Disaster area assessment method, system and equipment based on geological disaster exploration and medium

Abstract

The invention relates to a disaster area assessment method, system and device based on geological disaster exploration and a medium. The method comprises the following steps: analyzing multi-source exploration data to obtain a unit spatial-temporal characteristic tuple and a physical incidence matrix; performing graph neural network calculation processing based on the unit spatial-temporal feature tuple and the physical incidence matrix to obtain a dynamic risk assessment result, a multi-level attention weight and a physical state deviation amount; and based on the dynamic risk assessment result, the multi-level attention weight and the physical state deviation amount, performing report generation processing to obtain a geological disaster dynamic risk assessment report. According to the method, by means of multi-source data analysis, graph neural network calculation and the like, the accuracy, dynamism and interpretability of geological disaster assessment are improved, and the disaster risk accurate prevention and control and decision support capability is enhanced.

Description

Technical Field

[0001] This invention belongs to the field of computer technology, and in particular relates to methods, systems, equipment and media for disaster area assessment based on geological disaster investigation. Background Technology

[0002] Geological hazard assessment is a key technical means to ensure geological environmental safety and reduce disaster losses. With the development of exploration technology, multi-source exploration data such as remote sensing data, topographic and geological data, meteorological monitoring data, and engineering monitoring data have been widely used in the field of geological hazard assessment. However, existing geological hazard assessment methods usually rely on a single data source or traditional physical models for risk analysis, resulting in insufficient integration of multi-source data and unsystematic extraction of spatiotemporal features. This makes it difficult to form unit spatiotemporal feature tuples that comprehensively reflect the evolution of disasters, and lacks quantitative characterization of the physical correlations between multi-source data, leading to the missing construction of physical correlation matrices and affecting the accuracy of assessment results. Furthermore, existing assessment methods often use traditional statistical models or simple data-driven models, failing to fully combine the spatial correlation modeling capabilities of graph neural networks with the constraints of physical mechanisms. This makes it difficult to effectively capture the dynamic evolution characteristics of geological hazard risks and generate high-precision dynamic risk assessment results. In addition, existing assessment reports mostly only output risk level results, lacking in-depth analysis of the causes of risks, resulting in insufficient interpretability and practicality of the assessment reports, making it difficult to meet the needs of precise disaster risk prevention and control and decision support in engineering practice. Summary of the Invention

[0003] Therefore, it is necessary to provide disaster area assessment methods, systems, equipment and media based on geological disaster investigation to address the above-mentioned technical problems, aiming to improve the accuracy of assessment results and the precision of dynamic risk assessment, and enhance the ability to dynamically monitor and warn of disaster risks.

[0004] Firstly, this application provides a disaster area assessment method based on geological hazard investigation, including:

[0005] The multi-source exploration data is analyzed to obtain the spatiotemporal feature tuples and physical correlation matrices of the units;

[0006] Based on the unit spatiotemporal feature tuples and physical correlation matrix, graph neural network computation is performed to obtain dynamic risk assessment results, multi-level attention weights and physical state deviation.

[0007] Based on the dynamic risk assessment results, multi-level attention weights, and physical state deviations, a report generation process is performed to obtain a dynamic risk assessment report for geological disasters.

[0008] In one embodiment, multi-source exploration data is parsed to obtain unit spatiotemporal feature tuples and physical correlation matrices, including:

[0009] Spatiotemporal gridding alignment of remote sensing data, topographic data, geological data, meteorological monitoring data, and engineering monitoring data from multi-source exploration data is performed to obtain unified multi-source raw data. Based on the unified multi-source raw data, static and dynamic features are separated, and the static features are standardized and extracted to obtain the static background feature matrix of each spatial unit. The dynamic features are temporally regularized to obtain the dynamic observation feature tensor of each spatial unit.

[0010] Based on the preset seepage mechanics equations, physical analysis processing is performed on the time-series rainfall data in the dynamic observation feature tensor to obtain the hydrophysical feature tensor of each spatial unit.

[0011] Inter-unit runoff is extracted from hydrophysical feature tensors. Combined with geological structure similarity information in topographic data, physical correlation is calculated on the inter-unit runoff to obtain the physical correlation and construct a physical correlation matrix.

[0012] Based on the signal decomposition model, physical analysis processing is performed on the temporal deformation data in the dynamic observation feature tensor to obtain the deformation physical feature tensor of each spatial unit; the deformation physical feature tensor contains deformation residual components.

[0013] The static background feature matrix, hydrophysical feature tensor, and deformation physical feature tensor are concatenated by feature dimensions to obtain the unit spatiotemporal feature tuple.

[0014] In one embodiment, the unit spatiotemporal feature tuple includes temporal information of multiple time steps, and the temporal information of each time step includes a static background feature matrix, a hydrophysical feature tensor, and a deformation physical feature tensor.

[0015] Based on the unit spatiotemporal feature tuples and physical correlation matrix, graph neural network computation is performed to obtain dynamic risk assessment results, multi-level attention weights, and physical state deviations, including:

[0016] Traverse each time step of the unit spatiotemporal feature tuple, construct the adjacency relationship of the graph structure corresponding to each time step based on the physical correlation matrix, and perform feature initialization processing on each node in the corresponding graph structure based on the unit spatiotemporal feature tuple of the current time step to obtain the initialized node features and the corresponding initialized graph structure for each time step.

[0017] For each time step, the initialized graph structure is subjected to multi-layer physical constraint graph convolution processing to obtain the optimized node features of the final layer output; the multi-layer physical constraint graph convolution processing corresponds to the following steps:

[0018] Message aggregation processing is performed on the current node features and neighboring node features in the initialized graph structure of the current layer to obtain aggregated features; wherein, in the first layer processing, the current node features are the initialized node features;

[0019] The aggregated features are linearly transformed to obtain intermediate node features; a physical verifier is constructed based on a preset physical equation, and the physical consistency of the intermediate node features is verified by the physical verifier to obtain the physical state deviation of the current layer.

[0020] The physical state deviation and intermediate node features are weighted and fused to obtain updated node features. The updated node features are then used as the current node features of the next layer. The physical constraint graph convolution process is repeated until each layer has undergone physical constraint graph convolution to obtain the optimized node features of the final layer output at the corresponding time step.

[0021] Based on the optimized node features of the final layer output in each time step and the physical state deviation of each layer, factor attention weights are calculated to obtain the factor attention weight vector corresponding to each node in each time step.

[0022] During the message aggregation process at each time step, spatial correlation is calculated based on the current node features of the current layer and the deformation physical feature tensor of the corresponding time step to obtain the spatial correlation of the current layer. The physical correlation matrix and the spatial correlation of the current layer are weighted and fused to obtain the dynamic spatial attention weight of the current layer.

[0023] The optimized node features of the final layer corresponding to each time step are subjected to temporal feature fusion processing. In the temporal feature fusion processing, historical hydrological state is extracted from the hydrophysical feature tensor of each time step. Based on the historical hydrological state and the optimized node features output by the final layer of the corresponding time step, the temporal attention weight is calculated to obtain the temporal attention weight corresponding to each time step.

[0024] Risk probability mapping is performed on the optimized node features of the final layer corresponding to each time step to generate dynamic risk assessment results for each time step.

[0025] At each time step, the corresponding factor attention weight vector, time attention weight, and dynamic spatial attention weight of each layer are integrated to obtain multi-level attention weights. The physical state deviation of each layer corresponding to each time step is summarized to form the physical state deviation.

[0026] In one embodiment, a report generation process is performed based on the dynamic risk assessment results, multi-level attention weights, and physical state deviations to obtain a dynamic geological hazard risk assessment report, including:

[0027] Obtain and compare the dynamic risk assessment results corresponding to each time step according to the preset risk threshold to identify high-risk units;

[0028] For each high-risk unit, the factor attention weight vector at the corresponding time step is extracted. Combined with the feature dimension of the temporal information at the corresponding time step in the unit's spatiotemporal feature tuple, the contribution of the dominant factor is quantified and analyzed to obtain the contribution analysis chart of the dominant factor.

[0029] For each high-risk unit, the dynamic spatial attention weight of the corresponding time step is extracted, and the spatial influence range and spatial influence intensity are visualized to obtain a spatial influence correlation heatmap; the temporal attention weight of the corresponding time step of the high-risk unit is extracted, and the key influence moment is quantitatively analyzed to obtain a key influence moment curve.

