multi-scale and multi-source data of a research area, dividing the research area into N grid units, taking the grid units as graph nodes, and obtaining an edge set corresponding to each time step; calculating the weight of each connection edge between two graph nodes, and constructing a comprehensive adjacency matrix of each graph node at different time steps; constructing a multi-source spatio-temporal graph data matrix according to the standardized data set of each grid unit, introducing the comprehensive adjacency matrix, and obtaining a global spatio-temporal feature vector of the graph node; constructing a physical information neural network, inputting the global spatio-temporal feature vector into the physical information neural network, and predicting the real-time geological disaster occurrence probability of the position of each graph node in the research area; and evaluating the real-time geological disaster risk of the research area according to the geological disaster occurrence probability. The application provides a geological disaster risk prediction and evaluation method based on the coupling of a spatio-temporal graph convolution network and a physical information neural network.

Description

Technical Field

[0001] This invention relates to the field of geological disaster monitoring and analysis, and specifically to a method for predicting and evaluating geological disaster risks. Background Technology

[0002] Geological disasters refer to geological processes (phenomena) formed under the influence of natural or human factors that cause damage and loss to human life, property, and the environment. Once a geological disaster occurs, it generally causes serious casualties and economic losses. Therefore, the monitoring and predictive analysis of geological disaster risks are particularly important. Current technologies mostly use only single or a few types of monitoring data, failing to effectively integrate multi-source heterogeneous data from ground monitoring, meteorology, hydrology, and geological exploration, leading to highly biased evaluation results. The occurrence of geological disasters is highly complex and uncertain. Existing physical and mechanical methods and statistical analysis methods have significant technical shortcomings. Physical and mechanical methods require a large number of high-precision physical and mechanical parameters of soil and rock, and struggle to consider the impact of complex geological conditions and human activities, resulting in low computational efficiency and inability to meet real-time early warning needs. Statistical analysis methods rely on historical disaster data, making it difficult to capture the nonlinear characteristics and spatiotemporal heterogeneity of geological disaster evolution, limiting prediction accuracy, and failing to quantify uncertainty. Summary of the Invention

[0003] To address the aforementioned shortcomings of existing technologies, this invention provides a geological disaster risk prediction and evaluation method to solve problems such as difficulty in obtaining physical model parameters and insufficient extraction of spatiotemporal features in existing technologies.

[0004] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows: A method for predicting and assessing geological disaster risks is provided, comprising: Step S1: Collect multi-scale, multi-source data of the study area and preprocess it to obtain a standardized dataset; divide the study area into... N Each grid cell serves as a graph node, and the set of edges is obtained based on the connection relationships between the graph nodes. Step S2: Based on the type of connecting edge between two graph nodes in the edge set, calculate the weight of each type of connecting edge between the two graph nodes, and construct the comprehensive adjacency matrix for each graph node; Step S3: Construct a multi-source spatiotemporal graph data matrix based on the standardized dataset of each grid cell, construct a multi-scale spatiotemporal graph convolutional network, and introduce a comprehensive adjacency matrix to extract multi-scale spatiotemporal features from the multi-source spatiotemporal graph data matrix to obtain the global spatiotemporal feature vector of the graph node; Step S4: Construct a physical information neural network, including a data-driven branch and a physical constraint branch. Input the global spatiotemporal feature vector into the physical information neural network to predict the real-time probability of geological disasters occurring at the location of each graph node in the study area. Step S5: Evaluate the real-time geological disaster risk in the study area based on the probability of geological disaster occurrence.

[0005] Further, step S1 includes: Step S11: Collect microscale geological data, mesoscale geological data, and macroscale environmental data of the study area to obtain multiscale multi-source data; preprocess the multiscale multi-source data and map the multiscale multi-source data to the [0,1] interval to obtain a standardized dataset of multiscale multi-source data; Step S12: Divide the study area into N Each grid cell serves as a graph node. v n This yields the graph node set. , For the first N 1. Graph nodes; based on the connections between graph nodes, obtain the edges connecting two graph nodes, and thus obtain the edge set. , n Number the nodes in the graph; The connection relationships between graph nodes include spatial adjacency, geological connection, and hydrological connectivity between two grid cells; the types of connecting edges include spatial adjacency edges, geological connection edges, and hydrological connectivity edges.

[0006] Further, step S2 includes: Step S21: Based on the edge set Determine the types of edges connecting nodes in two graphs and calculate the weight of each type of edge. , , The weight of spatially adjacent edges. The weight of the geological structure edge, The weight of the hydrologically connected edge; For spatial adjacent edges: based on the nodes of the two graphs Euclidean distance between Calculate the weight of spatially adjacent edges. ; ;in, Spatial scale parameter; Regarding geological structural edges: based on the nodes in the two diagrams The difference in distance between the nearest geological structural zone or fault Calculate the weight of geological structural edges ; ;in, Parameters at the geological structural scale; For hydrological connectivity edges: based on the nodes in the two graphs Length of water flow path between Calculate the weight of hydrologically connected edges ; ;in, For hydrological scale parameters; Step S22: Based on the weights of each connecting edge Construct the spatial adjacency weight matrix for each graph node Geological structure weight matrix Hydrological connectivity weight matrix Calculate the comprehensive adjacency matrix for each graph node. ; ; in, These are the weight coefficients for spatially adjacent edges, geological structural edges, and hydrologically connected edges, respectively.

