A complex high-position landslide risk dynamic identification system based on multi-source InSAR cooperation

Abstract

The present application relates to the technical field of geological disaster monitoring and early warning, in particular to a complex high-position landslide risk dynamic identification system based on multi-source InSAR cooperation, comprising: a data acquisition module for constructing a normalized feature vector representing the state of each monitoring node; an association modeling module for quantifying the physical influence degree between heterogeneous nodes; a state prediction module for iteratively updating the hidden state of each heterogeneous node using a spatio-temporal graph neural network model, and interpreting the hidden state as incremental displacement; a risk assessment module for calculating the risk precursor index of each heterogeneous node deviating from the normal evolution mode based on the hidden state, and determining and issuing graded early warning according to the preset threshold system; the present application effectively alleviates the overfitting problem of pure data-driven models when data is sparse, significantly improving the generalization ability and physical interpretability of the prediction results.

Description

Technical Field

[0001] This invention relates to the field of geological disaster monitoring and early warning technology, specifically to a dynamic identification system for complex high-altitude landslide risks based on multi-source InSAR collaboration. Background Technology

[0002] As geological disaster prevention and control work deepens, complex high-altitude landslides, due to their strong concealment and severe destructive power, require accurate dynamic risk identification to safeguard people's lives and property. However, landslides are complex geological bodies that evolve in time and space, and their risk identification faces enormous challenges. Traditional landslide risk identification methods mainly rely on monitoring data from a single source or interpret and then fuse multi-source data. This not only leads to the loss of source information and time delay, but also makes it difficult to fully depict the entire process of landslide gestation and evolution.

[0003] Existing technologies have the following limitations: First, limitations in data fusion. Traditional monitoring methods typically fuse data after calculating physical quantities such as displacement. This post-fusion paradigm severs the inherent correlation between data, resulting in information loss. Furthermore, different sensor data have varying modes, frequencies, and dimensions, making effective feature-level fusion difficult at the raw level. Second, limitations in correlation modeling. Existing models often ignore the complex physical mechanisms within landslides. For example, when constructing node correlations, they only consider spatial distance without fully incorporating key geological prior knowledge such as elevation differences and gravity effects, leading to models that cannot accurately reflect the transmission paths of stress and displacement. Third, limitations in state prediction. Purely data-driven prediction models are prone to overfitting when monitoring data is sparse or noisy, and their prediction results may violate basic physical laws, exhibiting insufficient generalization ability and interpretability.

[0004] Spatiotemporal graph neural network technology has shown great potential in processing data with complex spatial topology and temporal evolution characteristics. Deeply fusing multi-source monitoring data at the feature layer and embedding physical constraints to train the model can significantly improve the accuracy and reliability of predictions. In summary, existing technologies are insufficient in the deep fusion of multi-source heterogeneous data, spatiotemporal correlation modeling with geological prior knowledge, and prediction driven by both data and physics, making it difficult to achieve accurate and dynamic identification and early warning of complex high-altitude landslide risks. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a dynamic identification system for complex high-altitude landslide risks based on multi-source InSAR collaboration. Specifically, the technical solution of this invention includes:

[0006] The data acquisition module is used to acquire multi-source heterogeneous sensor data streams in the landslide monitoring area, and construct a normalized feature vector characterizing the state of each monitoring node based on the multi-source heterogeneous sensor data streams.

[0007] The association modeling module is used to discretize the landslide monitoring area into heterogeneous nodes and quantify the degree of physical influence between heterogeneous nodes based on a weighted adjacency matrix containing prior geological knowledge.

[0008] The state prediction module is used to iteratively update the hidden state of each heterogeneous node based on the normalized feature vector and the weighted adjacency matrix using a spatiotemporal graph neural network model, and interpret the hidden state as an incremental displacement.

[0009] The risk assessment module is used to calculate the risk precursor index of each heterogeneous node deviating from the normal evolution mode based on the hidden state, and to determine and issue graded warnings according to the preset threshold system.