[0030] For each high-risk unit, the physical state deviation at the corresponding time step is extracted, and combined with the preset physical constraint standard, physical consistency deviation analysis is performed to obtain the physical state consistency analysis diagram.

[0031] The dynamic risk assessment results, high-risk units, contribution analysis charts of dominant factors, spatial impact correlation heatmaps, key impact moment curves, and physical state consistency analysis charts corresponding to each time step are structured and integrated to obtain a dynamic risk assessment report for geological disasters.

[0032] In one embodiment, based on the seepage mechanics equations, the time-series rainfall data in the dynamic observation feature tensor is physically analyzed to obtain the hydrophysical feature tensor of each spatial unit, including:

[0033] The time-series rainfall data corresponding to each time step is extracted from the dynamic observation feature tensor, and the rainfall intensity is standardized to obtain the unit area rainfall intensity sequence corresponding to each time step.

[0034] Based on soil type data from geological data, combined with rock layer distribution data and porosity data from topographic data, soil hydraulic parameters for each spatial unit are determined; soil hydraulic parameters include saturated hydraulic conductivity, air intake suction, and pore size distribution index.

[0035] The pre-set seepage mechanics equation is used as the core control equation. The unit area rainfall intensity sequence corresponding to each time step is used as the infiltration boundary condition, and the soil hydraulic parameters are used as the initial parameters of the equation to construct a one-dimensional vertical seepage numerical model for each spatial unit. The pre-set seepage mechanics equation is the Richards seepage mechanics equation.

[0036] The one-dimensional vertical seepage numerical model is discretized based on the finite volume method to obtain the discretized seepage numerical model. The discretized seepage numerical model is then solved using an implicit difference scheme to obtain the soil moisture content distribution data and pore water pressure distribution data in the spatial cells corresponding to each time step.

[0037] Feature extraction processing was performed on the soil moisture content distribution data and pore water pressure distribution data corresponding to each time step to obtain the mean soil moisture content, peak pore water pressure and seepage velocity vector of each spatial unit.

[0038] Based on the seepage velocity vector of each spatial unit and the pore water pressure distribution data of adjacent spatial units, the hydraulic gradient between adjacent spatial units is calculated. Combined with the saturated hydraulic conductivity in the soil hydraulic parameters, the inter-unit runoff is calculated using Darcy's law to obtain the inter-unit runoff corresponding to each time step.

[0039] The mean soil moisture content, peak pore water pressure, seepage velocity vector, and inter-unit runoff at each time step are integrated along the time dimension to construct the hydrophysical characteristic tensor of each spatial unit.

[0040] In one embodiment, 6, according to the method of claim 2, the physical correlation degree is calculated by the following formula:

[0041]

[0042] in, For the first Spatial unit in time step The physical correlation degree has a value range of [0,1]. For the first Spatial unit in time step and radial flux; For the first The maximum radial flux between the spatial units at each time step; spatial unit and The geological structure similarity coefficient, with a value range of [0,1]; spatial unit and Spatial Euclidean distance; This is the distance attenuation coefficient; For the first Spatial unit in time step Peak pore water pressure; , and All are weighting coefficients, satisfying These correspond to the contribution weights of runoff, geological similarity, and pore water pressure, respectively. Secondly, this application also provides a disaster area assessment system based on geological hazard investigation, including:

[0043] The data parsing module is used to parse multi-source exploration data to obtain unit spatiotemporal feature tuples and physical correlation matrices;

[0044] The graph neural network analysis module is used to perform graph neural network calculations based on unit spatiotemporal feature tuples and physical correlation matrices to obtain dynamic risk assessment results, multi-level attention weights, and physical state deviations.

[0045] The report generation module is used to generate a dynamic risk assessment report for geological disasters based on the results of dynamic risk assessment, multi-level attention weights, and physical state deviation.

[0046] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the first aspect.

[0047] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the first aspect.

[0048] The aforementioned disaster assessment method, system, equipment, and media based on geological hazard investigation first analyze multi-source investigation data to obtain unit spatiotemporal feature tuples and physical correlation matrices. This effectively solves the problems of insufficient multi-source data integration and lack of physical correlation quantification in traditional methods, improving the comprehensiveness and correlation of disaster characteristic representation. Secondly, graph neural network computation is performed based on the unit spatiotemporal feature tuples and physical correlation matrices, integrating the advantages of physical mechanisms and spatial correlation modeling to improve the accuracy of dynamic risk assessment results. Finally, report generation is performed by combining dynamic risk assessment results, multi-level attention weights, and physical state deviations. This clarifies the key risk influencing factors and spatiotemporal correlation patterns, enhancing the decision support value of the assessment report and meeting the needs of precise disaster prevention and control in engineering practice. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1A flowchart of a disaster area assessment method based on geological hazard investigation is provided as an exemplary embodiment of the present invention;

[0051] Figure 2 A flowchart of a method for obtaining a dynamic risk assessment report of geological disasters is provided as an exemplary embodiment of the present invention;

[0052] Figure 3 A schematic diagram of the structure of a disaster area assessment system based on geological disaster exploration, provided as an exemplary embodiment of the present invention. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0054] In one embodiment, such as Figure 1 As shown, a disaster area assessment method based on geological hazard investigation is provided. This embodiment illustrates the application of this method to a terminal. It is understood that this method can also be applied to a server, and further to a system including both a terminal and a server, and implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0055] S101: Analyze multi-source exploration data to obtain unit spatiotemporal feature tuples and physical correlation matrices.

[0056] Specifically, the occurrence and evolution of geological disasters are the result of the combined effects of multiple factors, including topographic conditions, geological structure, meteorological factors, and engineering disturbances. Therefore, by acquiring multi-source exploration data, including remote sensing data, topographic data, geological data, meteorological monitoring data, and engineering monitoring data, a comprehensive understanding of the occurrence and evolution process of geological disasters can be achieved. However, the aforementioned multi-source exploration data suffers from problems such as heterogeneous data types, inconsistent spatiotemporal benchmarks, and implicit correlations. Directly inputting these data into the model will lead to distorted feature representations and loss of correlation information, thereby affecting the accuracy and rationality of the assessment results. Therefore, a unified spatiotemporal modeling framework can be established first, dividing spatial units with a preset spatial resolution and dividing time steps with fixed time intervals (such as daily levels, which can be adapted according to the disaster evolution rate) to ensure that all data have a consistent benchmark in the spatiotemporal dimensions. Subsequently, various types of data can be preprocessed, including data cleaning and format standardization, to eliminate data heterogeneity. Furthermore, static and dynamic features can be separated based on the attribute characteristics of the data. Static features refer to characteristics that do not change or change very little over time, such as geological lithology, topographic slope, and soil type. Dynamic features, on the other hand, refer to features that evolve over time, such as rainfall intensity, deformation displacement, and groundwater level. These two types of features can be integrated into spatial units to form a feature set for each spatial unit at each time step, i.e., a unit spatiotemporal feature tuple. This feature tuple comprehensively reflects the spatiotemporal evolution state of each spatial unit. Furthermore, based on the physical mechanisms of geological hazard evolution, the correlation strength between different spatial units can be quantified. For example, the correlation between adjacent units under the influence of topographic slope and geological structure, such as hydraulic conduction and soil stability, can be represented by a matrix to obtain a physical correlation matrix. This matrix provides physical-level correlation constraints for subsequent graph neural network modeling, further avoiding the misleading non-physical correlations that may occur in purely data-driven modeling.

[0057] S102: Based on the unit spatiotemporal feature tuples and physical correlation matrix, graph neural network calculation is performed to obtain dynamic risk assessment results, multi-level attention weights, and physical state deviation.

