Geological disaster susceptibility assessment method and device, electronic equipment, medium and product

Abstract

The present application relates to the technical field of avalanche risk assessment, and discloses a geological disaster susceptibility assessment method, device, electronic equipment, medium and product, the method comprising: acquiring multi-source heterogeneous data of a target area; according to the multi-source heterogeneous data, using a multi-layer limit equilibrium method to identify snow layer instability zones, and using a Copula function to deduce a multi-level chain disaster path; according to the snow layer instability zones and the multi-level chain disaster path, constructing a heterogeneous graph containing disaster-causing nodes and associated relationships; using a graph neural network to perform spatial deduction on the heterogeneous graph, and embedding a preset physical constraint loss to obtain the start probability, movement run and chain triggering probability of each node in the heterogeneous graph; according to the start probability, movement run and chain triggering probability, generating an avalanche susceptibility assessment result, the present application realizes the organic integration of mechanism models and data-driven models, ensures the physical consistency of the assessment result, and thus improves the accuracy and reliability of geological disaster susceptibility assessment.

Description

Technical Field

[0001] This invention relates to the field of avalanche risk assessment technology, specifically to methods, devices, electronic equipment, media, and products for assessing the susceptibility to geological disasters. Background Technology

[0002] High-altitude long-distance avalanches are one of the major natural disasters faced by linear engineering projects such as power transmission lines. Existing assessment methods suffer from problems such as low accuracy in identifying snow layer instability, lack of assessment of multi-causal coupling effects and multi-level chain disasters, and poor physical consistency due to the separation between mechanism models and data-driven models, making it difficult to meet the accurate assessment needs in complex high-altitude environments.

[0003] Existing technologies mostly assess the risk of triggering a single disaster, lacking quantitative methods for the cascading effects of avalanches, landslides, debris flows, and glacial lake outburst floods. Furthermore, mechanistic models and data-driven models are independent of each other, resulting in insufficient physical consistency of assessment results. In addition, the assessment models have fixed spatiotemporal scales, failing to balance comprehensive investigation with detailed monitoring of key areas. Summary of the Invention

[0004] This invention provides a method, device, electronic equipment, medium, and product for assessing the susceptibility to geological disasters, in order to solve the problems of low accuracy in identifying snow layer instability, lack of assessment of multi-causal coupling effects and multi-level chain disasters, and poor physical consistency caused by the separation of mechanism models and data-driven models in the existing technology.

[0005] In a first aspect, the present invention provides a method for assessing the susceptibility to geological hazards, the method comprising: Acquire multi-source heterogeneous data of the target region; Based on multi-source heterogeneous data, the multi-level limit equilibrium method was used to identify the snow layer instability zone, and the Copula function was used to deduce the multi-level chain disaster path; Based on the snow layer instability zone and multi-level chain disaster pathways, a heterogeneous graph containing disaster-causing nodes and their relationships is constructed. A graph neural network is used to perform spatial extrapolation on the heterogeneous graph, and a preset physical constraint loss is embedded to obtain the start probability, movement distance and chain trigger probability of each node in the heterogeneous graph. Avalanche susceptibility assessment results are generated based on the initiation probability, running distance, and chain triggering probability.

[0006] This invention provides a geological hazard susceptibility assessment method. It identifies snow layer instability zones using a multi-level limit equilibrium method and uses a Copula function to extrapolate multi-level chain disaster paths, achieving accurate identification of snow layer instability and complete assessment of disaster chains. By constructing a heterogeneous graph containing disaster-causing nodes and their relationships, and employing a graph neural network for spatial extrapolation, while embedding pre-set physical constraint losses, it achieves an organic integration of mechanistic models and data-driven models, ensuring the physical consistency of the assessment results. This improves the accuracy and reliability of geological hazard susceptibility assessment and provides technical support for the risk quantification of multi-level chain disasters.

[0007] In one optional implementation, the snow layer instability zone is identified using a multi-layer limit equilibrium method based on multi-source heterogeneous data, including: The multi-layer snow profile of the target area is obtained, and the snow layer stability coefficient is calculated by the multi-layer limit equilibrium method based on the multi-layer snow profile. The snow layer stability coefficient is based on the Mohr-Coulomb strength criterion and combined with the multi-weak layer combination effect and the tension-shear composite instability mode. It is determined by the snow layer cohesion, normal stress, pore water pressure and internal friction angle. When the snow layer stability coefficient is less than the preset value, the multi-layer snow body profile is identified as the snow layer instability zone.

[0008] In the above technical solution, by obtaining multi-layer snow body profiles and using the multi-layer limit equilibrium method to calculate the snow layer stability coefficient, the snow layer stability identification is extended from a single snow layer to a multi-layer snow body structure. By combining the combined effect of multiple weak layers and the tensile-shear composite instability mode, the limitations of traditional methods that only consider a single slip surface or a single instability mode are overcome. This makes the identification results closer to the actual multi-layer snow body failure mechanism of high-altitude long-distance avalanches, thereby significantly improving the accuracy of snow layer instability zone identification and providing more reliable source area information support for subsequent avalanche susceptibility assessment.

[0009] In one optional implementation, a Copula function is used to deduce multi-level chain disaster paths, including: The trigger probabilities of multiple factors are obtained, and the joint probability distribution of multiple factors is constructed using the Copula function to obtain the comprehensive trigger probability of multiple factors. Based on the combined triggering probability of multiple factors, the instability probability and superimposed impact range of various types of disaster-causing objects downstream are determined step by step along the terrain connection direction, and the multi-level chain disaster path is deduced.

[0010] In the above technical solution, by obtaining the trigger probability of multiple factors and constructing a joint probability distribution using the Copula function, the comprehensive trigger probability of multiple factors is quantified, overcoming the problem of missing multi-factor coupling effect in traditional methods. At the same time, by judging the instability probability and superimposed impact range of downstream disaster-causing objects step by step along the terrain connection direction, the quantitative inference of multi-level chain disaster propagation path is realized, overcoming the limitation of traditional methods that only assess the initiation risk of a single disaster and ignore the cascading effect, and providing complete technical support for the whole-chain risk assessment of avalanche-landslide-debris flow-glacial lake outburst.

[0011] In one optional implementation, a heterogeneous graph containing disaster-causing nodes and their relationships is constructed based on the snow layer instability zone and multi-level chain disaster pathways, including: The grid unit nodes are determined by dividing the target area into grids; The source node location in the heterogeneous graph is determined based on the snow layer instability zone, and the triggering relationship edge in the heterogeneous graph is determined based on the multi-level chain disaster path; Assign corresponding attribute features to source region nodes and grid cell nodes to construct a heterogeneous graph containing disaster-causing nodes and their relationships; the attribute features include terrain features, snow cover features, meteorological features, geological features and deformation features.

[0012] In the above technical solution, the target area is divided into grid unit nodes, and the snow layer instability zone and multi-level chain disaster path are mapped to the source area node position and trigger relationship edge in the heterogeneous graph, respectively. At the same time, the nodes are assigned multi-dimensional attribute features such as terrain, snow cover, meteorology, geology and deformation, which realizes the structured integration of disaster identification results and multi-source heterogeneous data. This provides a unified data carrier for subsequent graph neural network inference and overcomes the problem of lack of effective connection between mechanism analysis results and data-driven models in traditional methods.

[0013] In one optional implementation, determining the source node location in the heterogeneity map based on the snow layer instability region includes: The spatial extent of the snow layer instability zone is mapped to grid cell nodes, and the grid cell nodes that fall within the spatial extent are marked as avalanche source zone nodes.

[0014] In the above technical solution, by mapping the spatial range of the snow layer instability zone to grid cell nodes and marking the grid cell nodes falling within this range as avalanche source zone nodes, the instability identification results are accurately mapped to the graph structure nodes, providing a clear starting node position for graph neural network inference and ensuring the accuracy of source zone spatial positioning in subsequent inferences.

[0015] In one optional implementation, determining the triggering relationship edges in the heterogeneous graph based on multi-level chain disaster paths includes: Along the propagation direction of the multi-level chain disaster path, triggering relationship edges are established between the upper-level disaster node and the lower-level disaster node. The direction of the triggering relationship edges represents the cascading propagation direction of the disaster.

[0016] In the above technical solution, by establishing directional triggering relationship edges between the upper-level disaster nodes and the lower-level disaster nodes along the propagation direction of the multi-level chain disaster path, the chain inference results are quantified into an edge structure that can be processed by a graph neural network. This allows the cascading propagation direction of the disaster to be embedded into the heterogeneous graph with a clear topological relationship, providing a topological basis for the calculation of the chain triggering probability in subsequent spatial inference.

[0017] In one optional implementation, constructing a heterogeneous graph containing disaster-causing nodes and their relationships further includes: Based on the distribution location of at least one disaster-bearing entity among transmission towers, line corridors, main traffic lines, engineering camps, and residential areas within the target area, the disaster-bearing entity nodes are determined. Add the disaster-bearing node to the heterogeneous graph.

[0018] In the above technical solution, by determining the distribution location of at least one disaster-bearing body among transmission towers, line corridors, traffic arteries, engineering camps, and residential areas as disaster-bearing body nodes and adding them to a heterogeneous graph, the assessment object is expanded from natural disaster-causing nodes to disaster-bearing body nodes. This achieves a unified graphical representation of the disaster impact range and the spatial distribution of disaster-bearing bodies, providing a node-level connection basis for the subsequent transformation of assessment results into engineering protection recommendations.

[0019] In one alternative implementation, the disaster-bearing node is added to the heterogeneous graph, including: Based on the geographical location of the disaster-bearing body, create a disaster-bearing body node at the corresponding location in the heterogeneous graph, and use the disaster-bearing body type, vulnerability level, and spatial range as the initial attributes of the disaster-bearing body node; Based on the spatial overlap between the disaster-bearing body node and the grid unit node at its location, spatial association edges are established between the disaster-bearing body node and the grid unit node to construct a heterogeneous graph.

[0020] In the above technical solution, by assigning the disaster-bearing body type, vulnerability level, and spatial range as initial attributes to the disaster-bearing body nodes, the disaster-bearing body nodes are loaded with the basic attribute information required for risk assessment during the heterogeneous graph construction stage. At the same time, spatial association edges are established based on the spatial overlap relationship between the disaster-bearing body nodes and the grid unit nodes in their location, realizing the effective connection between the disaster-bearing body and the surrounding environment in the heterogeneous graph, and providing a topological basis for the effective transmission of assessment results to the disaster-bearing body in subsequent simulations.

[0021] In one optional implementation, a graph neural network is used to perform spatial extrapolation on the heterogeneous graph, and a preset physical constraint loss is embedded to obtain the activation probability, movement distance, and chain triggering probability of each node in the heterogeneous graph, including: Using the attribute features of each node in the heterogeneous graph as the initial embedding and the connection relationship of each edge in the heterogeneous graph as the propagation path, a graph neural network is used to map the heterogeneous graph from the spatial domain to the spectral domain. Learnable convolutional kernel operations are performed in the spectral domain to capture global spatial dependencies, and then the graph is inversely transformed back to the spatial domain to iteratively update the features of each node. The data fitting loss is calculated, and the total loss function is obtained by weighted summation of the data fitting loss and the preset physical constraint loss. This is the derivation process of the constraint graph neural network. The preset physical constraint loss includes: activation constraint loss, motion constraint loss, and chain triggering constraint loss. Based on the updated characteristics of each node, output the start probability, running distance, and chain trigger probability of each node.

[0022] In the above technical solution, by using the attribute features of each node in the heterogeneous graph as the initial embedding and the edge connection relationship as the propagation path, a learnable convolution kernel operation is performed in the spectral domain to capture global spatial dependencies, overcoming the limitation of traditional spatial domain convolution that can only aggregate local neighborhood information. At the same time, by using the weighted sum of data fitting loss, activation constraint loss, motion constraint loss, and chain triggering constraint loss as the total loss function, the inference process is subject to the dual constraints of data-driven and physical laws, realizing the fusion of the mechanism model and the data-driven model, ensuring the physical consistency of the output results, and directly outputting the activation probability, motion range, and chain triggering probability of each node, providing complete risk quantification parameters for subsequent evaluation.

[0023] In one alternative implementation, the graph neural network includes a graph Fourier transform network and an inverse graph Fourier transform network, iteratively updating the features of each node, including: In each layer of propagation, the features of each node in the previous layer are aggregated by neighbor analysis and then mapped from the spatial domain to the spectral domain through a graph Fourier transform network. In the spectral domain, these features are multiplied with learnable convolutional kernels, and then returned to the spatial domain through an inverse graph Fourier transform network to obtain the updated node features for the current layer. This layer-by-layer propagation captures global spatial dependencies. The formula for each layer of propagation is as follows: ; in, Let i be the feature vector of node i in the (l+1)th layer. For the graph Fourier transform, This is the inverse graph Fourier transform. Let l be a learnable convolution kernel in the spectral domain. For element-wise product, For activation function, The weight of the terrain-connected edges. To perform weighted feature aggregation on all neighboring nodes j of node i.

