Geological disaster hidden danger identification method based on InSAR and semi-supervised learning

Abstract

The invention relates to the technical field of geological disaster hidden danger recognition, in particular to a geological disaster hidden danger recognition method based on InSAR and semi-supervised learning, which comprises the steps of dividing hydrological sensitivity of a monitoring area according to a regional underground water unit supply schematic diagram, and setting an InSAR data sampling frequency; acquiring earth surface time sequence displacement data, including accumulated displacement, displacement rate and displacement acceleration, by using a set sampling frequency for preliminarily identifying a potential hidden danger area; geological radar detection is carried out in the potential hidden danger area, underground structure features are evaluated through a geological radar profile map, and a multi-source feature vector is constructed; according to the method, the hydrological sensitivity of the monitoring area is divided according to the regional groundwater unit supply schematic diagram, and the InSAR data sampling frequency is set, so that the problem that the high-risk area is not monitored in time due to the fact that the hydrological sensitivity change is not considered because the conventional monitoring frequency mostly adopts a fixed sampling interval is solved.

Description

Technical Field

[0001] This invention relates to the field of geological hazard identification technology, and in particular to a geological hazard identification method based on InSAR and semi-supervised learning. Background Technology

[0002] Geological hazard identification is a crucial step in engineering construction, urban planning, and infrastructure operation and maintenance in mountainous areas. Existing monitoring technologies mainly include leveling, total station monitoring, surface crack inspection, ground-penetrating radar detection, and satellite remote sensing technology. Among these, InSAR technology is widely used because it can monitor millimeter-level surface deformation over large areas.

[0003] Traditional monitoring frequencies mostly use fixed sampling intervals, which, because they do not take into account changes in hydrological sensitivity, result in untimely monitoring of high-risk areas. Summary of the Invention

[0004] To overcome the above shortcomings, this invention provides a geological hazard hazard identification method based on InSAR and semi-supervised learning, aiming to improve the problem that traditional monitoring frequencies mostly adopt fixed sampling intervals, which lead to untimely monitoring of high-risk areas due to the failure to consider changes in hydrological sensitivity.

[0005] In a first aspect, the present invention provides the following technical solution: a method for identifying geological hazard risks based on InSAR and semi-supervised learning, comprising: Based on the regional groundwater unit recharge diagram, the hydrological sensitivity of the monitoring area is divided, and the InSAR data sampling frequency is set. The surface time-series displacement data, including cumulative displacement, displacement rate, and displacement acceleration, is obtained by using a set sampling frequency to preliminarily identify potential hazard areas. Ground-penetrating radar (GPR) detection is conducted in potential hazard areas. The underground structural characteristics are assessed using GPR profile maps, and multi-source feature vectors are constructed. Using total stations deployed at high-risk points, local precise displacement data, including cumulative displacement, displacement rate, and displacement acceleration, are obtained and used as labeled samples to input semi-supervised learning models. The surface time-series displacement data, GPR structural features, total station labeled data and environmental factors are input into the semi-supervised learning model to train the model and predict the probability of hidden dangers or risk level in unlabeled areas. The model is updated regularly based on new monitoring data, and a risk distribution map is output to achieve dynamic identification and early warning of potential geological hazards. At the same time, the layout of monitoring points and the data collection cycle can be adjusted.

[0006] By adopting the above technical solution, the hydrological sensitivity of the monitoring area is divided by the regional groundwater unit recharge diagram and the InSAR data sampling frequency is set, thereby obtaining the surface time-series displacement data and identifying potential hidden danger areas. This improves the problem that traditional monitoring frequencies mostly use fixed sampling intervals, which do not take into account changes in hydrological sensitivity, resulting in untimely monitoring of high-risk areas.

[0007] Preferably, the hydrological sensitivity assessment of the defined monitoring area includes the following steps: The monitoring area was rasterized to obtain hydrogeological parameters for each monitoring unit, including rainfall. Infiltration coefficient Soil permeability Catchment Index and groundwater level changes ; Normalize each parameter to the 0,1 range to obtain the sub-indices. The formula for calculating the rainfall infiltration index is as follows: Soil permeability index is The watershed index is Groundwater level sensitivity index is ; According to the preset weight vector Calculate the comprehensive hydrological sensitivity score : and ;in , , , The weights of each sub-indicator are determined using the Analytic Hierarchy Process (AHP) and satisfy the consistency ratio CR < 0.1. Based on ratings The range of values ​​is used to divide the monitoring area. For highly sensitive areas, This is a moderately sensitive area. This is a low-sensitivity area.

[0008] Preferably, setting the InSAR data sampling frequency includes the following steps: Hydrological sensitivity scoring ,use Determine the sampling period, where To minimize the sampling period, 37 days is used. The longest sampling period is 3060 days. As a regulating factor, It is also used to control the nonlinear mapping relationship between sensitivity and sampling period; When the InSAR displacement rate of the monitoring unit Exceeding the threshold or displacement acceleration Exceeding the threshold At that time, through Shorten the sampling period, among which For adjustment coefficients, The reference displacement rate is used to achieve adaptive dynamic adjustment of the sampling period.

