Fine-grained multi-element geological environment associated landslide susceptibility evaluation method
By constructing a knowledge graph of landslide susceptibility and using the graph neural network model TerraRiskNet, landslide susceptibility is analyzed in a fine-grained manner. This solves the problem of integrating multi-source data in complex geological environments, improves the accuracy and applicability of landslide susceptibility analysis, and provides scientific risk warnings.
Patent Information
- Application Number
- CN202510987110.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-17
- Publication Date
- 2025-10-28
AI Technical Summary
Existing landslide susceptibility analysis methods struggle to effectively integrate multi-source heterogeneous data in complex geological environments, leading to inaccurate evaluation results, especially with insufficient generalization performance in cases of environmental heterogeneity and sample scarcity.
A fine-grained, multi-element geological environment-related landslide susceptibility assessment method is adopted. By constructing a landslide susceptibility knowledge graph, using topographic triangles as basic units, and combining it with the graph neural network model TerraRiskNet, multi-factor normalized edge weights and time decay factor constraints are applied to adjust the local heterogeneity of nodes and optimize the knowledge graph embedding, thereby realizing landslide susceptibility assessment.
It improves the accuracy and applicability of landslide susceptibility analysis, effectively solves the problem of landslide susceptibility assessment under complex environmental conditions, and provides a scientific basis for risk warning.
Smart Images
Figure CN120853338A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of geological disaster risk analysis and prediction technology, and in particular to a method for evaluating landslide susceptibility based on fine-grained multi-factor geological environment correlation. Background Art
[0002] Accurate identification and quantitative prediction of landslide-prone areas have become a hot topic and a challenge in the field of landslide research. Traditional landslide proneness assessment methods based on statistical and physical models often struggle to effectively address the prediction requirements in complex geological environments. While machine learning-based assessment methods possess strong nonlinear feature mining capabilities, they still suffer from insufficient generalization performance when faced with environmental heterogeneity and scarce samples. Therefore, effectively utilizing geological environmental correlations to improve the accuracy of landslide proneness assessment has become a key issue that needs to be addressed.
[0003] In complex mountainous environments, landslide hazards typically exhibit significant concealment, suddenness, and uncertainty, leading to marked limitations in the assessment capabilities of traditional methods. Landslide formation is influenced by a combination of factors, including geological structure, topographic features, climatic conditions, and human activities. These factors vary significantly in time and space; therefore, effectively integrating this multi-source, heterogeneous data presents a major challenge for landslide prediction.
[0004] Current landslide susceptibility analysis methods mostly use equal grids to divide the study area, representing the characteristics within each grid by taking an average value. Furthermore, the grids are relatively independent and lack correlation. This is seriously inconsistent with real-world scenarios, leading to inaccurate evaluation results. Summary of the Invention
[0005] The purpose of this invention is to address the shortcomings of the prior art by providing a fine-grained, multi-element geological environment correlation landslide susceptibility assessment method. This method utilizes elevation-based topographic triangles for fine-grained study area division, which can effectively solve the problem of abrupt changes in features within the grid.
[0006] To achieve the above objectives, the present invention adopts the following technical solution: This invention provides a method for evaluating landslide susceptibility based on fine-grained multi-factor geological environment correlation, comprising the following steps: S1. Constructing a landslide susceptibility knowledge graph: Based on landslide disaster examples and landslide disaster assessments, construct the ontology layer of the knowledge graph. Extract preprocessed text from the fine-tuned UIE model and link entities. Integrate the knowledge through the TransR model to construct the landslide susceptibility knowledge graph, and store it in the form of a Neo4j graph database. S2. Construction of graph structure; S3 and TerraRiskNet for landslide susceptibility assessment; S4. Construction of multi-factor normalized edge weights; S5, Adjustment of edge weights under time decay factor constraints; S6. Adaptive adjustment of local heterogeneity at nodes; S7. Optimization of knowledge graph updates.
