Glacier collapse secondary disaster risk intelligent assessment and early warning method
By integrating multi-source data with graph neural networks, a joint characterization system for the three-dimensional dynamic deformation field and internal structure of glaciers is constructed. Combined with the risk evolution equation of multi-physics field coupling, the problems of single data and lagging early warning in the assessment of secondary disasters caused by glacier collapse are solved, and high-precision and real-time early warning capabilities are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NORTHWEST INST OF ECO ENVIRONMENT & RESOURCES CAS
- Filing Date
- 2026-04-14
- Publication Date
- 2026-05-12
AI Technical Summary
Existing methods for assessing the risk of secondary disasters caused by glacier collapse rely on a single data source, which cannot simultaneously acquire information on glacier surface deformation and internal structural evolution. They also lack cross-modal data alignment and semantic fusion, resulting in assessment results that lack physical consistency and long-term interpretability, making it difficult to achieve accurate and real-time early warning.
By integrating multi-source heterogeneous remote sensing data with in-situ sensor observation data, a joint characterization system for the three-dimensional dynamic deformation field and internal structure of glaciers is constructed. A spatiotemporal attention-driven graph neural network is used for multi-scale state perception. Combined with a risk evolution equation coupled with multi-physics fields, a risk level map and early warning instructions with spatiotemporal positioning capabilities are generated.
It enables precise, real-time, and proactive early warning of glacier collapses and their induced disasters, improves the accuracy and timeliness of identifying unstable areas, provides a scientific basis for emergency decision-making, and reduces casualties and economic losses.
Smart Images
Figure CN122020268A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the interdisciplinary field of artificial intelligence and geological disaster prevention and control, specifically involving a method for intelligent assessment and early warning of secondary disaster risks caused by glacier collapse. Background Technology
[0002] With the intensification of global warming, high-altitude glaciers are melting at an accelerated pace, leading to frequent glacier collapses and their induced secondary disasters such as landslides and debris flows, posing a serious threat to the lives and property of downstream residents and infrastructure. Glacier hazard risk assessment has become a crucial link in the geological disaster prevention and emergency response system. Traditional risk assessment methods mainly rely on historical disaster statistics, topographic analysis, and ground monitoring data from limited locations, making it difficult to comprehensively depict the dynamic evolution of the glacier system. In recent years, the development of remote sensing technology has promoted large-scale glacier change monitoring, but single data sources have significant limitations: optical satellite imagery is easily obscured by clouds and cannot penetrate the ice body, making it difficult to obtain information on its internal structure; ground sensors are costly to deploy and have limited coverage, making it difficult to achieve real-time perception across the entire region. Therefore, there is an urgent need to construct an intelligent assessment framework that can integrate multi-source heterogeneous observation data, reconstruct the three-dimensional dynamic evolution process of glaciers, and embed physical mechanisms.
[0003] The core of intelligent risk assessment for secondary disasters caused by glacier collapse lies in the accurate identification and prediction of glacier deformation fields, fissure development, ice displacement rates, and potential instability zones. This requires the system to not only possess high spatiotemporal resolution for external morphological perception but also to deeply analyze the stress distribution and rheological properties within the glacier. Currently, satellite remote sensing can provide long-term land cover change information, LiDAR systems mounted on UAVs can acquire 3D point cloud data with centimeter-level accuracy, and emerging computer vision technologies such as neural radiation fields (NeRF) have shown great potential for reconstructing continuous geometry and appearance from a sparse perspective. However, existing methods often use these technologies in isolation, lacking effective alignment and semantic fusion mechanisms for cross-modal data, and failing to embed physical priors such as glacier dynamic equations into the reconstruction and prediction models, resulting in assessment results lacking physical consistency and long-term interpretability.
[0004] Existing technologies generally suffer from three major shortcomings in glacier hazard risk assessment: First, the data perception dimension is limited, making it impossible to simultaneously acquire information on glacier surface deformation and internal structural evolution, resulting in a lag in the identification of critical instability states. Secondly, most three-dimensional reconstruction models are based on pure data and ignore the constraints of glacier rheology, which can easily lead to non-physical interpretations during extrapolation and prediction, thus reducing the reliability of early warning. Third, multi-source data fusion remains at the level of feature splicing or simple weighting, failing to establish a unified spatiotemporally consistent representation space, making it difficult to support high-precision dynamic risk quantification. Especially against the backdrop of frequent extreme weather events and accelerated glacier retreat, these problems make existing systems inadequate to meet the urgent need for accurate early warnings of sudden collapse events several hours to days in advance. Therefore, there is an urgent need to develop an intelligent assessment and early warning method for glacier disaster risks that integrates multimodal perception, embeds physical mechanisms, and supports dynamic three-dimensional reconstruction. Summary of the Invention
[0005] This invention provides an intelligent assessment and early warning method for the risk of secondary disasters caused by glacier collapse. It constructs a joint characterization system of the three-dimensional dynamic deformation field and internal structural evolution characteristics of glaciers by integrating multi-source heterogeneous remote sensing data, in-situ sensor observation data and high-resolution digital elevation models. On this basis, a spatiotemporal attention-driven graph neural network is used to perform multi-scale state perception and evolution trend modeling of key unstable areas of glaciers. Combined with hydrological and meteorological coupling factors, a risk evolution equation driven by multi-physics fields is constructed. Finally, a risk level map of secondary disasters caused by glacier collapse with spatiotemporal positioning capabilities and graded early warning instructions are generated, realizing accurate, real-time and proactive early warning of glacier collapse and its induced landslides, debris flows and other secondary disasters.