[0009] Preferably, the step of acquiring the temporal displacement data of the land surface using a set sampling frequency includes the following steps: Multi-temporal synthetic aperture radar images are acquired according to the sampling period, and the images of each temporal phase are registered. Construct a small baseline interferometric network that satisfies the vertical baseline threshold and the time baseline threshold, and generate multiple interferograms; By unwrapping the interferometric phase and correcting for atmospheric and orbital errors, the unwrapped phase vector is obtained. ; Establish a system of interferometric equations ,in Design matrix for the relationship between the interference pair and time. This represents the displacement at each time point relative to the reference time point. To interfere with the quantity, The number of observation times; The time-series shift vector is solved using the weighted least squares method: ;in This is a weight matrix, with weights... Set according to the square of the coherence of the corresponding interference pair; The obtained displacement vectors at each time point Perform quadratic polynomial fitting over time. Obtain displacement rate and displacement acceleration .

[0010] Preferably, the construction of the multi-source feature vector includes the following steps: Extract cumulative displacement, displacement rate, displacement acceleration, and local displacement gradient from InSAR time series results; The mean intensity of reflected waves, fracture density, cavity size, and medium attenuation coefficient were extracted from the ground-penetrating radar profile. Environmental factors to be acquired include slope, aspect, rainfall, and changes in groundwater level; After normalizing the above features, they are concatenated according to the monitoring unit number to form a feature vector.

[0011] Preferably, the process of acquiring precise local displacement data and constructing labeled samples using a total station includes the following steps: Total station reflector prisms are deployed at potentially high-risk locations, and monitoring data is collected at these locations according to a preset monitoring cycle. 3D coordinates ; Calculate relative to the reference time Three-dimensional displacement vector And its modulus length is obtained as the cumulative displacement. ; Displacement rate and acceleration are calculated using the finite difference method: in For adjacent observation periods; The obtained label vector Multi-source feature vectors of corresponding monitoring units Spatial registration is performed to create labeled training sample pairs. , which serves as the supervisory signal input for the semi-supervised learning model.

[0012] Preferably, the semi-supervised learning model is a classification model based on graph convolutional networks, and the input for model construction includes the following steps: An adjacency matrix is ​​constructed by combining feature similarity and spatial distance. And obtained by symmetric normalization ;in For the first Monitoring unit multi-source feature vectors, Geographical distance, These are the tradeoff coefficient and the scaling parameter, respectively. feature matrix and Input an L-layer graph convolutional network and output the predicted probability distribution. .

[0013] Preferably, the training process of the model includes the following steps: To supervise cross-entropy loss Graph smoothing regularization term and pseudo-label loss This constitutes the total loss. ;in For loss weights; After pre-training on the initial labeled set, an iterative pseudo-label strategy is used to add the high-confidence prediction results of unlabeled samples to the training set until convergence, and the probability of hidden danger and risk level are output.

[0014] Preferably, the step of generating a risk distribution map based on the updated model output includes: Obtain the predicted probability value of potential hazards for each monitoring unit. ; The thresholds for low, medium, and high risk classification were calculated using the quantile method. in Represents the set of probabilities of potential hazards of quantiles; Risk classification of monitoring units based on thresholds: when Determined to be low risk, when Determined to be of medium risk, when Determined to be high risk; The classification results are visualized to generate a dynamic risk distribution map, which serves as a basis for decision-making regarding adjustments to the layout of monitoring points and the data collection cycle.

[0015] Secondly, this invention provides the following technical solution: a geological hazard identification system based on InSAR and semi-supervised learning, comprising: The hydrological sensitivity and sampling setting module is used to divide the hydrological sensitivity of the monitoring area according to the regional groundwater unit recharge diagram, and set the InSAR data sampling frequency; The InSAR displacement monitoring and hazard identification module is used to acquire time-series displacement data of the ground surface using a set sampling frequency, including cumulative displacement, displacement rate, and displacement acceleration, for the preliminary identification of potential hazard areas. The radar detection and feature construction module is used to conduct ground-penetrating radar detection in potential hazard areas, assess underground structural features through ground-penetrating radar profile maps, and construct multi-source feature vectors. The total station data acquisition and label generation module is used to acquire local precise displacement data, including cumulative displacement, displacement rate, and displacement acceleration, using total stations deployed at high-risk points. These data are then used as labeled samples to input into a semi-supervised learning model. The semi-supervised learning training and prediction module is used to input surface time-series displacement data, GPR structural features, total station labeled data and environmental factors into the semi-supervised learning model, train the model and predict the probability of hidden dangers or risk levels in unlabeled areas. The model update and dynamic early warning module is used to update the model regularly based on new monitoring data, output risk distribution maps, realize dynamic identification and early warning of geological disaster hazards, and adjust the layout of monitoring points and data collection cycle.

[0016] Thirdly, the invention provides the following technical solution: a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-mentioned method for identifying geological hazards based on InSAR and semi-supervised learning.

[0017] Fourthly, the present invention provides the following technical solution: a readable storage medium storing a computer program, wherein the computer program, when executed by a processor, implements the above-mentioned method for identifying geological hazard risks based on InSAR and semi-supervised learning.

[0018] The present invention has the following beneficial effects: 1. In this invention, by dividing the hydrological sensitivity of the monitoring area according to the regional groundwater unit recharge diagram and setting the InSAR data sampling frequency, the surface time-series displacement data is obtained and potential hidden danger areas are identified. This improves the problem that traditional monitoring frequencies mostly use fixed sampling intervals, which do not take into account changes in hydrological sensitivity, resulting in untimely monitoring of high-risk areas.

