Geological disaster hidden danger checking method and system based on multi-source remote sensing image fusion

Abstract

The application discloses a kind of geological disaster hidden danger checking method and system based on multi-source remote sensing image fusion, it is related to geological disaster meteorological service field.The method comprises: obtaining meteorological precipitation data, elevation model, basic geographic vector and synthetic aperture radar data;Hydrological disaster high-risk candidate area is generated by fusing meteorological and terrain data;Microwave backscattering coefficient is extracted to extract soil moisture supersaturation object;Rigid ground object anchor point is extracted using vector data and non-rigid mask is generated, reset parameter to remove vegetation environmental noise;Supersaturation object and rigid anchor point are respectively used as source and target node to construct adjacency matrix, input graph neural network model to output spatial penetration deterioration coefficient;Calculate hydrostatic pressure in combination with elevation difference and generate normal penetration force vector, substitute into mechanics model to output stability degradation index.The application overcomes optical failure and radar vegetation noise interference, realizes the precise mechanics early warning of all-weather high-risk moisture object before disaster.

Description

Technical Field

[0001] This invention relates to the field of geological disaster monitoring technology, and in particular to a method and system for verifying potential geological disaster hazards based on multi-source remote sensing image fusion. Background Technology

[0002] Geological hazard investigation is a crucial foundation for geological hazard prevention and control. With the development of remote sensing technology, the use of multi-source remote sensing data to identify potential geological hazards such as landslides and debris flows can improve the coverage of investigation operations in large-scale mountainous areas.

[0003] During flood season verification in mountainous areas, heavy rainfall is often accompanied by cloud and fog obscuring the view, significantly reducing the usability of optical and thermal infrared remote sensing images. Synthetic Aperture Radar (SAR) utilizes microwaves for active observation, which is less affected by clouds and fog. However, the swaying of vegetation canopies, changes in canopy water content, and volume scattering caused by wind and rain introduce phase and amplitude noise. Directly identifying full-area SAR images would easily increase computational load and could potentially misjudge vegetation changes as surface anomalies.

[0004] Existing technologies suffer from insufficient integration between hydrological candidate areas, radar inversion results, and geotechnical indices, making it difficult to quantify and verify areas with high soil moisture saturation and their impact on downstream man-made structures before significant surface damage occurs. Therefore, an intelligent identification and verification method is needed that can limit the scope of high-risk verification under cloud and rain cover conditions, suppress vegetation-related environmental noise, and generate interpretable mechanical indices. Summary of the Invention

[0005] To address the issue that the accuracy and computational efficiency of existing technologies are affected by cloud and rain obstruction and vegetation noise during the flood season, this invention provides a method and system for verifying geological hazard risks based on multi-source remote sensing image fusion.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a method for verifying geological hazard risks based on multi-source remote sensing image fusion, comprising:

[0007] S1. Obtain meteorological precipitation grid sequence data, digital elevation model data, basic geographic information vector data of the target verification area, and synthetic aperture radar complex image data of the flood season whose imaging time is closest to the last time of the meteorological precipitation grid sequence data.

[0008] S2. The meteorological precipitation grid sequence data and the digital elevation model data are fused with topographic features to calculate and generate high-risk candidate areas for hydrological disasters.

[0009] S3. Extract the microwave backscattering coefficient of the synthetic aperture radar complex image data during the flood season in the candidate area of ​​high risk of hydrological disasters, invert the volume water content parameter based on the microwave backscattering coefficient, and extract the soil moisture saturation objects.

[0010] S4. Using the basic geographic information vector data, extract rigid feature anchor points and vegetation changeable areas, generate a non-rigid mask based on the vegetation changeable areas, reset the volumetric water content parameter within the coverage area of ​​the non-rigid mask to the background reference value, and update the soil moisture high saturation object according to the reset volumetric water content parameter to remove vegetation environmental noise features.

[0011] S5. Extract the spatial features of the soil moisture-saturated object updated in step S4 as the source node, extract the geometric features of the rigid ground anchor point as the target node, construct a directed adjacency matrix based on the source node and the target node, input the directed adjacency matrix and the node feature matrix into a pre-trained graph neural network model, and output the spatial permeability deterioration coefficient.

[0012] S6. Calculate the foundation hydrostatic pressure based on the elevation difference between the source node and the target node, multiply the foundation hydrostatic pressure by the spatial permeability deterioration coefficient to obtain the equivalent pore water pressure, input the equivalent pore water pressure into a preset mechanical evaluation model, generate a stability degradation index, and output the hidden danger verification results.

[0013] Furthermore, step S2 specifically includes:

[0014] S21. Divide the slope into units based on the digital elevation model data, and extract the slope matrix and catchment area matrix of each slope unit.

[0015] S22, according to the formula Calculate the runoff hazard of each slope unit; where, Let i be the runoff hazard level of the i-th slope unit. For the catchment area, For slope, For precipitation intensity, α is the infiltration rate, m and n are the scale calibration coefficients, and m and n are the exponents calibrated from historical samples.

[0016] S23. Extract the slope units with runoff hazard levels greater than the preset runoff hazard threshold as the high-risk candidate areas for hydrological disasters.

[0017] Furthermore, step S3 specifically includes:

[0018] S31. Perform radiometric calibration, speckle filtering and terrain correction on the synthetic aperture radar complex image data during the flood season to obtain the microwave backscattering coefficient, and use the water cloud model to separate the vegetation canopy component in the microwave backscattering coefficient to obtain the bare soil backscattering matrix.

[0019] S32. Input the bare soil backscattering matrix, radar incident angle and surface roughness into the empirical multinomial model calibrated by the soil moisture sampling points of the target verification area, and solve for the volumetric water content parameters corresponding to each pixel.