[0006] This invention provides a method for intelligent assessment and early warning of secondary disaster risks caused by glacier collapse, which includes: Acquire multi-source observation data of the glacier region, including synthetic aperture radar interferometry data, lidar point cloud data, ground-based synthetic aperture radar monitoring data, ice seismometer recording data, geothermal gradient sensor data, and meteorological and hydrological station observation data; Based on the synthetic aperture radar interferometry data and lidar point cloud data, a high-precision three-dimensional deformation field on the glacier surface is constructed. The three-dimensional deformation field includes displacement components along the line of sight, elevation change rate in the vertical direction, and displacement vector in the horizontal direction. Based on the ground-based synthetic aperture radar monitoring data and ice seismometer recording data, the depth of crack development, crack propagation rate and spatial distribution density of ice micro-fracture events inside the glacier were retrieved. Based on the geothermal gradient sensor data, the internal temperature profile of the glacier is calculated, and the thickness of the meltwater layer at the ice-rock interface and the intensity of latent heat release during phase change are determined. Using the three-dimensional deformation field, fracture development parameters, ice micro-fracture event density, and meltwater layer thickness as node features, a spatiotemporal graph structure is constructed with key glacier regions as nodes and physical adjacency relationships and mechanical conduction paths as edges. A spatiotemporal attention-driven graph neural network model is deployed on the spatiotemporal graph structure. The model dynamically updates the instability probability of each node through an autoregressive gating mechanism in the time dimension and an anisotropic message passing mechanism in the spatial dimension. The cumulative rainfall, daily average temperature, snowmelt rate and upstream water flow from meteorological and hydrological station observation data are introduced as external driving variables to construct a multi-physics coupled risk evolution equation. The risk evolution equation is defined as the time derivative of the instability probability being equal to the weighted sum of the deformation energy accumulation term, the crack propagation acceleration term, the meltwater lubrication effect term and the external hydrological disturbance term. Based on the numerical solution of the risk evolution equation, a spatiotemporal continuous risk level map covering the entire glacier basin is generated, and the risk level is divided into four levels: low risk, medium risk, high risk and extremely high risk. When the risk level of any grid cell reaches the high risk or extremely high risk threshold, a warning instruction of the corresponding level is triggered. The warning instruction includes the risk location coordinates, the expected collapse time window, the potential impact range, and recommended emergency response measures.
[0007] Preferably, the synthetic aperture radar interferometry data is acquired by a spaceborne platform at a frequency of no more than 6 days for revisiting the data, and its coherence threshold is set to no less than 0.6 to ensure that the deformation inversion accuracy is better than 5 mm.
[0008] Preferably, the lidar point cloud data is collected by an airborne platform during cloudless periods, with a point cloud density of no less than 8 points per square meter. After denoising, registration, and filtering, it is used to generate a digital surface model with a resolution of 0.5 meters.
[0009] Preferably, the ground-based synthetic aperture radar monitoring data is continuously acquired by a fixed radar system at a sampling frequency of once every 10 minutes, with a range resolution of 0.3 meters and an azimuth resolution of 0.1 degrees, used to capture sub-hourly deformation dynamics in the glacier front region.
[0010] Preferably, the data recorded by the ice seismometer is acquired by a three-component seismic detector array deployed on the glacier surface and near the subglacial passage, with a sampling frequency of 100 Hz. Microseismic events with a dominant frequency between 10 Hz and 50 Hz are extracted by short-time Fourier transform, and the location of the event is determined by a time-difference positioning algorithm.
[0011] Preferably, the geothermal gradient sensor is deployed along the depth direction of the glacier borehole at a spacing of 5 meters, with a measurement range of -30 degrees Celsius to +10 degrees Celsius and an accuracy of ±0.1 degrees Celsius, and is used to construct a vertical temperature gradient profile.
[0012] Preferably, the node division of the spatiotemporal graph structure adopts an adaptive meshing strategy, dynamically adjusting the mesh size according to the spatial variation coefficient of deformation rate gradient and crack density. In the area of severe deformation, the mesh side length does not exceed 20 meters, and in the stable area, the mesh side length does not exceed 100 meters.
[0013] Furthermore, the spatiotemporal attention-driven graph neural network model contains three graph convolutional layers, each of which integrates a temporal gating unit and a spatial attention weight matrix. The temporal gating unit adopts a long short-term memory network structure, and the spatial attention weight matrix is calculated by combining the Euclidean distance between nodes, the elevation difference, and the historical co-deformation correlation coefficient.
[0014] Preferably, in the multiphysics coupled risk evolution equation, the deformation energy accumulation term is defined as the square of the ratio of the current deformation rate to the historical maximum deformation rate; the crack propagation acceleration term is defined as the absolute value of the time derivative of the crack density; the meltwater lubrication effect term is defined as the product of the meltwater layer thickness and the freezing point temperature deviation; and the external hydrological disturbance term is defined as the weighted sum of the cumulative rainfall in the past 72 hours and the number of hours when the daily average temperature exceeds 0 degrees Celsius.