[0058] Specifically, the evolution of geological hazards exhibits significant spatial correlations (e.g., soil instability in one area may trigger a chain reaction in adjacent areas) and spatiotemporal dynamics (e.g., the cumulative effect of rainfall over time leads to a gradual increase in risk level). However, traditional assessment methods such as statistical models and static physical models struggle to effectively model these complex spatiotemporal relationships, resulting in assessment results that fail to reflect the dynamic evolution of risk and lack interpretability. Graph neural networks, as a deep learning model specifically designed for processing correlated data, can naturally model the relationships between spatial units through the structure of nodes and edges. Furthermore, by combining the physical constraints provided by the physical correlation matrix, they can further ensure that the model learning process conforms to the physical laws of geological hazard evolution, avoiding overfitting or non-physical predictions caused by purely data-driven approaches. For example,

[0059] The basic structure of a graph neural network can be constructed first, with each spatial unit as a node in the graph, the spatiotemporal feature tuples of the unit as the feature input of the node, and the physical association matrix as the adjacency matrix of the graph, used to define the association strength and interaction mode between nodes. Subsequently, the interaction and updating of node features can be achieved through a message passing mechanism. Each node aggregates the feature information of its neighboring nodes and performs a nonlinear transformation based on its own features, thereby capturing the association influence between spatial units. During computation, the graph neural network model can perform temporal fusion of node features at each time step, fully considering the temporal cumulative effect of dynamic features such as rainfall and deformation, to achieve dynamic risk assessment. Furthermore, to improve the interpretability of the assessment results, the model can simultaneously extract multi-level attention weights during training and inference, including attention weights in the feature dimension (reflecting the contribution of various features to risk assessment), attention weights in the spatial dimension (reflecting the influence strength of adjacent units on the risk of the target unit), and attention weights in the temporal dimension (reflecting the contribution of different time steps to the current risk level). Furthermore, to verify the physical rationality of the assessment results, the model can also introduce a physical state verification mechanism. By comparing the features output by the model with the theoretical features derived from the physical laws of geological hazards, the deviation of the physical state is calculated. This deviation measures the degree of fit between the assessment results and physical laws; the smaller the deviation, the stronger the physical rationality of the assessment results. Finally, through the output layer of the graph neural network, such as a fully connected layer combined with a Sigmoid or Softmax activation function, the final node features can be mapped to the assessment results (such as risk level, probability of occurrence, etc.) of the corresponding spatial units at each time step, i.e., the dynamic risk assessment results.

[0060] S103: Based on the dynamic risk assessment results, multi-level attention weights, and physical state deviation, report generation processing is performed to obtain a dynamic risk assessment report for geological disasters.

[0061] Specifically, in engineering practice, the core requirement for geological hazard assessment is not only to obtain the risk level, but also to clarify the key causes of the risk, the spatiotemporal evolution law, and the reliability of the assessment results, so as to formulate targeted prevention and control measures and emergency plans.

[0062] Therefore, the dynamic risk assessment results can first be visualized spatiotemporally using heatmaps, time-series curves, etc., to clearly present the risk level distribution and evolution trend of each spatial unit at different time steps, enabling decision-makers to intuitively grasp the spatiotemporal dynamic characteristics of the risk. Subsequently, in-depth analysis of risk causes can be conducted based on multi-level attention weights. For example, attention weights in the feature dimension can identify key features that play a dominant role in risk assessment, such as rainfall intensity and terrain slope being major risk triggers in a certain area. Furthermore, attention weights in the spatial dimension can clarify the risk propagation path and key affected areas, such as the risk of a high-risk unit being mainly affected by the soil stability of two adjacent units. In addition, attention weights in the time-series dimension can pinpoint key time nodes in risk evolution, such as a significant increase in risk level after three consecutive days of rainfall. Illustratively, reliability analysis of the assessment results can also be performed based on the deviation of physical state, marking areas with large deviations and indicating potential data anomalies or directions for model optimization, providing a reference for the use of the assessment results. Finally, by structuring and integrating the above content, a dynamic risk assessment report for geological disasters can be formed, which includes core modules such as dynamic risk spatiotemporal distribution, analysis of key risk causes, risk evolution patterns, and verification of the reliability of assessment results.

[0063] The aforementioned method first analyzes multi-source exploration data to obtain unit spatiotemporal feature tuples and physical correlation matrices, achieving systematic integration of multi-source data and accurate extraction of spatiotemporal features, thus improving the quality of basic data for geological hazard assessment. Secondly, based on the unit spatiotemporal feature tuples and physical correlation matrices, graph neural networks are used for computation to obtain dynamic risk assessment results, multi-level attention weights, and physical state deviations. This overcomes the problems of missing physical correlations and insufficient capture of dynamic evolution features in traditional methods, enhancing the accuracy and dynamism of the assessment results. Finally, by integrating the dynamic risk assessment results, multi-level attention weights, and physical state deviations into a report, not only can the risk level be output, but the causes of risk can also be deeply analyzed, enhancing the interpretability and practicality of the assessment report and meeting the needs of engineering practice for precise disaster risk prevention and control and decision support.

[0064] In one embodiment, multi-source exploration data is parsed to obtain unit spatiotemporal feature tuples and physical correlation matrices, including:

[0065] Spatiotemporal gridding alignment of remote sensing data, topographic data, geological data, meteorological monitoring data, and engineering monitoring data from multi-source exploration data is performed to obtain unified multi-source raw data. Based on the unified multi-source raw data, static and dynamic features are separated, and the static features are standardized and extracted to obtain the static background feature matrix of each spatial unit. The dynamic features are temporally regularized to obtain the dynamic observation feature tensor of each spatial unit.

[0066] Based on the preset seepage mechanics equations, physical analysis processing is performed on the time-series rainfall data in the dynamic observation feature tensor to obtain the hydrophysical feature tensor of each spatial unit.

[0067] Inter-unit runoff is extracted from hydrophysical feature tensors. Combined with geological structure similarity information in topographic data, physical correlation is calculated on the inter-unit runoff to obtain the physical correlation and construct a physical correlation matrix.

[0068] Based on the signal decomposition model, physical analysis processing is performed on the temporal deformation data in the dynamic observation feature tensor to obtain the deformation physical feature tensor of each spatial unit; the deformation physical feature tensor contains deformation residual components.

[0069] The static background feature matrix, hydrophysical feature tensor, and deformation physical feature tensor are concatenated by feature dimensions to obtain the unit spatiotemporal feature tuple.

[0070] Specifically, when performing spatiotemporal gridding alignment processing on remote sensing data, topographic data, geological data, meteorological monitoring data, and engineering monitoring data from multi-source exploration data, a unified spatial reference coordinate system and temporal sampling benchmark can be established first. For example, the spatial reference coordinate system can adopt the National Geodetic Coordinate System 2000, and the temporal sampling benchmark can be adapted according to the data type. Static data, such as topographic and geological data, can be based on a single sampling, while dynamic data, such as meteorological and engineering monitoring data, can be uniformly standardized into time-series data with fixed time intervals, such as 1 hour / sample. Subsequently, a regular grid partitioning method is used to divide the study area into several equal-sized spatial units, and spatial interpolation processing can be performed on each type of data based on bilinear interpolation, so that all data are mapped to the spatial units of the unified grid. In addition, timestamp alignment and missing value imputation can be performed on the time-series data simultaneously. That is, linear interpolation is used to imput short-term missing data, and LSTM model prediction is used to imput long-term missing data. Finally, a unified multi-source raw data can be obtained. This data is stored in a three-dimensional structure of spatial unit-time step-data type, which can ensure the consistency of data from different sources in the spatiotemporal dimension.

[0071] Based on unified multi-source raw data, static and dynamic features can be separated according to the temporal stability of data attributes. Static features are those that do not change over time or whose annual rate of change is less than a preset threshold (e.g., 0.1%), including geological lithology codes, topographic slope, soil type, and rock stratum depth. Dynamic features, on the other hand, are those that change significantly over time, including rainfall intensity, deformation displacement, groundwater level, and soil moisture. The static features can then be standardized using the Z-score normalization method, thereby eliminating dimensional differences between different static features and obtaining the static background feature matrix for each spatial unit. The matrix dimension is "number of spatial units × number of static features". For the dynamic features, a time-series data resampling method can be used to unify dynamic features from different sampling frequencies into a time-series sequence with a preset time step. A Savitzky-Golay filter is then used to smooth and denoise the resampled time-series data, resulting in the dynamic observation feature tensor for each spatial unit. The tensor dimension is "number of spatial units × number of time steps × number of dynamic features".

[0072] Based on a pre-defined seepage mechanics equation, time-series rainfall data in a dynamic observation feature tensor can be transformed into characteristic parameters reflecting soil hydrological conditions through hydrophysical mechanisms. For example, time-series rainfall data for each time step can be extracted from the dynamic observation feature tensor, representing the cumulative rainfall for each spatial unit at the corresponding time step. Subsequently, a rainfall infiltration and soil moisture transport model is constructed based on the seepage mechanics equation. This model simulates the infiltration process of rainfall in soil pores, changes in moisture distribution, and the evolution of pore water pressure. Finally, a set of hydrophysical parameters for each spatial unit at each time step is output. Organizing this parameter set according to the dimension of "spatial unit - time step - hydrophysical parameter" yields the hydrophysical feature tensor for each spatial unit.

