Intelligent landslide susceptibility prediction method and system based on graph convolutional network

Abstract

The invention relates to the technical field of geological disasters and geographic information, in particular to a landslide susceptibility intelligent prediction method and system based on a graph convolutional network, and the method comprises the following steps: obtaining a landslide regulation factor, and evaluating and selecting the landslide regulation factor to obtain a landslide susceptibility evaluation factor; performing correlation analysis and feature importance sorting on the landslide susceptibility evaluation factors to obtain landslide susceptibility evaluation optimization factors; in combination with the landslide susceptibility evaluation optimization factor, constructing a slope unit structure chart based on contour line features of the slope; and establishing a graph convolutional neural network structure, and combining the slope unit structure graph and the graph convolutional neural network structure to realize intelligent prediction of landslide susceptibility. According to the invention, through a graph convolutional neural network structure, a structure graph is constructed in combination with contour lines, so that terrain association is accurately described, grid limitation is broken through, a nonlinear relation is captured, landslide prediction precision is effectively improved, and an intelligent scheme is provided for landslide disaster early warning.

Description

Technical Field

[0001] This invention relates to the field of geological disaster and geographic information technology, and in particular to an intelligent prediction method and system for landslide susceptibility based on graph convolutional networks. Background Technology

[0002] In recent years, machine learning models have been widely used in landslide susceptibility assessment due to their ability to effectively capture the relationship between landslides and disaster-causing factors, and have achieved good predictive results thanks to their ability to process large-scale data. However, in most machine learning-based susceptibility assessment studies, the models usually use little preprocessing or directly extract feature information automatically from the raw data, forming so-called "black box" models that ignore the spatial correlation and mutual influence between adjacent grids or slope units.

[0003] In overcoming the limitations of traditional machine learning methods, neural network methods, especially Graph Convolutional Networks (GCNs), capture the topological relationships and feature associations of graph-structured data. They possess strong versatility, topological representation capabilities, and low-pass filtering properties, effectively improving prediction accuracy in landslide susceptibility research. However, current applications of GCN models in landslide susceptibility assessment still rely on the center points of evaluation units such as slope units or topographic units as nodes for mapping. These mapping methods have limitations in accurately extracting key topographic features of slope units, often failing to comprehensively capture micro-topographic features within slope units that influence landslide occurrence, such as slope, aspect, and curvature. The spatial autocorrelation characteristics of multiple influencing factors of landslides still need further exploration. How to more accurately capture topographic changes and landslide precursors has become a key issue in landslide susceptibility assessment research.

[0004] From the perspective of evaluation unit division, the mainstream evaluation units are grid units and slope units. In engineering practice, slope units are mainly used. A representative value of a certain feature (such as slope) is selected by statistical methods to replace the feature of the entire slope for subsequent model prediction. From the perspective of the application of machine learning in landslide susceptibility assessment, traditional machine learning methods such as decision trees, random forests, and support vector machines often evaluate a single grid or a single slope independently.

[0005] Existing landslide susceptibility assessment methods suffer from the following technical shortcomings: First, they often use statistical methods to select representative values ​​of a certain feature (such as slope) to represent the characteristics of the entire slope. However, slope units cover a large area of ​​the land surface and contain rich geographical information. Using only a single point to represent the slope is insufficient to reflect its environmental characteristics and cannot fully capture the micro-topographic features within the slope unit that affect landslide occurrence, such as slope, aspect, and curvature. Second, they evaluate individual grids or individual slopes independently, ignoring the spatial correlation and mutual influence between adjacent grids or slope units, which limits the accuracy and reliability of the evaluation results. Finally, they require a large number of landslide samples or detailed geological field surveys to obtain a large number of accurate parameters, making them extremely difficult to apply on a large scale. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a method and system for intelligent prediction of landslide susceptibility based on graph convolutional networks.