[0010] Preferably, the multi-source heterogeneous sensor data stream includes raw interferometric phase stream data and coherence stream data, which are acquired by InSAR satellites; the multi-source heterogeneous sensor data stream also includes three-dimensional real-time displacement vector stream, which is acquired by GNSS stations; the multi-source heterogeneous sensor data stream also includes pulsed rainfall data, which is acquired by rain gauges.

[0011] Preferably, the association modeling module is used to combine the Gaussian attenuation effect based on spatial distance and the terrain influence effect based on elevation difference to determine the weighted adjacency matrix;

[0012] Among them, the terrain impact effect is simulated by nonlinear transformation to simulate the asymmetry of the impact under gravity, so as to enhance the influence of upstream nodes on downstream nodes.

[0013] Preferably, the state prediction module iteratively updates the hidden state of each heterogeneous node, including:

[0014] By performing graph convolution operations, the hidden states of each neighboring node at the previous time step are weighted and aggregated to obtain the comprehensive influence representing the spatial neighborhood transmission.

[0015] The combined influence and the normalized feature vector of the heterogeneous node at the current moment are fed into the gate control loop unit to calculate the hidden state at the current moment.

[0016] Preferably, the spatiotemporal graph neural network model is trained using a hybrid loss function; the hybrid loss function includes prediction accuracy loss and physical regularization term.

[0017] Preferably, physical regularization terms are used to impose physical constraints during model training;

[0018] The physical constraint is that the deformation rate of the landslide should not slow down when effective rainfall occurs.

[0019] Preferably, the risk assessment module calculates a risk precursor index, including:

[0020] Using stable historical monitoring data, a set of hidden state vectors representing the normal state is generated through a trained spatiotemporal graph neural network model.

[0021] Based on the hidden state vector representing the normal state, the average hidden state and standard deviation of each node are calculated.

[0022] The risk precursor index is obtained by measuring the Euclidean distance between the current hidden state and the average hidden state, and then normalizing it using the standard deviation.

[0023] Preferred tiered early warning systems include:

[0024] When the risk precursor index is less than or equal to the first preset threshold, it is judged as a normal state and no warning is issued.

[0025] When the risk precursor index is greater than the first preset threshold and less than or equal to the second preset threshold, a warning at the attention level is issued.

[0026] When the risk precursor index exceeds the second preset threshold, a risk level warning is issued.

[0027] Compared with the prior art, the present invention has the following beneficial effects:

[0028] 1. This system performs feature-level fusion at the data source, constructing a unified normalized feature vector in real time from heterogeneous data streams such as InSAR, GNSS, and rainfall. This avoids the information loss and time delay caused by the traditional method of solving and fusing data first, providing high-quality input for accurate prediction and improving the real-time performance of risk identification.

[0029] 2. By constructing a weighted adjacency matrix that incorporates prior geological knowledge, this system not only considers the spatial distance between nodes but also innovatively incorporates the terrain effect caused by elevation differences. This method can simulate the asymmetric effects under gravity, making the model more realistically reflect the transmission path of stress and displacement within the landslide and improving the physical realism of the model.

[0030] 3. This system adopts a dual-driven training paradigm of data and physics. A physical regularization term is introduced into the model training to impose constraints on physical laws such as the landslide deformation rate should not slow down during effective rainfall. This effectively alleviates the overfitting problem that is prone to occur in pure data-driven models when data is sparse, and significantly improves the generalization ability and physical interpretability of the prediction results.

[0031] 4. The risk assessment of this system no longer relies on a single displacement threshold. Instead, it calculates the risk precursor index by measuring the deviation between the high-dimensional hidden state inside the model and the normal benchmark. Since the hidden state integrates spatiotemporal information, it can more comprehensively represent the overall state of the node, thereby enabling earlier and more sensitive capture of instability precursors and improving the sensitivity and foresight of the early warning. Attached Figure Description

[0032] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0033] Figure 1 This is a structural diagram of the system of the present invention. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0035] Example 1:

[0036] Please see Figure 1 A dynamic identification system for complex high-altitude landslide risks based on multi-source InSAR collaboration, comprising:

[0037] The data acquisition module is used to acquire multi-source heterogeneous sensor data streams in the landslide monitoring area, and construct a normalized feature vector characterizing the state of each monitoring node based on the multi-source heterogeneous sensor data streams.