[0019] 2. In this invention, by conducting ground-penetrating radar detection in potential hazard areas and extracting underground structural features, and then constructing multi-source feature vectors and fusing them with InSAR data, the problem of traditional single deformation monitoring being unable to identify underground cavities and fracture zones, and the lack of underground structural information leading to incomplete hazard identification is improved.

[0020] 3. In this invention, InSAR displacement features, ground-penetrating radar features, total station labeled data and environmental factors are input into a semi-supervised learning model for training, thereby predicting the probability of hidden dangers and risk levels in unlabeled areas. This improves the problem that traditional manual inspections mostly rely on point monitoring experience for judgment, and the lack of automated intelligent analysis results in low risk identification efficiency. Attached Figure Description

[0021] Figure 1 This is a flowchart of a geological hazard hazard identification method based on InSAR and semi-supervised learning proposed in this invention. Figure 2 This is a system architecture diagram of a geological hazard identification system based on InSAR and semi-supervised learning proposed in this invention. Detailed Implementation

[0022] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Example 1 In the first embodiment of the present invention, the present invention provides a method for identifying geological hazard risks based on InSAR and semi-supervised learning, such as... Figure 1 As shown, the process includes the following steps: based on the regional groundwater unit recharge diagram, the hydrological sensitivity of the monitoring area is divided, and the InSAR data sampling frequency is set; Furthermore, the hydrological sensitivity classification of monitoring areas includes the following steps: The monitoring area was rasterized to obtain hydrogeological parameters for each monitoring unit, including rainfall. Infiltration coefficient Soil permeability Catchment Index and groundwater level changes ; Normalize each parameter to the 0,1 range to obtain the sub-indices. The formula for calculating the rainfall infiltration index is as follows: Soil permeability index is The watershed index is Groundwater level sensitivity index is ; According to the preset weight vector Calculate the comprehensive hydrological sensitivity score : and ;in , , , The weights of each sub-indicator are determined using the Analytic Hierarchy Process (AHP) and satisfy the consistency ratio CR < 0.1. Based on ratings The range of values ​​is used to divide the monitoring area. For highly sensitive areas, This is a moderately sensitive area. This is a low-sensitivity area.

[0024] Specifically, the area to be monitored is rasterized to facilitate subsequent sensitivity calculations based on monitoring units. Hydrogeological parameters, including rainfall, are extracted from each raster unit. Infiltration coefficient Soil permeability Catchment Index and changes in groundwater level Rainfall and groundwater level changes can be obtained from regional hydrological station observation data and groundwater dynamic monitoring data. Infiltration coefficient and soil permeability can be given based on regional hydrogeological survey results or experimental parameters. Catchment index is calculated based on digital elevation model. Then, each parameter is normalized to make the indicators of different dimensions comparable. The normalization method adopts minimum-maximum normalization or standard deviation normalization. The normalized sub-indices are labeled as follows: Wherein: the rainfall infiltration index is expressed by the formula The calculated soil permeability index is: The water level index is The groundwater level sensitivity index is After obtaining the normalized sub-indices, they are processed according to the preset weight vector. Calculate the comprehensive hydrological sensitivity score , The weights satisfy the constraints. The weights are determined using the Analytic Hierarchy Process (AHP), which involves constructing a judgment matrix, calculating eigenvalues ​​and eigenvectors to obtain the weight allocation, and verifying the consistency ratio. satisfy The consistency requirement ensures that the weighting is reasonable and reliable. Finally, based on the obtained comprehensive sensitivity score... The monitoring areas are classified as follows: areas with a value higher than 0.66 are classified as high-sensitivity areas, those between 0.33 and 0.66 are classified as medium-sensitivity areas, and those lower than 0.33 are classified as low-sensitivity areas. The classification results are used to dynamically adjust the InSAR sampling frequency in the future, so that high-sensitivity areas can obtain a higher monitoring frequency, thereby improving the timeliness and effectiveness of hazard identification.

[0025] The above steps enable the quantitative evaluation and spatial zoning of the hydrological sensitivity of the monitoring area. Multi-source hydrogeological information, such as rainfall infiltration, soil permeability, water catchment conditions, and groundwater level changes, is integrated into a single sensitivity score, and spatial differences are expressed by high, medium, and low zoning. This provides a basis for the differentiated setting of subsequent InSAR sampling frequencies, allowing for more intensive monitoring of highly sensitive areas and earlier identification of high-risk hazards, thereby improving the pertinence and timeliness of geological disaster monitoring.

[0026] Setting the InSAR data sampling frequency includes the following steps: Hydrological sensitivity scoring ,use Determine the sampling period, where To minimize the sampling period, 37 days is used. The longest sampling period is 3060 days. As a regulating factor, It is also used to control the nonlinear mapping relationship between sensitivity and sampling period; When the InSAR displacement rate of the monitoring unit Exceeding the threshold or displacement acceleration Exceeding the threshold At that time, through Shorten the sampling period, among which For adjustment coefficients, The reference displacement rate is used to achieve adaptive dynamic adjustment of the sampling period.