[0020] S33. The product of the measured porosity of the target verification area and the preset saturation coefficient is used as the high saturation judgment threshold. Continuous pixel areas with volumetric water content parameters not less than the high saturation judgment threshold are extracted as the soil moisture high saturation objects.

[0021] Furthermore, step S4 specifically includes:

[0022] S41. Parse the basic geographic information vector data, convert the geometric polygons with the attribute of artificial structures into the rigid ground feature anchor points, and extract the geometric polygons with the attribute of vegetation coverage area as the vegetation change zone.

[0023] S42. The vegetation change zone is rasterized, and a mathematical morphological dilation operation is performed on the rasterization result. After deducting the rigid ground feature anchor points and their buffer zones, the non-rigid mask is generated.

[0024] S43. Perform coordinate registration and resolution alignment between the non-rigid mask and the spatial distribution matrix of the volumetric water content parameter, reset the volumetric water content parameter within the coverage area of ​​the non-rigid mask to the background reference value, and re-extract the soil moisture high saturation object based on the reset volumetric water content parameter.

[0025] Furthermore, step S5 specifically includes:

[0026] S51. Extract the three-dimensional spatial centroid coordinates and the mean internal volume water content of the soil moisture-saturated object to construct the source node, extract the three-dimensional geometric center coordinates of the rigid ground feature anchor point to construct the target node, and dimensionally align the source node features and target node features to obtain the node feature matrix.

[0027] S52. When the elevation of the source node is greater than the elevation of the target node, a directed edge is established from the source node to the target node, and the edge connection weight is calculated by using a Gaussian kernel function based on the spatial Euclidean distance between the two to construct the directed adjacency matrix.

[0028] S53. Aggregate node features through at least two graph convolutional layers of the graph neural network model, and output the spatial penetration deterioration coefficient with a value range of 0 to 1 through the fully connected layer and Sigmoid activation function at the end of the network.

[0029] Furthermore, step S6 specifically includes:

[0030] S61, according to Calculate the foundation hydrostatic pressure; wherein, The aforementioned basic hydrostatic pressure, The density of water, The elevation difference between the source node and the target node;

[0031] S62, according to Calculate the equivalent pore water pressure, and according to... Calculate the current shear strength; where, The equivalent pore water pressure, The space permeability deterioration coefficient is mentioned above. The current shear strength, For effective cohesion, For the total normal stress, The effective internal friction angle;

[0032] S63, according to Calculate the baseline shear strength, and according to Calculate the stability degradation index; wherein, The reference shear strength is... This refers to the stability degradation metric.

[0033] Furthermore, prior to step S5, the following steps are also included:

[0034] S71. Obtain historical heavy rainfall sequences, radar inversion state matrices, and corresponding actual instability records, and construct a training sample set containing positive samples that have experienced instability and negative samples that have not experienced instability.

[0035] S72. Standardize the node features in the training sample set, calculate the predicted instability probability in the forward propagation of the graph neural network model, and construct a joint loss function containing a binary cross-entropy term and a regularization penalty term based on the sample labels.

[0036] S73. The parameters of the graph neural network model are iteratively updated using the backpropagation algorithm based on the gradient of the joint loss function until the validation set loss meets the preset convergence condition, thereby obtaining the pre-trained graph neural network model.

[0037] Furthermore, after step S6, the following steps are also included:

[0038] S81. The stability degradation index and its associated three-dimensional coordinate data and observation time are time-series encoded and entered into the hidden danger feature tracking database;

[0039] S82. Using the differential autoregressive moving average model, based on the time series of stability degradation indicators in the hidden danger feature tracking database, predict the value of the stability degradation indicator for the next observation period.

[0040] S83. When the predicted value of the stability degradation index exceeds the preset warning threshold, output evacuation warning information including the location of the hidden danger and the warning level.

[0041] Another objective of this invention is to provide a geological hazard hazard verification system based on multi-source remote sensing image fusion, the system comprising:

[0042] The multimodal data access engine is configured to access meteorological observation data sources, terrain data sources, basic geographic information vector data sources, and radar satellite data sources to acquire meteorological precipitation grid sequence data, digital elevation model data, basic geographic information vector data, and synthetic aperture radar complex image data during the flood season.

[0043] The hydrological pre-constraint engine is connected in communication with the multimodal data access engine. It is used to obtain the slope and catchment area based on the digital elevation model data, and calculate the runoff hazard by combining the precipitation intensity and infiltration rate, and delineate the high-risk candidate areas for hydrological disasters.

[0044] A microwave physical inversion device is used to preprocess the synthetic aperture radar complex image data during the flood season in the candidate area of ​​high risk of hydrological disasters, extract the microwave backscattering coefficient, invert the volume water content parameter, and generate a soil moisture high saturation object.

[0045] A spatial topology analysis and graph computation engine includes a mask isolator and a graph neural network. The mask isolator is used to generate a non-rigid mask based on the basic geographic information vector data and reset the volumetric water content parameter within the coverage area of ​​the non-rigid mask to the background baseline value. The graph neural network is used to construct a directed adjacency matrix based on the rigid ground feature anchor points and the soil moisture highly saturated objects, and output the spatial permeability deterioration coefficient.

[0046] The mechanical disaster assessment processor, which is connected in communication with the spatial topology analysis and graph calculation engine, is used to calculate the foundation hydrostatic pressure and equivalent pore water pressure based on the elevation difference between the soil moisture highly saturated object and the rigid ground anchor point, and output the stability degradation index using the Mohr-Coulomb failure criterion.

[0047] Furthermore, the spatial topology analysis and graph computation engine also includes a node feature encoder and a tensor decoder, wherein:

[0048] The node feature encoder includes a centroid extraction module and a feature alignment module. The centroid extraction module is used to extract the three-dimensional spatial centroid coordinates and the average volumetric water content of the soil moisture-saturated object, as well as the three-dimensional geometric center coordinates of the rigid ground feature anchor point. The feature alignment module is used to convert the source node features and the target node features into node feature vectors with the same dimension.