[0038] The association modeling module is used to discretize the landslide monitoring area into heterogeneous nodes and quantify the degree of physical influence between heterogeneous nodes based on a weighted adjacency matrix containing prior geological knowledge.

[0039] The state prediction module is used to iteratively update the hidden state of each heterogeneous node based on the normalized feature vector and the weighted adjacency matrix using a spatiotemporal graph neural network model, and interpret the hidden state as an incremental displacement.

[0040] The risk assessment module is used to calculate the risk precursor index of each heterogeneous node deviating from the normal evolution mode based on the hidden state, and to determine and issue graded warnings according to the preset threshold system.

[0041] This invention provides a dynamic identification system for complex high-altitude landslide risks based on multi-source InSAR collaboration. The system aims to achieve accurate and dynamic identification and early warning of complex high-altitude landslide risks through multi-source data fusion, spatiotemporal correlation modeling, and prediction by embedding physical mechanisms. The system includes a data acquisition module, a correlation modeling module, a state prediction module, and a risk assessment module.

[0042] The purpose of the data acquisition module is to acquire multi-source heterogeneous sensor data of the monitoring area in real time and asynchronously, and transform it into a unified and standardized data structure to provide high-quality input for subsequent cross-modal analysis and prediction. In this embodiment, the module acquires multi-source heterogeneous sensor data streams of the landslide monitoring area through corresponding sensors. To address the differences in modality, frequency, and dimensions among different data streams, the module constructs a normalized feature vector characterizing the state of each monitoring node based on these data streams. This is achieved by, at any given time point... The raw or semi-processed data from each sensor is mapped to the corresponding nodes in the monitoring graph using a preset transformation function. Normalized eigenvectors This process is achieved through the following formula:

[0043]

[0044] in, : is the first in the figure Each monitoring node in time The normalized feature vectors are used as real-time, multimodal inputs to the subsequent spatiotemporal graph neural network model and are generated by this data acquisition module.

[0045] : for the first The original interferometric phase values ​​of each node at this moment are of floating-point data and are obtained from InSAR satellite sensors.

[0046] : This is the coherence value corresponding to the phase value. The data type is floating-point, and its source is collected by InSAR satellite sensors.

[0047] : This is a three-dimensional real-time displacement vector measured by GNSS stations. The data type is a three-dimensional vector, and its source is collected by ground GNSS stations. For nodes without GNSS coverage, this parameter can be estimated based on data from nearby GNSS stations through spatial interpolation methods, such as Kriging interpolation or inverse distance weighted interpolation, to ensure the uniformity of the dimension of the feature vector.

[0048] : To influence the first The regional rainfall data for each node is a floating-point number, and its source is collected from ground rain gauges.

[0049] This is a pre-defined normalization function for various heterogeneous data. Its purpose is to map raw data with different dimensions to similar numerical intervals, thereby eliminating dimensional differences and ensuring the convergence and stability of model training. It is pre-determined based on statistical analysis of historical data from the monitored area. For example, the min-max normalization method can be used to normalize any raw data... Convert to ,in and These are the minimum and maximum values ​​of the corresponding physical quantities obtained from historical data statistics;

[0050] The purpose of the correlation modeling module is to organize discrete monitoring nodes into a network structure that reflects the geological and physical characteristics of the landslide, providing a structured path for the effective dissemination of spatiotemporal information. In this embodiment, the module discretizes the entire landslide monitoring area into a dynamic graph composed of heterogeneous nodes. ,in For a set of nodes, This module represents a set of edges; nodes are defined based on their data source, such as InSAR pixel nodes, GNSS station nodes, etc.; this module is based on a weighted adjacency matrix that incorporates prior geological knowledge. Quantify the degree of physical influence between heterogeneous nodes; elements in the matrix The value represents the node For nodes The degree of physical impact;