[0007] To achieve the above objectives, in a first aspect, the present invention provides an intelligent prediction method for landslide susceptibility based on graph convolutional networks. The method includes the following steps: obtaining landslide adjustment factors; evaluating and selecting the landslide adjustment factors to obtain landslide susceptibility evaluation factors; performing correlation analysis and feature importance ranking on the landslide susceptibility evaluation factors to obtain optimized landslide susceptibility evaluation factors; constructing a slope unit structure map based on the contour line features of the slope using the optimized landslide susceptibility evaluation factors; establishing a graph convolutional neural network structure; and combining the slope unit structure map and the graph convolutional neural network structure to achieve intelligent prediction of landslide susceptibility. This invention achieves landslide susceptibility prediction by fusing landslide evaluation factors with spatial topological relationships through graph convolutional neural networks; the factor optimization mechanism removes redundant features through correlation analysis, improving the quality of model input; the slope unit structure map constructed based on contour lines accurately depicts the spatial correlation of terrain, breaking through the limitations of traditional raster data; the graph convolutional neural network structure effectively captures the nonlinear relationships between nodes, improving prediction accuracy; it combines spatial expressiveness and computational efficiency, providing an intelligent solution for geological disaster early warning, and has important practical value for disaster prevention and mitigation.

[0008] Optionally, the step of obtaining landslide adjustment factors, evaluating and selecting landslide susceptibility evaluation factors, includes: the landslide susceptibility evaluation factors include topographic factors, geological factors, human engineering activity factors, and triggering factors; the topographic factors include elevation, slope, aspect, curvature, surface roughness, and surface relief; the geological factors include lithology and distance from faults; the human engineering activity factors include distance from roads; and the triggering factors include distance from water systems and rainfall. This invention, through systematic classification and screening of landslide influencing factors, integrates multi-dimensional key elements, comprehensively covering the main causes of landslides, and improving the scientific rigor and relevance of factor selection by eliminating redundant features through evaluation and optimization; it enhances the ability of subsequent prediction models to characterize landslide causes, providing structured and highly relevant feature inputs for landslide susceptibility assessment.

[0009] Optionally, the step of performing correlation analysis and feature importance ranking on the landslide susceptibility evaluation factors to obtain optimized landslide susceptibility evaluation factors includes: performing correlation analysis on the landslide susceptibility evaluation factors based on Pearson correlation coefficients to obtain non-redundant features in the landslide susceptibility evaluation factors; obtaining variable importance scores for the landslide susceptibility evaluation factors based on the Gini index; ranking feature importance based on the variable importance scores to obtain feature ranking results; and combining the non-redundant features and the feature ranking results to obtain the optimized landslide susceptibility evaluation factors. This invention eliminates redundant features through Pearson correlation coefficient analysis, avoids multicollinearity interference between factors, and reduces model complexity; it accurately identifies key influencing factors by combining feature importance ranking with the Gini index, improving the model's focus on core elements; and making the input features more representative.

[0010] Optionally, the step of constructing a slope unit structure diagram based on the contour line characteristics of the slope, in conjunction with the landslide susceptibility evaluation optimization factor, includes: generating the contour lines based on a digital elevation model, and generating nodes on the contour lines at equal intervals to form an initial node set; traversing the nodes in the initial node set one by one to obtain the two nearest lower-level nodes between each node and the next layer of contour lines, so as to establish node connection edges; obtaining the node steepness of adjacent contour lines, and obtaining the lower-level node attributes based on the node steepness; and constructing the slope unit structure diagram by combining the initial node set, the node connection edges, and the node attributes. This invention accurately depicts the geometric features of slope terrain through contour line generation and node processing; the equidistant node layout ensures uniform spatial coverage and avoids feature omission; topological relationships are constructed based on the connection of nearest nodes to effectively capture terrain continuity and structural hierarchy; the node steepness attribute quantifies slope variation characteristics and directly relates to key factors of landslide risk; the constructed slope unit structure map combines spatial topological information and geological attribute characteristics, preserving terrain details while strengthening the expression of risk-related elements, thereby improving the ability to characterize and predict the spatial distribution of landslide susceptibility.

[0011] Optionally, obtaining the node steepness of adjacent contour lines and obtaining lower-level node attributes based on the node steepness includes: in three-dimensional space, the node steepness satisfies the following relationship: in, For node steepness, The three-dimensional coordinates of the first node are: The three-dimensional coordinates of the second node are used; a node steepness threshold is set. When the node steepness is equal to or greater than the node steepness threshold, the lower-level node is classified as a steep slope node; when the node steepness is less than the node steepness threshold, the lower-level node is classified as a flowing node. This invention calculates the vertical change rate between nodes using three-dimensional coordinates to quantify the steepness of the terrain; it sets a threshold to classify steep slope nodes and flowing nodes, accurately distinguishing the steepness characteristics of the terrain, transforming the geometric attributes of the terrain into calculable feature identifiers, and endowing the slope unit structure map with steepness semantic information, thereby improving the spatial sensitivity and accuracy of landslide susceptibility prediction.