[0049] The tensor decoder is located at the end of the graph neural network and includes a fully connected layer and a sigmoid activation function, used to map the node aggregation features output by the graph convolutional layer to the spatial penetration deterioration coefficient with a value range of 0 to 1.

[0050] The beneficial effects of this invention are as follows:

[0051] 1. This invention uses precipitation data and digital elevation model data to define high-risk candidate areas for hydrological disasters, and processes synthetic aperture radar complex image data during the flood season within the candidate areas, which can reduce the amount of radar data processing across the entire area and improve the targeting of verification.

[0052] 2. This invention extracts artificial structures and vegetation cover areas from basic geographic information vector data, generating rigid feature anchors and non-rigid masks, which can suppress environmental noise caused by vegetation canopy swaying and changes in canopy water content, and reduce the interference of abnormal echoes in vegetation areas on the verification results.

[0053] 3. This invention converts objects with high soil moisture saturation and rigid ground anchor points into a directed graph structure, uses a graph neural network to obtain the spatial permeability deterioration coefficient, and combines this coefficient with the foundation hydrostatic pressure calculated from the elevation difference to obtain the equivalent pore water pressure that can be input into the geotechnical mechanics model, thereby outputting a stability degradation index with clear physical meaning. Attached Figure Description

[0054] Figure 1 This is a flowchart illustrating the geological hazard hazard verification method based on multi-source remote sensing image fusion provided in an embodiment of the present invention.

[0055] Figure 2 This is a block diagram of the architecture modules of the geological hazard hazard verification system based on multi-source remote sensing image fusion provided in an embodiment of the present invention;

[0056] Figure 3 The following is a scenario profile and system execution diagram for a specific application example provided in the embodiments of the present invention. Detailed Implementation

[0057] The present invention will be further described below with reference to the accompanying drawings and embodiments. Components with the same reference numerals have the same meaning. The embodiments are used to illustrate the technical solutions of the present invention and do not constitute a limitation on the scope of protection.

[0058] like Figure 1 As shown, this application provides a method for verifying geological hazard risks based on multi-source remote sensing image fusion, including the following steps:

[0059] Step S1: Obtain meteorological precipitation grid sequence data, digital elevation model data, basic geographic information vector data, and synthetic aperture radar complex image data of the flood season with the imaging time closest to the last moment of the meteorological precipitation grid sequence data for the target verification area.

[0060] Meteorological precipitation grid sequence data is used to characterize the spatial precipitation intensity over time intervals. Digital elevation model (DEM) data is used to extract slope, elevation, and catchment area. Basic geographic information vector data includes at least a layer of man-made structures and a layer of vegetation cover. Synthetic aperture radar (SAR) composite image data from the flood season retains amplitude and phase information. All data are uniformly transformed to the same projected coordinate system and resampled and registered according to the spatial resolution of the DEM data. The time difference between the radar imaging time and the last moment of the precipitation sequence should not exceed the preset matching duration; if no data meets the conditions, this verification will not be performed.

[0061] In one embodiment, the system acquires hourly precipitation grid data of the target slope over the past 24 hours, digital elevation model data with a spatial resolution of 10 meters, basic geographic information vector data including retaining structures and vegetation cover areas, and acquires C-band synthetic aperture radar complex image data of the flood season with the imaging time closest to the end time of the precipitation sequence.

[0062] Step S2: The meteorological precipitation grid sequence data and digital elevation model data are fused with topographic features to calculate and generate high-risk candidate areas for hydrological disasters.

[0063] The system divides slopes into units based on digital elevation model data, calculates the slope and catchment area of ​​each unit, and then maps the precipitation intensity and infiltration rate at the same time and spatial location to the slope units to obtain runoff hazard. This process limits subsequent radar inversion to areas with higher runoff hazard.

[0064] In one embodiment, the natural gullies in the upper part of the slope have a large catchment area, and the rainfall intensity during the corresponding period exceeds the soil infiltration rate. The system classifies the gullies and their downstream impact areas as high-risk candidate areas for hydrological disasters based on runoff hazard thresholds.

[0065] Step S3: Extract the microwave backscattering coefficients from the synthetic aperture radar complex image data during the flood season in the high-risk candidate areas for hydrological disasters, invert the volumetric water content parameters based on the microwave backscattering coefficients, and extract objects with high soil moisture saturation.

[0066] Radar composite image data undergoes sequential orbit correction, thermal noise removal, radiometric calibration, speckle filtering, and terrain correction to obtain geocoded backscattering coefficients. A water cloud model is used to separate the scattering components from the vegetation canopy and bare soil. An empirical polynomial model calibrated by sample points is used to convert the bare soil backscattering coefficients into volumetric water content parameters. Radar inversion results characterize the volumetric water content within the effective detection depth range near the surface.

[0067] In one embodiment, the system obtains consecutive pixels with abnormal volumetric water content within a candidate region and calculates a high saturation threshold based on the porosity measured from the soil sample and a preset saturation coefficient. Consecutive pixel regions reaching this threshold are extracted as objects with high soil moisture saturation.

[0068] Step S4: Extract rigid feature anchor points and vegetation changeable areas using basic geographic information vector data. Generate a non-rigid mask based on the vegetation changeable areas. Reset the volumetric water content parameters within the coverage area of ​​the non-rigid mask to the background baseline value. Update the soil moisture high saturation objects according to the reset volumetric water content parameters to remove vegetation environmental noise features.

[0069] In the man-made structure layer, geometric polygons such as retaining walls, paved roads, and building foundations are rasterized to form rigid feature anchors. Geometric polygons such as trees, shrubs, and high-coverage grasslands in the vegetation layer form vegetation variability zones. Morphological dilation is performed on these vegetation variability zones, and the rigid feature anchors and their buffers are subtracted to obtain a non-rigid mask. This generation logic avoids classifying all areas outside the rigid feature anchors as vegetation variability zones.