[0051] The purpose of the state prediction module is to integrate the node's own characteristics with the influence of its neighborhood to predict the future state evolution of each monitoring point within the landslide body. In this embodiment, this module is based on the normalized feature vector generated by the data acquisition module. The weighted adjacency matrix generated by the association modeling module A spatiotemporal graph neural network model is used to iteratively update the hidden states of each heterogeneous node. Hidden state It is a high-dimensional vector for monitoring nodes. exist The abstract representation of the physical state at any given moment contains a comprehensive result of historical information, current input, and the influence of spatial neighborhood. This module interprets this high-dimensional, abstract hidden state into incremental displacements with clear physical meaning through a fully connected output layer. ;

[0052] The purpose of the risk assessment module is to perform real-time analysis of the internal states predicted by the model, dynamically quantify the risk of landslides evolving from a stable state to an unstable state, and provide graded early warnings. In this embodiment, this module is based on the hidden states output by the state prediction module. By measuring the degree of deviation from the normal state distribution, the risk precursor index of each heterogeneous node from the normal evolutionary pattern is calculated. Risk precursor index It is a dimensionless scalar whose value directly reflects the degree of abnormality of the current state of the monitoring node; the module judges the calculated risk precursor index based on a preset threshold system with clear statistical significance and issues corresponding graded warnings.

[0053] This embodiment constructs an end-to-end dynamic identification system, realizing a complete technical closed loop from multi-source heterogeneous data acquisition, physical correlation modeling, spatiotemporal state prediction to risk assessment and early warning. The system performs feature-level fusion directly at the raw data level, avoiding the information loss and time delay caused by the traditional paradigm of solving first and then fusing. Through the graph structure containing geological a priori information, the model can capture the complex physical mechanisms within the landslide. By utilizing a spatiotemporal graph neural network, it effectively integrates temporal evolution and spatial influence. Finally, by detecting anomalies in the model's internal state, it achieves dynamic and quantitative assessment of landslide risk. This system significantly improves the real-time performance, accuracy, and interpretability of risk identification for complex high-altitude landslides.

[0054] Example 2:

[0055] The multi-source heterogeneous sensor data stream includes raw interferometric phase stream data and coherent stream data, which are acquired by InSAR satellites; the multi-source heterogeneous sensor data stream also includes three-dimensional real-time displacement vector stream, which is acquired by GNSS stations; the multi-source heterogeneous sensor data stream also includes pulsed rainfall data, which is acquired by rain gauges.

[0056] Based on the implementation method described in Example 1, this embodiment defines the specific composition of the multi-source heterogeneous sensor data stream in order to ensure the comprehensiveness and complementarity of the input data, so as to more accurately capture the deformation and inducing factors of the landslide body.

[0057] The multi-source heterogeneous sensor data stream includes: raw interferometric phase stream data and coherence stream data acquired by InSAR satellites; three-dimensional real-time displacement vector stream acquired by GNSS stations; and pulsed rainfall data acquired by rain gauges.

[0058] By explicitly combining the wide-area monitoring capabilities of InSAR, the high-precision fixed-point observation capabilities of GNSS, and the quantification capabilities of rain gauges for core triggering factors, this system achieves comprehensive observation of the landslide deformation-cause chain. This synergy of multi-source data enables the normalization of feature vectors... It can more completely depict the physical state of the monitoring node at any given time, thus providing richer and more reliable information input for the state prediction module, directly improving the prediction accuracy and robustness of the subsequent risk identification model.

[0059] Example 3:

[0060] The association modeling module is used to combine the Gaussian decay effect based on spatial distance and the terrain influence effect based on elevation difference to determine the weighted adjacency matrix;

[0061] Among them, the terrain impact effect is simulated by nonlinear transformation to simulate the asymmetry of the impact under gravity, so as to enhance the influence of upstream nodes on downstream nodes.