[0007] Furthermore, S2 specifically refers to: Based on the existing landslide susceptibility knowledge graph, establish the relationships between geographical unit entities to form a graph structure to express the environmental relevance of different regions; The original terrain data is transformed into a terrain triangle network, and the maximum Z tolerance method is used to simplify the terrain data to generate terrain TIN data. Each TIN triangle unit is used as the basic unit for landslide susceptibility assessment. For each node in the topographic triangle, the gradient needs to be calculated on each evaluation factor. For soil and rock types and vegetation types, each category is converted into a binary vector. For two adjacent nodes... and The weights of the edges are established using the weighted gradient method: ; in, A vector representing the type of soil and rock mass at node 𝑖; A vector representing the type of soil and rock mass at node 𝑗; A vector representing the vegetation type at node 𝑖; A vector representing the vegetation type at node 𝑗; It is a node and nodes Spatial distance between them; and It is the coefficient of the contribution of soil and rock type and vegetation type to edge weight; After calculating the weight of each edge, an adjoining matrix is constructed.
[0008] Furthermore, S3 specifically refers to: The constructed adjacency matrix is input into TerraRiskNet. The graph convolutional layer learns the features of nodes by weighted aggregation of the neighbor information of each node. For the _th ... Layer node features Use GCN to update node features to obtain updated node features. : ; For activation functions; It is a symmetric normalized degree matrix; for =A + I adjacency matrix A plus identity matrix A; for The degree matrix; For the first Layer weights; Batch normalization is added after the graph convolutional layer to accelerate training and stabilize the learning process: ; in, and The scaling and offset parameters are for learning; To prevent division by zero of constants; The mean of the features; The variance is a characteristic.
[0009] To avoid the training difficulties of multi-layer networks, residual connections are added. Adding this to the processed output creates a deeper learning capability: ; in, For input features; The features are processed by graph convolution, batch normalization, and activation function. Calculate the attention coefficient between each pair of nodes: ; in, For nodes and nodes Attention coefficient; For transpose; For connection operation; For nodes The set of neighboring nodes; This is the weight matrix; and They are nodes and nodes The input feature vector; The update of each node is represented as follows: ; For Dropout of node features, the formula is: ; in, The input feature matrix; Let be a mask matrix consisting of binary random variables, each element of which is dropped independently with probability P, where P is the dropout ratio of Dropout.
[0010] Furthermore, S4 specifically includes: The formula for calculating edge weight is: ; in, For nodes In the The values that can be taken on each factor; For the first The weighting coefficients of each factor; To prevent division by zero errors involving tiny positive numbers.
[0011] Furthermore, S5 specifically involves proposing a time-weighted decay mechanism for dynamic factors such as rainfall and soil moisture that change significantly over time. The calculation formula is as follows: ; in, Initial edge weights for multiple factors; This represents the time difference between related data of adjacent nodes. This is the time decay coefficient.
[0012] Furthermore, S6 specifically involves defining a node local heterogeneity index to characterize the heterogeneity within different geological units. : ; in, Represents a node The set of adjacent nodes; Edge weights are adjusted based on local anisotropy: .
[0013] Furthermore, S7 specifically involves: defining projection constraints for knowledge fusion based on the TransR model, assuming entity embedding is... Relational embedding is Then the projected vector is calculated as follows: ; in, A projection matrix specific to the relation.
[0014] The beneficial effects of this invention are as follows: the fine-grained multi-element geological environment correlation landslide susceptibility evaluation method proposed in this invention can effectively incorporate geological environment correlation into the landslide susceptibility analysis method, thereby improving the accuracy of landslide susceptibility analysis results; This method constructs a knowledge graph for landslide susceptibility assessment, uses topographic triangles as the smallest evaluation unit, establishes a multi-factor related landslide susceptibility analysis network, and uses a graph neural network model to analyze landslide susceptibility. It can effectively solve the problems of accuracy and applicability in landslide susceptibility assessment under complex environmental conditions. Attached Figure Description
[0015] Figure 1 The process of constructing a knowledge graph of landslide susceptibility; Figure 2 This describes the process of constructing a graph structure based on environment associations. Detailed Implementation
[0016] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0017] A fine-grained, multi-factor geological environment-related landslide susceptibility assessment method includes the following steps: Please see Figure 1 S1. Constructing a landslide susceptibility knowledge graph: Based on landslide disaster examples and landslide disaster assessments, construct the ontology layer of the knowledge graph. Extract preprocessed text from the fine-tuned UIE model and link entities. Integrate the knowledge through the TransR model to construct the landslide susceptibility knowledge graph, and store it in the form of a Neo4j graph database. S2. Construction of graph structure; S3 and TerraRiskNet for landslide susceptibility assessment; S4. Construction of multi-factor normalized edge weights; S5, Adjustment of edge weights under time decay factor constraints; S6. Adaptive adjustment of local heterogeneity at nodes; S7. Optimization of knowledge graph updates.