[0070] In one embodiment, the system identifies the retaining wall as a rigid feature anchor point and the surrounding woodland and shrub cover area as a vegetation-vulnerable zone. After morphological dilation of one pixel and subtraction of the retaining wall buffer, a non-rigid mask is generated, and the volumetric water content parameter within the mask area is assigned a background baseline value; the system then re-extracts objects with high soil moisture saturation based on this, and the mask pixels do not participate in subsequent map calculations.

[0071] Step S5: Extract the spatial features of the soil moisture-saturated objects updated in step S4 as source nodes, extract the geometric features of the rigid ground anchor points as target nodes, construct a directed adjacency matrix based on the source nodes and target nodes, input the directed adjacency matrix and node feature matrix into the pre-trained graph neural network model, and output the spatial permeability deterioration coefficient.

[0072] Each source node corresponds to a denoised, highly saturated soil moisture object, and each target node corresponds to a rigid feature anchor point. Source and target nodes share the same feature dimension, and source nodes with higher elevations can point to target nodes with lower elevations. Edge connection weights decrease as the Euclidean distance in 3D space increases. The graph neural network aggregates features from adjacent nodes through graph convolutional layers and outputs a spatial permeability degradation coefficient in the range of 0 to 1 through fully connected layers and a sigmoid activation function.

[0073] In one embodiment, the system constructs the soil moisture-saturated object at the gully as the source node and the retaining wall below as the target node. A directed edge is established based on the three-dimensional spatial distance and elevation relationship between the two. The graph neural network outputs a spatial permeability degradation coefficient of 0.82, indicating that the current node combination corresponds to a high degree of permeability degradation.

[0074] Step S6: Calculate the foundation hydrostatic pressure based on the elevation difference between the source node and the target node. Multiply the foundation hydrostatic pressure by the spatial permeability deterioration coefficient to obtain the equivalent pore water pressure. Input the equivalent pore water pressure into the preset mechanical evaluation model to generate the stability degradation index and output the hidden danger verification results.

[0075] The foundation hydrostatic pressure is calculated as the product of the unit weight of water and the positive elevation difference. The spatial permeability deterioration coefficient is a dimensionless coefficient, which is multiplied by the foundation hydrostatic pressure to obtain the equivalent pore water pressure. The equivalent pore water pressure is substituted into the Mohr-Coulomb failure criterion to calculate the current shear strength of the soil and rock mass adjacent to the target node. The stability degradation index represents the percentage reduction in current shear strength relative to the reference shear strength.

[0076] In one embodiment, the system calculates the foundation hydrostatic pressure based on the elevation difference between the soil moisture-saturated object and the retaining wall, then calculates the equivalent pore water pressure based on the spatial permeability deterioration coefficient, and combines the effective cohesion, effective internal friction angle and total normal stress from the field survey or design data to calculate the stability degradation index.

[0077] Furthermore, the specific calculation process of step S2 includes:

[0078] Step S21: Divide the slope into units based on the digital elevation model data and calculate the slope of each unit. and catchment area The slope is represented by the maximum rate of change of elevation or the average slope within the slope unit, and the catchment area is represented by the sum of the grid areas upstream that flow into the slope unit.

[0079] In one embodiment, the system calculates that the average slope of the gully slope unit is 35 degrees and the upstream catchment area is approximately 20,000 square meters.

[0080] Step S22, according to Calculate the runoff hazard of the i-th slope element. . The precipitation intensity corresponding to the slope unit and the observation time. The infiltration rate is determined based on soil type or field infiltration tests, where α is the scale calibration coefficient, and m and n are exponents. The max function is used to limit the effective precipitation term to 0 when the precipitation intensity is not greater than the infiltration rate.

[0081] Runoff hazard is used for candidate area screening and is not directly equivalent to surface runoff with fixed physical dimensions. α, m, and n are calibrated based on historical precipitation, water accumulation, or instability records of the target area. In one embodiment, historical samples are divided into calibration and validation sets by time, and the parameter combination is determined by a comprehensive index of candidate area recall and false alarm rate.

[0082] Step S23: Extract slope units with runoff hazard greater than the preset runoff hazard threshold, and merge spatially adjacent slope units into high-risk candidate areas for hydrological disasters.

[0083] The preset runoff hazard threshold is determined based on the subject operating characteristic curves of historical samples or engineering experience, and the calibration results of the same dimensions are used in different seasons. In one embodiment, the runoff hazard of a gully slope unit exceeds the threshold, and therefore it is included in the subsequent radar inversion range.

[0084] Furthermore, the specific inversion process of step S3 includes:

[0085] Step S31: After preprocessing the radar complex image data, separate the backscattering components of vegetation and bare soil according to the water cloud model formula. The calculation formula is as follows: ,in, To observe the backscattering coefficient, The vegetation canopy scattering component, For bare soil scattering components, This represents the two-way transmittance of vegetation.

[0086] Vegetation two-way transmittance The calculation formula is: ,in, The vegetation attenuation coefficient is determined by sampling points. This is a vegetation descriptive quantity (calculated from vegetation index or vegetation water content). This is the radar incident angle.

[0087] In one embodiment, the system calculates vegetation descriptive quantities using vegetation indices matched to the radar imaging time period. And calibrate the vegetation attenuation coefficient based on the sample points in the target area. The backscattering matrix of bare soil is obtained by separating it from the observed backscattering coefficients.