[0062] Based on the implementation method described in Example 1, this example describes how the association modeling module determines the weighted adjacency matrix. Specific limitations were imposed, which allowed the quantification of the correlation between nodes to incorporate prior geological knowledge that conforms to the characteristics of landslide dynamics.

[0063] The association modeling module calculates the adjacency matrix elements using a comprehensive model. The model combines a Gaussian attenuation effect based on spatial distance with a terrain influence effect based on elevation differences; the terrain influence effect is handled by a nonlinear transformation function to account for the elevation differences between nodes. Its underlying logic lies in simulating the asymmetry of the effects of gravity, thereby significantly enhancing the influence of upstream nodes on downstream nodes;

[0064] The two effects mentioned above are fused using the following formula to determine the weighted adjacency matrix. elements in :

[0065]

[0066] in, : Quantized nodes in the adjacency matrix For nodes The weight values ​​affecting the intensity are floating-point numbers and are calculated and generated by this association modeling module.

[0067] : for nodes and The spatial distance between them is calculated based on the geographic coordinates of the nodes;

[0068] : This is the distance attenuation coefficient, a hyperparameter used to control the range of influence. Its setting is based on the characteristic scale of the landslide body in the geological survey report to ensure that the model conforms to the scale effect of the actual geological unit.

[0069] : for nodes and elevation difference Its source is calculated based on the digital elevation model data of the nodes;

[0070] : This is the elevation influence factor, and its dimension is the reciprocal of length, such as As a hyperparameter that adjusts the contribution of elevation difference to the asymmetry of influence weights, its value is optimized and determined by cross-validation of model performance on the validation set.

[0071] This embodiment transcends the limitations of traditional graph models that only consider spatial distance. By asymmetrically embedding the gravity effect caused by elevation differences into the adjacency matrix, the constructed landslide physical correlation network topology more realistically reflects the transmission path of stress and displacement within the landslide, providing a more physically consistent information propagation framework for subsequent spatiotemporal graph neural networks and improving prediction accuracy.

[0072] Example 4:

[0073] The state prediction module iteratively updates the hidden states of each heterogeneous node, including:

[0074] By performing graph convolution operations, the hidden states of each neighboring node at the previous time step are weighted and aggregated to obtain the comprehensive influence representing the spatial neighborhood transmission.

[0075] The combined influence and the normalized feature vector of the heterogeneous node at the current moment are fed into the gate control loop unit to calculate the hidden state at the current moment.

[0076] Based on the implementation method described in Example 1, this example describes how the state prediction module iteratively updates the hidden states of each heterogeneous node. It provides a specific implementation process that integrates the powerful spatial information aggregation capability of graph convolutional networks with the excellent time-series information processing capability of gated recurrent units;

[0077] At each time step Through graph convolution operations, the target node is... Each neighbor node In the previous hidden state Perform weighted aggregation to obtain the representation space neighborhood and pass it to the node. The overall influence vector;

[0078] The combined influence vector and nodes The normalized eigenvector at the current moment The data is sent together to the gate-controlled recurrent unit (GRU), where, through its internal update and reset gate mechanisms, historical information is selectively forgotten and new information is incorporated, ultimately calculating the node's value. Hidden state at the current moment The entire iterative update process can be described by the following formula:

[0079]

[0080] in, : for nodes exist The hidden state after each time update, as a high-dimensional abstract representation of the physical state of the node at the current time, is calculated and generated by this state prediction module.