[0027] Specifically, based on the comprehensive hydrological sensitivity score of each monitoring unit. Through formula Determine the initial sampling period of this unit, where Indicates the shortest sampling period; Indicates the longest sampling period; parameter This is used to adjust the nonlinear relationship between sensitivity and sampling period, allowing high-sensitivity areas to obtain shorter sampling periods and low-sensitivity areas to obtain longer sampling periods, thereby achieving a rational allocation of monitoring resources. Furthermore, when the InSAR displacement rate of a certain unit is monitored... Exceeding the threshold or displacement acceleration Exceeding the threshold At that time, the sampling period of the unit is dynamically adjusted using the formula. Calculate the new sampling period, where For adjustment coefficients, The sampling period is shortened by using the reference displacement rate to control the magnitude of the shortening. This can increase the monitoring frequency when deformation intensifies, while avoiding resource waste caused by an excessively short sampling period. Through the above steps, the sampling frequency is made sensitive to both hydrological conditions and deformation state, enabling the monitoring system to respond promptly to the development trend of potential disasters.

[0028] By setting the sampling frequency as described above, the monitoring cycle can be configured differently based on the hydrological sensitivity score. This allows for higher frequency InSAR observations in highly sensitive areas and lower frequency observations in less sensitive areas, thereby optimizing the allocation of monitoring resources. At the same time, the sampling cycle is automatically shortened when the displacement rate or acceleration exceeds the set threshold, enabling a dynamic response to deformation anomalies. This improves the timeliness and sensitivity of monitoring data, allowing potential geological hazards to be identified and warned earlier, and enhancing the adaptability and risk prevention capabilities of the monitoring system.

[0029] The surface time-series displacement data, including cumulative displacement, displacement rate, and displacement acceleration, is obtained by using a set sampling frequency to preliminarily identify potential hazard areas. Furthermore, acquiring temporal displacement data of the Earth's surface using a set sampling frequency includes the following steps: Multi-temporal synthetic aperture radar images are acquired according to the sampling period, and the images of each temporal phase are registered. Construct a small baseline interferometric network that satisfies the vertical baseline threshold and the time baseline threshold, and generate multiple interferograms; By unwrapping the interferometric phase and correcting for atmospheric and orbital errors, the unwrapped phase vector is obtained. ; Establish a system of interferometric equations ,in Design matrix for the relationship between the interference pair and time. This represents the displacement at each time point relative to the reference time point. To interfere with the quantity, The number of observation times; The time-series shift vector is solved using the weighted least squares method: ;in This is a weight matrix, with weights... Set according to the square of the coherence of the corresponding interference pair; The obtained displacement vectors at each time point Perform quadratic polynomial fitting over time. Obtain displacement rate and displacement acceleration .

[0030] Specifically, this enables the precise acquisition of temporal displacement data of the land surface according to a set cycle. Through image registration, small baseline interferometric network construction, and phase correction, the basic quality of the data is improved. Weighted least squares solution improves the accuracy of displacement vectors, and quadratic fitting further obtains cumulative displacement, displacement rate, and acceleration. These data can effectively capture the characteristics of land surface deformation and provide accurate and dynamic deformation information support for the preliminary identification of potential hazard areas.

[0031] Ground-penetrating radar (GPR) detection is conducted in potential hazard areas. The underground structural characteristics are assessed using GPR profile maps, and multi-source feature vectors are constructed. Furthermore, constructing multi-source feature vectors includes the following steps: Extract cumulative displacement, displacement rate, displacement acceleration, and local displacement gradient from InSAR time series results; The mean intensity of reflected waves, fracture density, cavity size, and medium attenuation coefficient were extracted from the ground-penetrating radar profile. Environmental factors to be acquired include slope, aspect, rainfall, and changes in groundwater level; After normalizing the above features, they are concatenated according to the monitoring unit number to form a feature vector, which can optionally be obtained by dimensionality reduction through principal component analysis.

[0032] Specifically, ground-penetrating radar (GPR) detection in potential hazard areas can assess subsurface structural characteristics through GPR profiles, providing relevant data for constructing multi-source feature vectors. When constructing multi-source feature vectors, the input data includes cumulative displacement, displacement rate, displacement acceleration, and local displacement gradient extracted from InSAR time-series results acquired and processed at a set sampling frequency; mean reflected wave intensity, fracture density, cavity size, and medium attenuation coefficient extracted from the GPR profiles obtained from the aforementioned GPR detection; and slope and aspect data obtained from regional topographic data, rainfall data obtained from hydrological station observations, and groundwater dynamic monitoring data. The system acquires two types of environmental factors: groundwater level changes. After normalizing the features of these inputs, they are concatenated according to the monitoring unit number to form a feature vector. Alternatively, dimensionality reduction can be achieved through principal component analysis. This multi-source feature vector serves as a key input for the subsequent input of surface time-series displacement data, GPR structural features, total station labeled data, and environmental factors into the semi-supervised learning model. It provides comprehensive and multi-dimensional feature information for the training of the semi-supervised learning model, helping the model predict the probability or risk level of hidden dangers in unlabeled areas. At the same time, it supplements the lack of underground structural information in single InSAR deformation monitoring, making the identification of geological disaster hazards more comprehensive.