[0088] Step S32: Input the bare soil backscattering matrix into the empirical polynomial model to solve for the soil volumetric water content parameter. The calculation formula of the empirical polynomial model is as follows: ,in, This is the volumetric water content parameter. For surface roughness, These are the undetermined coefficients for the model. The specific method for determining the model coefficients is as follows: Soil samples with different water content states in the target area are collected. The measured volumetric water content of the samples is paired with the backscattering coefficient, radar incident angle, and surface roughness data of the bare soil at the same location and time. The model coefficients are obtained by fitting the samples using the least squares method.

[0089] Step S33: Set the high saturation judgment threshold Extract volumetric water content A continuous pixel region is used as an object with high soil moisture saturation. Soil porosity is obtained through the ring sampler method or soil bulk density data; The preset saturation coefficient ranges from 0.90 to 1.00. If the area of ​​the extracted continuous pixel region is smaller than the preset minimum area threshold, it is determined to be discrete noise and removed.

[0090] In one embodiment, porosity was measured from a field soil sample. The preset saturation coefficient β is 0.95, and the high saturation threshold is 0.418 cubic meters per cubic meter. A continuous pixel region with a retrieved volumetric water content of 0.42 cubic meters per cubic meter reaches this threshold and is therefore extracted as a high-saturation soil moisture object. Porosity is used to characterize the proportion of pore volume to total volume, and the high saturation threshold is determined accordingly.

[0091] Furthermore, the specific masking process in step S4 includes:

[0092] Step S41: Extract the geometric polygons of man-made structures and vegetation cover areas according to the attribute fields of the basic geographic information vector data. The geometric polygons of man-made structures are rasterized to form rigid feature anchor points, and the geometric polygons of vegetation cover areas are rasterized to form vegetation change zones.

[0093] Step S42: Perform mathematical morphological dilation on the vegetation variable area to obtain the vegetation influence extension area, and then subtract the rigid ground feature anchor points and their buffer zones from the vegetation influence extension area to generate a non-rigid mask.

[0094] Morphological dilation uses circular or rectangular structural elements, and the dilation radius is determined based on the registration error between the radar image and the underlying geographic information vector data. In one embodiment, if the registration error is no greater than one pixel, a one-pixel dilation radius is used; the width of the rigid feature anchor point buffer is one pixel.

[0095] Step S43: Reproject the non-rigid mask onto the volumetric water content parameter space and resample it to the same resolution using the nearest neighbor method. For pixels with a non-rigid mask value of 1, set the volumetric water content parameter to the background baseline value; for pixels with a mask value of 0, retain the original volumetric water content parameter. After resetting, extract the soil moisture highly saturated objects again based on the high saturation judgment threshold and the minimum area threshold.

[0096] Background reference values ​​are used to indicate masked pixels that are not involved in subsequent calculations and can be set to 0 or have no data marker. In one embodiment, the system sets the background reference value to 0 and excludes such pixels when extracting objects with high soil moisture saturation. This mask only suppresses vegetation-vulnerable areas determined by the underlying geographic information vector data and does not delete natural slope data not covered by the vegetation mask.

[0097] Furthermore, the specific graph construction and reasoning process in step S5 includes:

[0098] Step S51: Calculate the three-dimensional centroid coordinates for each soil moisture-saturated object. and the average internal volumetric water content Calculate the three-dimensional geometric center coordinates for each rigid ground feature anchor point. The source node feature vector is set to

[0099] The target node feature vector is set to The last digit indicates the node type. Each feature dimension is standardized using the mean and standard deviation of the training set.

[0100] Step S52: Calculate the source node With the target node The three-dimensional Euclidean distance between them When the source node elevation greater than the target node elevation At that time, establish from the source node Point to target node The directed edge.

[0101] The formula for calculating the connection weight of a directed edge is as follows: ,in, This is the distance scale parameter. When no directed edge is established. Then, self-connecting edges with a weight of 1 are added to each node. The directed adjacency matrix is ​​then normalized. Distance scale parameter. It is determined by the median distance between the source node and the corresponding unstable target node in the training samples.

[0102] Step S53: Input the node feature matrix and the normalized directed adjacency matrix into the graph neural network model.

[0103] In one embodiment, the graph neural network model sequentially includes a first graph convolutional layer with an input dimension of 5 and an output dimension of 16, a ReLU activation layer, a second graph convolutional layer with an output dimension of 8, a graph-level average pooling layer, a fully connected layer, and a sigmoid activation function. The sigmoid activation function outputs a spatial permeability degradation coefficient in the range of 0 to 1. The aforementioned layers, connections, input dimensions, and output dimensions allow the model to be constructed using publicly available data structures.

[0104] Furthermore, the specific mechanical evaluation process in step S6 includes:

[0105] Step S61: Calculate the elevation difference between the source node and the target node. and in accordance with Calculate the hydrostatic pressure at the base. Specific weight of water. Take 9.81 kN per cubic meter.

[0106] Step S62, according to Calculate the equivalent pore water pressure and substitute it into the Mohr-Coulomb failure criterion. Space permeability degradation coefficient. The dimensionless number represents the basic hydrostatic pressure. Both the equivalent pore water pressure u and the unit of pressure are kilopascals (kPa). The Mohr-Coulomb failure criterion is expressed as... .in, c is the current shear strength, and c' is the effective cohesion. The total normal stress of the soil and rock mass adjacent to the target node. This is the effective internal friction angle. c' and It is obtained from on-site survey reports, design data, or indoor geotechnical tests. The function is used to avoid calculating negative effective normal stress. In one embodiment, the elevation difference between the source node and the target node is 30 meters, and the foundation hydrostatic pressure is... The spatial permeability deterioration coefficient is 294.3 kPa. The value is 0.82, and the equivalent pore water pressure u is 241.3 kPa.

[0107] Step S63, according to Calculate the baseline shear strength without considering the equivalent pore water pressure, and according to... Calculate the stability degradation index D. The larger the value of D, the more significant the decrease in shear strength relative to the reference state.