[0073] Furthermore, inter-unit runoff can be extracted from the hydrophysical characteristic tensor. This inter-unit runoff represents the water exchange volume between adjacent spatial units due to differences in hydraulic potential energy, and this parameter directly reflects the hydrological correlation strength between adjacent units. Based on this inter-unit runoff, the physical correlation degree can also be calculated by combining geological structure similarity information from topographic data. Geological structure similarity information includes lithological consistency, stratum dip angle differences, and fault structure distribution overlap between adjacent spatial units. Therefore, this information can be quantified first. Lithological consistency can be assigned a value based on lithological type matching degree, such as 1 for complete matching, 0.5 for partial matching, and 0 for no matching. Stratum dip angle differences can be calculated using the inverse cosine function of the dip angle difference; the smaller the difference, the closer the quantified value is to 1. Fault structure distribution overlap can be calculated based on the proportion of the overlap length of the fault zone in adjacent units. Subsequently, a physical correlation degree calculation model can be constructed, with the following formula:

[0074]

[0075] in, For the first Spatial unit in time step The physical correlation degree of j, with a value range of [0,1]; For the first Spatial unit in time step and radial flux; For the first The maximum radial flux between spatial units at each time step; spatial unit and The geological structure similarity coefficient, with a value range of [0,1], is calculated based on the matching degree of rock lithology and fault distribution in the terrain data; spatial unit and Spatial Euclidean distance; This is the distance attenuation coefficient, with a value range of [0.1, 0.5], calibrated according to the complexity of the regional terrain (larger values ​​for mountainous areas and smaller values ​​for plains); For the first Spatial unit in time step Peak pore water pressure; , and All are weighting coefficients, satisfying These correspond to the contribution weights of runoff, geological similarity, and pore water pressure, respectively, and can be calibrated using the analytic hierarchy process (AHP) combined with actual disaster case data. Based on this formula, the physical correlation degree of all adjacent spatial units can be calculated, and a physical correlation matrix can be constructed. The matrix dimension is "number of spatial units × number of spatial units", and the matrix elements are... The matrix can be either 0 (adjacent units) or 0 (non-adjacent units), and it intuitively represents the distribution of the physical correlation strength between spatial units.

[0076] Specifically, empirical mode decomposition (EMD) or variational mode decomposition (VMD) can be used as the signal decomposition model to perform physical analysis on the temporal deformation data in the dynamic observation feature tensor. This allows the nonlinear and non-stationary temporal deformation data to be decomposed into several stationary intrinsic mode functions (IMFs) and a residual component. The temporal deformation data represents the surface displacement of each spatial unit at different time steps. After processing by the signal decomposition model, each IMF corresponds to different frequency components in the deformation data, such as short-term noise, seasonal deformation, and trend deformation. The residual component corresponds to the anomalous deformation components in the deformation data that cannot be decomposed into stationary modes, i.e., the deformation residual component. Integrating the decomposed IMF features with the deformation residual component yields the deformation physical feature tensor for each spatial unit. The tensor dimension is "number of spatial units × number of time steps × number of deformation features," which accurately characterizes the physical laws and anomalous features of the deformation process.

[0077] Furthermore, the static background feature matrix can be expanded into a three-dimensional tensor (dimensions of the number of spatial units × number of time steps × number of static features), where the static feature values ​​remain consistent at each time step. The expanded static feature tensor, hydrophysical feature tensor, and deformation physical feature tensor are then stacked according to their feature dimensions to obtain a unit spatiotemporal feature tuple. The dimension of this tuple is the number of spatial units × number of time steps × (number of static features + number of hydrophysical features + number of deformation physical features), which comprehensively integrates the static background attributes, dynamic hydrological state, and dynamic deformation state of the spatial unit, providing complete feature input for subsequent modeling.

[0078] In one embodiment, based on the seepage mechanics equations, physical analysis is performed on the time-series rainfall data in the dynamic observation feature tensor to obtain the hydrophysical feature tensor of each spatial unit, which may include:

[0079] S1021. Extract the time-series rainfall data corresponding to each time step from the dynamic observation feature tensor, and perform rainfall intensity standardization processing on the time-series rainfall data to obtain the unit area rainfall intensity sequence corresponding to each time step.

[0080] Specifically, time-series rainfall data consists of the cumulative rainfall depth for each spatial unit at each time step. This data extraction process can be performed by matching the spatial unit identifier with the time step identifier based on the index information of the dynamic observation feature tensor, ensuring accurate correspondence between the rainfall data for each spatial unit at each time step. Subsequently, when standardizing the time-series rainfall data for rainfall intensity, the cumulative rainfall depth can be first converted to rainfall intensity, for example:

[0081]

[0082] in, Rainfall intensity, This represents the cumulative rainfall depth at the corresponding time step. The duration of the time step is defined. Based on the area information of the spatial cells, rainfall intensity can be converted into rainfall intensity per unit area, i.e., the amount of rainfall per unit area per unit time. This results in a sequence of rainfall intensity per unit area for each time step, which provides data support for the infiltration boundary conditions of the subsequent seepage model.

[0083] S1022. Based on soil type data in geological data, combined with rock layer distribution data and porosity data in topographic data, determine the soil hydraulic parameters of each spatial unit; soil hydraulic parameters include saturated hydraulic conductivity, air intake suction and pore size distribution index.

[0084] Specifically, when determining the soil hydraulic parameters of each spatial unit based on soil type data from geological data, combined with rock layer distribution data and porosity data from topographic data, the soil type data in the geological data can first be classified and coded, such as sandy loam coded as 1, loam coded as 2, clay coded as 3, etc. Then, the initial hydraulic parameters for the corresponding soil type can be obtained by consulting a soil hydraulic parameter handbook based on the coding results, including reference values ​​for saturated hydraulic conductivity, air intake suction, and pore size distribution index. Subsequently, the saturated hydraulic conductivity is corrected based on the rock layer distribution data. If a permeable rock layer (such as sandstone) exists within the spatial unit, a correction factor of 1.2-1.5 can be multiplied from the initial saturated hydraulic conductivity; if an impermeable rock layer (such as shale) exists, a correction factor of 0.3-0.6 can be multiplied. Furthermore, the pore size distribution index can be calibrated based on porosity data; for example, the calibration formula could be:

[0085]

[0086] in, The calibrated pore size distribution index. The initial pore size distribution index. This represents the actual porosity of the spatial unit. To reference the standard porosity of soil types, this calibration ensures that the soil hydraulic parameters match the actual geological conditions of the spatial unit. The final determined soil hydraulic parameters include saturated hydraulic conductivity, air intake suction, and pore size distribution index.

[0087] S1023. The preset seepage mechanics equation is used as the core control equation. The unit area rainfall intensity sequence corresponding to each time step is used as the infiltration boundary condition. Soil hydraulic parameters are used as the initial parameters of the equation to construct a one-dimensional vertical seepage numerical model for each spatial unit. The preset seepage mechanics equation is the Richards seepage mechanics equation.

[0088] Specifically, the Richards equation can be expressed as:

[0089] in, Soil volumetric water content, For time, Vertical depth Unsaturated hydraulic conductivity The pore water pressure head is the equation describing the relationship between the rate of change of soil moisture content over time and the vertical seepage velocity, thus characterizing the transport of soil moisture during rainfall infiltration. When constructing the model, the rainfall intensity sequence per unit area corresponding to each time step can be used as the infiltration boundary condition; that is, the upper boundary is rainfall infiltration, and the lower boundary is an impermeable boundary or a free drainage boundary, determined according to the groundwater depth in the study area. Soil hydraulic parameters can be used as initial parameters for the equation, and initial conditions for the model can be set (such as initial soil moisture content being field capacity, and initial pore water pressure being the head value corresponding to atmospheric pressure), ultimately forming a one-dimensional vertical seepage numerical model for each spatial unit.

[0090] S1024. The one-dimensional vertical seepage numerical model is discretized based on the finite volume method to obtain the discretized seepage numerical model. The discretized seepage numerical model is solved by implicit difference scheme to obtain the soil moisture content distribution data and pore water pressure distribution data in the spatial unit corresponding to each time step.

[0091] Specifically, during discretization, the vertical depth direction of the spatial unit can first be divided into several control volumes, such as each control volume having a thickness of 0.1m. Each control volume serves as a discretized computational unit. Integrating the Richards equations within each control volume yields the integral form of the control equations.