[0018] Specifically, S2 is: Based on the existing landslide susceptibility knowledge graph, establish the relationships between geographical unit entities to form a graph structure to express the environmental relevance of different regions; The original terrain data is transformed into a terrain triangle network, and the maximum Z tolerance method is used to simplify the terrain data to generate terrain TIN data. Each TIN triangle unit is used as the basic unit for landslide susceptibility assessment. Specifically, the Maximum Z-Tolerance method extracts data features based on data accuracy. This method first generates an initial TIN surface. In each iteration, the point with the largest difference between the original terrain and the TIN surface is added to the TIN surface, and candidate edges for new polygons are determined based on the lowest mean elevation error, generating a simplified terrain TIN. The point E with the largest current error is about to be inserted; after inserting E into triangle ABD, the edges of ABD become candidate edges {AD, AB, BD}. There are no quadrilaterals containing AD or AB, and they can be removed from {AD, AB, BD}. Quadrilateral EBCD can be triangulated using two methods, using diagonals BD or EC; if BD produces the lowest mean elevation error, then there is a new triangle EBD and the old triangle CDB, and then the point with the largest error is found again. The above steps are repeated until the maximum elevation difference between the terrain TIN data and the original terrain data satisfies the threshold.
[0019] For each node in the topographic triangle, the gradient needs to be calculated on each evaluation factor. For soil and rock types and vegetation types, each category is converted into a binary vector. For two adjacent nodes... and The weights of the edges are established using the weighted gradient method: ; in, A vector representing the type of soil and rock mass at node 𝑖; A vector representing the type of soil and rock mass at node 𝑗; A vector representing the vegetation type at node 𝑖; A vector representing the vegetation type at node 𝑗; It is a node and nodes Spatial distance between them; and It is the coefficient of the contribution of soil and rock type and vegetation type to edge weight; After calculating the weight of each edge, an adjoining matrix is constructed.
[0020] By generating node embeddings using TerraRiskNet, the model first aggregates information from neighboring nodes using GCN to form stable feature representations. Then, a GAT layer is used to further optimize these features, focusing on those crucial for landslide susceptibility assessment. This multi-layered processing structure enables the model not only to effectively capture local geographic features but also to dynamically adjust the contribution of each feature according to the needs of global prediction.
[0021] Specifically, S3 is: The constructed adjacency matrix is input into TerraRiskNet. The graph convolutional layer learns the features of nodes by weighted aggregation of the neighbor information of each node. For the _th ... Layer node features Use GCN to update node features to obtain updated node features. : ; For activation functions; It is a symmetric normalized degree matrix; for =A + I adjacency matrix A plus identity matrix A; for The degree matrix; For the first Layer weights; Batch normalization is added after the graph convolutional layer to accelerate training and stabilize the learning process: ; in, and The scaling and offset parameters are for learning; To prevent division by zero of constants; The mean of the features; The variance is a characteristic.
[0022] To avoid the training difficulties of multi-layer networks, residual connections are added. Adding this to the processed output creates a deeper learning capability: ; in, For input features; The features are processed by graph convolution, batch normalization, and activation function. Calculate the attention coefficient between each pair of nodes: ; in, For nodes and nodes Attention coefficient; For transpose; For connection operation; For nodes The set of neighboring nodes; This is the weight matrix; and They are nodes and nodes The input feature vector; The update of each node is represented as follows: ; For Dropout of node features, the formula is: ; in, The input feature matrix; Let be a mask matrix consisting of binary random variables, each element of which is dropped independently with probability P, where P is the dropout ratio of Dropout.
[0023] The node representation of each evaluation unit is generated through iterative optimization and then processed by a fully connected classification layer to determine the landslide susceptibility level. During training, the sample set is randomly divided into a 70% training set and a 30% validation set to ensure the model's generalization ability. The node representations are processed by the fully connected layer, and the probability of a landslide occurring at each node is determined using a softmax function. Finally, the probability of each evaluation unit is graded to generate a landslide susceptibility assessment map, providing a scientific basis for landslide risk early warning and emergency management.
[0024] Specifically, S4 is: The formula for calculating edge weight is: ; in, For nodes In the The values that can be taken on each factor; For the first The weighting coefficients of each factor; To prevent division by zero errors involving tiny positive numbers.