[0007] Further, step S3 includes: Step S31: Construct a multi-source spatiotemporal graph data matrix based on the standardized dataset of each grid cell, expand the multi-source spatiotemporal graph data matrix according to the time step to obtain the spatial data matrix of each time step, and divide the comprehensive adjacency matrix into a multi-scale adjacency matrix set according to different data scales; Step S32: Normalize the adjacency matrix of each data scale to obtain a normalized adjacency matrix, construct a multi-scale spatiotemporal graph convolutional network including a multi-scale graph convolutional layer and a causal temporal convolutional layer, input the spatial data matrix of each time step into the multi-scale graph convolutional layer, introduce the normalized adjacency matrix, and extract the spatiotemporal feature tensor. Step S33: Input the spatiotemporal feature tensor into the causal temporal convolutional layer, perform causal temporal convolution on the temporal feature tensor of each graph node, and extract the enhanced spatiotemporal feature tensor by weighting the spatial attention weights of each graph node to its neighboring graph nodes. Step S34: Perform global average pooling and global max pooling on the enhanced spatiotemporal feature tensor to aggregate the information in the time dimension, obtain the aggregated features in the time dimension, and perform dimensionality reduction and nonlinear transformation through a fully connected layer to output the global spatiotemporal feature vector of the graph nodes.

[0008] Further, step S31 includes: Step S311: Construct a multi-source spatiotemporal graph data matrix based on the standardized dataset of each grid cell. , TThe time step for collecting multi-scale, multi-source data. F Data dimensions for each grid cell, It is the set of real numbers; Step S312: Convert the multi-source spatiotemporal graph data matrix Expanding by time step, we obtain the spatial data matrix for each time step. ; synthesize the adjacency matrix It is divided into a multi-scale adjacency matrix set according to different data scales. , K The number of data scales, For the first K Adjacency matrix of several data scales; For each data scale k From the comprehensive adjacency matrix Extract path hop count Comprehensive adjacency weight And by filtering out weak associations, a multi-scale adjacency matrix set is obtained. ; ; in, For the first k Adjacency matrix at several data scales Representing nodes in two graphs Path hop count between Data scale k The overall adjacency weight threshold is as follows. The baseline for path hop count is set for each data scale.

[0009] Further, step S32 includes: Step S321: Normalize the adjacency matrix for each data scale to obtain the normalized adjacency matrix. ; ; in, For the first k A degree matrix for each data scale. It is the identity matrix; Step S322: Construct a multi-scale spatiotemporal graph convolutional network, including multi-scale graph convolutional layers and causal temporal convolutional layers, and process the spatial data matrix of each time step. Input a multi-scale graph convolutional layer and introduce a normalized adjacency matrix. ,exist K Perform graph convolution operations on various data scales to extract local features at different spatial scales. ; ; in, lThe number of the graph convolutional layer. It is the ReLU activation function. For the first k On the data scale, the first l The weight matrix of each graph convolutional layer For the first k On the data scale, the first l The bias vector of each graph convolutional layer This represents the local features output by the convolutional layer in the previous image; Step S323: Combine local features at different spatial scales By splicing the data, we can obtain the time steps. t Multiscale spatial features ; ; in, This indicates a splicing operation. For the first K Local features at each data scale; Step S324: Set the time step t Multiscale spatial features Stack them in chronological order to form a spatiotemporal feature tensor. ; ; in, For time step T Multi-scale spatial characteristics.

[0010] Further, step S33 includes: Step S331: Convert the spatiotemporal feature tensor Perform a dimensionality transformation to convert it into a graph node-time-feature dimension, obtaining the time series feature tensor for each graph node. Then, in the causal temporal convolutional layer, causal temporal convolution is performed on the time-series feature tensor of each graph node to extract local features in the time dimension. ; ; in, This indicates a dimension transformation operation. p This refers to the number of the causal temporal convolutional layer. q Number the convolution kernels for each time step. Q The size of the convolution kernel at each time step. The position of the convolution kernel at time step of the previous causal temporal convolutional layer. The output, For the first p Layered causal temporal convolutional layer with convolutional kernels at each time step q The weight matrix, For the firstp Bias vector of a causal temporal convolutional layer; Step S332: Local features in the time dimension Converting back to time-graph node-feature dimension yields the local features of the transformed time dimension. The spatiotemporal feature tensors are then concatenated to obtain the spatiotemporal feature tensors. ; ; in, For the first P Local features of the temporal dimension output by a causal temporal convolutional layer; Step S333: Calculate the spatial attention weight of each graph node to its neighboring graph nodes. And utilize spatial attention weights spatiotemporal feature tensor By performing a weighted summation, we obtain the spatially weighted spatiotemporal feature tensor. ; ; in, For graph nodes i The set of adjacent graph nodes, This represents a spatial attention multilayer perceptron. For graph nodes i Adjacent graph nodes j The spatiotemporal feature tensor; Step S334: Based on the spatially weighted spatiotemporal feature tensor Calculate the attention weights over time Utilizing time attention weights spatiotemporal feature tensor We perform a weighted summation to obtain the time-weighted spatiotemporal feature tensor. ; ; in, For time step Spatial weighted spatiotemporal feature tensor; Step S335: Convert the time-weighted spatiotemporal feature tensor With spatiotemporal feature tensor Perform residual connections to obtain enhanced spatiotemporal feature tensors. ; .

[0011] Further, step S34 includes: Step S341: For the enhanced spatiotemporal feature tensor Perform global average pooling and global max pooling, aggregating information from the time dimension to obtain the average pooling output. and max pooling output ; ; in, For time step T Enhanced spatiotemporal feature tensor; Step S342: Output the average pooling. and max pooling output By concatenating the features, we obtain the features aggregated along the time dimension. ; ; Step S343: Process features through a fully connected layer Perform dimensionality reduction and nonlinear transformation to output the global spatiotemporal feature vectors of the graph nodes. ; ; in, These are the weight matrix and bias vector of the fully connected layer, respectively.