[0015] Preferably, the risk level map is generated using a four-dimensional interpolation method, which maps the risk probability values of discrete nodes to a regular geographic grid. The time dimension uses linear extrapolation, and the spatial dimension uses Kriging interpolation. The final output is a dynamic risk map with a resolution of 10 meters and a time step of 1 hour.
[0016] Preferably, the triggering conditions for the early warning instruction are as follows: when the risk probability of the same grid cell is greater than 0.7 within three consecutive time steps, it is determined to be high risk and a yellow warning is activated; when the risk probability is greater than 0.9 and the deformation rate exceeds 10 cm per day, it is determined to be extremely high risk and a red warning is activated.
[0017] This invention provides an intelligent assessment and early warning system for the risk of secondary disasters caused by glacier collapse, comprising: The multi-source data acquisition module is used to acquire synthetic aperture radar interferometry data, lidar point cloud data, ground-based synthetic aperture radar monitoring data, ice seismometer recording data, geothermal gradient sensor data, and meteorological and hydrological station observation data in the glacier region. The three-dimensional deformation field construction module is used to construct a high-precision three-dimensional deformation field on the glacier surface based on synthetic aperture radar interferometry data and lidar point cloud data. The internal structure inversion module is used to invert the depth of fracture development, fracture propagation rate and spatial distribution density of ice microfracture events in glaciers based on ground-based synthetic aperture radar monitoring data and ice seismometer recording data. The meltwater layer identification module is used to calculate the internal temperature profile of the glacier based on geothermal gradient sensor data, and to determine the thickness of the meltwater layer at the ice-rock interface and the intensity of latent heat release during phase change. The spatiotemporal graph construction module is used to construct a spatiotemporal graph structure with key glacier regions as nodes and physical adjacency relationships and mechanical transmission paths as edges, using three-dimensional deformation field, fracture development parameters, ice micro-fracture event density and meltwater layer thickness as node features. The instability probability prediction module is used to deploy a spatiotemporal attention-driven graph neural network model on the spatiotemporal graph structure and dynamically update the instability probability of each node. The risk evolution modeling module is used to introduce meteorological and hydrological station observation data and construct a risk evolution equation coupled with multiple physics fields. The risk map generation module is used to generate a spatiotemporally continuous risk level map covering the entire glacier basin based on the numerical solution of the risk evolution equation. The early warning instruction triggering module is used to trigger the corresponding level of early warning instruction when the risk level of any grid cell reaches the high risk or extremely high risk threshold.
[0018] Preferably, the multi-source data acquisition module simultaneously receives various types of observation data through three methods: satellite downlink, UAV relay communication, and wired sensor network, and performs timestamp alignment and coordinate system unification processing.
[0019] Preferably, the three-dimensional deformation field construction module uses the least squares phase unwrapping algorithm to process the interference phase and combines it with lidar point cloud for terrain phase correction to eliminate the influence of atmospheric delay and orbital error.
[0020] Preferably, the internal structure inversion module uses the texture entropy change rate of the synthetic aperture radar amplitude image to identify the width of the surface crack opening, and combines the source mechanism solution of the ice seismic event to infer the orientation of the deep shear zone.
[0021] Preferably, the meltwater layer identification module inverts the heat flux under the ice by solving the one-dimensional heat conduction equation, and determines whether the meltwater layer exists based on the location of the 0-degree Celsius isotherm.
[0022] Preferably, the edge weights in the spatiotemporal graph construction module are determined by the mechanical coupling coefficient between nodes, which is equal to the product of the reciprocal of the elastic modulus and the square of the distance.
[0023] Preferably, the graph neural network model of the instability probability prediction module uses historical collapse event samples as positive samples and stable period observation data as negative samples during the training phase, and the loss function is the focal loss function to alleviate the sample imbalance problem.
[0024] Preferably, the risk evolution modeling module uses the finite difference method to discretize and solve the risk evolution equation, with a time step of 1 hour and a spatial step consistent with the risk map grid.
[0025] Preferably, the risk map generation module supports interface with a geographic information system platform to output risk level raster data in standard GeoTIFF format.
[0026] Furthermore, the warning instruction triggering module pushes the warning information to the emergency management command center, local disaster prevention departments, and public mobile terminals through a dedicated communication link, and records the warning response log for post-event evaluation.