[0108] In one embodiment, the effective cohesion c' is 20 kPa, and the total normal stress is... The effective internal friction angle is 500 kPa. The angle is 30 degrees. Reference shear strength. Approximately 308.7 kPa, current shear strength The pressure is approximately 169.4 kPa, and the stability degradation index D is approximately 45.1%.

[0109] Furthermore, prior to step S5, a training process for the graph neural network model is included, comprising steps S71 to S73.

[0110] Step S71: Obtain historical heavy rainfall sequences, radar inversion state matrices, and corresponding actual instability records, and construct a training sample set using the same data processing methods as steps S1 to S5.

[0111] The training sample set includes both positive samples that experienced actual instability and negative samples that did not experience actual instability within the same observation window. Positive samples are labeled with 1, and negative samples are labeled with 0. The samples are divided into training, validation, and test sets in chronological order to avoid data from the same rainfall event being included in different datasets simultaneously.

[0112] Step S72: Standardize the node features, calculate the predicted instability probability in the forward propagation of the graph neural network model, and construct the joint loss function.

[0113] The joint loss function is expressed as ,in, For sample labels, The spatial permeability degradation coefficient is the output of the model, and W represents the weights to be trained. This is the regularization coefficient. Therefore, the model training objective and the continuous coefficients output during the inference phase remain consistent.

[0114] Step S73: Update the model parameters using the backpropagation algorithm and the adaptive moment estimation optimizer, and determine convergence based on the validation set loss.

[0115] In one embodiment, the ratio of training set, validation set, and test set is 7:2:1, the batch size is 32, the initial learning rate is 0.001, the regularization coefficient λ is 0.0001, and the maximum number of training epochs is 500. Training stops when the validation set loss does not decrease for 20 consecutive epochs, and the model parameters at which the validation set loss is lowest are saved.

[0116] The time-series tracking and early warning process includes steps S81 to S83. Step S81: The stability degradation index, three-dimensional coordinates, and observation time are combined to form a time-series record, which is then entered into the hidden danger feature tracking database.

[0117] The database uses rigid feature anchor points as the main index and stores stability degradation indicators in ascending order of observation time, so that multiple verification results of the same target node form a time series with equal time intervals or after interpolation.

[0118] Step S82: Using the differential autoregressive moving average model, predict the value of the stability degradation index for the next observation period based on the time series of the stability degradation index.

[0119] The differencing order is determined based on the stationarity test of the sequence, while the autoregressive and moving average orders are determined based on the Akaike Information Criterion. In one embodiment, the model predicts the stability degradation index for the next two-hour observation period based on continuous observation results.

[0120] Step S83: Compare the predicted value of the stability degradation index with the preset warning threshold. When the predicted value exceeds the preset warning threshold, output evacuation warning information including the three-dimensional coordinates of the target node, the predicted value of the stability degradation index, and the warning level.

[0121] The warning threshold is determined based on relevant engineering safety standards, design parameters, or management department contingency plans. The system outputs warning information for management personnel to review and handle, but does not replace on-site verification and engineering judgment with algorithm results.

[0122] like Figure 2 As shown, this application provides a geological hazard hazard verification system 100 based on multi-source remote sensing image fusion, including a multimodal data access engine 101. The multimodal data access engine 101 is used to access meteorological observation data sources, topographic data sources, basic geographic information vector data sources, and radar satellite data sources to acquire meteorological precipitation grid sequence data, digital elevation model data, basic geographic information vector data, and synthetic aperture radar complex image data during the flood season, and to complete coordinate system I, time matching, and spatial registration.

[0123] The hydrological pre-constraint engine 102 is connected to the multimodal data access engine 101 to divide slope units based on digital elevation model data, calculate slope, catchment area and runoff hazard, and delineate high-risk candidate areas for hydrological disasters.

[0124] The microwave physical inversion unit 103 is used to perform radiometric calibration, speckle filtering and terrain correction on radar complex image data within the candidate area. It uses a water cloud model to obtain the backscattering matrix of bare soil, uses an empirical polynomial model calibrated by soil samples to obtain volumetric water content parameters, and extracts soil moisture saturation objects.

[0125] The spatial topology analysis and graph computation engine 104 includes a mask isolator 1041 and a graph neural network 1042. The mask isolator 1041 generates a non-rigid mask and resets the volumetric water content parameters of vegetation-vulnerable zones. The graph neural network 1042 constructs a directed adjacency matrix and a node feature matrix based on denoised soil moisture-saturated objects and rigid feature anchors, and outputs a spatial permeability deterioration coefficient. The graph neural network 1042 includes a node feature encoder, two graph convolutional layers, a graph-level average pooling layer, and a tensor decoder.

[0126] The mechanical catastrophic assessment processor 105 is communicatively connected to the spatial topology analysis and graph calculation engine 104. It is used to calculate the foundation hydrostatic pressure and equivalent pore water pressure, and to calculate the current shear strength and stability degradation index using the Mohr-Coulomb failure criterion. The multimodal data access engine 101, the hydrological pre-constraint engine 102, the microwave physical inversion engine 103, the spatial topology analysis and graph calculation engine 104, and the mechanical catastrophic assessment processor 105 can be implemented by the processor executing program instructions stored in memory.

[0127] Specific application example: Combining Figure 2 and Figure 3 As shown, the geological hazard verification of a mountain slope during the flood season is taken as an example. During heavy rainfall, the multimodal data access engine 101 of the intelligent identification and verification system 100 acquires hourly precipitation grid data, digital elevation model data, basic geographic information vector data including retaining walls and vegetation coverage areas, and C-band flood season synthetic aperture radar complex image data.