[0012] Further, step S4 includes: Step S41: Construct a physical information neural network, including a data-driven branch and a physical constraint branch. The physical constraint branch uses the geomechanical control equations as physical constraints. The governing equations of geomechanics are: ; in, For the cohesion of the soil and rock mass, For normal stress, u Pore ​​water pressure, It is the internal friction angle. For shear stress, For geological disaster safety factor; when At that time, geological disasters are prone to occur. At that time, geological disasters are less likely to occur; Step S42: Convert the global spatiotemporal feature vector In the input data-driven branch, a fully connected neural network is used to predict the probability of geological disasters occurring at the locations of graph nodes. and displacement field ; ; in, These are fully connected neural networks for predicting the probability of geological disasters and displacement fields, respectively. Step S43: Physical constraint branch based on the predicted displacement field of the graph nodes Calculated predicted stress field According to the stress field Calculating the geological hazard safety factor using normal stress. ; ; in, This represents the elastic stiffness matrix of the soil and rock mass at the location of the node in the graph. Represents the displacement field gradient operator; Step S44: Based on the geological disaster safety factor Constructing the physical loss function ; ; in, The regularization coefficient is . This is a gradient operator based on the geological hazard safety factor of the nodes in the study area. Step S45: Set the physical loss threshold ,like Then determine the probability of geological disaster occurrence predicted by the fully connected neural network. If the accuracy requirements are met, the fully connected neural network outputs a convergent result; otherwise, the probability of geological disasters predicted by the fully connected neural network is not met. If the accuracy requirement is not met, the network parameters of the fully connected neural network are corrected, and the fully connected neural network is trained again until a converged fully connected neural network is output. Step S46: Obtain a converged physical information neural network, collect micro-scale geological data, meso-scale geological data, and macro-scale environmental data of the study area, execute steps S1-S4, and generate a global spatiotemporal feature vector. The neural network, which inputs convergent physical information, outputs the real-time probability of geological disasters occurring at the location of each graph node in the study area. .

[0013] The beneficial effects of this invention are as follows: This invention proposes a geological hazard risk prediction and evaluation method based on the coupling of spatiotemporal graph convolutional networks and physical information neural networks. It utilizes the powerful nonlinear fitting capability of data-driven models and ensures the geomechanical rationality of the prediction results through physical constraints. The multi-scale spatiotemporal graph convolutional network can simultaneously capture the microscopic rock and soil characteristics, mesoscopic slope characteristics, and macroscopic regional characteristics of geological hazard evolution, as well as the evolution laws at different time scales, significantly improving the comprehensiveness and accuracy of geological hazard-related feature extraction. The invention also proposes an uncertainty quantification method based on physical consistency, which can identify prediction results that violate physical laws and improve the reliability of geological hazard risk prediction results. Attached Figure Description

[0014] Figure 1This is a flowchart of a geological hazard risk prediction and assessment method. Detailed Implementation

[0015] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0016] like Figure 1 As shown, a geological hazard risk prediction and assessment method includes: Step S1: Collect multi-scale, multi-source data from different time steps in the study area, and preprocess the data to obtain a standardized dataset; divide the study area into... N Each grid cell serves as a graph node, and the set of edges is obtained based on the connection relationships between the graph nodes.

[0017] Step S1 specifically includes: Step S11: Collect microscale geological data, mesoscale geological data, and macroscale environmental data of the study area at different time steps to obtain multiscale multi-source data; preprocess the multiscale multi-source data and map the multiscale multi-source data to the [0,1] interval to obtain a standardized dataset of multiscale multi-source data. The microscale geological data in this embodiment include the physical and mechanical parameters of the soil and rock mass in the study area (internal friction angle, cohesion, unit weight, elastic modulus, Poisson's ratio), pore water pressure, soil moisture content, etc.

[0018] Mesoscale geological data includes slope morphology data (slope, aspect, height, shape), stratigraphic lithology data, geological structure data, fracture distribution data, and surface displacement data for the study area. Stratigraphic lithology data includes stratum dip angle, stratum strike, stratum dip gradient, stratum dip direction, depth of various stratigraphic interfaces, thickness of surface soil and rock, thickness of weathered layer, and thickness of overburden. Geological structure data includes fault strike length, fault fracture zone width, fault influence range, fault depth, fold strike, fold dip angle, fold curvature, and fold development intensity. Fracture distribution data includes fracture strike, fracture extension length, fracture width, fracture depth, fracture density, fracture strike dip angle, fracture opening, fracture displacement, and fracture vertical drop.

[0019] Macro-scale environmental data includes topographic and geomorphological data, water system distribution data, land use data, vegetation cover data, and meteorological and hydrological data for the study area. Topographic and geomorphological data includes elevation values, altitude, topographic relief, surface incision depth, surface incision density, and topographic slope variation coefficient for the study area. Water system distribution data includes water system direction, water system density, river network density, catchment area, drainage area, river channel slope, valley width, riverbed slope, surface runoff coefficient, water system permeability coefficient, and water area influence range for the study area. Land use data includes the proportion of land use types, construction land density, vegetation land proportion, and bare land coverage rate for the study area. Meteorological and hydrological data includes hourly rainfall, daily rainfall, cumulative rainfall, evaporation, sunshine duration, wind speed, wind direction, and atmospheric precipitation frequency.

[0020] Step S12: Divide the study area into N Each grid cell serves as a graph node. v n This yields the graph node set. , For the first N 1. Graph nodes; based on the connections between graph nodes, obtain the edges connecting two graph nodes, and thus obtain the edge set. , n Number the nodes in the graph.

[0021] The connections between nodes in the graph include spatial adjacency between two grid cells, geological connection, and hydrological connectivity.

[0022] If two grid cells i , j Spatially adjacent nodes represent corresponding graph nodes. There are spatial adjacent edges between them.