[0027] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention overcomes the technical limitations of traditional glacier monitoring, which relies on a single optical image data source. For the first time, it achieves deep fusion of multi-source heterogeneous data from synthetic aperture radar, lidar, ground-based radar, ice seismometers, and geothermal sensors, solving the core challenges of invisible changes in glacier internal structure and the inability to acquire three-dimensional deformation in real time. By constructing a spatiotemporal neural network model constrained by physical mechanisms, the glacier is treated as a dynamic system with mechanical transmission characteristics, rather than an isolated set of pixels, significantly improving the accuracy and timeliness of unstable area identification. A multi-physics coupled risk evolution equation is introduced, incorporating deformation, fissures, meltwater, and meteorological and hydrological factors into a unified mathematical framework, upgrading risk assessment from static threshold judgment to dynamic process simulation. The generated risk level map possesses clear spatiotemporal positioning capabilities and quantitative probabilistic support, allowing early warning commands to be accurate to specific grid units and time windows. This provides a scientific basis for emergency evacuation, engineering protection, and resource allocation, effectively reducing casualties and economic losses caused by secondary disasters from glacier collapses. Attached Figure Description
[0028] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a schematic diagram of the core principle framework of the spatiotemporal attention-driven graph neural network in this invention for multi-scale state perception and evolution trend modeling of key unstable regions of glaciers. Figure 3 This is a logical flowchart of the fusion of multi-source heterogeneous observation data and the joint characterization of the three-dimensional dynamic deformation field and internal structure of glaciers in this invention. Figure 4 This is a logical flowchart of the glacier spatiotemporal map structure constructed based on physical adjacency relationships and mechanical transmission paths in this invention; Figure 5 This is a flowchart illustrating the logical process of constructing and numerically solving the risk evolution equation driven by multi-physics fields in this invention. Detailed Implementation
[0029] Please refer to the attached document. Figures 1 to 5 This invention provides an intelligent assessment and early warning method for secondary disaster risks caused by glacier collapse. Its core lies in integrating multi-source heterogeneous remote sensing data, in-situ sensor observation data, and a high-resolution digital elevation model to construct a joint characterization system of the glacier's three-dimensional dynamic deformation field and internal structural evolution characteristics. Based on this, a spatiotemporal attention-driven graph neural network is used to perform multi-scale state perception and evolution trend modeling of key unstable areas of the glacier. Combined with hydrological and meteorological coupling factors, a risk evolution equation driven by multi-physics fields is constructed. Finally, a risk level map of secondary disaster risks with spatiotemporal positioning capabilities and graded early warning instructions are generated, enabling accurate, real-time, and proactive early warning of glacier collapses and their induced landslides, debris flows, and other secondary disasters.
[0030] The method includes the following steps: S1. Acquire multi-source observation data of the glacier region, including synthetic aperture radar interferometry data, lidar point cloud data, ground-based synthetic aperture radar monitoring data, ice seismometer recording data, geothermal gradient sensor data, and meteorological and hydrological station observation data. S2. Based on the synthetic aperture radar interferometry data and lidar point cloud data, construct a high-precision three-dimensional deformation field on the glacier surface. This three-dimensional deformation field includes displacement components along the line of sight, the vertical elevation change rate, and the horizontal displacement vector. S3. Based on the ground-based synthetic aperture radar monitoring data and ice seismometer recording data, invert the depth of fracture development, fracture propagation rate, and spatial distribution density of micro-fracture events within the glacier. S4. Based on the geothermal gradient sensor data, calculate the internal temperature profile of the glacier and determine the thickness of the meltwater layer at the ice-rock interface and the intensity of latent heat release during phase change. S5. Using the three-dimensional deformation field, fracture development parameters, ice micro-fracture event density, and meltwater layer thickness as node features, a spatiotemporal graph structure is constructed with key glacier regions as nodes and physical adjacency relationships and mechanical transmission paths as edges. S6. A spatiotemporal attention-driven graph neural network model is deployed on this spatiotemporal graph structure. This model dynamically updates the instability probability of each node through an autoregressive gating mechanism in the time dimension and an anisotropic message passing mechanism in the spatial dimension. S7. Accumulated rainfall, average daily temperature, snowmelt rate, and upstream water flow from meteorological and hydrological station observation data are introduced as external driving variables to construct a multi-physics coupled risk evolution equation. This risk evolution equation is defined as the time derivative of the instability probability being equal to the weighted sum of the deformation energy accumulation term, fracture propagation acceleration term, meltwater lubrication effect term, and external hydrological disturbance term. S8. Based on the numerical solution of the risk evolution equation, a spatiotemporal continuous risk level map covering the entire glacier basin is generated. The risk levels are divided into four levels: low risk, medium risk, high risk, and extremely high risk. S9. When the risk level of any grid cell reaches the high risk or extremely high risk threshold, a warning instruction of the corresponding level is triggered. The warning instruction includes the risk location coordinates, the expected collapse time window, the potential impact range, and the recommended emergency response measures.
[0031] In step S1, multi-source observation data of the glacier region are acquired. Synthetic Aperture Radar (SAR) interferometry data is acquired via a spaceborne platform at a revisit frequency of no more than 6 days, with a coherence threshold set to no less than 0.6 to ensure deformation inversion accuracy better than 5 mm. LiDAR point cloud data is collected via an airborne platform during cloudless periods, with a point cloud density of no less than 8 points per square meter. After denoising, registration, and filtering, it is used to generate a digital surface model with a resolution of 0.5 meters. Ground-based SAR monitoring data is continuously acquired via a fixed radar system at a sampling frequency of once every 10 minutes, achieving a range resolution of 0.3 meters and an azimuth resolution of 0.1 degrees, used to capture sub-hourly deformation dynamics in the glacier front region. Ice seismometer data is collected by a three-component seismic detector array deployed on the glacier surface and near subglacial passages, with a sampling frequency of 100 Hz. Microseismic events with dominant frequencies between 10 Hz and 50 Hz are extracted using short-time Fourier transform, and the event locations are determined using a time-difference-of-travel (TDOT) positioning algorithm. Geothermal gradient sensors are deployed along the depth of the glacier boreholes at 5-meter intervals, measuring from -30°C to +10°C with an accuracy of ±0.1°C, to construct vertical temperature gradient profiles. Meteorological and hydrological station observation data, including cumulative rainfall, daily average temperature, snowmelt rate, and upstream water flow, are collected hourly and transmitted in real-time to the central processing unit via wired or wireless communication links. All data undergoes timestamp alignment and coordinate system unification upon receipt to ensure spatiotemporal consistency in subsequent fusion analysis.