[0024] In the above technical solution, node features are mapped from the spatial domain to the spectral domain through graph Fourier transform. In the spectral domain, they are multiplied with learnable convolutional kernels to capture global spatial dependencies. Then, the inverse transform is used to return to the spatial domain, so that each layer of propagation can integrate long-distance correlation information between nodes in the global scope. This overcomes the limitation of traditional graph neural networks, which can only aggregate local neighborhood features and are difficult to capture the whole picture of the long-range motion of avalanches. At the same time, multi-scale abstraction and iterative update of features are realized through layer-by-layer propagation.

[0025] In one alternative implementation, the total loss function is calculated as follows: ; in, For the total loss function, For data fitting loss, For the initiation constraint loss based on the Moore-Coulomb criterion, The motion constraint loss is based on the Voellmy rheological model. For chain-triggered constraint loss based on limit equilibrium method, All are weighting coefficients.

[0026] In one optional implementation, a graph neural network is used to perform spatial extrapolation on the heterogeneous graph, and a preset physical constraint loss is embedded, further comprising: By using terrain features, snow cover features, and geological features as covariates, a spatial distribution function of rheological parameters in a graph neural network is constructed. Online stochastic gradient descent or Kalman filtering is used to dynamically update the rheological parameters over time. Adaptive calibration of physical parameters is performed based on spatial distribution function and dynamic time updates.

[0027] In the above technical solution, a spatial distribution function of rheological parameters is constructed by using terrain features, snow cover features, and geological features as covariates, enabling the rheological parameters to adaptively adjust with changes in spatial location. Simultaneously, online stochastic gradient descent or Kalman filtering is used to dynamically update the rheological parameters over time, allowing for real-time calibration as monitoring data is integrated. This combination of spatial distribution function and dynamic time updates achieves adaptive calibration of physical parameters, overcoming the limitations of traditional methods that use globally fixed rheological parameters, which restricts the accuracy of motion simulation under different terrain, snow cover, and geological conditions. This allows the rheological parameters to dynamically adapt to changes in spatial location and time, thereby improving the accuracy of calculating the running distance constraint loss.

[0028] In one optional implementation, an avalanche susceptibility assessment result is generated based on the initiation probability, running distance, and chain triggering probability, including: The comprehensive risk index of each node is determined based on the activation probability, running distance, and chain trigger probability of each node. Based on the comprehensive risk index, the susceptibility level of each node is determined according to the preset grading threshold, and the avalanche susceptibility assessment results of the target area are generated; the susceptibility levels include low susceptibility area, medium susceptibility area, high susceptibility area and extremely high susceptibility area.

[0029] In the above technical solution, the activation probability, running distance and chain trigger probability of each node are integrated into a comprehensive risk index. The comprehensive risk index is divided into four susceptibility levels: low, medium, high and extremely high, according to the preset grading threshold. This achieves the graded output of avalanche susceptibility assessment results, making the assessment results intuitively reflect the risk differences in different areas and providing a clear basis for differentiated graded management.

[0030] In one alternative implementation, the method further includes: High-risk sections are identified based on avalanche susceptibility assessment results, and the assessment of high-risk sections is automatically switched from kilometer-level daily scale to hundred-meter-level hourly scale. Based on the uncertainty of the assessment results quantified by Bayesian neural networks, a confidence level classification for each avalanche susceptibility assessment result is obtained. In the above technical solution, high-risk sections are identified based on the avalanche susceptibility assessment results, and the assessment of high-risk sections is automatically switched from a kilometer-level daily scale to a hundred-meter-level hourly scale, overcoming the problem of insufficient detailed information on high-risk sections caused by the use of a globally fixed assessment scale in traditional methods. Simultaneously, by quantifying the uncertainty of the assessment results using Bayesian neural networks, a confidence level classification for each assessment result is output, overcoming the problem of traditional assessments lacking quantitative indicators of reliability, thus providing dual support for differentiated risk management and decision-making confidence assessment.

[0031] In one alternative implementation, the method further includes: Based on the avalanche susceptibility assessment results, climate scenario data were obtained, and a time-domain generalized graph network was used to predict the spatiotemporal migration patterns of avalanche-prone areas in different future periods under different climate scenarios. A long-term adaptation map is generated based on the prediction results. The long-term adaptation map includes location information of high-risk corridor areas, location information of recommended avoidance areas, and location information of areas where protection can be strengthened.

[0032] The above technical solution overcomes the problem that traditional assessments only address current conditions and cannot support long-term planning by acquiring climate scenario data and using a time-domain generalized graph network to predict the spatiotemporal migration patterns of avalanche-prone areas in different future periods under different climate scenarios. At the same time, by generating a long-term adaptation map containing location information of high-risk corridor areas, suggested avoidance areas, and areas that can be strengthened for protection, it provides spatial decision-making references for long-term risk response under different climate scenarios.

[0033] Secondly, the present invention provides a geological hazard susceptibility assessment device, the device comprising: The data acquisition module is used to acquire multi-source heterogeneous data from the target area; The instability identification and risk quantification module is used to identify snow layer instability zones based on multi-source heterogeneous data using the multi-level limit equilibrium method, and to deduce multi-level chain disaster paths using the Copula function; The heterogeneous graph construction module is used to construct a heterogeneous graph containing disaster-causing nodes and their relationships based on snow layer instability zones and multi-level chain disaster paths. The hybrid inference module is used to perform spatial inference on heterogeneous graphs using graph neural networks and embeds a preset physical constraint loss to obtain the start probability, movement range and chain trigger probability of each node in the heterogeneous graph. The assessment module is used to generate avalanche susceptibility assessment results based on the initiation probability, running distance, and chain trigger probability.

[0034] Thirdly, the present invention provides an electronic device, comprising: a memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the computer instructions to perform the geological hazard susceptibility assessment method of the first aspect or any corresponding embodiment described above.

[0035] Fourthly, the present invention provides a computer-readable storage medium storing computer instructions for causing a computer to execute the geological hazard susceptibility assessment method of the first aspect or any corresponding embodiment described above.

[0036] Fifthly, the present invention provides a computer program product, including computer instructions for causing a computer to execute the geological hazard susceptibility assessment method of the first aspect or any corresponding embodiment described above. Attached Figure Description

[0037] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0038] Figure 1 This is a first structural block diagram of a geological hazard susceptibility assessment system according to an embodiment of the present invention; Figure 2 This is a second structural block diagram of a geological hazard susceptibility assessment system according to an embodiment of the present invention; Figure 3This is a schematic diagram of a cloud-edge collaborative deployment architecture according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the first process of the geological hazard susceptibility assessment method according to an embodiment of the present invention; Figure 5 This is a schematic diagram of the second process of the geological hazard susceptibility assessment method according to an embodiment of the present invention; Figure 6 This is a schematic diagram of the PC-HGNO (Physics-Constrained Heterogeneous Graph Neural Operator) hybrid modeling algorithm architecture and working principle according to an embodiment of the present invention; Figure 7 This is a schematic diagram of the third process of the geological hazard susceptibility assessment method according to an embodiment of the present invention; Figure 8 This is a structural block diagram of a geological hazard susceptibility assessment device according to an embodiment of the present invention; Figure 9 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0040] It is understood that before using the technical solutions disclosed in the various embodiments of the present invention, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in the present invention and their authorization should be obtained in accordance with relevant laws and regulations through appropriate means.

[0041] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0042] As an optional application scenario of this invention, such as Figure 1 As shown, this invention provides a geological hazard susceptibility assessment system, which further includes: a perception layer, a data layer, a cognition layer, a deduction layer, a support layer, and a decision-making layer; wherein: A risk-driven active perception and multi-source fusion module is set up in the perception layer to construct an integrated three-dimensional perception network of space-air-ground-depth (snow depth), increasing risk-driven collaborative scheduling capabilities and continuous monitoring capabilities along the entire line. This enables risk-driven adaptive dynamic scheduling of monitoring resources and continuous full-coverage monitoring along the transmission line, ultimately acquiring multi-source heterogeneous data of the target area. It consists of a space-based remote sensing observation unit, an airborne UAV detailed inspection unit, a ground-based in-situ monitoring network unit, a distributed fiber optic acoustic sensing unit, a multi-element data access unit, and a meta-reinforcement learning collaborative scheduling unit. Among them: The satellite-based remote sensing observation unit integrates data from Sentinel-1 / 2 SAR (Sentinel-1 / Sentinel-2 Synthetic Aperture Radar), optical satellites, Gaofen series satellites, ICESat-2 (Ice, Cloud, and Land Elevation Satellite-2), laser altimetry satellites, meteorological satellites, and hyperspectral satellites to retrieve snow cover, snow depth, snow surface temperature, glacier flow velocity, and surface deformation. Hyperspectral data is used to obtain microscopic physical parameters such as snow grain size and liquid water content; multi-angle remote sensing is used to obtain information on the anisotropic structure of snow. The satellite-based remote sensing observation unit is used to acquire large-scale, continuous coverage of basic data on snow cover, snow depth, snow surface temperature, glacier flow velocity, and surface deformation in the target area, providing regional-scale background information on the disaster-prone environment for avalanche susceptibility assessment.

[0043] Airborne UAV detailed inspection unit: based on oblique photography and airborne LiDAR (Light Detection and Ranging), point cloud density ≥50 points / m². 2 The system acquires sub-meter-level topographic and snow surface features in key sections, and identifies anomalies such as snow surface cracks, snow eaves, and moraine. The UAV is equipped with multispectral / hyperspectral sensors to monitor snow accumulation characteristics. The UAV radiosonde system is used to intensively detect the vertical profiles of atmospheric temperature and humidity.

[0044] Ground-based in-situ monitoring network unit: Deploys automatic snow profile monitoring equipment (based on snowplow-type or spiral automatic sampling mechanisms, combined with temperature chains, near-infrared moisture sensors, and dielectric constant sensors to achieve automated continuous profile measurement of snow layer thickness, density, temperature gradient, water content, and crystal morphology in each vertical layer), MIMO (Multiple-Input Multiple-Output) radar (detection of internal snow layer structure), snow layer stress sensing array, near-surface meteorological station (temperature, wind speed and direction, snowfall, radiation), snow depth radar, microwave radiometer (atmospheric temperature and humidity profile, 0-10km), wind profile radar (vertical distribution of upper-level wind field), and in-situ snow mechanical parameter testing equipment (snow layer shear strength, tensile strength, and compression creep parameters), outputting hourly snow layer physical and mechanical parameters.

[0045] Distributed Fiber Acoustic Sensing Unit: Utilizing existing OPGW (Optical Fiber Composite Overhead Ground Wire) or dedicated sensing cables along the transmission line, a joint monitoring system of Distributed Acoustic Sensing (DAS) and Brillouin Optical Time-Domain Analysis (BOTDA) is deployed. DAS monitors microseismic signals, internal snow layer fracturing signals, and ground vibrations generated by avalanche movement in real time with decimeter-level spatial resolution (measurement point spacing as small as 25cm), enabling real-time detection and location of avalanche events. BOTDA monitors temperature and static strain distribution along the line based on the Brillouin scattering principle, working in conjunction with DAS to achieve dynamic and static joint monitoring, acquiring information on line icing status and tower foundation deformation. The joint monitoring system is unaffected by visual obstruction and severe weather, achieving continuous, full-coverage monitoring along the transmission line.

[0046] Multi-element data access unit: Unifies access to numerical weather prediction, watershed hydrology, seismic network and geological structure data, completes precipitation phase identification and ground motion parameter calculation; accesses monitoring data on glacial lake area and water level changes based on optical / SAR (Synthetic Aperture Radar) imagery, glacier surface velocity monitoring data based on SAR offset tracking, and glacier terminus change monitoring data.

[0047] Meta-reinforcement learning collaborative scheduling unit: Adaptive dynamic scheduling of monitoring resources is achieved using the meta-reinforcement learning framework. A Markov Decision Process (MDP) is constructed: the state space S represents the current availability of each monitoring resource, historical observation coverage, current meteorological conditions, and regional risk level; the action space A represents satellite programming observation parameters, UAV autonomous flight path and parameters, radar scan cycle and range, and observation mode switching; the reward function R is designed as a task-adaptive reward mechanism. (1); in, To predict the width of the confidence interval (uncertainty). For resource consumption costs, For information gain (adding effective observation coverage). Task urgency (based on real-time risk assessment). These are dynamic weighting coefficients.

[0048] Proximal Policy Optimization (PPO) is employed as the inner reinforcement learning algorithm for task-level optimization, while the outer layer uses a Long Short-Term Memory (LSTM) network as a meta-learner to capture cross-task prior knowledge. Based on the snow condition anomaly identification results, encrypted UAV aerial surveys and frequency-increased sampling of ground stations are automatically triggered, while simultaneously outputting optimization schemes for filling observation blind spots.