[0023] If two grid cells i , j If they are located on both sides of the same geological structural zone or fault, then they represent the corresponding nodes in the two diagrams. There are geological structural boundaries between them; If two grid cells i , j If there is a hydrological connectivity relationship between the two graph nodes, it means that the corresponding nodes in the graphs are connected. There are hydrologically connected edges between them; edge set The elements in are , This represents the connecting edge between nodes in two graphs. m The types of connecting edges include spatial adjacency edges, geological structure edges, and hydrological connectivity edges.

[0024] Step S2: Based on the type of connecting edge between two graph nodes in the edge set, calculate the weight of each type of connecting edge between the two graph nodes, and construct the comprehensive adjacency matrix of each graph node.

[0025] Step S2 specifically includes: Step S21: Based on the edge set Determine the types of edges connecting nodes in two graphs and calculate the weight of each type of edge. , , The weight of spatially adjacent edges. The weight of the geological structure edge, The weight of the hydrologically connected edge; For spatial adjacent edges: based on the nodes of the two graphs Euclidean distance between Calculate the weight of spatially adjacent edges. ; ;in, Spatial scale parameters; spatial scale parameters The attenuation scale, which characterizes the interaction between geospatial units, determines the range of spatial association between different geomorphic units. The larger the value, the stronger the spatial influence between distant grid units. For example, the grid length of the grid unit can be taken as a spatial scale parameter. Regarding geological structural edges: based on the nodes in the two diagrams The difference in distance between the nearest geological structural zone or fault Calculate the weight of geological structural edges ; ;in, Geological structural scale parameters; geological structural scale parameters Characterize the spatial attenuation range of the influence of geological structural elements, for example, taking the topographic relief as a geological structural scale parameter; For hydrological connectivity edges: based on the nodes in the two graphs Length of water flow path between Calculate the weight of hydrologically connected edges ; ;in, Hydrological scale parameters; hydrological scale parameters The spatial scale characterizing the effects of water systems and hydrological connectivity elements is defined by setting the connectivity radius of the hydrological space as a hydrological scale parameter. The connectivity radius of the hydrological space is used to define the effective range of hydrological connectivity. Its value is usually determined based on the distribution of water systems, topographic slope, and hydrological response unit division results in the study area. It is used to measure the effective distance of interaction between two grid units through surface runoff or groundwater connection. If the connectivity radius of the hydrological space is large, it indicates that hydrological connectivity over a longer distance is considered (applicable to plains or highly permeable areas). If the connectivity radius of the hydrological space is small, it indicates that hydrological connectivity over a short distance is considered (applicable to mountainous areas or areas with poor fissure development). In this embodiment, for valley or river areas, the connectivity radius of the hydrological space is taken as the river network buffer distance, that is, a certain buffer distance on both sides of the river channel (such as 50m or 100m); for surface runoff areas, the connectivity radius of the hydrological space is taken as the length of the surface runoff confluence path, which is used to cut off excessively long confluence paths. If the length of the water flow path between two graph nodes exceeds the length of the surface runoff confluence path, then its hydrological connectivity is considered to be small and can be ignored.

[0026] Step S22: Based on the weights of each connecting edge Construct the spatial adjacency weight matrix for each graph node Geological structure weight matrix Hydrological connectivity weight matrix Calculate the comprehensive adjacency matrix for each graph node. ; ; in, These are the weight coefficients for spatially adjacent edges, geological structural edges, and hydrologically connected edges, respectively. This embodiment satisfies The value is flexibly selected based on different types of geological hazards, with the optimal value for landslide hazards. Landslide disasters (rock mass fracture, tectonic activity) Debris flow disaster (dominated by hydrology and gully confluence) Shallow soil landslides (dominated by hydrological seepage) .

[0027] This invention divides the study area into a spatiotemporal graph structure, which can effectively represent the spatial topological relationships and temporal evolution characteristics of geological hazards. It comprehensively considers three edge connection relationships: spatial adjacency, geological structure, and hydrological connectivity, which can more accurately reflect the propagation mechanism of geological hazards. The weight of each connection edge is calculated using a Gaussian kernel function, which can reflect the distance attenuation effect and conforms to the actual propagation law of geological hazards.

[0028] Step S3: Construct a multi-source spatiotemporal graph data matrix based on the standardized dataset of each grid cell, construct a multi-scale spatiotemporal graph convolutional network, introduce a comprehensive adjacency matrix, extract multi-scale spatiotemporal features from the multi-source spatiotemporal graph data matrix, and obtain the global spatiotemporal feature vector of the graph node.

[0029] Step S3 specifically includes: Step S31: Construct a multi-source spatiotemporal graph data matrix based on the standardized dataset of each grid cell, expand the multi-source spatiotemporal graph data matrix according to the time step to obtain the spatial data matrix of each time step, and divide the comprehensive adjacency matrix into a multi-scale adjacency matrix set according to different data scales.

[0030] Specifically: Step S311: Construct a multi-source spatiotemporal graph data matrix based on the standardized dataset of each grid cell. , T The time step for collecting multi-scale, multi-source data. F Data dimensions for each grid cell, It is the set of real numbers; Step S312: Convert the multi-source spatiotemporal graph data matrix Expanding by time step, we obtain the spatial data matrix for each time step. ; synthesize the adjacency matrix It is divided into a multi-scale adjacency matrix set according to different data scales. , K The number of data scales, For the first K Adjacency matrix of several data scales; This embodiment K =3, the data scale includes micro, meso, and macro scales; at the micro scale, the spatial adjacency weight, geological structure weight, and hydrological connectivity weight between all directly connected graph nodes are used to construct the micro-scale adjacency matrix. radius at the microscale This represents the weight between directly connected graph nodes, i.e., 1-hop neighbor graph nodes; the radius at the mesoscale. This indicates that a mesoscale adjacency matrix is ​​constructed by taking the spatial adjacency weights, geological structure weights, and hydrological connectivity weights between nodes in a 3-hop neighbor graph. The mesoscale reflects the correlation between local watersheds / slope groups; the macroscale radius... This indicates that a macro-scale adjacency matrix is ​​constructed by taking the spatial adjacency weights, geological structure weights, and hydrological connectivity weights between nodes in a 5-hop neighbor graph. The macro scale reflects the relationship between regional tectonics and water system patterns.