[0032] In step S2, a high-precision three-dimensional deformation field of the glacier surface is constructed. First, a least-squares phase unwrapping algorithm is performed on the synthetic aperture radar interferometry data to obtain the phase change sequence along the radar line of sight. Then, the terrain phase is corrected using a digital surface model generated from lidar point clouds to eliminate the influence of atmospheric delay and orbital error. Next, the line-of-sight displacement component is decomposed into two orthogonal components in the vertical and horizontal directions. The vertical elevation change rate is obtained by differential calculation using multi-temporal interferometry, while the horizontal displacement vector is estimated using an azimuth offset tracking algorithm. Finally, the deformation information in the three directions is fused to form a three-dimensional deformation field covering the entire glacier surface, with a spatial resolution of 10 meters and a temporal resolution of 6 days.
[0033] In step S3, parameters for the development of internal glaciers are retrieved. Ground-based synthetic aperture radar (SAR) monitoring data is used to identify the width of surface crack openings in the glacier front region, based on the texture entropy change rate of the radar amplitude image—when cracks open, local scattering characteristics change abruptly, leading to a significant increase in texture entropy. Data recorded by glaciers is used to detect micro-fracture events within the ice mass. Energy peaks in the 10 Hz to 50 Hz frequency band are extracted using short-time Fourier transform and marked as valid microseismic events. Using the travel time difference between P-waves and S-waves, combined with the known P-wave velocity (approximately 3800 m / s) and S-wave velocity (approximately 1900 m / s) of the ice mass, a double-difference localization algorithm is used to determine the three-dimensional spatial coordinates of microseismic events. The crack development depth is defined as the deepest boundary of the concentrated distribution of microseismic events. The crack propagation rate is calculated by dividing the change in the crack front position over a continuous time period by the time interval. The spatial distribution density of ice mass micro-fracture events is expressed as the number of microseismic events per unit volume.
[0034] In step S4, the internal temperature profile of the glacier is calculated and the meltwater layer parameters are determined. A geothermal gradient sensor is deployed at one measuring point every 5 meters along the borehole depth to record temperature values in real time. The measured temperature sequence is arranged by depth and fitted with a one-dimensional linear or nonlinear function to obtain a continuous temperature profile. The location of the ice-rock interface is determined by a sensor at the bottom of the borehole. The meltwater layer thickness is defined as the vertical distance from the 0°C isotherm in the temperature profile to the ice-rock interface. The latent heat release intensity of the phase change is obtained by calculating the heat flux through the 0°C isotherm per unit time. The heat flux is calculated using Fourier's law, i.e., thermal conductivity multiplied by the temperature gradient. The thermal conductivity of the ice body is taken as 2.2 W / m / Kelvin.
[0035] In step S5, a spatiotemporal map structure of the glacier is constructed. First, the glacier region is adaptively meshed, and the mesh size is dynamically adjusted according to the spatial variation coefficient of deformation rate gradient and fracture density: in regions where the deformation rate is greater than 5 cm per day or the fracture density is greater than 0.5 events per cubic meter, the mesh edge length does not exceed 20 meters; in other stable regions, the mesh edge length does not exceed 100 meters; the center point of each mesh is defined as a node of the graph structure; the node feature vector consists of the average three-dimensional deformation field component, fracture development depth, fracture propagation rate, microfracture event density, and meltwater layer thickness within the mesh; the edge connection relationship is determined by both physical adjacency and mechanical transmission path: any two nodes are spatially adjacent and the elevation difference is less than 50 meters, then an edge connection is established; in addition, if there is a historical co-deformation correlation coefficient greater than 0.7 between two nodes, then an edge is forcibly established; the edge weight is determined by the mechanical coupling coefficient, which is equal to the product of the reciprocal of the ice elastic modulus and the square of the Euclidean distance between nodes, and the ice elastic modulus is taken as 900 MPa.
[0036] In step S6, a spatiotemporal attention-driven graph neural network model is deployed to dynamically update the instability probability. This model consists of three graph convolutional layers, each integrating a temporal gating unit and a spatial attention weight matrix. The temporal gating unit employs a long short-term memory network structure, taking as input the node features of the current and previous time steps and outputting a state update gating signal along the temporal dimension. The spatial attention weight matrix... Calculated in the following way: ; For nodes With nodes Euclidean distance, For nodes With nodes The absolute value of the elevation difference, For nodes With nodes The Pearson correlation coefficient of the two-node deformation sequences over the past 24 hours. For nodes With nodes Euclidean distance, For nodes With nodes The absolute value of the elevation difference, For nodes With nodes The Pearson correlation coefficient of the two-node deformation sequences over the past 24 hours. This is a learnable nonlinear mapping function. During message passing, each node aggregates the weighted sum of features from its neighbors and merges it with its own historical state to generate a new node embedding. Finally, a fully connected layer maps the node embedding to an instability probability value, ranging from 0 to 1.