[0128] The hydrological pre-constraint engine 102 delineates the gully and its downstream impact zone as high-risk candidate areas for hydrological disasters based on the gully slope, catchment area, rainfall intensity, and infiltration rate. The microwave physical inversion device 103 inverts the volumetric water content parameters within the candidate areas. The field soil sample porosity is 0.44, the saturation coefficient is 0.95, and the high saturation threshold is 0.418 cubic meters per cubic meter; a continuous area with an inverted volumetric water content of 0.42 cubic meters per cubic meter is extracted as a high-saturation soil moisture object.

[0129] The mask isolator 1041 extracts the retaining wall as a rigid feature anchor point and the woodland and shrub cover areas as vegetation-vulnerable areas. After morphological dilation and rigid feature buffer subtraction, a non-rigid mask is generated. The volumetric water content parameter within the mask coverage area is set as the background reference value, thereby reducing the impact of vegetation environmental noise on the extraction results of highly saturated objects.

[0130] The spatial topology analysis and graph computing engine 104 constructs the three-dimensional spatial centroid coordinates and average volumetric water content of the soil moisture-saturated object as source nodes, and constructs the three-dimensional geometric center coordinates of the retaining wall as target nodes. After establishing directed edges based on the three-dimensional spatial distance and elevation relationship, the graph neural network 1042 outputs a spatial permeability deterioration coefficient of 0.82.

[0131] The mechanical disaster assessment processor 105 calculates the foundation hydrostatic pressure as 294.3 kPa based on a 30-meter elevation difference and the equivalent pore water pressure as 241.3 kPa based on the spatial permeability deterioration coefficient. Under the conditions of an effective cohesion of 20 kPa, a total normal stress of 500 kPa, and an effective internal friction angle of 30 degrees, the calculated stability degradation index is approximately 45.1%. The system outputs the hazard verification results, including the target node coordinates and the stability degradation index, and outputs corresponding warning information when the index or its predicted value exceeds the preset warning threshold.

[0132] The above embodiments illustrate the technical solution of the present invention. Equivalent substitutions or modifications made by those skilled in the art without departing from the concept of the present invention should all fall within the protection scope of the present invention.

Claims

1. A method for verifying geological hazard risks based on multi-source remote sensing image fusion, characterized in that, Includes the following steps: S1. Obtain meteorological precipitation grid sequence data, digital elevation model data, basic geographic information vector data of the target verification area, and synthetic aperture radar complex image data of the flood season whose imaging time is closest to the last time of the meteorological precipitation grid sequence data. S2. The meteorological precipitation grid sequence data and the digital elevation model data are fused with topographic features to calculate and generate high-risk candidate areas for hydrological disasters. S3. Extract the microwave backscattering coefficient of the synthetic aperture radar complex image data during the flood season in the candidate area of ​​high risk of hydrological disasters, invert the volume water content parameter based on the microwave backscattering coefficient, and extract the soil moisture saturation objects. S4. Using the basic geographic information vector data, extract rigid feature anchor points and vegetation changeable areas, generate a non-rigid mask based on the vegetation changeable areas, reset the volumetric water content parameter within the coverage area of ​​the non-rigid mask to the background reference value, and update the soil moisture high saturation object according to the reset volumetric water content parameter to remove vegetation environmental noise features. S5. Extract the spatial features of the soil moisture-saturated object updated in step S4 as the source node, extract the geometric features of the rigid ground anchor point as the target node, construct a directed adjacency matrix based on the source node and the target node, input the directed adjacency matrix and the node feature matrix into a pre-trained graph neural network model, and output the spatial permeability deterioration coefficient. S6. Calculate the foundation hydrostatic pressure based on the elevation difference between the source node and the target node, multiply the foundation hydrostatic pressure by the spatial permeability deterioration coefficient to obtain the equivalent pore water pressure, input the equivalent pore water pressure into a preset mechanical evaluation model, generate a stability degradation index, and output the hidden danger verification results.

2. The geological hazard hazard verification method based on multi-source remote sensing image fusion according to claim 1, characterized in that, Step S2 specifically includes: S21. Divide the slope into units based on the digital elevation model data, and extract the slope matrix and catchment area matrix of each slope unit. S22, according to the formula Calculate the runoff hazard of each slope unit; where, Let i be the runoff hazard level of the i-th slope unit. For the catchment area, For slope, For precipitation intensity, α is the infiltration rate, m and n are the scale calibration coefficients, and m and n are the exponents calibrated from historical samples. S23. Extract the slope units with runoff hazard levels greater than the preset runoff hazard threshold as the high-risk candidate areas for hydrological disasters.

3. The geological hazard hazard verification method based on multi-source remote sensing image fusion according to claim 1, characterized in that, Step S3 specifically includes: S31. Perform radiometric calibration, speckle filtering and terrain correction on the synthetic aperture radar complex image data during the flood season to obtain the microwave backscattering coefficient, and use the water cloud model to separate the vegetation canopy component in the microwave backscattering coefficient to obtain the bare soil backscattering matrix. S32. Input the bare soil backscattering matrix, radar incident angle and surface roughness into the empirical multinomial model calibrated by the soil moisture sampling points of the target verification area, and solve for the volumetric water content parameters corresponding to each pixel. S33. The product of the measured porosity of the target verification area and the preset saturation coefficient is used as the high saturation judgment threshold. Continuous pixel areas with volumetric water content parameters not less than the high saturation judgment threshold are extracted as the soil moisture high saturation objects.

4. The geological hazard hazard verification method based on multi-source remote sensing image fusion according to claim 1, characterized in that, Step S4 specifically includes: S41. Parse the basic geographic information vector data, convert the geometric polygons with the attribute of artificial structures into the rigid ground feature anchor points, and extract the geometric polygons with the attribute of vegetation coverage area as the vegetation change zone. S42. The vegetation change zone is rasterized, and a mathematical morphological dilation operation is performed on the rasterization result. After deducting the rigid ground feature anchor points and their buffer zones, the non-rigid mask is generated. S43. Perform coordinate registration and resolution alignment between the non-rigid mask and the spatial distribution matrix of the volumetric water content parameter, reset the volumetric water content parameter within the coverage area of ​​the non-rigid mask to the background reference value, and re-extract the soil moisture high saturation object based on the reset volumetric water content parameter.