[0081] This is a gated loop unit function, whose function is to merge the current input and historical information to update the state;

[0082] : for nodes exist The external input feature vector at time t is provided by the data acquisition module;

[0083] : This is a graph convolution operation, its function is to aggregate data from the neighborhood. Historical status information, among which The source is the association modeling module. This is the hidden state from the previous moment. It is a weight matrix that a model needs to learn during training, used to perform linear transformations on the hidden states;

[0084] By combining graph convolution with GRU, a spatiotemporal information processing unit was constructed. Graph convolution operations effectively capture the spatial interactions within the landslide mass at each time step, while GRU handles the temporal dependence of the deformation process, thus enabling more accurate prediction of landslide dynamics driven by both spatial effects and temporal accumulation at each time step. At that time, the initial hidden state of each node It can be initialized as a zero vector;

[0085] To achieve the final displacement prediction, the state prediction module further uses a decoder, such as a fully connected output layer, to output the hidden state at the current moment. Mapped to physically interpretable incremental displacement predictions This process can be represented by the following formula: ,in This represents a multilayer perceptron or a single fully connected layer, whose weights are learned during model training.

[0086] Example 5:

[0087] The spatiotemporal graph neural network model is trained using a hybrid loss function, which includes prediction accuracy loss and physical regularization term.

[0088] Physical regularization terms are used to impose physical constraints during model training;

[0089] The physical constraint is that the deformation rate of the landslide should not slow down when effective rainfall occurs.

[0090] This embodiment optimizes the training mechanism of the spatiotemporal graph neural network model, specifying that it uses a hybrid loss function for training, aiming to ensure that the model's prediction results not only have high data fitting accuracy, but also conform to established physical laws.

[0091] The hybrid loss function consists of a prediction accuracy loss and a physical regularization term. The prediction accuracy loss is a data-driven term used to measure the difference between the model's predicted values ​​and the actual observed values. The physical regularization term is a knowledge-driven term, which applies explicit physical constraints during model training, encoding prior physical knowledge as a differentiable penalty term. In this embodiment, the applied physical constraint is: when effective rainfall occurs, the deformation rate of the landslide should not slow down. The complete hybrid loss function... Defined by the following formula:

[0092]

[0093] in, : This is the total loss function, which serves as the optimization objective for model training and is defined in this embodiment;

[0094] : The total number of nodes participating in the calculation;

[0095] : represents the square of the Euclidean distance, used to calculate the prediction accuracy loss; where The incremental displacement predicted by the model. These are the actual incremental displacement observations;

[0096] : This is a reference displacement, which is used to make the prediction accuracy loss term dimensionless to ensure that the dimensions of each component of the hybrid loss function are consistent. Its source can be determined statistically based on the displacement amplitude range of historical monitoring data.

[0097] : This is a weighting coefficient, a hyperparameter used to balance the model's fitting accuracy to the observed data with its adherence to physical laws. Its value is determined through cross-validation.

[0098] To correct the linear unit function, a one-way penalty is implemented here: a penalty term is generated only when physical constraints are violated;

[0099] : Normalized rate ,in The reference rate is calculated from the predicted incremental displacement. The source is determined by historical data statistics or geological survey reports;

[0100] Normalized rainfall ,in For real-time rainfall, the critical rainfall threshold is... The source is determined by historical data statistics or geological survey reports;

[0101] By introducing a hybrid loss function that includes a physical regularization term, prior geological knowledge is injected into the model training process. This data- and physics-driven training paradigm effectively alleviates the overfitting problem that may occur in pure data-driven models when monitoring data is sparse or noisy, improves the model's generalization ability, and ensures that the model's prediction results are physically reasonable and interpretable.

[0102] Example 6:

[0103] The risk assessment module calculates a risk precursor index, including:

[0104] Using stable historical monitoring data, a set of hidden state vectors representing the normal state is generated through a trained spatiotemporal graph neural network model.

[0105] Based on the hidden state vector representing the normal state, the average hidden state and standard deviation of each node are calculated.

[0106] The risk precursor index is obtained by measuring the Euclidean distance between the current hidden state and the average hidden state, and then normalizing it using the standard deviation.

[0107] Based on the implementation method described in Example 1, this example describes how the risk assessment module calculates the risk precursor index. It provides a detailed implementation process, the core of which is to identify risk precursors by quantifying the deviation of the internal state vector of the model from the normal baseline;

[0108] In the initial stage of system deployment, a large amount of historical monitoring data that is marked as stable, such as monitoring data in which the cumulative displacement and deformation rate of nodes are lower than the preset geological safety threshold within a specific monitoring period, are input into the trained spatiotemporal graph neural network model to generate a set of hidden state vectors representing the normal state.