[0092]

[0093] in, and The first The first control unit in the first and the Moisture content at time step For time step, To control the body thickness, and The first The first control body and the first , No. The water conductivity of the interface of the control body. , , These represent the pore water pressure head for the corresponding control volume. Subsequently, when solving the discretized equation using an implicit difference scheme, a system of equations can be constructed and solved using the Newton-Raphson iterative method to obtain the soil moisture content distribution data and pore water pressure distribution data for each control volume within the spatial unit corresponding to each time step. Integrating these data according to vertical depth yields the soil moisture content distribution data and pore water pressure distribution data at the spatial unit level.

[0094] S1025. Perform feature extraction processing on the soil moisture content distribution data and pore water pressure distribution data corresponding to each time step to obtain the mean soil moisture content, peak pore water pressure and seepage velocity vector of each spatial unit.

[0095] Specifically, the average soil moisture content of each spatial unit can be obtained by calculating the arithmetic mean of the soil moisture content of all control volumes, i.e. , This represents the average soil moisture content. To control the number of bodies, For the first The soil moisture content of each control volume. The peak pore water pressure is the maximum pore water pressure in each control volume, which can be directly obtained from the pore water pressure distribution data. The seepage velocity vector can be calculated based on Darcy's law, where the expression for Darcy's law is:

[0096]

[0097] in, This represents the vertical seepage velocity; the negative sign indicates that the seepage direction is opposite to the direction of the head gradient. This refers to unsaturated hydraulic conductivity, whose value varies with soil volumetric water content. The higher the moisture content, the greater the hydraulic conductivity. 1 represents the vertical gradient of the pore water pressure head, characterizing the rate of change of the pore water pressure head along the vertical depth direction; 1 represents the gradient contribution corresponding to the gravity term, which is a fixed constant because gravity causes water to permeate vertically downwards. For pore water pressure head, The vertical depth is represented by the velocity vector. By calculating the seepage velocity of each control volume, the seepage velocity vector of the spatial unit can be obtained, which reflects the direction and intensity of soil moisture transport.

[0098] S1026. Based on the seepage velocity vector of each spatial unit and the pore water pressure distribution data of adjacent spatial units, calculate the hydraulic gradient between adjacent spatial units. Combined with the saturated hydraulic conductivity in the soil hydraulic parameters, calculate the inter-unit runoff using Darcy's law to obtain the inter-unit runoff corresponding to each time step.

[0099] Specifically, the hydraulic gradient is the ratio of the difference in pore water pressure head between adjacent spatial units to the unit spacing, and its formula can be:

[0100]

[0101] in, spatial unit With spatial units The hydraulic gradient between them and They are spatial units and Pore ​​water pressure head spatial unit and The center-to-center spacing. Combining the saturated hydraulic conductivity in the soil hydraulic parameters, the inter-unit runoff is calculated using Darcy's law, with the following formula:

[0102]

[0103] in, This refers to the radial flow rate between units. For saturated hydraulic conductivity, spatial unit and The contact area is calculated, and the inter-unit radial flux corresponding to each time step is obtained through this calculation, providing core parameters for the construction of the physical correlation matrix.

[0104] S1027. The mean soil moisture content, peak pore water pressure, seepage velocity vector and inter-unit runoff corresponding to each time step are integrated in a time-series dimension to construct the hydrophysical characteristic tensor of each spatial unit.

[0105] Specifically, the mean soil moisture content, peak pore water pressure, seepage velocity vector, and inter-unit runoff at each time step can be combined to form a hydrophysical feature vector (with the dimension of "number of hydrophysical parameters") for that time step. Then, these feature vectors are stacked sequentially according to the time steps to form a time-series feature sequence. Ultimately, this allows the construction of a hydrophysical feature tensor for each spatial unit. This tensor has the dimension of time step number × number of hydrophysical parameters, comprehensively encompassing the hydrophysical state information of the spatial unit at each time step.

[0106] In one embodiment, the unit spatiotemporal feature tuple includes temporal information of multiple time steps, and the temporal information of each time step includes a static background feature matrix, a hydrophysical feature tensor, and a deformation physical feature tensor.

[0107] Based on the unit spatiotemporal feature tuples and physical correlation matrix, graph neural network computation is performed to obtain dynamic risk assessment results, multi-level attention weights, and physical state deviations, including:

[0108] Traverse each time step of the unit spatiotemporal feature tuple, construct the adjacency relationship of the graph structure corresponding to each time step based on the physical correlation matrix, and perform feature initialization processing on each node in the corresponding graph structure based on the unit spatiotemporal feature tuple of the current time step to obtain the initialized node features and the corresponding initialized graph structure for each time step.

[0109] For each time step, the initialized graph structure is subjected to multi-layer physical constraint graph convolution processing to obtain the optimized node features of the final layer output; the multi-layer physical constraint graph convolution processing corresponds to the following steps S1031-S1033:

[0110] S1031. Perform message aggregation processing on the current node features and neighboring node features in the initialized graph structure of the current layer to obtain aggregated features; wherein, in the first layer processing, the current node features are the initialized node features;

[0111] S1032. Perform linear transformation on the aggregated features to obtain intermediate node features; construct a physical verifier based on the preset physical equation, and perform physical consistency verification on the intermediate node features through the physical verifier to obtain the physical state deviation of the current layer.

[0112] S1033. The physical state deviation and intermediate node features are weighted and fused to obtain the updated node features. The updated node features are used as the current node features of the next layer. The physical constraint graph convolution process is repeated until each layer has undergone physical constraint graph convolution process to obtain the optimized node features of the final layer output at the corresponding time step.

[0113] Based on the optimized node features of the final layer output in each time step and the physical state deviation of each layer, factor attention weights are calculated to obtain the factor attention weight vector corresponding to each node in each time step.

[0114] During the message aggregation process at each time step, spatial correlation is calculated based on the current node features of the current layer and the deformation physical feature tensor of the corresponding time step to obtain the spatial correlation of the current layer. The physical correlation matrix and the spatial correlation of the current layer are weighted and fused to obtain the dynamic spatial attention weight of the current layer.

[0115] The optimized node features of the final layer corresponding to each time step are subjected to temporal feature fusion processing. In the temporal feature fusion processing, historical hydrological state is extracted from the hydrophysical feature tensor of each time step. Based on the historical hydrological state and the optimized node features output by the final layer of the corresponding time step, the temporal attention weight is calculated to obtain the temporal attention weight corresponding to each time step.

[0116] Risk probability mapping is performed on the optimized node features of the final layer corresponding to each time step to generate dynamic risk assessment results for each time step.

[0117] At each time step, the corresponding factor attention weight vector, time attention weight, and dynamic spatial attention weight of each layer are integrated to obtain multi-level attention weights. The physical state deviation of each layer corresponding to each time step is summarized to form the physical state deviation.

[0118] Specifically, the multi-time-step temporal information contained in the unit spatiotemporal feature tuple can comprehensively depict the temporal dynamics of geological disaster evolution, while the physical correlation matrix clarifies the physical constraints between spatial units. Therefore, based on these two, graph neural network computation can achieve dynamic risk assessment and interpretability analysis through deep modeling in the spatiotemporal dimension and physical mechanism constraints, and output the physical state deviation for result verification. Illustratively, when traversing each time step of the unit spatiotemporal feature tuple and constructing the adjacency relationship of the graph structure corresponding to each time step based on the physical correlation matrix, since the hydrological and deformation states of spatial units differ at different time steps, the corresponding spatial correlation strength, although based on the physical correlation matrix, must also ensure that the graph structure of each time step independently represents the spatial interaction relationship at the current moment, avoiding cross-time-step correlation interference. For example, each spatial unit can be mapped as a node in the graph structure, and the elements in the physical correlation matrix can be directly used as the weights of the edges between corresponding nodes to represent the physical correlation strength between adjacent units. The edge weights of non-adjacent units are set to 0, thus constructing the graph structure adjacency relationship corresponding to each time step.

[0119] Subsequently, when initializing the features of each node in the graph structure based on the spatiotemporal feature tuples of the units at the current time step, we can first extract the static background feature matrix, hydrophysical feature tensor, and deformation physical feature tensor of each spatial unit at the current time step. These three types of features are then concatenated according to their feature dimensions to form an initial node feature vector. This vector is then subjected to Min-Max normalization to ensure that the feature values ​​are within the same numerical range. This process yields the initialized node features and the initialized graph structure for each time step, providing standardized input for subsequent graph convolution calculations.