5. The geological hazard hazard verification method based on multi-source remote sensing image fusion according to claim 1, characterized in that, Step S5 specifically includes: S51. Extract the three-dimensional spatial centroid coordinates and the mean internal volume water content of the soil moisture-saturated object to construct the source node, extract the three-dimensional geometric center coordinates of the rigid ground feature anchor point to construct the target node, and dimensionally align the source node features and target node features to obtain the node feature matrix. S52. When the elevation of the source node is greater than the elevation of the target node, a directed edge is established from the source node to the target node, and the edge connection weight is calculated by using a Gaussian kernel function based on the spatial Euclidean distance between the two to construct the directed adjacency matrix. S53. Aggregate node features through at least two graph convolutional layers of the graph neural network model, and output the spatial penetration deterioration coefficient with a value range of 0 to 1 through the fully connected layer and Sigmoid activation function at the end of the network.

6. The geological hazard hazard verification method based on multi-source remote sensing image fusion according to claim 1, characterized in that, Step S6 specifically includes: S61, according to Calculate the foundation hydrostatic pressure; wherein, The aforementioned basic hydrostatic pressure, The density of water, The elevation difference between the source node and the target node; S62, according to Calculate the equivalent pore water pressure, and according to... Calculate the current shear strength; where, The equivalent pore water pressure, The space permeability deterioration coefficient is mentioned above. The current shear strength, For effective cohesion, For the total normal stress, The effective internal friction angle; S63, according to Calculate the baseline shear strength, and according to Calculate the stability degradation index; wherein, The reference shear strength is... This refers to the stability degradation metric.

7. The geological hazard hazard verification method based on multi-source remote sensing image fusion according to claim 1, characterized in that, Before step S5, the following is also included: S71. Obtain historical heavy rainfall sequences, radar inversion state matrices, and corresponding actual instability records, and construct a training sample set containing positive samples that have experienced instability and negative samples that have not experienced instability. S72. Standardize the node features in the training sample set, calculate the predicted instability probability in the forward propagation of the graph neural network model, and construct a joint loss function containing a binary cross-entropy term and a regularization penalty term based on the sample labels. S73. The parameters of the graph neural network model are iteratively updated using the backpropagation algorithm based on the gradient of the joint loss function until the validation set loss meets the preset convergence condition, thereby obtaining the pre-trained graph neural network model.

8. The method for verifying geological hazard risks based on multi-source remote sensing image fusion according to claim 1, characterized in that, Following step S6, the following is also included: S81. The stability degradation index and its associated three-dimensional coordinate data and observation time are time-series encoded and entered into the hidden danger feature tracking database; S82. Using the differential autoregressive moving average model, based on the time series of stability degradation indicators in the hidden danger feature tracking database, predict the value of the stability degradation indicator for the next observation period. S83. When the predicted value of the stability degradation index exceeds the preset warning threshold, output evacuation warning information including the location of the hidden danger and the warning level.

9. A geological hazard hazard verification system based on multi-source remote sensing image fusion, used to implement the geological hazard hazard verification method based on multi-source remote sensing image fusion as described in any one of claims 1 to 8, characterized in that, The system includes: The multimodal data access engine is configured to access meteorological observation data sources, terrain data sources, basic geographic information vector data sources, and radar satellite data sources to acquire meteorological precipitation grid sequence data, digital elevation model data, basic geographic information vector data, and synthetic aperture radar complex image data during the flood season. The hydrological pre-constraint engine is connected in communication with the multimodal data access engine. It is used to obtain the slope and catchment area based on the digital elevation model data, and calculate the runoff hazard by combining the precipitation intensity and infiltration rate, and delineate the high-risk candidate areas for hydrological disasters. A microwave physical inversion device is used to preprocess the synthetic aperture radar complex image data during the flood season in the candidate area of ​​high risk of hydrological disasters, extract the microwave backscattering coefficient, invert the volume water content parameter, and generate a soil moisture high saturation object. A spatial topology analysis and graph computation engine includes a mask isolator and a graph neural network. The mask isolator is used to generate a non-rigid mask based on the basic geographic information vector data and reset the volumetric water content parameter within the coverage area of ​​the non-rigid mask to the background baseline value. The graph neural network is used to construct a directed adjacency matrix based on the rigid ground feature anchor points and the soil moisture highly saturated objects, and output the spatial permeability deterioration coefficient. The mechanical disaster assessment processor, which is connected in communication with the spatial topology analysis and graph calculation engine, is used to calculate the foundation hydrostatic pressure and equivalent pore water pressure based on the elevation difference between the soil moisture highly saturated object and the rigid ground anchor point, and output the stability degradation index using the Mohr-Coulomb failure criterion.

10. The geological hazard hazard verification system based on multi-source remote sensing image fusion according to claim 9, characterized in that, The spatial topology analysis and graph computation engine also includes a node feature encoder and a tensor decoder, wherein: The node feature encoder includes a centroid extraction module and a feature alignment module. The centroid extraction module is used to extract the three-dimensional spatial centroid coordinates and the average volumetric water content of the soil moisture-saturated object, as well as the three-dimensional geometric center coordinates of the rigid ground feature anchor point. The feature alignment module is used to convert the source node features and the target node features into node feature vectors with the same dimension. The tensor decoder is located at the end of the graph neural network and includes a fully connected layer and a sigmoid activation function, used to map the node aggregation features output by the graph convolutional layer to the spatial penetration deterioration coefficient with a value range of 0 to 1.

Images