[0033] Using total stations deployed at high-risk points, local precise displacement data, including cumulative displacement, displacement rate, and displacement acceleration, are obtained and used as labeled samples to input semi-supervised learning models. Furthermore, acquiring precise local displacement data and constructing labeled samples using a total station includes the following steps: Total station reflector prisms are deployed at potentially high-risk locations, and monitoring data is collected at these locations according to a preset monitoring cycle. 3D coordinates ; Calculate relative to the reference time Three-dimensional displacement vector And its modulus length is obtained as the cumulative displacement. ; Displacement rate and acceleration are calculated using the finite difference method: in For adjacent observation periods; The obtained label vector Multi-source feature vectors of corresponding monitoring units Spatial registration is performed to create labeled training sample pairs. , which serves as the supervisory signal input for the semi-supervised learning model.

[0034] Specifically, total station reflecting prisms are deployed at potentially high-risk points, and monitoring data is collected at these points according to a preset monitoring cycle. 3D coordinates These three-dimensional coordinates are the spatial location data of the monitoring points obtained directly from total station observations; a reference time was selected. Calculate monitoring points At any moment Relative to the reference time Three-dimensional displacement vector The magnitude of the three-dimensional displacement vector is then calculated and used as a monitoring point. At any moment cumulative displacement ;by For the time step, the finite difference method is used, through the formula... Calculate monitoring points At any moment displacement rate Through formula Calculate monitoring points At any moment displacement acceleration The obtained result includes cumulative displacement. Displacement rate Displacement acceleration label vector Compared with the multi-source feature vectors constructed earlier for the corresponding monitoring units Spatial registration is performed to create labeled training sample pairs. The labeled training samples serve as the supervisory signal input for the semi-supervised learning model, providing it with training data containing precise displacement labels. This allows the model to optimize parameters through supervised learning, improving the accuracy of predicting the probability or risk level of hazards in unlabeled areas. This addresses the prediction bias that may exist when training a model solely based on unlabeled data, and enhances the reliability of subsequent training and prediction processes for the semi-supervised learning model.

[0035] The surface time-series displacement data, GPR structural features, total station labeled data and environmental factors are input into the semi-supervised learning model to train the model and predict the probability of hidden dangers or risk level in unlabeled areas. The semi-supervised learning model is a classification model based on graph convolutional networks. The input for building the model includes the following steps: An adjacency matrix is ​​constructed by combining feature similarity and spatial distance. And obtained by symmetric normalization ;in For the first Monitoring unit multi-source feature vectors, Geographical distance, These are the tradeoff coefficient and the scaling parameter, respectively. feature matrix and enter Layered graph convolutional network, outputting predicted probability distribution .

[0036] Specifically, surface temporal displacement data, GPR structural features, total station annotation data, and environmental factors are input into a semi-supervised learning model to train the model and predict the probability of hidden dangers or risk levels in unlabeled areas. Surface temporal displacement data is obtained by acquiring multi-temporal synthetic aperture radar (SAR) images at a set sampling frequency, followed by registration, small baseline interferometry network construction, interferometric phase unwrapping, atmospheric orbit error correction, and weighted least squares solution. GPR structural features are extracted from SAR profiles after SAR detection in potential hazard areas. Total station annotation data is obtained by deploying total station reflecting prisms at high-risk points and collecting three-dimensional coordinates according to a preset monitoring cycle, calculating cumulative displacement, displacement rate, and displacement acceleration. Environmental factors are obtained by acquiring slope, aspect, rainfall, and groundwater level changes. These data are integrated to form the first... Monitoring unit multi-source feature vector ,all Together they constitute the feature matrix The semi-supervised learning model is a classification model based on graph convolutional networks. When constructing this model, an adjacency matrix is ​​first built by combining feature similarity and spatial distance. For the weighting coefficient, and These are scale parameters, and all three are set based on the geological conditions and data distribution of the monitoring area. For the first Monitoring unit and the The geographical distance of the monitoring units is calculated using their spatial coordinates. After constructing an adjacency matrix, it is symmetrically normalized to obtain the result. Then the feature matrix and enter A layered graph convolutional network, after computation, outputs a layer weight matrix. With the Monitoring unit in Layer feature output The predicted probability distribution is obtained by combining the softmax function. , This represents the probability of potential hazards in each monitoring unit; subsequent adjustments will be based on... The quantile method was used to calculate the thresholds for low, medium, and high risk classification. Based on these thresholds, the monitoring units were classified into risk categories, and risk distribution maps were generated. To provide a basis for subsequent model updates based on new monitoring data, this process provides comprehensive feature information to the model through multi-source data fusion. The construction of the adjacency matrix can capture the feature similarity and spatial correlation between monitoring units, and the graph convolutional network outputs... It can accurately predict the probability of hidden dangers in unmarked areas, effectively make up for the lack of information in single data types, improve the comprehensiveness and accuracy of geological disaster hazard identification, and lay the foundation for dynamic identification and early warning of geological disaster hazards.

[0037] The training process of the model includes the following steps: To supervise cross-entropy loss Graph smoothing regularization term and pseudo-label loss This constitutes the total loss. ;in For loss weights; After pre-training on the initial labeled set, an iterative pseudo-label strategy is used to add the high-confidence prediction results of unlabeled samples to the training set until convergence, and the probability of hidden danger and risk level are output.