[0037] In step S7, the risk evolution equation of multiphysics coupling is constructed. Let... For a certain grid cell at time... If the probability of instability is given, then its time derivative is... / It consists of a weighted sum of 4 terms: ;in, , , , These are the learnable weight coefficients, with initial values set to 0.4, 0.3, 0.2, and 0.1, respectively.
[0038] Deformation energy accumulation term Defined as the current deformation rate With the historical maximum deformation rate The square of the ratio, i.e. .
[0039] Crack propagation acceleration term Defined as the absolute value of the time derivative of the fracture density, i.e. ,in This represents the fracture density.
[0040] Water-lubricating effect term Defined as the thickness of the meltwater layer Deviation from freezing point temperature The product of , This is the average daily temperature at the glacier surface, therefore .
[0041] External hydrological disturbance items Defined as the cumulative rainfall in the past 72 hours Hours with average daily temperature above 0 degrees Celsius The weighted sum, i.e. .
[0042] The equation is solved by discretization using the finite difference method, with a time step of 1 hour and the initial condition being the initial instability probability output by the graph neural network.
[0043] In step S8, a spatiotemporally continuous risk level map is generated. The risk probability values of discrete grid nodes are mapped to a regular geographic grid through four-dimensional interpolation: the spatial dimension uses ordinary kriging interpolation with a range of 50 meters and a nugget effect of 0.05; the temporal dimension uses linear extrapolation to predict the risk evolution over the next 12 hours; the final output is a dynamic risk map with a resolution of 10 meters and a time step of 1 hour; the risk level classification criteria are as follows: <0.3 indicates low risk, 0.3≤ <0.7 indicates medium risk, 0.7≤ <0.9 indicates high risk. A value of ≥0.9 indicates an extremely high risk.
[0044] In step S9, an early warning instruction is triggered. When the risk probability of the same grid cell exceeds 0.7 for three consecutive time steps (i.e., three hours), it is classified as high risk, and a yellow warning is activated. When the risk probability exceeds 0.9 and the deformation rate exceeds 10 cm per day, it is classified as extremely high risk, and a red warning is activated. The early warning instruction includes the latitude and longitude coordinates of the risk location, the expected collapse time window (extrapolated based on the time when the risk probability reaches its peak), the potential impact range (derived through collapse volume estimation and terrain diffusion model simulation), and recommended emergency response measures (such as evacuation routes, traffic control areas, and engineering reinforcement recommendations). The early warning information is pushed to the emergency management command center, local disaster prevention departments, and public mobile terminals via a dedicated communication link, and an early warning response log is recorded for post-event assessment.
[0045] The system upon which the above method relies includes a multi-source data acquisition module, a three-dimensional deformation field construction module, an internal structure inversion module, a meltwater layer identification module, a spatiotemporal map construction module, an instability probability prediction module, a risk evolution modeling module, a risk map generation module, and an early warning command triggering module.
[0046] The multi-source data acquisition module receives synthetic aperture radar interferometry data via satellite downlink, receives lidar point cloud data via UAV relay communication, and receives data from ground-based radar, ice seismometer, geothermal sensor, and meteorological and hydrological stations via wired sensor network. It also performs nanosecond-level timestamp alignment and WGS84 coordinate system unification.
[0047] The three-dimensional deformation field construction module uses the least squares phase unwrapping algorithm to process the interference phase and combines it with lidar point cloud for terrain phase correction to eliminate the influence of atmospheric delay and orbital error, and outputs a three-dimensional deformation vector field.
[0048] The internal structure inversion module uses the texture entropy change rate of synthetic aperture radar amplitude images to identify the width of surface crack openings, and combines the source mechanism solution of ice seismic events to infer the orientation of deep shear zones, outputting a three-dimensional crack distribution model.
[0049] The meltwater layer identification module inverts the heat flux under ice by solving the one-dimensional heat conduction equation, and determines whether the meltwater layer exists based on the location of the 0-degree Celsius isothermal surface, outputting the meltwater layer thickness and heat release intensity.
[0050] The edge weights in the spatiotemporal graph construction module are determined by the mechanical coupling coefficient between nodes. The mechanical coupling coefficient is equal to the product of the reciprocal of the elastic modulus and the square of the distance. The graph structure is stored in the in-memory database in the form of an adjacency matrix.
[0051] The graph neural network model of the instability probability prediction module uses historical collapse event samples as positive samples and stable period observation data as negative samples during the training phase. The loss function is the focal loss function, which focuses on difficult-to-classify samples and alleviates the problem of imbalance between positive and negative samples.
[0052] The risk evolution modeling module uses the finite difference method to discretize and solve the risk evolution equation. The time step is 1 hour, and the spatial step is consistent with the risk map grid. The solution process is executed in parallel on a GPU accelerator.