[0049] A snow-topography coupled data governance and knowledge graph module is set up in the data layer to perform full-process governance of multi-source heterogeneous data, and to build a coupled data foundation with a unified spatiotemporal benchmark and a disaster domain knowledge graph. It consists of a data quality control and intelligent fusion unit, a spatiotemporal registration and indexing unit, a snow-topography coupled database unit, a historical avalanche case knowledge base unit, and a metadata and shared service unit. Among them: The data quality control and intelligent fusion unit employs a temporal deep learning model for outlier identification and nonlinear interpolation. It utilizes a semantic difference fusion Transformer (SDFTransformer) architecture for heterogeneous data fusion, comprising a dual-branch, multi-level fusion structure: the feature enhancement stage extracts directional context through a dual-axis attention module to maintain structural continuity; the feature learning stage reduces spectral variation in optical images and suppresses speckle noise in SAR images using a region-aware visual Transformer; and the fusion stage integrates optical and SAR features through a difference-guided cross-modal hierarchical fusion module. A Bayesian weighted fusion algorithm outputs a snow depth and snow density fusion product with uniform accuracy.

[0050] Spatiotemporal registration and indexing unit: used to unify coordinate system, time reference and spatial resolution, complete spatial registration with 3km×3km grid as reference and complete temporal alignment with hourly time step; build a spatiotemporal integrated indexing mechanism to support millisecond-level data retrieval by region and time phase.

[0051] Snow-topography coupled database unit: used to classify and store five sub-databases: topography, snow cover, meteorology, geology, and disaster information, and to establish the coupling relationship between snow layer parameters and topographic features.

[0052] Historical avalanche case knowledge base unit: used for structured storage of high-level remote avalanche historical events, constructing an avalanche knowledge graph with event-environment associations, and supporting model calibration and verification.

[0053] Metadata and Shared Services Unit: Used for unified management of data metadata, providing hierarchical access control for data retrieval, subscription, and standardized interface services.

[0054] A static baseline analysis module for high-altitude avalanche-prone environments is set up at the cognitive layer. This module is used to extract detailed avalanche-prone factors from high-altitude terrain with large elevation differences and generate a static baseline susceptibility benchmark map. It consists of a detailed terrain factor extraction unit, an underlying surface feature analysis unit, an avalanche-prone environment sensitivity zoning unit, and a static baseline susceptibility preliminary assessment unit. Among them: Detailed topographic factor extraction unit: used to extract high-level key topographic factors such as slope, aspect, relative elevation difference, snow catchment area, gully longitudinal gradient, checkpoint width, and topographic roughness based on a high-precision DEM (Digital Elevation Model).

[0055] Underlying surface feature analysis unit: used to invert vegetation cover, surface roughness, soil and rock type, distribution of glaciers and moraines, and quantify the impact of the underlying surface on snow stability and motion resistance.

[0056] Disaster-prone environment sensitivity zoning unit: used to classify disaster-prone environment sensitivity levels and identify potential avalanche initiation areas using information volume models and hierarchical analysis.

[0057] Static baseline susceptibility preliminary assessment unit: used to combine the long-term trend of permafrost degradation, calculate the static baseline susceptibility index based on topography and underlying surface factors, and form a baseline map for dynamic assessment.

[0058] The cognitive layer also includes a module for dynamic evolution and instability identification of snow layer structures. This module is used to adapt to the special high-altitude environment, simulate the entire life cycle evolution of snow layers, and identify the instability of multi-layered snow bodies. It consists of a snow accumulation and windblown snow dynamics simulation unit, a snow layer degradation and deep frost development simulation unit, a snow ablation and seepage process simulation unit, a snow layer mechanical parameter closed-loop calibration unit, and a multi-layered snow body combination instability identification unit. Among them: Snow accumulation and blowing snow dynamics simulation unit: used to introduce the groove and narrow tube effect correction algorithm and the steep cliff collapse-re-accumulation algorithm to simulate the spatial redistribution of snow accumulation under the combined effects of snowfall and blowing snow.

[0059] Snow layer degradation and deep frost development simulation unit: used to optimize energy balance parameters under high-altitude, strong ultraviolet and low-pressure environments based on the SNOWPACK physical model (an open-source, one-dimensional, physical process-based multi-layer energy balance snow layer model developed by the Swiss Federal Institute for Snow and Avalanche Research), and to simulate snow crystal degradation and deep frost weak layer development.

[0060] Snow melting and seepage process simulation unit: used to calculate snow melting rate, snow layer liquid water content and seepage front location using energy balance snow melting model and unsaturated seepage algorithm.

[0061] Snow layer mechanical parameter closed-loop calibration unit: used to correct mechanical parameters such as snow layer shear strength, snow layer cohesion and internal friction angle hourly based on in-situ monitoring data.

[0062] Multi-layer snow mass combination instability identification unit: The multi-layer limit equilibrium method is adopted, considering the tension-shear combined instability mode and the combined effect of multiple weak layers. The formula for calculating the snow layer stability coefficient FS is as follows: (2); in, This is the snow layer stability coefficient. For snow layer cohesion, For snow density, The slope angle, It is the acceleration due to gravity. Snow depth Pore ​​water pressure, The internal friction angle is used. When the snow layer stability coefficient FS < 1, snow layer instability is determined to have started, and potential snow layer instability zones are identified.

[0063] The cognitive layer also includes a multi-inducing, coupled chain-like triggering risk quantification module, used to quantify the nonlinear coupling effects of multiple triggering factors and extrapolate multi-level chain disasters; it consists of a meteorological induced factor quantification assessment unit, a seismic ground motion triggering risk unit, an ice avalanche / rockfall impact triggering unit, a multi-inducing factor nonlinear coupling unit, and a multi-level chain-like triggering extrapolation unit. Among them: Meteorological trigger quantitative assessment unit: Based on the threshold method and logistic regression model, calculate the trigger probability of single meteorological triggers such as blizzard, temperature rise, rainfall, and strong wind.

[0064] Seismic triggering risk unit: The Newmark (Newmark Displacement Method) is used to calculate the permanent displacement of the snow layer under seismic motion, outputting the probability and spatial extent of earthquake-triggered avalanches. Real-time seismic monitoring data is integrated, and the probability of earthquake-triggered avalanches is assessed based on a joint criterion of peak ground acceleration, slope, and snow cover conditions.

[0065] Ice avalanche / rockfall impact triggering unit: Based on impact dynamics and debris flow motion simulation, the impact load of high-altitude ice avalanche and rockfall on snow accumulation is quantified, and the probability of impact-triggered avalanche and potential initiation zone are calculated.

[0066] Multi-inducing Nonlinear Coupled Unit: The Copula function is introduced to characterize the joint probability distribution of multiple inducing factors, quantify the synergistic weakening and antagonistic effects among the inducing factors, and output the comprehensive trigger probability and contribution decomposition results of multiple inducing factors. The Copula function is a statistical tool used to construct the joint probability distribution of multivariate random variables. Its core function is to separate the individual characteristics of multiple variables from their correlations and model them separately.

[0067] Multi-level chain-triggered simulation unit: based on avalanche impact force Based on the slope limit equilibrium criterion, the impact of avalanches on downstream slopes, gullies, and glacial lakes is determined, and the spatiotemporal propagation path of avalanche-landslide-debris flow-glacial lake outburst disaster (i.e., multi-level chain disaster path) is deduced. Among them, the potential path of lag effect and the superimposed impact range of glacial lake outburst disaster need to be considered.

[0068] A dynamic simulation module driven by a hybrid mechanism and data is set up in the extrapolation layer. It incorporates a physical constraint heterogeneous graph neural operator engine to uniformly represent multiple types of disaster-causing nodes and their relationships using heterogeneous graphs. Long-distance spatial extrapolation is achieved through graph neural networks (including Fourier spectral domain graph convolution), and three layers of physical constraint loss functions—snow layer initiation, motion rheology, and chain triggering—are embedded to ensure the physical consistency of the evaluation results. The module consists of an automatic avalanche motion path identification unit, a physical constraint heterogeneous graph neural operator unit (i.e., a PC-HGNO hybrid extrapolation unit), a motion phase dynamic evolution unit, a terrain constraint loss correction unit, an avalanche-air wave coupling evaluation unit, and a physical parameter adaptive calibration unit. Among these: Automatic avalanche movement path identification unit: used to extract the movement paths of the main trench and branches based on flow direction analysis and minimum resistance path algorithm, and output the longitudinal profile feature parameters of the path.

[0069] The PC-HGNO (Physics-Constrained Heterogeneous Graph Neural Operator) hybrid inference unit is used to call the Physically Constrained Heterogeneous Graph Neural Operator, input heterogeneous graph data and physical parameters, and output the start probability, predicted range and chain trigger probability of each node, so as to realize near real-time high-precision simulation of avalanche motion speed, thickness and impact force.

[0070] The dynamic evolution unit of the moving phase state is used to calculate in real time the changes in the proportion of media involved in snow liquefaction, rock and soil entrainment, and ice fracturing during the motion process, and dynamically adjusts the rheological parameters in the graph neural network. and .

[0071] Terrain constraint loss correction unit: used to supplement the calculation of local energy loss due to ditch turns, narrowing of the pass, and slope roughness, and to correct the predicted values ​​of long-range movement distance and impact force.

[0072] Avalanche-airwave coupling assessment unit: used to simultaneously calculate the impact force of dense snow flow and the overpressure distribution of powder snow wave based on the two-layer flow theory, and comprehensively assess the combined threat to disaster-bearing bodies at high and low altitudes.

[0073] Physical parameter adaptive calibration unit: used for PC-HGNO-based physical parameter spatial heterogeneity modeling and online adaptive calibration mechanism, realizing spatial adaptation and time dynamic updating of rheological parameters.

[0074] A multi-scale adaptive dynamic evaluation and uncertainty quantification module is set up at the decision-making level to construct a differentiated and adaptive evaluation system, quantify the uncertainty of the entire chain, and realize model self-learning iteration. It consists of a differentiated indicator system construction unit, a multi-model fusion evaluation unit, a spatiotemporal scale adaptive evaluation unit, a full-chain uncertainty quantification unit, and a transfer learning and domain adaptation unit. Among them: Differentiated indicator system construction unit: used to dynamically adjust the weights and grading thresholds of assessment indicators for different disaster-bearing entities such as transportation arteries, power transmission corridors, engineering camps, and residential areas.

[0075] Multi-model fusion evaluation unit: used to fuse statistical models, physical mechanism models and PC-HGNO hybrid models, and output a comprehensive susceptibility index through DS (Dempster-Shafer Theory) evidence theory.

[0076] Spatiotemporal scale adaptive assessment unit: used to support automatic switching between routine daily assessment at the kilometer level across the entire region and detailed hourly assessment at the hundred-meter level in high-risk sections, and to update the 24-hour prediction results every 3 hours.

[0077] The end-to-end uncertainty quantification unit is used to transmit and decompose three types of errors: data input, model parameters, and model structure. It outputs the confidence level of the evaluation results based on a Bayesian neural network.

[0078] Transfer learning and domain adaptation unit: Used to pre-train models with data from global avalanche-prone areas, and transfer knowledge to data-scarce regions through domain adversarial networks to improve generalization in low-sample scenarios.

[0079] The decision-making level also includes a simulation visualization and tiered early warning decision-making module, used to construct a 3D simulation interactive platform to achieve accurate targeted early warning, scenario-based simulation, and engineering risk auxiliary decision-making. It consists of a 3D simulation construction unit, a scenario-based simulation interactive unit, a tiered early warning threshold control unit, an accurate targeted early warning push unit, and an engineering risk and disaster assessment unit. Among them: 3D simulation building unit: Based on a high-precision 3D geographic model of the power transmission corridor, it integrates DEM (Digital Elevation Model) / DSM (Digital Surface Model), 3D model of the power transmission line, surrounding terrain, vegetation and water system elements, and spatial distribution of monitoring equipment to build a digital mirror that maps and synchronizes with the physical world in real time.

[0080] Contextualized simulation and interactive unit: Supports customizable scenario parameters such as snowfall, temperature rise, and earthquake, and calculates and compares avalanche susceptibility, movement process, and impact range under multiple scenarios in real time.

[0081] Tiered early warning threshold control unit: Combining the distribution of disaster-bearing bodies and the level of susceptibility, it sets tiered early warning triggering rules and automatic cancellation logic, and outputs four-color early warning (blue / yellow / orange / red). Each level of early warning is accompanied by uncertainty range and ranking of the contribution of key disaster-causing factors.

[0082] Accurately target early warning push unit: Establish an impact range-affected objects-push channel mapping mechanism to push early warning information in a hierarchical and classified manner.

[0083] Engineering Risk and Disaster Assessment Unit: Based on the vulnerability model of the disaster-bearing body, calculate the failure probability and potential losses, and output suggestions for emergency response, resource allocation and engineering protection.