[0038] Specifically, to supervise cross-entropy loss Graph smoothing regularization term False label loss and weight regularization term Constituting the total loss ,in An initial annotation set based on total station annotation data provides a supervisory signal, enabling the model to learn the correlation between the characteristics of known high-risk points and potential hazards. Through the adjacency matrix Constraining the prediction probability of adjacent monitoring units and The consistency aligns with the spatial correlation of geological hazards. To explore the supervisory value of high-confidence prediction results in unlabeled samples. To prevent model overfitting, , , As loss weights, the influence of each loss term is adjusted according to the model training accuracy requirements. After pre-training on the initial labeled set to enable the model to initially grasp the rules of hazard identification, an iterative pseudo-label strategy is adopted to add the high-confidence prediction results of unlabeled samples to the training set, continuously supplementing training data and optimizing model parameters until the model converges to ensure prediction stability. The final output of hazard probability and risk level provides core judgment basis for subsequent generation of risk distribution maps, dynamic identification of geological disaster hazards and early warning. At the same time, it makes full use of unlabeled data to reduce dependence on a large number of labeled samples and improve the model's generalization ability to identify hazards in the entire monitoring area.

[0039] The model is updated regularly based on new monitoring data, and a risk distribution map is output to achieve dynamic identification and early warning of potential geological hazards. At the same time, the layout of monitoring points and the data collection cycle can be adjusted. The steps for generating a risk distribution map based on the updated model output include: Obtain the predicted probability value of potential hazards for each monitoring unit. ; The thresholds for low, medium, and high risk classification were calculated using the quantile method. in Represents the set of probabilities of potential hazards of quantiles; Risk classification of monitoring units based on thresholds: when Determined to be low risk, when Determined to be of medium risk, when Determined to be high risk; The classification results are visualized to generate a dynamic risk distribution map, which serves as a basis for decision-making regarding adjustments to the layout of monitoring points and the data collection cycle.

[0040] Specifically, the model is updated periodically based on new monitoring data, and the updated model outputs a predicted probability value of potential hazards for each monitoring unit. ,based on The thresholds for low, medium, and high risk classification were calculated using the quantile method. Represents the set of probabilities of potential hazards of Quantiles, specifically the threshold values ​​are , , Then, based on these thresholds, the monitoring units are classified into risk categories. When it is judged as low risk, When it is determined to be medium risk, When a risk is identified as high-risk, the risk classification results are then visualized to generate a dynamic risk distribution map. Figure 1 On the one hand, it can realize the dynamic identification and early warning of geological disaster hazards, and timely grasp the spatial distribution and changing trends of hazards. On the other hand, it serves as a decision-making basis for adjusting the layout of monitoring points and the data collection cycle, so that monitoring resources can be allocated to high-risk areas and the data collection frequency can be optimized. At the same time, by regularly updating the model with new monitoring data, the timeliness and accuracy of hazard identification can be guaranteed, and the problems of risk misjudgment or untimely warning caused by model lag can be avoided.