[0109] Based on this set of normal state vectors, for each monitoring node Statistical analysis was performed separately to calculate the average hidden state. with standard deviation ;

[0110] During the real-time monitoring phase, the hidden state at the current moment is measured. The average hidden state of this node's history The Euclidean distance between them, and the standard deviation of the node's history. Normalizing this distance yields a dimensionless risk precursor index. The calculation process is defined by the following formula:

[0111]

[0112] in, : A risk precursor index, dynamically quantified nodes exist The degree to which it deviates from its normal evolutionary pattern at any given time is calculated by this risk assessment module;

[0113] : for nodes At the present moment The hidden state vector is generated in real time by the state prediction module;

[0114] : for nodes The average hidden state, which serves as the baseline center for the normal state of the node, is obtained by statistically averaging the hidden states under a large amount of stable historical data.

[0115] : for nodes The standard deviation of the Euclidean norm of the hidden state vector, as a normalized scalar measure, measures the statistical significance of deviation. It is obtained by statistically calculating the Euclidean norm of the hidden state vector under a large amount of stable historical data.

[0116] : is a very small positive constant, its purpose is to prevent the denominator from being zero due to the standard deviation being zero in calculations;

[0117] This embodiment deepens the basis of risk identification from a single physical output to the high-dimensional hidden state within the model. Since the hidden state integrates spatiotemporal information and multi-source input, it can more comprehensively characterize the overall state of the node than any single physical quantity. This allows for earlier and more sensitive capture of weak abnormal signals during the process of landslides changing from quantitative to qualitative change, thus improving the sensitivity and foresight of the early warning.

[0118] Example 7:

[0119] Tiered early warning includes:

[0120] When the risk precursor index is less than or equal to the first preset threshold, it is judged as a normal state and no warning is issued.

[0121] When the risk precursor index is greater than the first preset threshold and less than or equal to the second preset threshold, a warning at the attention level is issued.

[0122] When the risk precursor index exceeds the second preset threshold, a risk level warning is issued.

[0123] Based on the implementation method described in Example 1, this example specifically implements the judgment logic of graded early warning, transforming the continuously changing risk precursor index into discrete early warning levels with clear guidance for handling.

[0124] The hierarchical early warning system in this embodiment establishes a three-level response mechanism based on a first preset threshold and a second preset threshold:

[0125] When the risk precursor index When the value is less than or equal to the first preset threshold, the monitoring node is determined to be in a normal evolution mode and no warning is issued; in this embodiment, the first preset threshold can be set to 1.5;

[0126] When the risk precursor index When the value is greater than the first preset threshold and less than or equal to the second preset threshold, the node's state is determined to have a statistically significant deviation, and a warning at the attention level is issued; in this embodiment, the second preset threshold can be set to 3.0;

[0127] When the risk precursor index When the value exceeds the second preset threshold, the node's state is determined to have deviated significantly, and a risk warning is issued. The setting of the second preset threshold of 3.0 can be derived from the three sigma criterion in statistics.

[0128] By establishing a threshold system with clear statistical significance, the automatic conversion from quantitative risk index to graded early warning is realized. Compared with the traditional method of relying on experience to set fixed physical quantity thresholds, this early warning logic has stronger adaptability and theoretical basis because the risk precursor index it evaluates is itself a normalized, dimensionless statistical quantity, which makes the early warning system more universal for landslides with different geological conditions.