[0120] Specifically, multi-layer physical constraint graph convolution is performed on the initialized graph structure corresponding to each time step. This not only captures the multi-level correlation influence between spatial units through deep iteration, but also ensures that the model learning process conforms to the physical laws of geological disaster evolution by incorporating physical constraints. When performing message aggregation, feature information from all neighboring nodes is collected for each node in the current layer. This information is then weighted and summed based on the edge weights (physical correlation strength) between nodes to obtain the aggregated features. The calculation formula is as follows:

[0121]

[0122] in, For the target node The aggregation characteristics, For the target node The set of neighboring nodes, For the target node With neighboring nodes Edge weights (physical correlation) between them Neighboring nodes The current node's characteristics, during the first layer of processing, This involves initializing node features. When performing linear transformations on aggregated features, a fully connected layer can be used to perform dimensional mapping and nonlinear transformations on the aggregated features to obtain intermediate node features. The activation function of the fully connected layer can be the ReLU function to enhance the model's nonlinear fitting ability.

[0123] In a schematic illustration, when constructing a physical verifier based on preset physical equations, the preset physical equations can be selected from the core equilibrium equations corresponding to the geological hazard type. For example, a soil stability equilibrium equation is used for landslides, and a fluid dynamics equilibrium equation is used for debris flows. The core function of this physical verifier is to calculate the deviation between intermediate node features and theoretical physical values. When performing physical consistency verification on intermediate node features through the physical verifier, the theoretical physical values ​​of the node features can first be calculated based on the preset physical equations, combined with environmental parameters at the current time step (such as soil moisture content and pore water pressure). Then, the absolute difference between the intermediate node features and the theoretical physical values ​​is calculated to obtain the physical state deviation of the current layer. This deviation directly reflects the physical rationality of the intermediate features. Furthermore, the physical state deviation and intermediate node features can be weighted and fused. The fusion weight can be determined based on the normalized result of the physical state deviation. The larger the deviation, the larger the fusion weight, to strengthen the corrective effect of physical constraints on unreasonable features. For example, the fusion formula can be:

[0124]

[0125] in, For the updated node features, The fusion weights (ranging from 0.3 to 0.5, calibrated according to the disaster type) are used for fusion. Features of intermediate nodes This is the theoretical physical value. This represents the deviation of the physical state of the current layer. The updated node features are used as the current node features of the next layer. The above message aggregation, linear transformation, physical consistency verification, and weighted fusion processing are repeated until the preset number of layers are processed, resulting in the optimized node features of the final layer output at the corresponding time step. These features contain both spatial multi-level correlation information and conform to the laws of physical evolution.

[0126] Specifically, factor attention weights are used to quantify the contribution of three types of factors—static background features, hydrophysical features, and deformation physical features—to the risk assessment results. For example, the correlation score between the feature corresponding to each factor and the feature of the final optimization node can be calculated first. This score is obtained by calculating the Pearson correlation coefficient. Then, the correlation score is weighted and fused with the mean of the deviation of the physical state at each layer to obtain the factor importance score. Finally, the factor importance score is normalized using the Softmax function to obtain the factor attention weight vector corresponding to each node at each time step. The dimension of this vector is consistent with the number of factor categories, and the sum of the element values ​​is 1, which intuitively reflects the contribution ratio of each type of factor.

[0127] Furthermore, when calculating the dynamic spatial attention weights during the message aggregation processing at each time step, spatial correlation can be calculated first based on the current node features of the current layer and the deformation physical feature tensor of the corresponding time step. This captures the synergy between the current layer node and its neighboring nodes in terms of feature response and deformation state. For example, first calculate the cosine similarity of the current node features of the target node and its neighboring nodes, then calculate the Euclidean distance between the deformation residual components in their deformation physical feature tensors. By weighting and fusing the cosine similarity and the reciprocal of the Euclidean distance, the spatial correlation can be obtained. When weighting and fusing the physical correlation matrix with the spatial correlation of the current layer, the weight coefficient of the physical correlation matrix can be set to... For example, if the value range is 0.5-0.7, it is determined based on the reliability of geological data, while the weighting coefficient for spatial correlation is 1- By weighted summation, the dynamic spatial attention weight of the current layer can be obtained. This weight retains the fixed association based on physical mechanism and incorporates the dynamic association of the current layer feature evolution, which makes the interaction weight between nodes more in line with the actual evolution state.

[0128] Specifically, by performing temporal feature fusion processing on the optimized node features of the final layer corresponding to each time step, the temporal evolution pattern of the risk assessment results can be captured, and feature noise from a single time step can be eliminated. For example, temporal feature fusion processing can employ a temporal attention mechanism, and in this process, historical hydrological states can be extracted from the hydrophysical feature tensor of each time step. These historical hydrological states include cumulative parameters such as the average soil moisture content and peak pore water pressure of the previous N time steps. Subsequently, temporal attention weights can be calculated based on the historical hydrological states and the optimized node features output from the final layer of the corresponding time step. That is, the similarity between the hydrological states of the current time step and the historical time steps is first calculated, and then the contribution score of the optimized node features of the historical time steps to the current risk assessment is calculated (determined by the mapping error between historical features and current features). By multiplying the hydrological state similarity and the contribution score, the temporal importance score can be obtained. After Softmax normalization, the temporal attention weight corresponding to each time step can be obtained. This weight can highlight the key time steps that have a significant impact on the current risk.

[0129] Specifically, a mapping model can be constructed using fully connected layers combined with a sigmoid activation function to perform risk probability mapping on the optimized node features of the final layer corresponding to each time step. The input to the fully connected layer is a weighted fusion vector of the optimized node features of the final layer and the temporal attention weights, and the output is a probability value between 0 and 1, which represents the probability of disaster occurrence for the corresponding spatial unit at the current time step. By organizing the probability values ​​of all spatial units according to the spatial unit index and the time step index, dynamic risk assessment results corresponding to each time step can be generated.

[0130] Finally, multi-level attention weights can be integrated at each time step. This involves using the temporal attention weight as the global weight and multiplying it element-wise with the factor attention weight vector and the dynamic spatial attention weights of each layer. This yields an integrated multi-level attention weight tensor, which comprehensively includes attention information across the factor, spatial, and temporal dimensions. When summarizing the physical state deviations of each layer at each time step, a summary vector can be constructed using the maximum and mean of the physical state deviations at each layer. This vector reflects both the worst-case scenario of physical consistency and the overall deviation level, ultimately forming the physical state deviation, providing data support for verifying the reliability of subsequent evaluation results.

[0131] In one embodiment, such as Figure 2 As shown, based on the dynamic risk assessment results, multi-level attention weights, and physical state deviation, a report generation process is performed to obtain a dynamic risk assessment report for geological hazards, including:

[0132] S201: Obtain and compare the dynamic risk assessment results corresponding to each time step with the preset risk threshold to identify high-risk units;

[0133] S203: For each high-risk unit, extract the factor attention weight vector of the corresponding time step, combine it with the feature dimension of the time series information of the corresponding time step in the unit spatiotemporal feature tuple, and perform quantification analysis of the contribution of the dominant factor to obtain the dominant factor contribution analysis chart.

[0134] S203: For each high-risk unit, extract the dynamic spatial attention weight of the corresponding time step, perform spatial influence range and spatial influence intensity visualization processing to obtain a spatial influence correlation heatmap; extract the temporal attention weight of the corresponding time step of the high-risk unit, perform quantitative analysis of key influence moments to obtain a key influence moment curve.

[0135] S204: For each high-risk unit, extract the physical state deviation at the corresponding time step, combine it with the preset physical constraint standard, perform physical consistency deviation analysis processing, and obtain the physical state consistency analysis diagram.

[0136] S205: The dynamic risk assessment results, high-risk units, contribution analysis charts of dominant factors, spatial impact correlation heatmaps, key impact moment curves, and physical state consistency analysis charts corresponding to each time step are structured and integrated to obtain a dynamic risk assessment report for geological disasters.