[0025] Specifically, S5 proposes a time-weighted decay mechanism for dynamic factors such as rainfall and soil moisture that change significantly over time. The calculation formula is as follows: ; in, Initial edge weights for multiple factors; This represents the time difference between related data of adjacent nodes. This is the time decay coefficient.
[0026] Specifically, S6 involves defining a nodal local heterogeneity index to characterize the heterogeneity within different geological units. : ; in, Represents a node The set of adjacent nodes; Edge weights are adjusted based on local anisotropy: .
[0027] Specifically, S7 involves defining projection constraints for knowledge fusion based on the TransR model, where the entity embedding is... Relational embedding is Then the projected vector is calculated as follows: ; in, A projection matrix specific to the relation.
[0028] A fine-grained, multi-factor geological environment-related landslide susceptibility assessment method is proposed. First, knowledge is extracted from unstructured texts on the internet, including scientific literature, landslide disaster reports, and online encyclopedias. A fine-tuned UIE model is then used to construct a landslide disaster knowledge graph for landslide susceptibility assessment. The UIE model solves the information extraction tasks of entities, relationships, and events simultaneously using only one model. This framework breaks down the information extraction task into two elements: points and relationships. Hints built based on points and relationships are called Structural Graph Instructors (SSIs) and are linked to the text. The sequence of SSIs and text is input into the UIE model, and the output is a Structured Extraction Language (SEL), thus enabling unified modeling of various complex information extraction tasks from large-scale texts. The UIE model learns the probabilities of the text itself and uses these probabilities to predict output entities and relationships, thereby reducing the need for large-scale supervised text extraction and, consequently, large supervised datasets. In this study, the efficiency of knowledge extraction is significantly improved by fine-tuning the UIE model used for joint entity-relation extraction. The ontology layer of the knowledge graph is constructed based on "landslide disaster examples" and "landslide disaster assessment". The relationships between entities are extracted through the UIE model, and the extracted knowledge is integrated using the TransR model to construct a knowledge graph for landslide susceptibility analysis.
[0029] Please see Figure 2 Secondly, by utilizing the abstract expressive power of graph structures, a graph network constrained by geological environment is established based on the knowledge graph of landslide susceptibility. Topographic triangles are used as evaluation unit nodes, and the changes in geological environment are comprehensively reflected by the gradient of the changes in evaluation factors among evaluation units, thus demonstrating the geological environment relationship between each evaluation unit in geographic space.
[0030] The constructed graph structure is input into the landslide susceptibility assessment model TerraRiskNet. TerraRiskNet efficiently combines GCN, Batch Normalization, Residual Connection, Self-Attention, and Dropout layers to fully leverage the advantages of each component. First, the GCN layer aggregates information from neighboring nodes through node features and convolutional operations to form stable local feature representations, helping the model effectively capture local features in the geographic environment and correlate these features with the probability of landslide occurrence. By learning low-dimensional representations of nodes, GCN can preserve the spatial topology of the graph while extracting key geographic and environmental features.
[0031] To further enhance feature representation capabilities, the model introduces residual connections, allowing the output of each layer to be directly added to the input, thus avoiding potential information loss or attenuation during multi-layer transmission. Furthermore, batch normalization is applied to the output of each layer, effectively accelerating model training and improving its stability.
[0032] Next, TerraRiskNet further refines node features through a self-attention mechanism, assigning different attention weights to each node and its neighbors. Through multi-head self-attention, the model can capture complex relationships between nodes and identify which geological features or environmental conditions are more important in specific contexts, thereby finely adjusting the model's focus.
[0033] Finally, Dropout is used to prevent overfitting and enhance the model's generalization ability. Through this hierarchical structure, TerraRiskNet successfully combines the stability of GCN with the flexibility of the self-attention mechanism, enabling the model to not only capture local geographical features but also dynamically adjust the contribution of each feature in the global prediction, thereby effectively improving the accuracy and reliability of landslide susceptibility assessment.
[0034] The embodiments described above are merely illustrative of implementation methods of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be defined by the appended claims.