[0031] The hop count between graph nodes is defined as follows: If two graph nodes are directly connected by at least one edge (including any one or more of spatial adjacency edges, geological structure edges, or hydrological connectivity edges), then these two graph nodes are 1-hop neighbors. If it takes two connecting edges (i.e., one intermediate graph node) to travel from graph node A to graph node B, then they are 2-hop neighbors; and so on. In other words, the hop count is determined by the shortest path length between the two graph nodes in the graph structure, and the edge type does not affect the hop count.

[0032] For each data scale k From the comprehensive adjacency matrix Extract path hop count Comprehensive adjacency weight And by filtering out weak associations, a multi-scale adjacency matrix set is obtained. ; ; in, For the first k Adjacency matrix of several data scales Representing nodes in two graphs Path hop count between Data scale k The overall adjacency weight threshold is as follows. The path hop count baseline is set for each data scale. .

[0033] Step S32: Normalize the adjacency matrix of each data scale to obtain a normalized adjacency matrix, construct a multi-scale spatiotemporal graph convolutional network including a multi-scale graph convolutional layer and a causal temporal convolutional layer, input the spatial data matrix of each time step into the multi-scale graph convolutional layer, introduce the normalized adjacency matrix, and extract the spatiotemporal feature tensor.

[0034] Specifically: Step S321: Normalize the adjacency matrix for each data scale to obtain the normalized adjacency matrix. ; ; in, For the first k A degree matrix for each data scale. It is the identity matrix; the symmetric normalized adjacency matrix can guarantee the stability and convergence of graph convolution.

[0035] Step S322: Construct a multi-scale spatiotemporal graph convolutional network, including multi-scale graph convolutional layers and causal temporal convolutional layers, and process the spatial data matrix of each time step. Input a multi-scale graph convolutional layer and introduce a normalized adjacency matrix. ,existK Perform graph convolution operations on various data scales to extract local features at different spatial scales. ; ; in, l The number of the graph convolutional layer. It is the ReLU activation function. For the first k On the data scale, the first l The weight matrix of each graph convolutional layer For the first k On the data scale, the first l The bias vector of each graph convolutional layer This represents the local features output by the convolutional layer in the previous image; Step S323: Combine local features at different spatial scales By splicing the data, we can obtain the time steps. t Multiscale spatial features ; ; in, This indicates a splicing operation. For the first K Local features at each data scale; Step S324: Set the time step t Multiscale spatial features Stack them in chronological order to form a spatiotemporal feature tensor. ; ; in, For time steps T Multi-scale spatial characteristics; Multi-scale graph convolution can simultaneously capture features from different spatial scales: microscale captures local features of a single slope, mesoscale captures features of a valley basin, and macroscale captures features of regional geological structures. Multi-scale feature stitching can preserve information from different data scales, avoiding the problem of insufficient feature extraction at a single scale.

[0036] Step S33: Input the spatiotemporal feature tensor into the causal temporal convolutional layer, perform causal temporal convolution on the temporal feature tensor of each graph node, and extract the enhanced spatiotemporal feature tensor by weighting the spatial attention weights of each graph node to its neighboring graph nodes.

[0037] Specifically: Step S331: Convert the spatiotemporal feature tensor Perform a dimensionality transformation to convert it into a graph node-time-feature dimension, obtaining the time series feature tensor for each graph node. Then, in the causal temporal convolutional layer, causal temporal convolution is performed on the time-series feature tensor of each graph node to extract local features in the time dimension. ; ; in, This indicates a dimension transformation operation. p This refers to the number of the causal temporal convolutional layer. q Number the convolution kernels for each time step. Q The size of the convolution kernel at each time step. The position of the convolution kernel at time step of the previous causal temporal convolutional layer. The output, For the first p Layered causal temporal convolutional layers with convolutional kernels at each time step q The weight matrix, For the first p Bias vector of a causal temporal convolutional layer; Step S332: Local features in the time dimension Converting back to time-graph node-feature dimension yields the local features of the transformed time dimension. Then, the components are spliced ​​together to obtain the spatiotemporal feature tensor. ; ; in, For the first P Local features of the temporal dimension output by a causal temporal convolutional layer; Causal temporal convolution can guarantee the causality of time series and avoid interference from future information on past predictions; performing temporal convolution independently on each graph node can capture the differences in temporal evolution characteristics at different spatial locations; multi-layer temporal convolution can extract features at different time scales, such as short-term rainfall changes and long-term deformation trends.

[0038] Step S333: Calculate the spatial attention weight of each graph node to its neighboring graph nodes. And utilize spatial attention weights spatiotemporal feature tensor By performing a weighted summation, we obtain the spatially weighted spatiotemporal feature tensor. ; ; in, For graph nodes i The set of adjacent graph nodes, This represents a spatial attention multilayer perceptron. For graph nodes i Adjacent graph nodes j The spatiotemporal feature tensor; Step S334: Based on the spatially weighted spatiotemporal feature tensor Calculate the attention weights over time Utilizing time attention weights spatiotemporal feature tensor We perform a weighted summation to obtain the time-weighted spatiotemporal feature tensor. ; ; in, For time steps Spatial weighted spatiotemporal feature tensor; Step S335: Convert the time-weighted spatiotemporal feature tensor With spatiotemporal feature tensor Perform residual connections to obtain enhanced spatiotemporal feature tensors. ; .