[0041] Example 2: During the construction and operation of a highway tunnel in a mountainous area, the complex hydrogeological conditions of the region have led to potential geological hazards such as rock instability, surface subsidence, and underground cavity development in shallow sections and near fault fracture zones. Traditional monitoring methods, using fixed-frequency InSAR observations, cannot promptly capture deformation anomalies in highly hydrogeologically sensitive areas. Furthermore, relying solely on surface deformation data lacks information on underground structures, resulting in incomplete hazard identification. Additionally, manual inspections combined with point-based total station monitoring are inefficient and lack timely dynamic updates and early warnings of hazard risks, easily leading to construction safety accidents or road damage during operation. To address these issues, this invention employs a geological hazard identification system based on InSAR and semi-supervised learning, the structure of which is as follows: Figure 2 As shown. The specific implementation process of this system is as follows: The hydrological sensitivity and sampling setting module works as follows: Based on the groundwater recharge diagram of the tunnel area, it extracts the rainfall data for each monitoring unit. Infiltration coefficient Soil permeability Catchment Index and changes in groundwater level After normalizing these parameters, the weights are determined using the Analytic Hierarchy Process (AHP) to calculate the comprehensive hydrological sensitivity score. ,according to , , Divide the area into high, medium, and low sensitivity zones, and then base the scores on the classification. The InSAR data sampling frequency is dynamically set by displacement threshold; the effect is to obtain a short sampling period of about 37 days in the high hydrologically sensitive area of ​​the tunnel, and a longer sampling period in the low sensitive area, so as to avoid the monitoring lag in high-risk areas caused by traditional fixed sampling and optimize the allocation of monitoring resources. The InSAR displacement monitoring and hazard identification module works as follows: It acquires multi-temporal synthetic aperture radar images of the area above and around the tunnel at the set sampling frequency. After image registration, small baseline interferometric network construction, interferometric phase unwrapping, and atmospheric / orbital error correction, the surface temporal displacement vector is solved using the weighted least squares method. Then, the cumulative displacement, displacement rate, and displacement acceleration are obtained through quadratic polynomial fitting. The goal is to preliminarily identify potential hazard areas where the surface displacement rate above the tunnel exceeds the threshold, providing a clear scope for subsequent focused detection. The radar detection and feature construction module works as follows: Ground-penetrating radar (GPR) is used to detect potential hazards in the initially identified areas. The distribution of the subsurface medium is analyzed using GPR profiles, and the mean reflected wave intensity, fracture density, cavity size, and medium attenuation coefficient are extracted. Simultaneously, environmental factors such as slope, aspect, rainfall, and groundwater level changes in the area are acquired. These features are then normalized and concatenated with displacement features extracted by InSAR to construct multi-source feature vectors for each monitoring unit. The effect achieved is to supplement the information on the underground structure around the tunnel, solve the problem that traditional single deformation monitoring cannot cover underground hidden dangers, and provide comprehensive and multi-dimensional features for subsequent model training. The total station data acquisition and tag generation module works as follows: Total station reflecting prisms are deployed at high-risk points identified by InSAR, and the three-dimensional coordinates of the monitoring points are acquired at preset intervals. Calculate relative to the reference time The magnitude of the three-dimensional displacement vector is used as the cumulative displacement. Displacement rate calculated using the finite difference method With acceleration , label vector Multi-source feature vectors of corresponding monitoring units Spatial registration to form labeled training sample pairs The goal is to provide high-precision local displacement annotation data, which provides reliable supervision signals for semi-supervised learning models and avoids prediction bias caused by the lack of accurate labels. Semi-supervised learning training and prediction module operation: The above-mentioned surface temporal displacement data, GPR structural features, total station labeled data, and environmental factors are input into a semi-supervised learning model based on graph convolutional networks. First, an adjacency matrix is ​​constructed by combining feature similarity and spatial distance and then normalized. Then, the feature matrix and the normalized adjacency matrix are input into the model. Layered graph convolutional networks, through the total loss function The model is trained and iteratively incorporated with high-confidence predictions from unlabeled samples until the model converges. Finally, the probability of hidden dangers and the risk level of unlabeled units in the entire tunnel monitoring area are output. The effect achieved is to make full use of limited labeled data and a large amount of unlabeled data to realize the probability prediction of hidden dangers in the entire tunnel area, improve the coverage and accuracy of hidden danger identification, and solve the problems of low efficiency and incomplete coverage of traditional manual inspections. The model update and dynamic early warning module works by periodically incorporating new monitoring data from tunnel construction (such as new SAR images and ground-penetrating radar data from the excavation stage) into the system update model to obtain the predicted probability values ​​of potential hazards for each monitoring unit after the update. Calculated using the quantile method , , As a risk threshold, a dynamic risk distribution map is generated after classification according to the threshold. Based on the distribution map, monitoring points are added in high-risk areas and the data collection cycle is shortened. The effect is to realize the dynamic identification and early warning of potential geological hazards in tunnels, to keep abreast of the changing trends of hazards during the excavation process, to optimize the layout of monitoring points and the data collection cycle, and to effectively prevent the occurrence of disasters such as rock collapse and surface subsidence.

[0042] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A geological disaster hidden danger identification method based on InSAR and semi-supervised learning, characterized in that, The method comprises the following steps: According to the groundwater recharge schematic diagram of the regional groundwater unit, the hydrological sensitivity of the monitoring area is divided, and the InSAR data sampling frequency is set; The surface time series displacement data is obtained by using the set sampling frequency, including cumulative displacement, displacement rate and displacement acceleration, which are used for preliminary identification of potential hidden trouble area; Geological radar detection is carried out in the potential hidden trouble area, the underground structure characteristics are evaluated through the geological radar profile, and a multi-source feature vector is constructed; The local accurate displacement data is obtained by using the total station instrument arranged at the high risk point, including cumulative displacement, displacement rate and displacement acceleration, which are used as labeled sample input into the semi-supervised learning model; The surface time series displacement data, GPR structure characteristics, total station instrument labeled data and environmental factors are input into the semi-supervised learning model, the model is trained, and the hidden trouble probability or risk level of the unlabeled area is predicted; The model is updated regularly according to the new monitoring data, and the risk distribution map is output, so that the dynamic identification and early warning of geological disaster hidden trouble are realized, and the monitoring point layout and data acquisition cycle can be adjusted.

2. The method according to claim 1, wherein, The hydrological sensitivity of the monitoring area comprises the following steps: The monitoring area is rasterized to obtain hydrogeological parameters of each monitoring unit, including rainfall , infiltration coefficient , soil permeability , catchment index , and groundwater level change ; Each parameter is normalized to the interval 0, 1 to obtain sub-indexes ; wherein the rainfall infiltration index is calculated by the formula: ; the soil permeability index is ; the catchment index is ; and the groundwater level sensitivity index is ; According to a preset weight vector , calculate the comprehensive hydrological sensitivity score : and ; wherein , , , respectively correspond to the weight of each sub-index, the weight is determined by the analytic hierarchy process (AHP), and the consistency ratio CR is less than 0.

1. According to the score divides the monitoring area into value ranges, is a high sensitive area, is a medium sensitive area, is a low sensitive area.