[0137] Specifically, the preset risk threshold can be optimized using ROC curve analysis, taking into account the historical geological disaster records, engineering safety standards, and disaster loss tolerance of the study area. For example, firstly, geological disaster cases from the past N years can be collected in the study area. The dynamic risk assessment results (historical simulation values) of the spatial units corresponding to the cases at each time step before the disaster occurred can be extracted. Using the occurrence of the disaster as a positive sample and the absence of the disaster as a negative sample, an ROC curve can be plotted, and the risk probability value corresponding to the largest area under the curve can be selected as the preset risk threshold. Subsequently, when comparing the dynamic risk assessment results at each time step with the threshold, it can be determined whether the risk probability value of each spatial unit is greater than or equal to the preset risk threshold. If it is satisfied, it is marked as a high-risk unit, and the spatial coordinate index, corresponding time step identifier, and specific risk probability value of the unit are recorded to form a list of high-risk units, ensuring the accurate location and temporal traceability of high-risk units.

[0138] Specifically, when performing a quantitative analysis of the dominant factor contribution for each high-risk unit, the factor attention weight vector at the corresponding time step can be extracted. This vector contains the contribution proportions of three types of factors: static background features, hydrophysical features, and deformation physical features. Subsequently, the feature dimensions of the time-series information at the corresponding time step in the unit's spatiotemporal feature tuple can be combined. Since the feature dimensions clearly define the specific parameters included in each type of factor (e.g., static background features include parameters such as topographic slope and geological lithology, while hydrophysical features include parameters such as soil moisture content and pore water pressure), the mean attention weight of each type of factor can be calculated first. Then, this mean can be weighted and fused with the standard deviation of the parameters under the corresponding feature dimension to obtain the dominant contribution score of each type of factor. Based on the dominant contribution score, a bar chart can be constructed to analyze the dominant factor contribution. The horizontal axis represents the factor category, and the vertical axis represents the contribution score. The chart can also be labeled with the core parameter names and corresponding weight proportions of each type of factor, allowing decision-makers to intuitively grasp the core disaster-causing factors and key parameters of the high-risk unit.

[0139] To illustrate, when visualizing the spatial impact range and intensity for each high-risk unit, the dynamic spatial attention weight at the corresponding time step can be extracted first. This weight represents the dynamic correlation strength between the high-risk unit and its neighboring units. The spatial impact range can be defined centered on the high-risk unit, selecting neighboring units whose dynamic spatial attention weights are greater than a preset correlation threshold (determined based on the mean of the physical correlation matrix) to constitute the impact range. The impact range is then divided into three levels according to the weight values: core impact area, secondary impact area, and potential impact area. The spatial impact intensity can be visualized using a heatmap. The spatial baseline of the heatmap is consistent with the spatial unit grid of the evaluation model, and the color depth represents the magnitude of the dynamic spatial attention weight; the darker the color, the stronger the correlation. Simultaneously, the spatial coordinates of the high-risk unit and the boundary lines of the impact area levels are overlaid on the heatmap to form a spatial impact correlation heatmap, thus clearly presenting the risk propagation path and impact boundary of the high-risk unit. In addition, the time attention weights of the corresponding time steps of high-risk units can be extracted for quantitative analysis of key impact moments. The time attention weights reflect the contribution intensity of each time step to the current high-risk state. Therefore, a key impact moment curve can be plotted with the time step as the horizontal axis and the attention weight value as the vertical axis. The time step corresponding to the peak of the curve is the key impact moment. At the same time, key triggering conditions such as hydrological and physical characteristics (such as cumulative rainfall) and deformation and physical characteristics (such as cumulative displacement) corresponding to the moment can be marked in the chart, thereby clarifying the temporal evolution nodes of the high-risk state.

[0140] In a schematic example, when analyzing the physical state consistency deviation for each high-risk unit, the physical state deviation at the corresponding time step can be extracted. This deviation reflects the degree of agreement between the assessment results and the physical evolution laws of geological hazards. Subsequently, a preset physical constraint standard can be used. This standard is a theoretical deviation range derived from the core physical equations of geological hazards, such as the soil stability coefficient deviation threshold for landslides and the fluid shear stress deviation threshold for debris flows. Specific values ​​can be determined through indoor physical model tests or numerical simulation calibration. During analysis, the ratio of the physical state deviation of the high-risk unit to the preset physical constraint standard can be calculated. If the ratio is less than 1, it is judged as "physically consistent"; if the ratio is greater than or equal to 1, it is judged as "physically deviating," and the specific parameters corresponding to the deviation (such as the magnitude of the deviation of pore water pressure from the theoretical value) are recorded. Subsequently, a physical state consistency analysis chart can be constructed using a line graph. The horizontal axis of the chart represents the spatial coordinate index of high-risk units, and the vertical axis represents the ratio of the deviation to the standard. At the same time, a baseline with a ratio of 1 is drawn to visually mark high-risk units with large physical deviations. The possible causes of the deviations, such as data acquisition errors or sudden changes in local geological conditions, can be explained in the chart notes to provide a basis for judging the reliability of the assessment results.

[0141] Finally, using time steps as the framework, the dynamic risk assessment results of each time step can be integrated into a dynamic risk spatiotemporal distribution module, combining dynamic risk assessment results with a list of high-risk units. This module allows users to switch between time steps to view the risk evolution process. Secondly, the contribution analysis charts of the dominant factors for each high-risk unit and the spatial impact correlation heatmaps can be integrated to form a key risk cause analysis module, clearly identifying the core disaster-causing factors and their spatial impact range. Subsequently, the key impact moment curves of all high-risk units can be integrated to extract common evolutionary patterns, such as the fact that most high-risk units are triggered after X consecutive rainfall time steps, forming a risk evolution pattern module. Finally, integrating the physical state consistency analysis chart and the explanation of deviation causes creates a module for verifying the reliability of the assessment results. Organizing these modules logically, using a standardized engineering document format, and supporting interactive association between spatial coordinates and charts (e.g., clicking on the coordinates of a high-risk unit will jump to the corresponding analysis chart), ultimately generates a dynamic geological disaster risk assessment report. This report not only meets the engineering decision-making needs for precise risk positioning but also provides a clear target direction for the formulation of prevention and control measures through multi-dimensional analysis.

[0142] Based on the same inventive concept, this application also provides a disaster area assessment system based on geological hazard investigation for implementing the disaster area assessment method based on geological hazard investigation described above. The solution provided by this system is similar to the solution described in the above method. Therefore, the specific limitations of one or more embodiments of the disaster area assessment system based on geological hazard investigation provided below can be found in the limitations of the disaster area assessment method based on geological hazard investigation described above, and will not be repeated here.

[0143] In one exemplary embodiment, such as Figure 3 As shown, a disaster area assessment system 300 based on geological hazard investigation is provided, including:

[0144] Data parsing module 301 is used to parse multi-source exploration data to obtain unit spatiotemporal feature tuples and physical correlation matrices;

[0145] The graph neural network analysis module 302 is used to perform graph neural network calculations based on the unit spatiotemporal feature tuples and physical correlation matrix to obtain dynamic risk assessment results, multi-level attention weights, and physical state deviations.

[0146] The report generation module 303 is used to generate a report based on the dynamic risk assessment results, multi-level attention weights, and physical state deviations, to obtain a dynamic risk assessment report for geological disasters.

[0147] In one exemplary embodiment, the present invention also provides a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the disaster area assessment method based on geological disaster investigation of this application. A multi-core processor is preferred to improve the system's parallel processing capability. The memory provides sufficient temporary storage space to support program execution and data processing. The memory capacity should be large enough to accommodate large amounts of data and computational tasks.

[0148] In one exemplary embodiment, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the disaster area assessment method based on geological hazard investigation of this application. The computer-readable storage medium may include: a read-only memory, a random access memory, a solid-state drive, or an optical disk, etc.

[0149] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A disaster area assessment method based on geological hazard investigation, characterized in that, The method includes: The multi-source exploration data is analyzed to obtain the spatiotemporal feature tuples and physical correlation matrices of the units; Based on the spatiotemporal feature tuples of the unit and the physical correlation matrix, graph neural network calculations are performed to obtain dynamic risk assessment results, multi-level attention weights, and physical state deviations. Based on the dynamic risk assessment results, the multi-level attention weights, and the physical state deviation, a report generation process is performed to obtain a dynamic risk assessment report for geological disasters.