[0053] The risk map generation module supports integration with geographic information system platforms, outputting risk level raster data in standard GeoTIFF format, including timestamps, coordinate reference systems, and risk probability values.
[0054] The warning command triggering module pushes warning information through three redundant links: Beidou short message, 4G / 5G cellular network and satellite communication, to ensure communication reliability in extreme environments, and automatically records the warning release time, the recipient's confirmation status and subsequent handling feedback.
[0055] This embodiment, through the above-described method and system, achieves intelligent assessment and early warning of the risk of secondary disasters caused by glacier collapses throughout the entire process, all elements, and all time and space. It solves the problem of early warning failure caused by traditional methods due to single data, static models, and delayed response, and provides technical support for disaster prevention and mitigation in high-altitude areas.
Claims
1. A method for intelligent assessment and early warning of secondary disaster risks from glacier collapses, characterized in that, include: Acquire multi-source observation data of the glacier region; Construct a spatiotemporal graph structure with key glacier regions as nodes and physical adjacency relationships and mechanical transmission paths as edges; A spatiotemporal attention-driven graph neural network model is deployed on the spatiotemporal graph structure. The model dynamically updates the instability probability of each node through an autoregressive gating mechanism in the time dimension and an anisotropic message passing mechanism in the spatial dimension. The multi-source observation data includes meteorological and hydrological station observation data. The cumulative rainfall, daily average temperature, snowmelt rate and upstream water flow from the meteorological and hydrological station observation data are introduced as external driving variables to construct a multi-physics coupled risk evolution equation. The risk evolution equation is defined as the time derivative of the instability probability being equal to the weighted sum of the deformation energy accumulation term, the crack propagation acceleration term, the meltwater lubrication effect term and the external hydrological disturbance term. Based on the numerical solution of the risk evolution equation, a spatiotemporal continuous risk level map covering the entire glacier basin is generated, and the risk level is divided into four levels: low risk, medium risk, high risk and extremely high risk. When the risk level of any grid cell reaches the high risk or extremely high risk threshold, a warning instruction of the corresponding level is triggered. The warning instruction includes the risk location coordinates, the expected collapse time window, the potential impact range, and recommended emergency response measures.
2. The intelligent assessment and early warning method for secondary disaster risks caused by glacier collapse according to claim 1, characterized in that, The multi-source observation data also includes synthetic aperture radar interferometry data, lidar point cloud data, ground-based synthetic aperture radar monitoring data, ice seismometer recorded data, and geothermal gradient sensor data; Based on the synthetic aperture radar interferometry data and lidar point cloud data, a high-precision three-dimensional deformation field on the glacier surface is constructed. The three-dimensional deformation field includes displacement components along the line of sight, elevation change rate in the vertical direction, and displacement vector in the horizontal direction. Based on the ground-based synthetic aperture radar monitoring data and ice seismometer recording data, the depth of crack development, crack propagation rate and spatial distribution density of ice micro-fracture events inside the glacier were retrieved. Based on the geothermal gradient sensor data, the internal temperature profile of the glacier is calculated, and the thickness of the meltwater layer at the ice-rock interface and the intensity of latent heat release during phase change are determined. Using the three-dimensional deformation field, fracture development parameters, ice micro-fracture event density, and meltwater layer thickness as node features, a spatiotemporal graph structure is constructed with key glacier regions as nodes and physical adjacency relationships and mechanical transmission paths as edges.
3. The intelligent assessment and early warning method for secondary disaster risks caused by glacier collapse according to claim 2, characterized in that, Based on the synthetic aperture radar interferometry data and lidar point cloud data, a high-precision three-dimensional deformation field of the glacier surface is constructed, including: The least-squares phase unwrapping algorithm is performed on the synthetic aperture radar interferometry data to obtain the phase change sequence along the line of sight; The digital surface model generated by the lidar point cloud is used to correct the terrain phase in order to eliminate the effects of atmospheric delay and orbital error; The line-of-sight displacement component is decomposed into the vertical elevation change rate and the horizontal displacement vector. Deformation information from three directions is integrated to form a three-dimensional deformation field with a spatial resolution of 10 meters and a temporal resolution of 6 days.
4. The intelligent assessment and early warning method for secondary disaster risks caused by glacier collapse according to claim 3, characterized in that, Based on the ground-based synthetic aperture radar monitoring data and ice seismometer recording data, the depth of fracture development, fracture propagation rate, and spatial distribution density of ice microfracture events within the glacier are retrieved, including: Identifying surface crack opening width based on the rate of change of texture entropy in synthetic aperture radar amplitude images; Microseismic events with dominant frequencies between 10 Hz and 50 Hz were extracted from ice seismometer data using short-time Fourier transform. By combining the travel time difference between P-waves and S-waves with the P-wave velocity and S-wave velocity of ice, the three-dimensional spatial coordinates of microseismic events are determined using a double-difference positioning algorithm. The deepest boundary where microseismic events are concentrated is defined as the fracture development depth. The fracture propagation rate is obtained by dividing the change in the position of the fracture tip by the time interval within a continuous time period. The number of microseismic events per unit volume is defined as the spatial distribution density of ice body microfracture events.