[0084] The support layer includes a cloud-edge collaborative operation and maintenance (LOO&M) and model self-evolution module to achieve full lifecycle management of the system, adapt to weak network environments in high-altitude areas, and support continuous model self-evolution. It consists of a user and permission management unit, a model algorithm lifecycle management unit, an automated scheduling and O&M monitoring unit, a cloud-edge collaborative deployment unit, and a climate change adaptation and self-evolution unit. A schematic diagram of the cloud-edge collaborative deployment architecture is shown below. Figure 3 As shown, where: User and permission management unit: Based on the RBAC (Role-Based Access Control) model, it implements hierarchical permission management and unified identity authentication, and records all operation logs.

[0085] Model algorithm lifecycle management unit: Enables containerized deployment of models, version control, performance quantification dashboards, and canary releases.

[0086] Automated scheduling and operation and maintenance monitoring unit: orchestrates automated workflows for data access, model calculation, and result publication, and monitors the system's operating status in real time and issues anomaly alerts.

[0087] Cloud-edge collaborative deployment unit: It adopts an architecture that combines cloud-based full-domain computing with edge-based local lightweight assessment: The cloud is responsible for full-domain large-scale computing, big data storage and PC-HGNO model training, while the edge carries local lightweight assessment algorithms, local data access and offline operation capabilities. It supports independent operation of the edge in the event of a network outage, and automatically synchronizes data and early warning results after the network is restored.

[0088] Climate Change Adaptation and Self-Evolution Unit: Coupled with CMIP6 (Coupled Model Intercomparison Project Phase 6) climate scenarios (SSP1-2.6, SSP2-4.5, SSP5-8.5), predict the migration patterns of future avalanche-prone areas; based on new events and early warning feedback, the model achieves continuous self-evolution through incremental learning (Elastic Weight Consolidation (EWC) method).

[0089] This invention provides a workflow for a geological hazard susceptibility assessment system, as follows: Figure 2 As shown: Step 1: Multi-source sensing and data governance (data infrastructure layer): Execution modules: Module 1, Proactive Perception and Multi-Source Fusion; and Module 2, Data Governance and Knowledge Graph. Core actions: Acquire monitoring data from various front-end physical sensors, perform data cleaning and standardization, construct a knowledge graph, and output standardized and usable feature data in a unified manner.

[0090] Step Two: Static Background and Dynamic Evolution Analysis (Dual Baseline Layer): Execution modules: Module 3, static background analysis, and Module 4, snow layer evolution and instability identification are executed in parallel.

[0091] Core operations: On one side, a baseline map of the region's geography / environment is generated based on feature data. On the other side, changes in the snow layer are dynamically tracked, and the snow layer stability coefficient of the snow layer structure is output, providing real-time status information for subsequent risk simulations.

[0092] Step 3: Chain-based deduction and mechanism-data hybrid simulation (core computing layer): Execution Modules: Module 5, Chain-Triggered Risk Quantification to Module 6, Mechanism-Data Hybrid Dynamics Simulation.

[0093] Core Action: Introducing the PC-HGNO engine at the heart of the system. First, it calculates the "chain path" of the disaster and the initial trigger probability; then, combining physical mechanisms and big data, it performs complex dynamic simulations to generate accurate susceptibility / path / chain probability.

[0094] Step 4: Multidimensional Convergence and Adaptive Uncertainty Assessment (Comprehensive Judgment Layer): Execution Module: Module 7, Adaptive Evaluation and Uncertainty Quantification.

[0095] The core process involves aggregating the background map, stability coefficients, chain paths, and simulated probabilities generated in the first three steps. Based on the quantification of uncertainty, a comprehensive assessment is performed, ultimately outputting a visualized risk zoning.

[0096] Step 5: Visualization and Tiered Early Warning Execution (Decision Application Layer): Execution Module: Module 8, Simulation Visualization and Hierarchical Early Warning Decision Making.

[0097] Core function: Visualize the risk assessment results using 3D / 2D simulation, and automatically trigger different levels of early warning instructions based on preset risk thresholds, pushing them to terminal devices or command centers.

[0098] Step Six: Cloud-Edge Collaborative Operation and Maintenance and Continuous Self-Evolution (Closed-Loop Layer Throughout the Entire System): Execution Module: Module Nine, Cloud-Edge Collaborative Operation and Maintenance and Self-Evolution.

[0099] Core actions: Providing operational support through a dotted-line feedback mechanism; providing data governance feedback for step one; and providing model update feedback and threshold optimization for the core engine in step three. This ensures the system can be continuously optimized during actual operation, achieving true intelligent self-evolution.

[0100] How to achieve high-precision dynamic assessment and full-chain closed-loop early warning of avalanche risk in complex high-altitude environments, improve the efficiency of sensing resource utilization and monitoring coverage, achieve high assessment accuracy and physical consistency, complete the integrated simulation and second-level extrapolation of the entire avalanche initiation-movement-chain cascade process, and support cloud-edge collaborative offline operation and long-term disaster prevention planning on a decades-scale scale are urgent problems to be solved.

[0101] This invention provides a method for assessing the susceptibility of geological hazards. By using the multi-level limit equilibrium method to identify snow layer instability zones and combining it with the Copula function to deduce multi-level chain disaster paths, a heterogeneous graph is constructed and a graph neural network with embedded physical constraint loss is used for spatial deduction to generate the initiation probability, movement distance, and chain trigger probability of each node, thereby improving the accuracy and physical consistency of avalanche susceptibility assessment.

[0102] According to an embodiment of the present invention, a method for assessing the susceptibility to geological hazards is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0103] This embodiment provides a method for assessing geological hazard susceptibility, which can be used in the aforementioned electronic device, wherein the electronic device is equipped with a geological hazard susceptibility assessment system. Figure 4 This is a flowchart of a geological hazard susceptibility assessment method according to an embodiment of the present invention, such as... Figure 4 As shown, the process includes the following steps: Step S401: Obtain multi-source heterogeneous data of the target area.

[0104] The target area refers to the geographical area where an avalanche risk assessment needs to be conducted, such as a 2000km stretch of a 500kV transmission corridor in a certain region. 2 scope.

[0105] Multi-source heterogeneous data refers to a collection of data in various formats and types obtained from different sources, which is collected through the risk-driven active perception and multi-source fusion module of the geological hazard susceptibility assessment system.

[0106] Step S402: Based on multi-source heterogeneous data, the multi-level limit equilibrium method is used to identify the snow layer instability zone, and the Copula function is used to deduce the multi-level chain disaster path.

[0107] Among them, the multi-layer limit equilibrium method is a method based on the Mohr-Coulomb strength criterion, combined with the multi-weak layer combination effect and the tension-shear composite instability mode, which identifies the snow layer instability zone by calculating the combined stability coefficient of the multi-layer snow body profile.

[0108] The snow layer instability zone refers to the area where the stability coefficient is determined to be less than a preset threshold and there is a risk of instability initiation after the snow layer stability coefficient is calculated by the multi-layer limit equilibrium method.

[0109] A multi-level chain disaster path refers to a disaster cascading propagation path derived by assessing the instability probability and cumulative impact range of various downstream disaster-causing objects along the terrain connectivity direction based on the comprehensive triggering probability of multiple factors. These disaster-causing objects include landslides, debris flow channels, and glacial lakes.

[0110] Specifically, after acquiring heterogeneous data from multiple sources, including satellite-based remote sensing, UAV aerial surveys, ground-based in-situ monitoring, and distributed fiber optic sensing, through the risk-driven active perception and multi-source fusion module, the snow layer structure dynamic evolution and instability identification module calculates the combined stability coefficient based on multi-layer snow body profiles using the multi-layer limit equilibrium method. Based on the Mohr-Coulomb strength criterion and combined with the combined effect of multiple weak layers and the tension-shear composite instability mode, areas with combined stability coefficients less than a preset threshold are identified as snow layer instability zones. Simultaneously, the multi-inducing coupled chain-triggered risk quantification module obtains the multi-inducing trigger probability, constructs a multi-inducing joint probability distribution using the Copula function to obtain the comprehensive trigger probability, and progressively identifies the instability probability and superimposed impact range of downstream disaster-causing objects along the terrain connectivity direction, deducing multi-level chain disaster paths. Finally, the snow layer instability zone is used as the source region node location, and the multi-level chain disaster path is used as the trigger relationship edge, providing structured input information for heterogeneous graph construction.

[0111] Step S403: Based on the snow layer instability zone and multi-level chain disaster paths, construct a heterogeneous graph containing disaster-causing nodes and their relationships.

[0112] In this context, disaster-causing nodes refer to graph nodes in the heterogeneous graph used to characterize the various participating elements of the disaster system, including: grid cell nodes, source region nodes, and disaster-bearing body nodes. Grid cell nodes are basic spatial units determined by dividing the target area into grids; source region nodes are determined based on the snow layer instability zone and are obtained by marking grid cell nodes falling within the spatial range of the instability zone, used to indicate the avalanche initiation location; disaster-bearing body nodes are optional nodes, determined based on the distribution location of at least one disaster-bearing body among transmission towers, line corridors, transportation arteries, engineering camps, and residential areas, and have initial attributes such as disaster-bearing body type, vulnerability level, and spatial range.

[0113] Association relationships refer to the edges in a heterogeneous graph used to represent the topological connections between disaster-causing nodes. They define the logical paths of disaster propagation or spatial dependence and include triggering relationship edges and spatial association edges. Triggering relationship edges are determined based on multi-level chain disaster paths and are established between higher-level and lower-level disaster nodes along the disaster propagation direction, with their direction representing the cascading propagation direction of the disaster. Spatial association edges are established based on the spatial overlap relationship between disaster-bearing nodes and the grid cell nodes in their location, and are used to connect disaster-bearing nodes with the grid cell nodes in their location.

[0114] A heterogeneous graph (also known as a heterogeneous network or heterogeneous information network) is a graph data structure that can represent multiple entity types and their complex relationships.

[0115] Specifically, the snow-topography coupled data governance and knowledge graph module or the mechanism and data hybrid driven dynamics simulation module, based on the snow layer instability zone output by the snow layer structure dynamic evolution and instability identification module and the multi-level chain disaster path output by the multi-causal coupled chain trigger risk quantification module, establishes grid unit nodes based on the grid units divided in the target area. The spatial range of the snow layer instability zone is mapped to the grid unit nodes to determine the location of the source area node. At the same time, along the propagation direction of the multi-level chain disaster path, trigger relationship edges are established between the upper-level disaster node and the lower-level disaster node. Corresponding topographic, snow cover, meteorological, geological and deformation attributes are assigned to each node, thereby constructing a heterogeneous graph containing the association between disaster-causing nodes and trigger relationship edges.

[0116] Step S404: Use a graph neural network to perform spatial extrapolation on the heterogeneous graph and embed a preset physical constraint loss to obtain the start probability, movement distance and chain trigger probability of each node in the heterogeneous graph.

[0117] Graph neural networks refer to deep learning models that use graph Fourier transform to map the heterogeneous graph from the spatial domain to the spectral domain, perform learnable convolutional kernel operations in the spectral domain to capture global spatial dependencies, then inversely transform back to the spatial domain, iteratively update the features of each node, and embed a preset physical constraint loss for inference, including: Graph Fourier Transform Networks: Map heterogeneous graphs from the spatial domain to the spectral domain, and perform product operations with learnable convolutional kernels in the spectral domain to capture global spatial dependencies.

[0118] Inverse Graph Fourier Transform Network: Maps spectral domain features back to the spatial domain to obtain the updated node features of the current layer.

[0119] Learnable convolution kernel: A parameter matrix that performs element-wise multiplication with the graph Fourier transform features in the spectral domain to extract global spatial features.

[0120] Activation function: performs a nonlinear transformation on the updated node features.

[0121] The total loss function is composed of a weighted sum of the data fitting loss and the preset physical constraint loss. The preset physical constraint loss includes the initiation constraint loss, motion constraint loss, and chain-triggered constraint loss, which are used to constrain the physical consistency of the deduction process.

[0122] Preset physical constraint loss is a physical consistency constraint mechanism embedded in the graph neural network inference process. It is used to ensure that the model output results such as the start probability, motion range, and chain trigger probability conform to the real physical laws of avalanche evolution.

[0123] The initiation probability refers to the likelihood of avalanche initiation at each node (especially source region nodes and grid cell nodes) in the heterogeneous graph, with a value ranging from 0 to 1. This probability is determined by the snow layer stability coefficient (denoted by FS). When FS < 1, the initiation probability is higher, and when FS ≥ 1, the initiation probability is lower. The initiation constraint loss (based on the Mohr-Coulomb strength criterion) ensures that the prediction of the initiation probability is consistent with the snow layer mechanical stability criterion.

[0124] The avalanche travel distance refers to the horizontal distance or longitudinal extension range along the terrain after the avalanche starts from the source area, measured in kilometers or meters, reflecting the spatial range that the avalanche may affect. This value is constrained by the Voellmy rheological model (the Voellmy rheological model constraint is the physical basis of the motion constraint loss in this invention, used to constrain the avalanche travel distance predicted by the graph neural network to conform to physical laws). The motion constraint loss ensures that the relative error between the model-predicted travel distance and the theoretical travel distance calculated by the physical equations is within a reasonable range.