[0039] This invention can automatically identify key areas of disaster evolution through spatial attention mechanism, such as high-risk areas near faults and the leading edge of slopes; and can automatically identify key time nodes of disaster evolution through temporal attention mechanism, such as the moment of heavy rainfall and the moment of accelerated deformation; residual connection can alleviate the gradient vanishing problem of deep networks and improve the training stability and convergence speed of the model.

[0040] Step S34: Perform global average pooling and global max pooling on the enhanced spatiotemporal feature tensor to aggregate the information in the time dimension, obtain the aggregated features in the time dimension, and perform dimensionality reduction and nonlinear transformation through a fully connected layer to output the global spatiotemporal feature vector of the graph nodes.

[0041] Specifically: Step S341: For the enhanced spatiotemporal feature tensor Perform global average pooling and global max pooling, aggregating information from the time dimension to obtain the average pooling output. and max pooling output ; ; in, For time step T Enhanced spatiotemporal feature tensor; Step S342: Output the average pooling. and max pooling output By concatenating the features, we obtain the features aggregated along the time dimension. ; ; Global average pooling captures the overall trend in the time dimension, while global max pooling captures extreme events in the time dimension. Feature concatenation can comprehensively utilize the advantages of both average pooling and max pooling.

[0042] Step S343: Process features through a fully connected layer Perform dimensionality reduction and nonlinear transformation to output the global spatiotemporal feature vectors of the graph nodes. ; ; in, These are the weight matrix and bias vector of the fully connected layer, respectively.

[0043] Multi-scale graph convolution can capture features at different spatial scales simultaneously, while causal temporal convolution can effectively capture the causal relationships of time series and avoid the leakage of future information. Introducing a spatiotemporal attention mechanism can adaptively learn the importance weights of different spatiotemporal locations, highlighting key areas and time nodes in disaster evolution. Multi-scale feature fusion can comprehensively utilize information at different scales, improving the comprehensiveness and discriminative ability of feature representation.

[0044] Step S4: Construct a physical information neural network, including a data-driven branch and a physical constraint branch. Input the global spatiotemporal feature vector into the physical information neural network to predict the real-time probability of geological disasters at the location of each graph node in the study area.

[0045] Step S4 specifically includes; Step S41: Construct a physical information neural network, including a data-driven branch and a physical constraint branch. The physical constraint branch uses the geomechanical control equations as physical constraints. The geomechanical governing equations are: ; in, For the cohesion of the soil and rock mass, For normal stress, u Pore ​​water pressure, It is the internal friction angle. For shear stress, For geological disaster safety factor; when At that time, geological disasters are prone to occur. At that time, geological disasters are less likely to occur; Step S42: Convert the global spatiotemporal feature vector In the input data-driven branch, a fully connected neural network is used to predict the probability of geological disasters occurring at the locations of graph nodes. and displacement field ; ; in, These are fully connected neural networks for predicting the probability of geological disasters and displacement fields, respectively. Step S43: Physical constraint branch based on the predicted displacement field of the graph nodes Calculated predicted stress field According to the stress field Calculating the geological hazard safety factor using normal stress. ; ; in, This represents the elastic stiffness matrix of the soil and rock mass at the location of the node in the graph. The displacement field gradient operator represents the rate of change of spatial displacement of the soil and rock mass. It consists of the spatial partial derivatives of horizontal and vertical displacements. The displacement gradient is the core physical quantity of the strain tensor of the soil and rock mass, and it quantifies the degree of deformation of the soil and rock mass through the spatial change of the displacement field.

[0046] Step S44: Based on the geological disaster safety factor Constructing the physical loss function ; ; in, The regularization coefficient is . This is a gradient operator based on the geological hazard safety factor at the location of the nodes in the study area. It characterizes the spatial change rate of the geological hazard safety factor in the study area, ensures that the geological hazard safety factor meets the spatial continuity, suppresses abrupt noise generated by the model prediction, and avoids unreasonable stability abrupt change areas. This invention introduces physical constraint branches to limit the prediction results of fully connected neural networks, making the prediction results conform to the laws of geomechanics, suppressing non-physical anomaly outputs, realizing the coupling of data-driven and physical mechanisms, and improving the scientificity and reliability of geological disaster prediction and evaluation.

[0047] Step S45: Set the physical loss threshold ,like Then determine the probability of geological disaster occurrence predicted by the fully connected neural network. If the accuracy requirements are met, the fully connected neural network outputs a convergent result; otherwise, the probability of geological disasters predicted by the fully connected neural network is not met. If the accuracy requirements are not met, the network parameters of the fully connected neural network are corrected, and the fully connected neural network is trained until the output of the fully connected neural network converges. Step S46: Obtain a converged physical information neural network, collect micro-scale geological data, meso-scale geological data, and macro-scale environmental data of the study area, execute steps S1-S18, and generate a global spatiotemporal feature vector. The neural network, which inputs convergent physical information, outputs the real-time probability of geological disasters occurring at the location of each graph node in the study area. .

[0048] Step S5: Evaluate the real-time geological disaster risk in the study area based on the probability of geological disaster occurrence.

[0049] Specifically, based on the probability of geological disasters occurring Evaluate the real-time geological hazard risk and probability of geological hazard occurrence in the study area. The larger the value, the greater the risk of geological disasters at the location of each node in the map, and vice versa. At the same time, the more locations in the study area with high geological disaster risk, the greater the risk of large-scale geological disasters, and vice versa.