5. The intelligent assessment and early warning method for secondary disaster risks caused by glacier collapse according to claim 4, characterized in that, Based on the geothermal gradient sensor data, the internal temperature profile of the glacier is calculated, and the thickness of the meltwater layer at the ice-rock interface and the intensity of latent heat release during phase change are determined, including: The temperature values measured by the geothermal gradient sensors deployed every 5 meters along the borehole depth are fitted with a one-dimensional temperature profile based on the depth. The vertical distance from the 0°C isotherm to the ice-rock interface in the temperature profile is defined as the meltwater layer thickness. The heat flux through the 0°C isothermal surface is calculated based on Fourier's law, and this heat flux is defined as the latent heat release intensity of phase change.
6. The intelligent assessment and early warning method for secondary disaster risks caused by glacier collapse according to claim 5, characterized in that, Using the three-dimensional deformation field, fracture development parameters, ice microfracture event density, and meltwater layer thickness as nodal features, a spatiotemporal graph structure is constructed with key glacier regions as nodes and physical adjacency relationships and mechanical transmission paths as edges, including: Based on the spatial variation coefficient of deformation rate gradient and fracture density, the glacier region is adaptively meshed. In areas where the deformation rate is greater than 5 cm per day or the fracture density is greater than 0.5 events per cubic meter, the mesh side length is set to no more than 20 meters, and in other areas, the mesh side length is set to no more than 100 meters. Each grid center point is defined as a node, and the node feature vector consists of the average three-dimensional deformation field components, fracture development depth, fracture propagation rate, microfracture event density, and meltwater layer thickness within that grid. If two nodes are spatially adjacent and their elevation difference is less than 50 meters, or their historical collaborative deformation correlation coefficient is greater than 0.7, then an edge connection is established. The edge weights are set as mechanical coupling coefficients, which are equal to the product of the reciprocal of the elastic modulus of the ice body and the square of the Euclidean distance between nodes.
7. The intelligent assessment and early warning method for secondary disaster risks caused by glacier collapse according to claim 6, characterized in that, A spatiotemporal attention-driven graph neural network model is deployed on the spatiotemporal graph structure to dynamically update the instability probability of each node, including: A three-layer graph convolutional layer is used, with each layer integrating a temporal gating unit based on a long short-term memory network and a spatial attention weight matrix based on Euclidean distance between nodes, elevation difference, and historical cooperative deformation correlation coefficient. A new node embedding is generated by aggregating the weighted sum of neighbor node features through an anisotropic message passing mechanism and merging it with its own historical state. The fully connected layer maps the node embeddings to instability probability values ranging from 0 to 1.
8. The intelligent assessment and early warning method for secondary disaster risks caused by glacier collapse according to claim 7, characterized in that, By introducing cumulative rainfall, daily average temperature, snowmelt rate, and upstream water flow from meteorological and hydrological station observation data as external driving variables, a multi-physics coupled risk evolution equation is constructed, including: The time derivative of the instability probability is expressed as a weighted sum of the deformation energy accumulation term, the crack propagation acceleration term, the meltwater lubrication effect term, and the external hydrological disturbance term; The deformation energy accumulation term is defined as the square of the ratio of the current deformation rate to the historical maximum deformation rate; The crack propagation acceleration term is defined as the absolute value of the time derivative of the crack density; The meltwater lubrication effect term is defined as the product of the meltwater layer thickness and the freezing point temperature deviation; The external hydrological disturbance term is defined as the weighted sum of the cumulative rainfall over the past 72 hours and the number of hours when the average daily temperature exceeds 0 degrees Celsius.
9. The intelligent assessment and early warning method for secondary disaster risks caused by glacier collapse according to claim 8, characterized in that, Based on the numerical solution of the aforementioned risk evolution equation, a spatiotemporal continuous risk level map covering the entire glacier basin is generated, including: The risk evolution equation is discretized and solved using the finite difference method with a time step of 1 hour. The initial condition is the initial instability probability output by the graph neural network. The risk probability values of discrete nodes are mapped to a regular geographic grid with a spatial resolution of 10 meters through ordinary kriging interpolation, with a range of 50 meters and a nugget effect of 0.
05. Linear extrapolation is used in the time dimension to generate risk evolution predictions for the next 12 hours; Risk levels are classified based on risk probability values: <0.3 indicates low risk, 0.3≤ <0.7 indicates medium risk, 0.7≤ <0.9 indicates high risk. A value of ≥0.9 indicates an extremely high risk.
10. The intelligent assessment and early warning method for secondary disaster risks caused by glacier collapse according to claim 9, characterized in that, When the risk level of any grid cell reaches the high-risk or extremely high-risk threshold, the corresponding level of warning instruction is triggered, including: When the risk probability of the same grid cell is greater than 0.7 within three consecutive time steps, it is judged as high risk and a yellow alert is activated. When the risk probability is greater than 0.9 and the deformation rate exceeds 10 centimeters per day, it is judged as extremely high risk and a red alert is activated. The warning instruction includes the latitude and longitude coordinates of the risk location, the expected collapse time window extrapolated based on the peak time of the risk probability, the potential impact range obtained through collapse volume estimation and terrain diffusion model simulation, and recommended emergency response measures.