Claims
1. A fine-grained, multi-factor geological environment correlation method for landslide susceptibility assessment, characterized in that, The following steps are involved: S1. Constructing a landslide susceptibility knowledge graph: Based on landslide disaster examples and landslide disaster assessments, construct the ontology layer of the knowledge graph. Extract preprocessed text from the fine-tuned UIE model and link entities. Integrate the knowledge through the TransR model to construct the landslide susceptibility knowledge graph, and store it in the form of a Neo4j graph database. S2. Construction of graph structure; S3 and TerraRiskNet for landslide susceptibility assessment; S4. Construction of multi-factor normalized edge weights; S5, Adjustment of edge weights under time decay factor constraints; S6. Adaptive adjustment of local heterogeneity at nodes; S7. Optimization of knowledge graph updates.
2. The method for evaluating landslide susceptibility based on fine-grained multi-factor geological environment correlation according to claim 1, characterized in that, Specifically, S2 is: Based on the existing landslide susceptibility knowledge graph, establish the relationships between geographical unit entities to form a graph structure to express the environmental relevance of different regions; The original terrain data is transformed into a terrain triangle network, and the maximum Z tolerance method is used to simplify the terrain data to generate terrain TIN data. Each TIN triangle unit is used as the basic unit for landslide susceptibility assessment. For each node in the topographic triangle, the gradient needs to be calculated on each evaluation factor. For soil and rock types and vegetation types, each category is converted into a binary vector. For two adjacent nodes... and The weights of the edges are established using the weighted gradient method: ； in, A vector representing the type of soil and rock mass at node 𝑖; A vector representing the type of soil and rock mass at node 𝑗; A vector representing the vegetation type at node 𝑖; A vector representing the vegetation type at node 𝑗; It is a node and nodes Spatial distance between them; and It is the coefficient of the contribution of soil and rock type and vegetation type to edge weight; After calculating the weight of each edge, an adjoining matrix is constructed.
3. The method for evaluating landslide susceptibility based on fine-grained multi-factor geological environment correlation according to claim 2, characterized in that, Specifically, S3 is: The constructed adjacency matrix is input into TerraRiskNet. The graph convolutional layer learns the features of nodes by weighted aggregation of the neighbor information of each node. For the _th ... Layer node features Use GCN to update node features to obtain updated node features. : ； For activation functions; It is a symmetric normalized degree matrix; for =A + I adjacency matrix A plus identity matrix A; for The degree matrix; For the first Layer weights; Batch normalization is added after the graph convolutional layer to accelerate training and stabilize the learning process: ； in, and The scaling and offset parameters are for learning; To prevent division by zero of constants; The mean of the features; The variance of the feature; To avoid the training difficulties of multi-layer networks, residual connections are added. Adding this to the processed output creates a deeper learning capability: ； in, For input features; The features are processed by graph convolution, batch normalization, and activation function. Calculate the attention coefficient between each pair of nodes: ； in, For nodes and nodes Attention coefficient; For transpose; For connection operation; For nodes The set of neighboring nodes; This is the weight matrix; and They are nodes and nodes The input feature vector; The update of each node is represented as follows: ； For Dropout of node features, the formula is: ； in, The input feature matrix; Let be a mask matrix consisting of binary random variables, each element of which is dropped independently with probability P, where P is the dropout ratio of Dropout.
4. The method for evaluating landslide susceptibility based on fine-grained multi-factor geological environment correlation according to claim 3, characterized in that, Specifically, S4 is: The formula for calculating edge weight is: ； in, For nodes In the The values that can be taken on each factor; For the first The weighting coefficients of each factor; To prevent division by zero errors involving tiny positive numbers.
5. The method for evaluating landslide susceptibility based on fine-grained multi-factor geological environment correlation according to claim 4, characterized in that, Specifically, S5 proposes a time-weighted decay mechanism for dynamic factors such as rainfall and soil moisture that change significantly over time. The calculation formula is as follows: ； in, Initial edge weights for multiple factors; This represents the time difference between related data of adjacent nodes. This is the time decay coefficient.
6. The method for evaluating landslide susceptibility based on fine-grained multi-factor geological environment correlation according to claim 5, characterized in that, Specifically, S6 involves defining a nodal local heterogeneity index to characterize the heterogeneity within different geological units. : ； in, Represents a node The set of adjacent nodes; Edge weights are adjusted based on local anisotropy: 。 7. The method for evaluating landslide susceptibility based on fine-grained multi-factor geological environment correlation according to claim 6, characterized in that, Specifically, S7 involves defining projection constraints for knowledge fusion based on the TransR model, where the entity embedding is... Relational embedding is Then the projected vector is calculated as follows: ； in, A projection matrix specific to the relation.