[0129] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A dynamic identification system for complex high-altitude landslide risks based on multi-source InSAR collaboration, characterized in that, include: The data acquisition module is used to acquire multi-source heterogeneous sensor data streams in the landslide monitoring area, and construct a normalized feature vector characterizing the state of each monitoring node based on the multi-source heterogeneous sensor data streams. The association modeling module is used to discretize the landslide monitoring area into heterogeneous nodes and quantify the degree of physical influence between heterogeneous nodes based on a weighted adjacency matrix containing prior geological knowledge. The state prediction module is used to iteratively update the hidden state of each heterogeneous node based on the normalized feature vector and the weighted adjacency matrix using a spatiotemporal graph neural network model, and interpret the hidden state as an incremental displacement. The risk assessment module is used to calculate the risk precursor index of each heterogeneous node deviating from the normal evolution mode based on the hidden state, and to determine and issue graded warnings according to the preset threshold system. The association modeling module is used to combine the Gaussian decay effect based on spatial distance and the terrain influence effect based on elevation difference to determine the weighted adjacency matrix; Among them, the terrain impact effect is simulated by nonlinear transformation to simulate the asymmetry of the impact under gravity, so as to enhance the influence of upstream nodes on downstream nodes; The association modeling module calculates the adjacency matrix elements using a comprehensive model. The terrain impact effect is handled by a nonlinear transformation function to address the elevation difference between nodes. Its underlying logic lies in simulating the asymmetry of the effects of gravity, thereby significantly enhancing the influence of upstream nodes on downstream nodes; The two effects mentioned above are fused using the following formula to determine the weighted adjacency matrix. elements in : in, : Quantized nodes in the adjacency matrix For nodes The weight values ​​that affect the intensity; : for nodes and Spatial distance between them; : This is the distance attenuation coefficient, used as a hyperparameter to control the range of influence; : for nodes and elevation difference ; : This is the elevation influence factor, and its dimension is the reciprocal of length, such as .

2. The complex high-altitude landslide risk dynamic identification system based on multi-source InSAR collaboration according to claim 1, characterized in that, The multi-source heterogeneous sensor data stream includes raw interferometric phase stream data and coherent stream data, which are acquired by InSAR satellites; the multi-source heterogeneous sensor data stream also includes three-dimensional real-time displacement vector stream, which is acquired by GNSS stations; the multi-source heterogeneous sensor data stream also includes pulsed rainfall data, which is acquired by rain gauges.

3. The dynamic identification system for complex high-altitude landslide risks based on multi-source InSAR collaboration as described in claim 1, characterized in that, The state prediction module iteratively updates the hidden states of each heterogeneous node, including: By performing graph convolution operations, the hidden states of each neighboring node at the previous time step are weighted and aggregated to obtain the comprehensive influence representing the spatial neighborhood transmission. The combined influence and the normalized feature vector of the heterogeneous node at the current moment are fed into the gate control loop unit to calculate the hidden state at the current moment.

4. The complex high-altitude landslide risk dynamic identification system based on multi-source InSAR collaboration according to claim 1, characterized in that, The spatiotemporal graph neural network model is trained using a hybrid loss function, which includes prediction accuracy loss and physical regularization term.

5. The complex high-altitude landslide risk dynamic identification system based on multi-source InSAR collaboration according to claim 4, characterized in that, Physical regularization terms are used to impose physical constraints during model training; The physical constraint is that the deformation rate of the landslide should not slow down when effective rainfall occurs.

6. The complex high-altitude landslide risk dynamic identification system based on multi-source InSAR collaboration according to claim 1, characterized in that, The risk assessment module calculates a risk precursor index, including: Using stable historical monitoring data, a set of hidden state vectors representing the normal state is generated through a trained spatiotemporal graph neural network model. Based on the hidden state vector representing the normal state, the average hidden state and standard deviation of each node are calculated. The risk precursor index is obtained by measuring the Euclidean distance between the current hidden state and the average hidden state, and then normalizing it using the standard deviation.

7. The dynamic identification system for complex high-altitude landslide risks based on multi-source InSAR collaboration according to claim 1, characterized in that, Tiered early warning includes: When the risk precursor index is less than or equal to the first preset threshold, it is judged as a normal state and no warning is issued. When the risk precursor index is greater than the first preset threshold and less than or equal to the second preset threshold, a warning at the attention level is issued. When the risk precursor index exceeds the second preset threshold, a risk level warning is issued.

Images