Claims

1. A method for predicting and assessing geological disaster risks, characterized in that, include: Step S1: Collect multi-scale, multi-source data of the study area and preprocess it to obtain a standardized dataset; The study area is divided into N Each grid cell serves as a graph node, and the set of edges is obtained based on the connection relationships between the graph nodes. Step S2: Based on the type of connecting edge between two graph nodes in the edge set, calculate the weight of each type of connecting edge between the two graph nodes, and construct the comprehensive adjacency matrix for each graph node; Step S3: Construct a multi-source spatiotemporal graph data matrix based on the standardized dataset of each grid cell, construct a multi-scale spatiotemporal graph convolutional network, and introduce a comprehensive adjacency matrix to extract multi-scale spatiotemporal features from the multi-source spatiotemporal graph data matrix, thereby obtaining the global spatiotemporal feature vector of the graph node; Step S4: Construct a physical information neural network, including a data-driven branch and a physical constraint branch. Input the global spatiotemporal feature vector into the physical information neural network to predict the real-time probability of geological disasters occurring at the location of each graph node in the study area. Step S5: Evaluate the real-time geological disaster risk in the study area based on the probability of geological disaster occurrence.

2. The geological hazard risk prediction and assessment method according to claim 1, characterized in that, Step S1 includes: Step S11: Collect microscale geological data, mesoscale geological data, and macroscale environmental data of the study area to obtain multiscale multi-source data; preprocess the multiscale multi-source data and map the multiscale multi-source data to the [0,1] interval to obtain a standardized dataset of multiscale multi-source data; Step S12: Divide the study area into N Each grid cell serves as a graph node. v n This yields the graph node set. , For the first N 1. Graph nodes; based on the connections between graph nodes, obtain the edges connecting two graph nodes, and thus obtain the edge set. , n Number the nodes in the graph; The connection relationships between graph nodes include spatial adjacency, geological connection, and hydrological connectivity between two grid cells; the types of connecting edges include spatial adjacency edges, geological connection edges, and hydrological connectivity edges.

3. The geological hazard risk prediction and assessment method according to claim 2, characterized in that, Step S2 includes: Step S21: Based on the edge set Determine the types of edges connecting nodes in two graphs and calculate the weight of each type of edge. , , The weight of spatially adjacent edges. The weight of the geological structure edge, The weight of the hydrologically connected edge; For spatial adjacent edges: based on the nodes of the two graphs Euclidean distance between Calculate the weight of spatially adjacent edges. ; ;in, Spatial scale parameter; Regarding geological structural edges: based on the nodes in the two diagrams The difference in distance between the nearest geological structural zone or fault Calculate the weight of geological structural edges ; ;in, Parameters at the geological structural scale; For hydrological connectivity edges: based on the nodes in the two graphs Length of water flow path between Calculate the weight of hydrologically connected edges ; ;in, For hydrological scale parameters; Step S22: Based on the weights of each connecting edge Construct the spatial adjacency weight matrix for each graph node Geological structure weight matrix Hydrological connectivity weight matrix Calculate the comprehensive adjacency matrix for each graph node. ; ； in, These are the weight coefficients for spatially adjacent edges, geological structural edges, and hydrologically connected edges, respectively.

4. The geological hazard risk prediction and assessment method according to claim 3, characterized in that, Step S3 includes: Step S31: Construct a multi-source spatiotemporal graph data matrix based on the standardized dataset of each grid cell, expand the multi-source spatiotemporal graph data matrix according to the time step to obtain the spatial data matrix of each time step, and divide the comprehensive adjacency matrix into a multi-scale adjacency matrix set according to different data scales; Step S32: Normalize the adjacency matrix of each data scale to obtain a normalized adjacency matrix, construct a multi-scale spatiotemporal graph convolutional network including a multi-scale graph convolutional layer and a causal temporal convolutional layer, input the spatial data matrix of each time step into the multi-scale graph convolutional layer, introduce the normalized adjacency matrix, and extract the spatiotemporal feature tensor. Step S33: Input the spatiotemporal feature tensor into the causal temporal convolutional layer, perform causal temporal convolution on the temporal feature tensor of each graph node, and extract the enhanced spatiotemporal feature tensor by weighting the spatial attention weights of each graph node to its neighboring graph nodes. Step S34: Perform global average pooling and global max pooling on the enhanced spatiotemporal feature tensor to aggregate the information in the time dimension, obtain the aggregated features in the time dimension, and perform dimensionality reduction and nonlinear transformation through a fully connected layer to output the global spatiotemporal feature vector of the graph nodes.

5. The geological hazard risk prediction and assessment method according to claim 4, characterized in that, Step S31 includes: Step S311: Construct a multi-source spatiotemporal graph data matrix based on the standardized dataset of each grid cell. , T The time step for collecting multi-scale, multi-source data. F Data dimensions for each grid cell, It is the set of real numbers; Step S312: Convert the multi-source spatiotemporal graph data matrix Expanding by time step, we obtain the spatial data matrix for each time step. ; synthesize the adjacency matrix It is divided into a multi-scale adjacency matrix set according to different data scales. , K The number of data scales, For the first K Adjacency matrix of several data scales; For each data scale k From the comprehensive adjacency matrix Extract path hop count Comprehensive adjacency weight And by filtering out weak associations, a multi-scale adjacency matrix set is obtained. ; ； in, For the first k Adjacency matrix at several data scales Representing nodes in two graphs Path hop count between Data scale k The overall adjacency weight threshold is as follows. The baseline for path hop count is set for each data scale.