[0125] The chain-triggered probability refers to the likelihood that an avalanche, after reaching its downstream location, will trigger secondary disasters such as landslides, debris flows, and glacial lake outbursts. Its value ranges from 0 to 1. This probability is determined by the slope's ultimate equilibrium criterion (…). (Time instability) constraint, and ensure that the prediction of chain triggering probability is consistent with the slope safety factor through chain triggering constraint loss.

[0126] Specifically, the mechanism and data hybrid driving dynamics simulation module calls the physical constraint heterogeneous graph neural operator engine. It takes the heterogeneous graph constructed by the snow-topography coupled data governance and knowledge graph module as input, uses the attribute features of each node in the graph as the initial embedding, and the connection relationships of each edge as the propagation path. A graph Fourier transform is used to map the heterogeneous graph from the spatial domain to the spectral domain. Learnable convolutional kernel operations are performed in the spectral domain to capture global spatial dependencies, and then the graph is inversely transformed back to the spatial domain, iteratively updating the features of each node. During the simulation, the data fitting loss, the initiation constraint loss based on the Mohr-Coulomb strength criterion, the motion constraint loss based on the Voellmy rheological model, and the chain triggering constraint loss based on the limit equilibrium method are weighted and summed as the total loss function to constrain the physical consistency of the simulation process. Finally, based on the updated node features, the initiation probability, motion range, and chain triggering probability of each node are output, providing three key quantitative parameters for subsequent comprehensive risk index calculation and susceptibility level classification.

[0127] Step S405: Generate avalanche susceptibility assessment results based on the initiation probability, running distance, and chain trigger probability.

[0128] The avalanche susceptibility assessment result refers to the susceptibility level (including low-susceptibility, medium-susceptibility, high-susceptibility, and extremely high-susceptibility zones) of each node in the target area, generated based on three quantitative parameters: initiation probability, movement distance, and chain triggering probability, through comprehensive risk index calculation and preset classification thresholds. This level characterizes the avalanche risk level of each node under the current disaster-prone environment and triggering conditions. This result is the final output form of the system's assessment of the avalanche susceptibility of the target area.

[0129] Specifically, the multi-scale adaptive dynamic assessment and uncertainty quantification module integrates three quantitative parameters—node initiation probability, movement distance, and chain trigger probability—output by the mechanism and data hybrid driving dynamics simulation module into a comprehensive risk index for each node. Based on preset grading thresholds, this comprehensive risk index is divided into low-risk, medium-risk, high-risk, and extremely high-risk zones, generating an avalanche susceptibility assessment result for the target area. This result, using each node as the basic unit, forms a spatially distributed susceptibility zone, providing a basis for tiered early warning (four-color warning).

[0130] The geological hazard susceptibility assessment method provided in this embodiment identifies snow layer instability zones using the multi-level limit equilibrium method and uses Copula functions to deduce multi-level chain disaster paths, achieving accurate identification of snow layer instability and complete assessment of disaster chains. By constructing a heterogeneous graph containing disaster-causing nodes and their relationships, and using graph neural networks for spatial deduction, and embedding preset physical constraint losses, the method achieves the organic integration of mechanistic models and data-driven models, ensuring the physical consistency of assessment results. This improves the accuracy and reliability of geological hazard susceptibility assessment and provides technical support for the risk quantification of multi-level chain disasters.

[0131] This embodiment provides a method for assessing geological hazard susceptibility, which can be used in the aforementioned electronic device, wherein the electronic device is equipped with a geological hazard susceptibility assessment system. Figure 5 This is a flowchart of a geological hazard susceptibility assessment method according to an embodiment of the present invention, such as... Figure 5 As shown, the process includes the following steps: Step S501: Obtain multi-source heterogeneous data for the target area. For details, please refer to [link to relevant documentation]. Figure 4 Step S401 of the illustrated embodiment will not be described again here.

[0132] Step S502: Based on multi-source heterogeneous data, the multi-level limit equilibrium method is used to identify the snow layer instability zone, and the Copula function is used to deduce the multi-level chain disaster path.

[0133] Specifically, step S502 above includes: Step S5021: Obtain the multi-layer snow profile of the target area, and calculate the snow layer stability coefficient using the multi-layer limit equilibrium method based on the multi-layer snow profile. The snow layer stability coefficient is determined by the snow layer cohesion, normal stress, pore water pressure and internal friction angle, in combination with the multi-weak layer combination effect and the tension-shear composite instability mode. When the snow layer stability coefficient is less than the preset value, the multi-layer snow profile is identified as a snow layer instability zone.

[0134] Specifically, the snow layer stability coefficient is calculated according to the above formula (2), that is, the formula for calculating the snow layer stability coefficient FS: (2); in, This is the snow layer stability coefficient. For snow layer cohesion, For snow density, The slope angle, It is the acceleration due to gravity. Snow depth Pore ​​water pressure, It is the internal friction angle.

[0135] The preset value is set to 1. When the snow layer stability coefficient FS < 1, it is determined that the snow layer instability has started and potential snow layer instability zones are identified.

[0136] Step S5022: Obtain the trigger probability of multiple factors, and use the Copula function to construct the joint probability distribution of multiple factors to obtain the comprehensive trigger probability of multiple factors; based on the comprehensive trigger probability of multiple factors, determine the instability probability and superimposed impact range of each type of disaster-causing object in the downstream along the terrain connection direction, and deduce the multi-level chain disaster path.

[0137] Specifically, the meteorological triggering quantification unit, the seismic ground motion triggering risk unit, and the ice avalanche / rockfall impact triggering unit of the multi-inducing coupled chain-triggered risk quantification module calculate the triggering probability of each individual trigger. The multi-inducing nonlinear coupling unit uses the Copula function to construct the joint probability distribution of each trigger to characterize the coupling effect between multiple triggers and obtain the comprehensive triggering probability of multiple triggers. Based on this comprehensive triggering probability, the multi-level chain-triggered deduction unit determines the instability probability and superimposed impact range of various types of disaster-causing objects such as downstream landslides, debris flow channels, and glacial lakes along the terrain connectivity direction, and deduces the spatiotemporal propagation direction of the multi-level chain disaster path of avalanche-landslide-debris flow-glacial lake outburst, and uses this path as the basis for determining the triggering relationship edges in the heterogeneous graph construction.

[0138] Step S503: Based on the snow layer instability zone and multi-level chain disaster paths, construct a heterogeneous graph containing disaster-causing nodes and their relationships.

[0139] Specifically, step S503 includes: Step S5031: Determine the grid cell nodes based on the grid division of the target area.

[0140] Specifically, the snow-topography coupled data governance and knowledge graph module or the mechanism and data hybrid driving dynamics simulation module divides the target area into regular grids according to the spatial range of the target area with a preset spatial resolution (e.g., 300m×300m). Each grid corresponds to a grid unit node, which serves as the basic spatial unit of the heterogeneous graph.

[0141] For example, a total of 22,340 grid nodes are divided at a resolution of 300m. Terrain connectivity is extracted based on DEM to construct edge connections, and node features and edge weights are initialized.

[0142] Step S5032: Determine the source node position in the heterogeneous graph based on the snow layer instability zone, and determine the trigger relationship edge in the heterogeneous graph based on the multi-level chain disaster path.

[0143] In an optional implementation, step S5032 includes: Step a1: Map the spatial extent of the snow layer instability zone to grid cell nodes, and mark the grid cell nodes that fall within the spatial extent as avalanche source zone nodes.

[0144] Specifically, the snow-topography coupled data governance and knowledge graph module or the mechanism and data hybrid driving dynamics simulation module maps the spatial range of the snow layer instability zone output by the snow layer structure dynamic evolution and instability identification module to the grid cell nodes divided by the target region. The grid cell nodes that fall within the range of the instability zone are marked as avalanche source nodes, thereby clarifying the starting position of avalanche initiation in the heterogeneous graph and providing clear source region spatial location for subsequent graph neural network inference.

[0145] Step a2: Along the propagation direction of the multi-level chain disaster path, establish trigger relationship edges between the upper-level disaster node and the lower-level disaster node. The direction of the trigger relationship edges represents the cascading propagation direction of the disaster.

[0146] Specifically, the snow-topography coupled data governance and knowledge graph module or the mechanism and data hybrid driving dynamics simulation module, based on the multi-level chain disaster path output by the multi-causal coupled chain-triggered risk quantification module, establishes triggering relationship edges between the upper-level disaster node (such as the avalanche source area node) and the lower-level disaster node (such as the landslide node, debris flow channel node, glacial lake node, etc.) along the propagation direction of the path. The direction of the edge represents the cascading propagation direction of the disaster, thereby quantifying the chain inference results into a directed edge structure in the heterogeneous graph.

[0147] Step S5033: Assign corresponding attribute features to source region nodes and grid cell nodes to construct a heterogeneous graph containing disaster-causing nodes and their relationships; the attribute features include terrain features, snow cover features, meteorological features, geological features and deformation features.

[0148] Specifically, the snow-topography coupled data governance and knowledge graph module uses the identified source region nodes and grid unit nodes as the basic node set of the heterogeneous graph. It extracts the attribute features corresponding to each node from the snow-topography coupled database constructed by this module and assigns values. Finally, the physical constraint heterogeneous graph neural operator (PC-HGNO) engine of the mechanism and data hybrid driving dynamics simulation module completes the construction of the heterogeneous graph.

[0149] Step S5034: Determine the disaster-bearing body nodes based on the distribution location of at least one of the disaster-bearing bodies among transmission towers, line corridors, traffic arteries, engineering camps, and residential areas within the target area; add the disaster-bearing body nodes to the heterogeneous graph.

[0150] In an optional implementation, step S5034 includes: Step b1: Based on the geographical location of the disaster-bearing body, create a disaster-bearing body node at the corresponding location in the heterogeneous graph, and use the disaster-bearing body type, vulnerability level, and spatial range as the initial attributes of the disaster-bearing body node.

[0151] The simulation visualization and hierarchical early warning decision-making module or the snow-topography coupled data governance and knowledge graph module creates disaster-bearing body nodes in the corresponding spatial locations of at least one disaster-bearing body among transmission towers, line corridors, transportation arteries, engineering camps, and residential areas within the target area. The disaster-bearing body type (such as transmission towers, line corridors, transportation arteries, etc.), vulnerability level (reflecting the vulnerability of the disaster-bearing body to avalanche impact), and spatial range (the geographical space occupied by the disaster-bearing body) are assigned as the initial attributes of the node. This allows the disaster-bearing body node to be loaded with the basic attribute information required for risk assessment during the heterogeneous graph construction stage, providing a node-level data foundation for the subsequent transmission of assessment results to the disaster-bearing body risk.

[0152] Step b2: Based on the spatial overlap relationship between the disaster-bearing body node and the grid unit node at its location, establish the spatial association edge between the disaster-bearing body node and the grid unit node to construct a heterogeneous graph.

[0153] The snow-topography coupled data governance and knowledge graph module or the mechanism and data hybrid driving dynamics simulation module establishes spatial association edges between the disaster-bearing node and the grid unit node based on the spatial overlap relationship between the geographical space occupied by the disaster-bearing node and the grid unit node where it is located. This enables the disaster-bearing node and the surrounding terrain environment to be effectively connected in the heterogeneous graph, providing a topological basis for the effective transmission of evaluation results (start-up probability, movement distance, chain trigger probability) from the grid unit node to the disaster-bearing node in the subsequent graph neural network inference, and completing the construction of the heterogeneous graph.

[0154] The architecture and working principle diagram of the PC-HGNO (Physics-Constrained Heterogeneous Graph Neural Operator) hybrid modeling algorithm are shown below. Figure 6 As shown, the Physics-Constrained Heterogeneous Graph Neural Operator (PC-HGNO) provided in this embodiment serves as an engine for hybrid modeling of mechanisms and data. It combines the ability to express the multi-factor associations of heterogeneous graphs, the efficient deduction capabilities of neural operators, and the mechanism constraint capabilities of physical equations.

[0155] For example, the construction of heterogeneous graph carriers includes: The study area is modeled as a heterogeneous information network, i.e., a heterogeneous graph. : Node set V: contains 7 types of nodes, including grid cell nodes, avalanche source area nodes, landslide body nodes, debris flow channel nodes, glacial lake nodes, glacier nodes, and disaster-bearing body nodes. Each type of node carries attribute features of the corresponding dimension.

[0156] Edge set E: contains five types of edges: terrain connectivity edges, wind direction related edges, trigger relationship edges, spatial adjacency edges, and temporal evolution edges. Different types of edges adopt independent weight calculation strategies.

[0157] Node attribute X: includes five major categories of features: terrain, snow cover, meteorology, geology, and deformation, which are updated sequentially with the monitoring data.

[0158] The formula for calculating edge weights (terrain-connected edges) is as follows: (3); in, The weight of the terrain-connected edges. Let i be the elevation difference from node i to j. The angle between the slope aspect and the line connecting the node. The average path resistance coefficient, This is the elevation difference attenuation coefficient.