3. The method according to claim 1, characterized in that, The setting of the InSAR data sampling frequency comprises the following steps: According to the hydrological sensitivity score , using determining a sampling period, wherein is the shortest sampling period, taken as 37 days, is the longest sampling period, taken as 3060 days, is an adjustment factor, and for controlling the non-linear mapping relationship between sensitivity and sampling period; When the InSAR displacement rate of the monitoring unit exceeds a threshold or the displacement acceleration exceeds a threshold , by shortening the sampling period, wherein is an adjustment coefficient, is a reference displacement rate for implementing adaptive dynamic adjustment of the sampling period.

4. The method according to claim 3, wherein, The surface time series displacement data obtained by using the set sampling frequency comprises the following steps: The multi-temporal synthetic aperture radar images are obtained according to the sampling period, and each temporal image is registered; A small baseline interference network meeting the vertical baseline threshold and the time baseline threshold is constructed, and a plurality of interference maps are generated; unwrapping the interferometric phase and atmospheric and orbital error correction to obtain an unwrapped phase vector ; establishing a system of interference equations wherein is a design matrix of the interference pairs in relation to the time instants, is the displacement amount of each time instant in relation to a reference time instant, is the number of interference pairs, is the number of observation time instants; The weighted least square method is used to solve the time series displacement vector: ; wherein is a weight matrix, and the weight is set according to the square of the coherence of the corresponding interference pair. The displacement vectors found at each time instant are A quadratic polynomial fit is performed in time The displacement velocity is obtained And the displacement acceleration .

5. The method according to claim 1, wherein, The construction of the multi-source feature vector comprises the following steps: The cumulative displacement, displacement rate, displacement acceleration and local displacement gradient are extracted from the InSAR time series results; The average reflection wave intensity, crack density, cavity size and medium attenuation coefficient are extracted from the geological radar profile; The environmental factors include slope, slope direction, rainfall and groundwater level change; After the above features are normalized, they are spliced according to the monitoring unit number to form a feature vector.

6. The method according to claim 1, wherein, The local accurate displacement data obtained by using the total station instrument and the construction of the labeled sample comprise the following steps: A total station reflector prism is arranged at a potential high-risk point, and three-dimensional coordinates of the monitoring point are collected according to a preset monitoring period ;​ Calculate relative to the reference time Three-dimensional displacement vector And its modulus length is obtained as the cumulative displacement. ; The displacement velocity and acceleration are calculated using finite difference method: where is the adjacent observation period; The resulting label vector The multi-source feature vector of the corresponding monitoring unit Spatial registration is performed to form a labeled training sample pair as a supervision signal input of the semi-supervised learning model.

7. The method according to claim 1, wherein, The semi-supervised learning model is a classification model based on graph convolution network, and the construction of the model input comprises the following steps: An adjacency matrix is constructed by combining feature similarity and spatial distance , and is normalized to obtain ; wherein is the first monitoring unit multi-source feature vector, is the geographical distance, is the weighting coefficient and scale parameter, respectively; characteristic matrix with input layer graph convolutional network, outputting a prediction probability distribution .

8. The method according to claim 7, characterized in that, The training process of the model comprises the following steps: with supervised cross-entropy loss , graph smoothing regularizer , and pseudo-label loss , to form the total loss ; wherein is the loss weight; After pre-training on the initial labeled set, the iterative pseudo-label strategy is adopted to add the high confidence prediction results of the unlabeled sample to the training set until convergence, and the hidden trouble probability and risk level are output.

9. The method according to claim 1, wherein, The step of generating the risk distribution map according to the output result of the updated model comprises: obtaining a hazard probability prediction value for each monitoring unit ; Thresholds for low, medium, and high risk classification are calculated using quantile methods wherein represents a set of hazard probabilities of quantiles Risk classifying the monitoring unit according to the threshold value: when low risk is determined, when medium risk is determined, when high risk is determined; The classification result is visualized to generate a dynamic risk distribution map, which is used as a decision basis for adjusting the monitoring point layout and data acquisition cycle. 10.A geological disaster hidden danger identification system based on InSAR and semi-supervised learning, characterized in that, The system for the geological disaster hidden trouble identification method based on InSAR and semi-supervised learning of any one of claims 1-9, the system comprises: A hydrological sensitivity and sampling setting module for dividing the hydrological sensitivity of the monitoring area according to the groundwater recharge schematic diagram of the regional groundwater unit, and setting the InSAR data sampling frequency; An InSAR displacement monitoring and hidden trouble identification module for obtaining the surface time series displacement data by using the set sampling frequency, including cumulative displacement, displacement rate and displacement acceleration, which are used for preliminary identification of potential hidden trouble area; a radar detection and feature construction module, configured to perform geological radar detection in a potential hidden danger area, evaluate underground structure features through a geological radar profile, and construct a multi-source feature vector; a total station collection and label generation module, configured to use a total station arranged at a high-risk point to obtain local accurate displacement data, including cumulative displacement, displacement rate, and displacement acceleration, and use the data as a labeled sample input into a semi-supervised learning model; a semi-supervised learning training and prediction module, configured to input surface time-series displacement data, GPR structure features, total station labeled data, and environmental factors into a semi-supervised learning model, train the model, and predict hidden danger probability or risk level in an unlabeled area; a model updating and dynamic early warning module, configured to regularly update the model according to new monitoring data, output a risk distribution map, realize dynamic identification and early warning of geological disaster hidden dangers, and simultaneously adjust monitoring point layout and data collection period.

Images