6. The geological hazard risk prediction and assessment method according to claim 5, characterized in that, Step S32 includes: Step S321: Normalize the adjacency matrix for each data scale to obtain the normalized adjacency matrix. ; ； in, For the first k A degree matrix for each data scale. It is the identity matrix; Step S322: Construct a multi-scale spatiotemporal graph convolutional network, including multi-scale graph convolutional layers and causal temporal convolutional layers, and process the spatial data matrix of each time step. Input a multi-scale graph convolutional layer and introduce a normalized adjacency matrix. ,exist K Perform graph convolution operations on various data scales to extract local features at different spatial scales. ; ； in, l The number of the graph convolutional layer. It is the ReLU activation function. For the first k On the data scale, the first l The weight matrix of each graph convolutional layer For the first k On the data scale, the first l The bias vector of each graph convolutional layer This represents the local features output by the convolutional layer in the previous image; Step S323: Combine local features at different spatial scales By splicing the data, we can obtain the time steps. t Multiscale spatial features ; ； in, This indicates a splicing operation. For the first K Local features at each data scale; Step S324: Set the time step t Multiscale spatial features Stack them in chronological order to form a spatiotemporal feature tensor. ; ； in, For time step T Multi-scale spatial characteristics.

7. The geological hazard risk prediction and assessment method according to claim 6, characterized in that, Step S33 includes: Step S331: Convert the spatiotemporal feature tensor Perform a dimensionality transformation to convert it into a graph node-time-feature dimension, obtaining the time series feature tensor for each graph node. Then, in the causal temporal convolutional layer, causal temporal convolution is performed on the time-series feature tensor of each graph node to extract local features in the time dimension. ; ； in, This indicates a dimension transformation operation. p This refers to the number of the causal temporal convolutional layer. q Number the convolution kernels for each time step. Q The size of the convolution kernel at each time step. The position of the convolution kernel at time step of the previous causal temporal convolutional layer. The output, For the first p Layered causal temporal convolutional layer with convolutional kernels at each time step q The weight matrix, For the first p Bias vector of a causal temporal convolutional layer; Step S332: Local features in the time dimension Converting back to time-graph node-feature dimension yields the local features of the transformed time dimension. The spatiotemporal feature tensors are then concatenated to obtain the spatiotemporal feature tensors. ; ； in, For the first P Local features of the temporal dimension output by a causal temporal convolutional layer; Step S333: Calculate the spatial attention weight of each graph node to its neighboring graph nodes. And utilize spatial attention weights spatiotemporal feature tensor By performing a weighted summation, we obtain the spatially weighted spatiotemporal feature tensor. ; ； in, For graph nodes i The set of adjacent graph nodes, This represents a spatial attention multilayer perceptron. For graph nodes i Adjacent graph nodes j The spatiotemporal feature tensor; Step S334: Based on the spatially weighted spatiotemporal feature tensor Calculate the attention weights over time Utilizing time attention weights spatiotemporal feature tensor We perform a weighted summation to obtain the time-weighted spatiotemporal feature tensor. ; ； in, For time steps Spatial weighted spatiotemporal feature tensor; Step S335: Convert the time-weighted spatiotemporal feature tensor With spatiotemporal feature tensor Perform residual connections to obtain enhanced spatiotemporal feature tensors. ; 。 8. The geological hazard risk prediction and assessment method according to claim 7, characterized in that, Step S34 includes: Step S341: For the enhanced spatiotemporal feature tensor Perform global average pooling and global max pooling, aggregating information from the time dimension to obtain the average pooling output. and max pooling output ; ； in, For time step T Enhanced spatiotemporal feature tensor; Step S342: Output the average pooling. and max pooling output By concatenating the features, we obtain the features aggregated along the time dimension. ; ； Step S343: Process features through a fully connected layer Perform dimensionality reduction and nonlinear transformation to output the global spatiotemporal feature vectors of the graph nodes. ; ； in, These are the weight matrix and bias vector of the fully connected layer, respectively.

9. The geological hazard risk prediction and assessment method according to claim 8, characterized in that, Step S4 includes: Step S41: Construct a physical information neural network, including a data-driven branch and a physical constraint branch. The physical constraint branch uses the geomechanical control equations as physical constraints. The geomechanical governing equations are: ; in, For the cohesion of the soil and rock mass, For normal stress, u Pore ​​water pressure, It is the internal friction angle. For shear stress, For geological disaster safety factor; when At that time, geological disasters are prone to occur. At that time, geological disasters are less likely to occur; Step S42: Convert the global spatiotemporal feature vector In the input data-driven branch, a fully connected neural network is used to predict the probability of geological disasters occurring at the locations of graph nodes. and displacement field ; ； in, These are fully connected neural networks for predicting the probability of geological disasters and displacement fields, respectively. Step S43: Physical constraint branch based on the predicted displacement field of the graph nodes Calculated predicted stress field According to the stress field Calculating the geological hazard safety factor using normal stress. ; ； in, This represents the elastic stiffness matrix of the soil and rock mass at the location of the node in the graph. Represents the displacement field gradient operator; Step S44: Based on the geological disaster safety factor Constructing the physical loss function ; ； in, The regularization coefficient is . This is a gradient operator based on the geological hazard safety factor of the nodes in the study area. Step S45: Set the physical loss threshold ,like Then determine the probability of geological disaster occurrence predicted by the fully connected neural network. If the accuracy requirements are met, the fully connected neural network outputs a convergent result; otherwise, the probability of geological disasters predicted by the fully connected neural network is not met. If the accuracy requirement is not met, the network parameters of the fully connected neural network are corrected, and the fully connected neural network is trained again until a converged fully connected neural network is output. Step S46: Obtain a converged physical information neural network, collect micro-scale geological data, meso-scale geological data, and macro-scale environmental data of the study area, execute steps S1-S4, and generate a global spatiotemporal feature vector. The neural network, which inputs convergent physical information, outputs the real-time probability of geological disasters occurring at the location of each graph node in the study area. .

Images