[0159] Step S504: Use a graph neural network to perform spatial extrapolation on the heterogeneous graph and embed a preset physical constraint loss to obtain the start probability, movement distance and chain trigger probability of each node in the heterogeneous graph.

[0160] Specifically, step S504 includes: Step S5041: Using the attribute features of each node in the heterogeneous graph as the initial embedding and the connection relationship of each edge in the heterogeneous graph as the propagation path, a graph neural network is used to map the heterogeneous graph from the spatial domain to the spectral domain. Learnable convolutional kernel operations are performed in the spectral domain to capture global spatial dependencies. Then, the graph is inversely transformed back to the spatial domain, and the features of each node are iteratively updated.

[0161] In one optional implementation, the graph neural network includes a graph Fourier transform network and an inverse graph Fourier transform network, and step S5041 above includes: Step c: In each layer of propagation, the features of each node in the previous layer are aggregated by neighbor analysis and then mapped from the spatial domain to the spectral domain through a graph Fourier transform network. In the spectral domain, these features are multiplied with learnable convolutional kernels, and then returned to the spatial domain through an inverse graph Fourier transform network to obtain the updated node features of the current layer. This layer-by-layer propagation captures global spatial dependencies. The formula for each layer of propagation is as follows: (4); in, Let i be the feature vector of node i in the (l+1)th layer. For the graph Fourier transform, This is the inverse graph Fourier transform. Let l be a learnable convolution kernel in the spectral domain. For element-wise product, For activation function, The weight of the terrain-connected edges. To perform weighted feature aggregation on all neighboring nodes j of node i.

[0162] This step introduces Fourier spectral domain graph convolution to replace traditional graph convolution, enabling long-range spatial dependency modeling in the spectral domain. The idea is to map spatial domain features to the spectral domain through graph Fourier transform, perform learnable convolution kernel operations in the spectral domain to capture global spatial relationships, and then map them back to the spatial domain through inverse transform.

[0163] Step S5042: Calculate the data fitting loss, and use the weighted sum of the data fitting loss and the preset physical constraint loss as the total loss function. This is the derivation process of the constraint graph neural network. The preset physical constraint loss includes: activation constraint loss, motion constraint loss and chain triggering constraint loss.

[0164] In one alternative implementation, the total loss function is calculated as follows: (5); in, For the total loss function, For data fitting loss, For the initiation constraint loss based on the Moore-Coulomb criterion, The motion constraint loss is based on the Voellmy rheological model. For chain-triggered constraint loss based on limit equilibrium method, These are all weighting coefficients, with empirical values ​​of 0.3, 0.3, and 0.2.

[0165] Start-up constraint loss Based on the Mohr-Coulomb strength criterion, the snow layer stability coefficient FS of nodes with high start-up probability is constrained to be less than 1. (6); in, The total number of nodes participating in the calculation can be either the number of nodes in the source region or the number of nodes in all grid cells in this implementation, where i is the index node. Let be the activation probability of node i predicted by the graph neural network, with a value ranging from 0 to 1. Let be the snow layer stability coefficient of the i-th node.

[0166] The snow layer stability coefficient FS is calculated using formula (2).

[0167] Motion constraint loss Based on the Voellmy rheological model, constrain the relative error between the predicted running distance and the theoretical running distance; (7); in, Let the avalanche motion distance of node i be predicted by the graph neural network. The theoretical running distance of node i is calculated based on the Voellmy rheological model. It is a very small positive number.

[0168] Theoretical running distance The governing equations are derived from the depth integral equations of motion: (8); in, Let t be the avalanche velocity and t be the time. The turbulence coefficient is... Let be the Coulomb friction coefficient. This is the turbulence resistance term.

[0169] Integrating the above equation yields the theoretical maximum running distance: (9); The above formula can be simplified to: (10); Among them, terminal speed , The coefficient of friction is Coulomb. The turbulence coefficient is... This is the acceleration due to gravity.

[0170] Chain constraint loss Based on the limit equilibrium method, the consistency of the criteria for slope instability under avalanche impact is constrained; (11); Among them, slope safety factor The calculation formula is as follows: (12); In the formula, For slope cohesion, For normal stress, This refers to the pore water pressure in the slope. Let W be the internal friction angle of the slope, and W be the weight of the slope. The force of the avalanche is A, and the windward area is A. It is the slope angle.

[0171] Avalanche impact force Calculation formula: (13); in, This is the drag coefficient (values ​​range from 1.0 to 1.3).

[0172] Step S5043: Based on the updated characteristics of each node, output the start probability, running distance, and chain trigger probability of each node.

[0173] Step S5044: Using terrain features, snow cover features, and geological features as covariates, construct the spatial distribution function of rheological parameters in the graph neural network; use online stochastic gradient descent or Kalman filtering to dynamically update the rheological parameters over time; and perform adaptive calibration of physical parameters based on the spatial distribution function and the dynamic time update.

[0174] Specifically, this step involves modeling the spatial heterogeneity of physical parameters and adaptive calibration.

[0175] Spatial heterogeneity modeling: Topographic features (slope, curvature, roughness), snow cover features (snow depth, density, temperature), and basement features (vegetation, lithology) are used as covariates to construct a spatial distribution function for rheological parameters. (14); (15); By learning the spatial distribution pattern of parameters through graph neural networks, spatial adaptation of parameters can be achieved.

[0176] Time-series dynamic adaptive calibration: Utilizing online stochastic gradient descent (SGD) or Kalman filtering, parameters are updated while maintaining real-time performance. , (16); (17); (18); in, Let be the Coulomb friction coefficient at time t. The initial or baseline value of the Coulomb friction coefficient is determined by formula (14). Let be the time-varying increment of the Coulomb friction coefficient at time t. Let be the turbulence coefficient at time t. The initial or baseline value of the turbulence coefficient is determined by formula (15). Let be the time-varying increment of the turbulence coefficient at time t. The time-varying increment of the Coulomb friction coefficient at time t+1 is given. The time-varying increment of the turbulence coefficient at time t+1 , For adaptive learning rate, , This represents the gradient of the motion constraint loss with respect to the parameters.

[0177] Step S505: Generate avalanche susceptibility assessment results based on the initiation probability, running distance, and chain trigger probability.

[0178] Specifically, step S505 includes: Step S5051: Determine the comprehensive risk index of each node based on the activation probability, running distance, and chain trigger probability of each node.

[0179] The differentiated index system construction unit of the multi-scale adaptive dynamic assessment and uncertainty quantification module dynamically adjusts the weights of each index according to the type of disaster-bearing body. The multi-model fusion assessment unit takes the three parameters of node initiation probability, movement range, and chain triggering probability output by the mechanism and data hybrid driving dynamics simulation module as input, and performs multi-source evidence fusion on the above three indicators through DS evidence theory to output a comprehensive risk index for each node. This index comprehensively reflects the overall risk level of each node in terms of initiation probability, movement impact range, and chain triggering.

[0180] Step S5052: Based on the comprehensive risk index, determine the susceptibility level of each node according to the preset classification threshold, and generate the avalanche susceptibility assessment results for the target area; the susceptibility levels include low susceptibility area, medium susceptibility area, high susceptibility area and extremely high susceptibility area.

[0181] The multi-scale adaptive dynamic assessment and uncertainty quantification module divides the comprehensive risk index of each node into four level intervals according to preset grading thresholds (such as the natural breakpoint method or quantile method), determining the susceptibility level of each node as low-susceptibility, medium-susceptibility, high-susceptibility, or extremely high-susceptibility zone. Finally, it generates a spatial distribution susceptibility zoning map of the target area with each node as the basic unit. At the same time, the grading warning threshold control unit of the simulation visualization and grading warning decision module outputs the corresponding four-color warning (blue / yellow / orange / red) based on the susceptibility level, providing a grading basis for subsequent high-risk section identification, automatic switching of multi-scale assessment, and engineering protection recommendations.

[0182] Step S506: Identify high-risk sections based on avalanche susceptibility assessment results, and automatically switch the high-risk sections from kilometer-level daily scale assessment to hundred-meter-level hourly scale assessment; quantify the uncertainty of the assessment results using a Bayesian neural network to obtain the confidence level of each avalanche susceptibility assessment result.

[0183] Specifically, the multi-scale adaptive dynamic assessment and uncertainty quantification module identifies continuous spatial ranges with high or extremely high susceptibility levels as high-risk sections based on the susceptibility levels of each node in the target area. It then automatically switches these high-risk sections from routine daily assessment at the kilometer level to detailed hourly assessment at the hundred-meter level (with 24-hour prediction results updated every 3 hours), achieving synergy between comprehensive surveys and key monitoring. Building upon this, the end-to-end uncertainty quantification unit, based on a Bayesian neural network, decomposes the uncertainty in the assessment process along three dimensions: data input, model parameters, and model structure. It outputs the confidence level of each assessment result, providing a reliability quantification index for graded early warning and decision-making confidence assessment.

[0184] Step S507: Based on the avalanche susceptibility assessment results, obtain climate scenario data and use a time-domain generalized graph network to predict the spatiotemporal migration patterns of avalanche-prone areas in different future periods under different climate scenarios; generate a long-term adaptation map based on the prediction results. The long-term adaptation map includes location information of high-risk corridor areas, location information of suggested avoidance areas, and location information of areas that can be strengthened for protection.

[0185] Specifically, the Climate Change Adaptation and Self-Evolution Unit uses the current avalanche susceptibility assessment results as a benchmark, and simultaneously incorporates multiple climate scenario data from CMIP6, such as SSP1-2.6, SSP2-4.5, and SSP5-8.5. It employs a time-domain generalized graph network to predict the spatiotemporal migration patterns of avalanche-prone areas under different climate scenarios in the future. Based on the prediction results, it generates a long-term adaptation map containing location information of high-risk corridor areas, suggested avoidance areas, and areas where protection can be strengthened. This provides spatial decision-making references for the long-term planning, line avoidance, and protection upgrades of linear projects such as power transmission lines under different climate scenarios.

[0186] The geological hazard susceptibility assessment method provided in this embodiment enables high-precision dynamic assessment and full-chain closed-loop early warning of avalanche risk in complex high-altitude environments. It improves the efficiency of sensing resource utilization and monitoring coverage, with high assessment accuracy and physical consistency. It completes integrated simulation and second-level extrapolation of the entire avalanche initiation-movement-chain cascade process, and supports cloud-edge collaborative offline operation and long-term disaster prevention planning on a multi-decade scale.

[0187] As one or more specific application embodiments of the present invention, combined with Figure 7 The present invention provides a more detailed description of the geological hazard susceptibility assessment method, such as... Figure 7 As shown, the specific process is as follows: Step 1, Active perception step of monitoring resources based on meta-reinforcement learning. LSTM is used as meta-learner to extract scheduling policy features from historical tasks as the initial policy network parameters of PPO. The task adaptive reward mechanism and dynamic action space are integrated to realize adaptive dynamic scheduling of monitoring resources in mixed normal and sudden task scenarios. Step 2, continuous monitoring of the entire line based on distributed fiber optic acoustic sensing: deploy a joint monitoring system of distributed acoustic sensing and Brillouin optical time domain analysis using the existing OPGW optical cable of the transmission line to acquire vibration, temperature and strain signals along the line in real time with decimeter-level spatial resolution. Step 3, Snow layer tension-shear composite instability identification step based on multi-layer limit equilibrium method, considering the multi-weak layer combination effect and tension-shear composite instability mode, and calculating the snow layer stability coefficient; Step 4: Based on the multi-causal nonlinear coupling and multi-level chain triggering deduction steps of the Copula function, characterize the joint probability distribution of multiple causes, quantify the synergistic and antagonistic effects between causes, and deduce the multi-level chain disaster path and superimposed impact of "avalanche-landslide-debris flow-glacial lake outburst". Step 5: Based on the mechanism and data hybrid deduction steps of the physical constraint heterogeneous graph neural operator, the heterogeneous graph uniformly carries multiple types of disaster-causing nodes and their relationships. Long-distance spatial deduction is achieved through Fourier spectral domain graph convolution. Three physical constraint loss functions of snow layer initiation, motion rheology, and chain triggering are embedded to ensure the physical consistency of the evaluation results. Step 6, multi-scale adaptive dynamic assessment step, supports automatic switching between the general scale of the whole domain and the detailed scale of high-risk sections, and quantifies the uncertainty of the whole link assessment; Step 7, which generates a long-term adaptation map based on the CMIP6 climate scenario, uses a time-domain generalized graph network to predict the migration patterns of avalanche-prone areas under different scenarios.

[0188] Application Example: Combined with an avalanche risk assessment of a 500kV transmission corridor in a certain region: 1. Study Area Overview and System Deployment: The study area is a 2000 km² area along a 500 kV transmission corridor in the lower reaches of a certain river in a certain region. Three ground-based in-situ snow cover monitoring stations (including automatic snow profile monitoring equipment, MIMO radar, and a weather station) are deployed in the study area, accessing periodic observation data from two SAR satellites and one optical satellite. One inspection drone is also deployed. A DAS and BOTDA joint monitoring system is deployed using the OPGW optical cable along the line. A comprehensive assessment service and a PC-HGNO engine are deployed in the cloud, and edge computing nodes are deployed at the three maintenance stations along the line.