2. The method according to claim 1, characterized in that, The process of parsing multi-source exploration data to obtain unit spatiotemporal feature tuples and physical correlation matrices includes: The remote sensing data, topographic data, geological data, meteorological monitoring data, and engineering monitoring data in the multi-source exploration data are subjected to spatiotemporal grid alignment processing to obtain unified multi-source raw data. Based on the unified multi-source raw data, static features and dynamic features are separated, and the static features are subjected to standardized extraction processing to obtain the static background feature matrix of each spatial unit. The dynamic features are subjected to temporal regularization processing to obtain the dynamic observation feature tensor of each spatial unit. Based on the preset seepage mechanics equations, the time-series rainfall data in the dynamic observation feature tensor is physically analyzed to obtain the hydrophysical feature tensor of each spatial unit. Inter-unit runoff is extracted from the hydrophysical feature tensor. Combined with the geological structure similarity information in the topographic data, the physical correlation degree of the inter-unit runoff is calculated to obtain the physical correlation degree, and the physical correlation matrix is ​​constructed. Based on the signal decomposition model, the temporal deformation data in the dynamic observation feature tensor is physically analyzed to obtain the deformation physical feature tensor of each spatial unit; the deformation physical feature tensor contains deformation residual components. The static background feature matrix, the hydrophysical feature tensor, and the deformation physical feature tensor are concatenated by feature dimensions to obtain the unit spatiotemporal feature tuple.

3. The method according to claim 1, characterized in that, The unit spatiotemporal feature tuple includes temporal information of multiple time steps, and the temporal information of each time step includes a static background feature matrix, a hydrophysical feature tensor, and a deformation physical feature tensor. The graph neural network computation based on the unit spatiotemporal feature tuples and the physical correlation matrix yields dynamic risk assessment results, multi-level attention weights, and physical state deviations, including: Traverse each time step of the unit spatiotemporal feature tuple, construct the adjacency relationship of the graph structure corresponding to each time step based on the physical correlation matrix, and perform feature initialization processing on each node in the corresponding graph structure based on the unit spatiotemporal feature tuple of the current time step to obtain the initialized node features and the corresponding initialized graph structure for each time step. For each time step, the initialized graph structure is subjected to multi-layer physical constraint graph convolution processing to obtain the optimized node features of the final layer output; the multi-layer physical constraint graph convolution processing corresponds to the following steps: The current node features and neighboring node features in the initialized graph structure of the current layer are subjected to message aggregation processing to obtain aggregated features; wherein, in the first layer processing, the current node features are the initialized node features; The aggregated features are linearly transformed to obtain intermediate node features; a physical verifier is constructed based on a preset physical equation, and the physical verifier is used to perform physical consistency verification on the intermediate node features to obtain the physical state deviation of the current layer. The physical state deviation and the intermediate node features are weighted and fused to obtain updated node features. The updated node features are used as the current node features of the next layer. The physical constraint graph convolution process is repeated until each layer has undergone physical constraint graph convolution process to obtain the optimized node features of the final layer output corresponding to the time step. Based on the optimized node features of the final layer output in each time step and the physical state deviation of each layer, factor attention weight calculation is performed to obtain the factor attention weight vector corresponding to each node in each time step. In the message aggregation process corresponding to each time step, spatial correlation degree calculation is performed based on the current node features of the current layer and the deformation physical feature tensor of the corresponding time step to obtain the spatial correlation degree of the current layer; the physical correlation matrix and the spatial correlation degree of the current layer are weighted and fused to obtain the dynamic spatial attention weight of the current layer. The optimized node features of the final layer corresponding to each time step are subjected to time-series feature fusion processing. In the time-series feature fusion processing, historical hydrological states are extracted from the hydrophysical feature tensor of each time step. Based on the historical hydrological states and the optimized node features output by the final layer of the corresponding time step, time attention weights are calculated to obtain the time attention weights corresponding to each time step. Risk probability mapping is performed on the optimized node features of the final layer corresponding to each time step to generate the dynamic risk assessment result corresponding to each time step. In each of the time steps, the corresponding factor attention weight vector, the time attention weight, and the dynamic spatial attention weight of each layer are integrated to obtain the multi-level attention weight, and the physical state deviation of each layer corresponding to each time step is summarized to form the physical state deviation.

4. The method according to claim 3, characterized in that, The process of generating a report based on the dynamic risk assessment results, the multi-level attention weights, and the physical state deviation yields a dynamic geological hazard risk assessment report, including: Acquire and compare the dynamic risk assessment results corresponding to each time step with a preset risk threshold to identify high-risk units; For each high-risk unit, the factor attention weight vector corresponding to the time step is extracted, and combined with the feature dimension of the temporal information corresponding to the time step in the unit spatiotemporal feature tuple, the contribution quantification analysis of the dominant factor is performed to obtain the contribution analysis chart of the dominant factor. For each high-risk unit, the dynamic spatial attention weight corresponding to the time step is extracted, and the spatial influence range and spatial influence intensity are visualized to obtain a spatial influence correlation heatmap; the time attention weight corresponding to the time step of the high-risk unit is extracted, and the key influence moment is quantitatively analyzed to obtain a key influence moment curve. For each high-risk unit, the physical state deviation at the corresponding time step is extracted, and combined with the preset physical constraint standard, physical consistency deviation analysis is performed to obtain a physical state consistency analysis diagram. The dynamic risk assessment results, high-risk units, contribution analysis charts of dominant factors, spatial impact correlation heatmaps, key impact moment curves, and physical state consistency analysis charts corresponding to each time step are structured and integrated to obtain the dynamic risk assessment report of geological disasters.

5. The method according to claim 2, characterized in that, The process of performing physical analysis on the time-series rainfall data in the dynamic observation feature tensor based on the seepage mechanics equations to obtain the hydrophysical feature tensor of each spatial unit includes: Extract the time-series rainfall data corresponding to each time step from the dynamic observation feature tensor, and perform rainfall intensity standardization on the time-series rainfall data to obtain the unit area rainfall intensity sequence corresponding to each time step; Based on the soil type data in the geological data, combined with the rock layer distribution data and porosity data in the topographic data, the soil hydraulic parameters of each spatial unit are determined; the soil hydraulic parameters include saturated hydraulic conductivity, air intake suction, and pore size distribution index. Using the preset seepage mechanics equation as the core control equation, the rainfall intensity sequence per unit area corresponding to each time step is used as the infiltration boundary condition, and the soil hydraulic parameters are used as the initial parameters of the equation to construct a one-dimensional vertical seepage numerical model for each spatial unit; the preset seepage mechanics equation is the Richards seepage mechanics equation. The one-dimensional vertical seepage numerical model is discretized based on the finite volume method to obtain the discretized seepage numerical model. The discretized seepage numerical model is then solved using an implicit difference scheme to obtain the soil moisture content distribution data and pore water pressure distribution data within the spatial unit corresponding to each time step. Feature extraction processing is performed on the soil moisture content distribution data and pore water pressure distribution data corresponding to each time step to obtain the mean soil moisture content, peak pore water pressure and seepage velocity vector of each spatial unit; Based on the seepage velocity vector of each spatial unit and the pore water pressure distribution data of adjacent spatial units, the hydraulic gradient between adjacent spatial units is calculated. Combined with the saturated hydraulic conductivity in the soil hydraulic parameters, the inter-unit runoff is calculated using Darcy's law to obtain the inter-unit runoff corresponding to each time step. The mean soil moisture content, peak pore water pressure, seepage velocity vector, and inter-unit runoff corresponding to each time step are integrated along a temporal dimension to construct the hydrophysical feature tensor of each spatial unit.

6. The method according to claim 2, characterized in that, The physical correlation degree is calculated using the following formula: in, For the first Spatial unit in time step The physical correlation degree has a value range of [0,1]. For the first Spatial unit in time step and radial flux; For the first The maximum radial flux between the spatial units at each time step; spatial unit and The geological structure similarity coefficient, with a value range of [0,1]; spatial unit and Spatial Euclidean distance; This is the distance attenuation coefficient; For the first Spatial unit in time step Peak pore water pressure; , and All are weighting coefficients, satisfying These correspond to the contribution weights of runoff, geological similarity, and pore water pressure, respectively.

7. A disaster area assessment system based on geological hazard investigation, characterized in that, The system includes: The data parsing module is used to parse multi-source exploration data to obtain unit spatiotemporal feature tuples and physical correlation matrices; The graph neural network analysis module is used to perform graph neural network calculations based on the spatiotemporal feature tuples of the unit and the physical correlation matrix to obtain dynamic risk assessment results, multi-level attention weights, and physical state deviations. The report generation module is used to generate a report based on the dynamic risk assessment results, the multi-level attention weights, and the physical state deviation, to obtain a dynamic risk assessment report for geological disasters.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images