[0189] 2. Data Governance and Baseline Analysis: A 10m resolution DEM of the study area was accessed, and topographic factors and underlying surface features were extracted to generate a static baseline susceptibility map. Sentinel-1 / 2 imagery from 2018 to 2024, UAV LiDAR topographic data, time-series data from ground meteorological and snow monitoring stations, and 1:50,000 geological maps were collected to construct a standardized feature dataset. A grid of 22,340 nodes was created at a 300m resolution. Topographic connectivity was extracted from the DEM to construct edge connections, and node features and edge weights were initialized. Historical avalanche events in the region over the past 30 years were reviewed, a case knowledge base was constructed, and initial model calibration and parameter calibration were completed.

[0190] 3. Training and Validation of the PC-HGNO Model: Using 326 historical high-altitude long-distance avalanche events as labels, the training and test sets were divided in a 7:3 ratio. Physical constraint weights λ1=0.3, λ2=0.3, and λ3=0.2 were set, and a progressive training strategy was adopted to train the PC-HGNO model. The AUC on the test set reached 0.935. The physical constraint term reduced the PDE residual by 62% compared to the pure data-driven model. The avalanche path prediction accuracy was improved by more than 18% compared to the pure data-driven model, and there were no physical inconsistencies where FS>1 was judged as high-risk, thus verifying the effectiveness of the physical constraints.

[0191] 4. Dynamic Rolling Assessment and Detailed Switching: The system receives hourly real-time and forecast data daily, driving snow layer evolution simulation and stability calculation. When the accumulated snowfall exceeds 30mm for three consecutive days, the perception and coordination scheduling unit automatically triggers UAVs to conduct aerial surveys of three high-risk snow accumulation areas, simultaneously increasing the sampling frequency of monitoring stations to 10 minutes per instance. After data fusion, the system automatically switches these three sections to detailed assessments at the 100-meter and 1-hour scales, updating the 24-hour forecast results every 3 hours.

[0192] 5. Multi-scenario Hybrid Simulation: During a period of heavy snowfall, the system calculated that the overall susceptibility of a certain gully section was extremely high, with a high risk of chain-triggered landslides. The PC-HGNO engine was used to conduct multi-scenario simulations: In the most probable scenario, the avalanche's maximum travel distance was 8.2 km and the maximum impact force was 145 kPa. Although it did not directly impact the tower locations, the powder snow blast could trigger wind-induced discharge of the conductors, requiring close monitoring. In the worst-case scenario, the avalanche triggered downstream slope instability and transformed into a debris flow, with a maximum travel distance of 11.6 km, burying two tower locations.

[0193] 6. Early warning and decision-making: The system automatically issues orange alerts to the operation and maintenance department, pushes a list of affected tower locations, conductor wind deflection risk warnings and emergency inspection routes; managers can interact with the simulation platform to simulate the protective effect of adding snow dams and optimize the handling plan.

[0194] 7. Output: Generate a high-altitude long-distance avalanche susceptibility zoning map of the study area, identify 12 high-risk corridor sections, of which 3 sections have an extremely high overall susceptibility level, and propose route detour suggestions and proactive prevention and control measures for sections adjacent to the route.

[0195] 8. Closed-loop iteration: After the event, the system will store the actual avalanche range, motion parameters and disaster impact verified on site into the database, automatically trigger PC-HGNO model parameter calibration and incremental learning, and complete self-evolutionary iteration; the early warning effectiveness post-evaluation module will automatically calculate the hit rate and lead time of this early warning and optimize the early warning threshold; the climate change adaptation module will be called regularly every year to predict the avalanche risk migration pattern along the corridor in the next 30 years based on SSP2-4.5 and SSP5-8.5 scenarios, and output suggestions for upgrading high-risk sections and avoiding routes.

[0196] In summary, this invention constructs an integrated three-dimensional perception network encompassing "space-ground-depth" by employing meta-reinforcement learning-driven active perception scheduling and distributed fiber optic acoustic sensing for full-line monitoring. Meta-reinforcement learning integrates a task-adaptive reward mechanism with a dynamic action space, effectively addressing mixed routine and unexpected task scenarios. The DAS and BOTDA joint monitoring system achieves continuous full-coverage monitoring along the entire route with a spatial resolution of 25cm, unaffected by visual obstruction or inclement weather, thus improving the utilization rate and monitoring coverage of high-altitude observation resources.

[0197] This invention proposes a physically constrained heterogeneous graph neural operator, PC-HGNO, which uses heterogeneous graphs to uniformly represent 7 types of disaster-causing nodes and 5 types of association relationships. It achieves long-distance spatial dependency modeling through Fourier spectral domain graph convolution and embeds a three-layer physical constraint loss function: "initiation-motion-chain". This algorithm combines the advantages of heterogeneous graph association representation, efficient neural operator deduction, and physical mechanism constraints. The PDE (Partial Differential Equation) residual is reduced by 62% compared to a purely data-driven model, and the evaluation accuracy is improved by more than 18%, fundamentally avoiding abnormal outputs that violate natural laws.

[0198] This invention comprehensively covers the entire disaster chain from snow layer instability to avalanche initiation, remote motion, and cascading events, achieving integrated assessment of avalanches, landslides, debris flows, and glacial lake outburst floods. Through multi-layered snow mass tension-shear composite instability identification, multi-causal nonlinear coupling of Copula functions, dynamic evolution of mixed-medium phase states, and terrain constraint loss correction, the accuracy of remote motion simulation is significantly improved. The Fourier spectral domain neural operator architecture compresses the dynamics extrapolation time from hours to seconds, meeting near-real-time early warning requirements.

[0199] This invention adopts a cloud-edge collaborative deployment architecture, with the cloud supporting large-scale computing and model training across the entire domain, while the edge supports independent offline operation in environments without internet access. It couples the CMIP6 scenario prediction with risk migration patterns under climate change, and combines incremental learning to achieve continuous self-evolution of the model. Multi-scale adaptive assessment and end-to-end uncertainty quantification, combined with a 3D simulation interactive platform, enable accurate targeted early warning and scenario-based extrapolation, comprehensively supporting long-term disaster prevention planning on a decades-long scale.

[0200] This embodiment also provides a geological hazard susceptibility assessment device, which is used to implement the above embodiments and preferred embodiments; details already described will not be repeated. As used below, the term "module" can refer to a combination of software and / or hardware that performs a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0201] This embodiment provides a geological hazard susceptibility assessment device, such as... Figure 8 As shown, it includes: Data acquisition module 801 is used to acquire multi-source heterogeneous data of the target area; The instability identification and risk quantification module 802 is used to identify snow layer instability zones based on multi-source heterogeneous data using the multi-level limit equilibrium method, and to deduce multi-level chain disaster paths using the Copula function; Heterogeneous graph construction module 803 is used to construct a heterogeneous graph containing disaster-causing nodes and their relationships based on snow layer instability zones and multi-level chain disaster paths; The hybrid inference module 804 is used to perform spatial inference on the heterogeneous graph using a graph neural network and embed a preset physical constraint loss to obtain the start probability, movement distance and chain trigger probability of each node in the heterogeneous graph. The assessment module 805 is used to generate avalanche susceptibility assessment results based on the initiation probability, running distance, and chain trigger probability.

[0202] In some optional implementations, the instability identification and risk quantification module 802 includes: The snow layer instability zone identification unit is used to acquire multi-layer snow body profiles of the target area. Based on these profiles, the multi-layer limit equilibrium method is used to calculate the snow layer stability coefficient. This coefficient is based on the Mohr-Coulomb strength criterion and incorporates the combined effects of multiple weak layers and the tensile-shear composite instability mode. It is determined by snow layer cohesion, normal stress, pore water pressure, and internal friction angle. When the snow layer stability coefficient is less than a preset value, the multi-layer snow body profile is identified as a snow layer instability zone. The multi-level chain disaster path simulation unit is used to acquire the trigger probabilities of multiple factors and uses a Copula function to construct the joint probability distribution of these factors, obtaining the comprehensive trigger probability. Based on this comprehensive trigger probability, the instability probability and cumulative impact range of downstream disaster-causing objects are determined level by level along the terrain connectivity direction, thus simulating the multi-level chain disaster path.

[0203] In some alternative implementations, the heterogeneous graph construction module 803 includes: The grid cell node determination unit is used to determine the grid cell nodes based on the grid division of the target area. The node location and triggering relationship edge determination unit is used to determine the source region node location in the heterogeneous graph based on the snow layer instability zone, and to determine the triggering relationship edge in the heterogeneous graph based on the multi-level chain disaster path; Heterogeneous graph construction unit is used to assign corresponding attribute features to source region nodes and grid unit nodes to construct a heterogeneous graph containing disaster-causing nodes and their relationships; attribute features include terrain features, snow cover features, meteorological features, geological features and deformation features.

[0204] In some optional implementations, the node location and triggering relationship edge determination unit includes: The source node determines the sub-cell, which is used to map the spatial range of the snow layer instability zone to the grid cell node, and marks the grid cell node that falls into the spatial range as the avalanche source zone node.

[0205] Triggering relationship edges determine the sub-units and are used to establish the propagation direction along the multi-level chain disaster path. Triggering relationship edges are established between the upper-level disaster node and the lower-level disaster node, and the direction of the triggering relationship edges represents the cascading propagation direction of the disaster.

[0206] In some alternative implementations, the heterogeneous graph construction module 803 further includes: The disaster-bearing body node determination unit is used to determine the disaster-bearing body nodes based on the distribution location of at least one of the disaster-bearing bodies among transmission towers, line corridors, transportation arteries, engineering camps and residential areas within the target area; and to add the disaster-bearing body nodes to the heterogeneous graph.

[0207] In some optional implementations, the disaster-bearing body node determination unit includes: The disaster-bearing body node creation and attribute assignment sub-unit is used to create disaster-bearing body nodes at corresponding locations in the heterogeneous graph based on the geographical location of the disaster-bearing body, and to use the disaster-bearing body type, vulnerability level, and spatial range as the initial attributes of the disaster-bearing body nodes. Heterogeneous graph construction sub-units are used to establish spatial association edges between disaster-bearing body nodes and grid unit nodes based on the spatial overlap relationship between disaster-bearing body nodes and their respective grid unit nodes, thereby constructing a heterogeneous graph.

[0208] In some alternative implementations, the hybrid inference module 804 includes: The iterative update unit is used to take the attribute features of each node in the heterogeneous graph as the initial embedding, take the connection relationship of each edge in the heterogeneous graph as the propagation path, use graph neural network to map the heterogeneous graph from the spatial domain to the spectral domain, perform learnable convolutional kernel operation in the spectral domain to capture global spatial dependencies, and then inversely transform back to the spatial domain to iteratively update the features of each node. The loss calculation unit is used to calculate the data fitting loss. It calculates the total loss function by weighting and summing the data fitting loss with the preset physical constraint loss. This is the derivation process of the constraint graph neural network. The preset physical constraint loss includes: activation constraint loss, motion constraint loss, and chain-triggered constraint loss. The output unit is used to output the start probability, movement distance, and chain trigger probability of each node based on the updated characteristics of each node.

[0209] In some optional implementations, the graph neural network includes a graph Fourier transform network and an inverse graph Fourier transform network, iteratively updating the features of each node. The iterative update unit is specifically used for: In each layer of propagation, the features of each node in the previous layer are aggregated by neighbor analysis and then mapped from the spatial domain to the spectral domain through a graph Fourier transform network. In the spectral domain, these features are multiplied with learnable convolutional kernels, and then returned to the spatial domain through an inverse graph Fourier transform network to obtain the updated node features for the current layer. This layer-by-layer propagation captures global spatial dependencies. The formula for each layer of propagation is as follows: ; in, Let i be the feature vector of node i in the (l+1)th layer. For the graph Fourier transform, This is the inverse graph Fourier transform. Let l be a learnable convolutional kernel in the spectral domain. For element-wise product, For activation function, The weight of the terrain-connected edges. To perform weighted feature aggregation on all neighboring nodes j of node i.

[0210] In some alternative implementations, the total loss function is calculated as follows: ; in, For the total loss function, For data fitting loss, For the initiation constraint loss based on the Moore-Coulomb criterion, The motion constraint loss is based on the Voellmy rheological model. For chain-triggered constraint loss based on limit equilibrium method, All are weighting coefficients.

[0211] In some optional implementations, the hybrid inference module 804 further includes: The adaptive calibration unit is used to construct the spatial distribution function of rheological parameters in the graph neural network by taking terrain features, snow cover features, and geological features as covariates; it uses online stochastic gradient descent or Kalman filtering to dynamically update the rheological parameters over time; and it performs adaptive calibration of physical parameters based on the spatial distribution function and the dynamic update over time.

[0212] In some alternative implementations, the evaluation module 805 includes: The risk index calculation unit is used to determine the comprehensive risk index of each node based on the activation probability, running distance, and chain trigger probability of each node. The assessment unit is used to determine the susceptibility level of each node based on the comprehensive risk index and a preset grading threshold, and to generate the avalanche susceptibility assessment results for the target area; the susceptibility levels include low susceptibility area, medium susceptibility area, high susceptibility area and extremely high susceptibility area.

[0213] In some alternative embodiments, the device further includes: The scale assessment and quantification module is used to identify high-risk sections based on avalanche susceptibility assessment results, and automatically switch the high-risk sections from kilometer-level daily scale assessment to hundred-meter-level hourly scale assessment; based on the uncertainty of the assessment results quantified by Bayesian neural network, the confidence level of each avalanche susceptibility assessment result is obtained.

[0214] In one alternative embodiment, the device further includes: The evolution prediction module is used to obtain climate scenario data based on the avalanche susceptibility assessment results, and use a time-domain generalized graph network to predict the spatiotemporal migration patterns of avalanche-prone areas in different future periods under different climate scenarios; based on the prediction results, a long-term adaptation map is generated, which includes location information of high-risk corridor areas, location information of suggested avoidance areas, and location information of areas that can be strengthened for protection.

[0215] The geological hazard susceptibility assessment device provided in this embodiment of the invention can execute the geological hazard susceptibility assessment method provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects for executing the method. Further functional descriptions of the various modules and units described above are the same as in the corresponding embodiments described above, and will not be repeated here.

[0216] Figure 9 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention.

[0217] The following is a detailed reference. Figure 9 This diagram illustrates a structural schematic suitable for implementing an electronic device according to embodiments of the present invention. The electronic device may include a processor (e.g., a central processing unit, graphics processor, etc.) 901, which can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 902 or a program loaded from memory 908 into random access memory (RAM) 903. The RAM 903 also stores various programs and data required for the operation of the electronic device. The processor 901, ROM 902, and RAM 903 are interconnected via a bus 904. An input / output (I / O) interface 905 is also connected to the bus 904.

[0218] Typically, the following devices can be connected to I / O interface 905: input devices 906 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 907 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; memory devices 908 including, for example, magnetic tapes, hard disks, etc.; and communication devices 909. Communication device 909 allows electronic devices to exchange data via wireless or wired communication with other devices. Although Figure 9 Electronic devices with various devices are shown, but it should be understood that it is not required to implement or have all of the devices shown, and more or fewer devices may be implemented or have instead.

[0219] In particular, according to embodiments of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device 909, or installed from a memory 908, or installed from a ROM 902. When the computer program is executed by the processor 901, it performs the functions defined in the geological hazard susceptibility assessment method of the embodiments of the present invention.

[0220] Figure 9 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments of the present invention.

[0221] This invention also provides a computer-readable storage medium. The methods described above according to embodiments of the invention can be implemented in hardware or firmware, or implemented as computer code that can be recorded on a storage medium, or implemented as computer code downloaded via a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and then stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code. When the software or computer code is accessed and executed by the computer, processor, or hardware, the geological hazard susceptibility assessment method shown in the above embodiments is implemented.

[0222] A portion of this invention can be applied as a computer program product, such as computer program instructions, which, when executed by a computer, can invoke or provide the methods and / or technical solutions according to the invention through the operation of the computer. Those skilled in the art will understand that the forms in which computer program instructions exist in a computer-readable medium include, but are not limited to, source files, executable files, installation package files, etc. Correspondingly, the ways in which computer program instructions are executed by a computer include, but are not limited to: the computer directly executing the instructions, or the computer compiling the instructions and then executing the corresponding compiled program, or the computer reading and executing the instructions, or the computer reading and installing the instructions and then executing the corresponding installed program. Here, the computer-readable medium can be any available computer-readable storage medium or communication medium accessible to a computer.

[0223] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A method for assessing the susceptibility to geological hazards, characterized in that, The method includes: Acquire multi-source heterogeneous data of the target region; Based on the multi-source heterogeneous data, the multi-level limit equilibrium method was used to identify the snow layer instability zone, and the Copula function was used to deduce the multi-level chain disaster path; Based on the snow layer instability zone and the multi-level chain disaster path, a heterogeneous graph containing disaster-causing nodes and their relationships is constructed. A graph neural network is used to perform spatial extrapolation on the heterogeneous graph, and a preset physical constraint loss is embedded to obtain the start probability, movement distance and chain trigger probability of each node in the heterogeneous graph. Based on the initiation probability, running distance, and chain trigger probability, an avalanche susceptibility assessment result is generated.

2. The method according to claim 1, characterized in that, The step of identifying snow layer instability zones using the multi-source heterogeneous data and the multi-level limiting equilibrium method includes: A multi-layer snow profile of the target area is obtained, and the snow layer stability coefficient is calculated using the multi-layer limit equilibrium method based on the multi-layer snow profile. The snow layer stability coefficient is determined by the snow layer cohesion, normal stress, pore water pressure and internal friction angle, in combination with the multi-weak layer combination effect and the tension-shear composite instability mode. When the snow layer stability coefficient is less than a preset value, the multi-layer snow body profile is identified as a snow layer instability zone.

3. The method according to claim 1, characterized in that, The method of using Copula functions to deduce multi-level chain disaster paths includes: The trigger probabilities of multiple factors are obtained, and the joint probability distribution of multiple factors is constructed using the Copula function to obtain the comprehensive trigger probability of multiple factors. Based on the combined triggering probability of the multiple factors, the instability probability and superimposed impact range of each type of disaster-causing object downstream are determined step by step along the terrain connection direction, and the multi-level chain disaster path is deduced.

4. The method according to claim 1, characterized in that, The construction of a heterogeneous graph containing disaster-causing nodes and their relationships, based on the snow layer instability zone and multi-level chain disaster pathways, includes: The grid unit nodes are determined by dividing the target area into grids; The source node position in the heterogeneous graph is determined based on the snow layer instability zone, and the triggering relationship edge in the heterogeneous graph is determined based on the multi-level chain disaster path; Assign corresponding attribute features to the source region nodes and grid unit nodes to construct a heterogeneous graph containing disaster-causing nodes and their relationships; the attribute features include terrain features, snow cover features, meteorological features, geological features and deformation features.

5. The method according to claim 4, characterized in that, The step of determining the source node position in the heterogeneity map based on the snow layer instability region includes: The spatial range of the snow layer instability zone is mapped to the grid cell nodes, and the grid cell nodes falling within the spatial range are marked as avalanche source zone nodes.

6. The method according to claim 4, characterized in that, The step of determining the triggering relationship edges in the heterogeneous graph based on the multi-level chain disaster path includes: Along the propagation direction of the multi-level chain disaster path, a triggering relationship edge is established between the upper-level disaster node and the lower-level disaster node, and the direction of the triggering relationship edge represents the cascading propagation direction of the disaster.

7. The method according to claim 4, characterized in that, The construction of the heterogeneous graph, which includes disaster-causing nodes and their relationships, also includes: Based on the distribution location of at least one disaster-bearing entity among transmission towers, line corridors, main traffic lines, engineering camps, and residential areas within the target area, the disaster-bearing entity nodes are determined. Add the disaster-bearing node to the heterogeneous graph.

8. The method according to claim 7, characterized in that, Adding the disaster-bearing node to the heterogeneous graph includes: Based on the geographical location of the disaster-bearing body, a disaster-bearing body node is created at the corresponding position in the heterogeneous graph, and the disaster-bearing body type, vulnerability level, and spatial range are used as the initial attributes of the disaster-bearing body node. Based on the spatial overlap relationship between the disaster-bearing node and the grid unit node at its location, a spatial association edge is established between the disaster-bearing node and the grid unit node to construct the heterogeneous graph.

9. The method according to claim 4, characterized in that, A graph neural network is used to perform spatial extrapolation on the heterogeneous graph, and a preset physical constraint loss is embedded to obtain the activation probability, movement distance, and chain trigger probability of each node in the heterogeneous graph, including: Using the attribute features of each node in the heterogeneous graph as the initial embedding and the connection relationship of each edge in the heterogeneous graph as the propagation path, a graph neural network is used to map the heterogeneous graph from the spatial domain to the spectral domain. Learnable convolutional kernel operations are performed in the spectral domain to capture global spatial dependencies, and then the graph is inversely transformed back to the spatial domain to iteratively update the features of each node. The data fitting loss is calculated, and the total loss function is obtained by weighted summation of the data fitting loss and the preset physical constraint loss. This is the derivation process of the constraint graph neural network. The preset physical constraint loss includes: activation constraint loss, motion constraint loss, and chain-triggered constraint loss. Based on the updated characteristics of each node, output the start probability, running distance, and chain trigger probability of each node.

10. The method according to claim 9, characterized in that, The graph neural network includes a graph Fourier transform network and an inverse graph Fourier transform network. The iterative update of the features of each node includes: In each layer of propagation, the features of each node in the previous layer are aggregated by neighbor analysis and then mapped from the spatial domain to the spectral domain through a graph Fourier transform network. In the spectral domain, these features are multiplied with learnable convolutional kernels, and then returned to the spatial domain through an inverse graph Fourier transform network to obtain the updated node features for the current layer. This layer-by-layer propagation captures global spatial dependencies. The formula for each layer of propagation is as follows: ； in, Let i be the feature vector of node i in the (l+1)th layer. For the graph Fourier transform, This is the inverse graph Fourier transform. Let l be a learnable convolution kernel in the spectral domain. For element-wise product, For activation function, The weight of the terrain-connected edges. To perform weighted feature aggregation on all neighboring nodes j of node i.

11. The method according to claim 9, characterized in that, The formula for calculating the total loss function is as follows: ； in, For the total loss function, For data fitting loss, For the initiation constraint loss based on the Moore-Coulomb criterion, The motion constraint loss is based on the Voellmy rheological model. For chain-triggered constraint loss based on limit equilibrium method, All are weighting coefficients.

12. The method according to claim 9, characterized in that, The step of using a graph neural network to perform spatial extrapolation on the heterogeneous graph and embedding a preset physical constraint loss also includes: By using terrain features, snow cover features, and geological features as covariates, a spatial distribution function of rheological parameters in a graph neural network is constructed. The rheological parameters are dynamically updated over time using online stochastic gradient descent or Kalman filtering. Adaptive calibration of physical parameters is performed based on the spatial distribution function and time-based dynamic updates.

13. The method according to claim 1, characterized in that, The avalanche susceptibility assessment result is generated based on the initiation probability, running distance, and chain triggering probability, including: The comprehensive risk index of each node is determined based on the activation probability, running distance, and chain trigger probability of each node. Based on the comprehensive risk index, the susceptibility level of each node is determined according to a preset grading threshold, and an avalanche susceptibility assessment result for the target area is generated; the susceptibility level includes low susceptibility area, medium susceptibility area, high susceptibility area and extremely high susceptibility area.

14. The method according to claim 1, characterized in that, The method further includes: Based on the avalanche susceptibility assessment results, high-risk sections are identified, and the assessment of high-risk sections is automatically switched from kilometer-level daily scale to hundred-meter-level hourly scale. Based on the uncertainty of the evaluation results quantified by the Bayesian neural network, the confidence level of each avalanche susceptibility evaluation result is obtained.

15. The method according to claim 1, characterized in that, The method further includes: Based on the avalanche susceptibility assessment results, climate scenario data were obtained, and a time-domain generalized graph network was used to predict the spatiotemporal migration patterns of avalanche-prone areas in different future periods under different climate scenarios. A long-term adaptation map is generated based on the prediction results. The long-term adaptation map includes location information of high-risk corridor areas, location information of suggested avoidance areas, and location information of areas that can be reinforced for protection.

16. A geological hazard susceptibility assessment device, characterized in that, The device includes: The data acquisition module is used to acquire multi-source heterogeneous data from the target area; The instability identification and risk quantification module is used to identify snow layer instability zones using the multi-level limit equilibrium method based on the multi-source heterogeneous data, and to deduce multi-level chain disaster paths using the Copula function; The heterogeneous graph construction module is used to construct a heterogeneous graph containing disaster-causing nodes and their relationships based on the snow layer instability zone and the multi-level chain disaster path. The hybrid inference module is used to perform spatial inference on the heterogeneous graph using a graph neural network and embed a preset physical constraint loss to obtain the start probability, movement distance and chain trigger probability of each node in the heterogeneous graph. The evaluation module is used to generate avalanche susceptibility assessment results based on the initiation probability, running distance, and chain triggering probability.

17. An electronic device, characterized in that, include: A memory and a processor are communicatively connected, the memory stores computer instructions, and the processor executes the geological hazard susceptibility assessment method according to any one of claims 1 to 15 by executing the computer instructions.

18. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to execute the geological hazard susceptibility assessment method according to any one of claims 1 to 15.

19. A computer program product, characterized in that, It includes computer instructions for causing a computer to execute the geological hazard susceptibility assessment method according to any one of claims 1 to 15.

Images