[0012] Optionally, the step of establishing a graph convolutional neural network structure and combining the slope unit structure map with the graph convolutional neural network structure to achieve intelligent prediction of landslide susceptibility includes: establishing a slope unit dataset for the study area; obtaining a slope unit subgraph by combining the slope unit dataset and the slope unit structure map; extracting node features from the slope unit subgraph to obtain feature vectors; constructing a feature matrix based on the feature vectors; and obtaining a landslide probability distribution by combining the graph convolutional neural network structure and the feature matrix to achieve intelligent prediction of landslide susceptibility. This invention integrates spatial topological relationships by constructing slope unit subgraphs, fuses multi-dimensional evaluation factors by combining feature matrices, and utilizes a graph convolutional neural network structure to effectively aggregate node features and neighbor information, capture the spatial dependence and nonlinear relationship of landslide risk, deeply integrate structured spatial features and attribute features, improve the ability to characterize landslide susceptibility under complex terrain conditions, achieve high-precision, spatially subdivided probability prediction, provide intelligent and quantitative decision support for landslide disaster early warning, and enhance the scientific nature and timeliness of disaster prevention and mitigation.

[0013] Optionally, obtaining the landslide probability distribution by combining the graph convolutional neural network structure and the feature matrix includes: the graph convolutional neural network structure includes an input layer, multiple convolutional layers, a max-pooling layer, and an output layer; the feature matrix is ​​input into the graph convolutional neural network structure through the input layer; the feature matrix is ​​convolved using the multiple convolutional layers to obtain an output feature matrix; global graph embedding is performed on the output feature matrix based on the max-pooling layer, and a binary classification prediction of landslide-non-landslide is performed based on the output feature matrix to obtain the landslide probability distribution; the landslide probability distribution is output based on the output layer. This invention achieves landslide probability prediction through a layered graph convolutional architecture, effectively integrating spatial structure learning and attribute feature analysis, improving the representation ability of the graph convolutional neural network structure for complex terrain patterns, and achieving efficient and accurate quantitative assessment of landslide susceptibility.

[0014] Optionally, the step of using the multiple convolutional layers to perform convolution processing on the feature matrix to obtain the output feature matrix includes: in, For the first The feature matrix output after processing by the convolutional layer For activation function, for The degree matrix, This is the corrected adjacency matrix. For the first The feature matrix output after processing by the convolutional layer For the first The learnable weight matrix of the convolutional layer. This invention effectively captures spatial dependencies through multiple convolutional layers, improving the ability of graph convolutional neural network structures to express complex terrain features and enhancing the sensitivity of landslide susceptibility prediction.

[0015] Optionally, the graph convolutional neural network structure includes: training the graph convolutional neural network structure by introducing a cross-entropy loss function, wherein the cross-entropy loss function satisfies the following relationship: in, This represents the cross-entropy loss value. For the total sample size, For index variables, For the first The true label of each sample For the graph convolutional neural network structure, the first... This invention calculates the probability value of a sample being predicted as belonging to the "landslide" category. By quantifying the difference between the predicted probability and the true label, it effectively guides the optimization of graph convolutional model parameters, improves the ability of the graph convolutional neural network structure to distinguish between landslides and non-landslides, and enhances the spatial feature representation ability through the cross-entropy loss function, effectively improving the prediction accuracy of the landslide probability distribution.

[0016] Secondly, this invention provides an intelligent landslide susceptibility prediction system based on graph convolutional networks. The system executes the intelligent landslide susceptibility prediction method based on graph convolutional networks provided by this invention. The system includes an input device, an output device, a processor, and a memory, all interconnected. The memory stores a computer program, which includes program instructions, and the processor is configured to call these instructions. This invention achieves intelligent landslide susceptibility prediction through high-performance hardware collaboration, supports efficient data processing, integrates spatial topology learning and attribute feature analysis, accurately captures nonlinear topographic relationships, improves prediction accuracy, enhances the scientific rigor and timeliness of disaster prevention decisions, and provides an intelligent solution for landslide disaster early warning. Attached Figure Description

[0017] Figure 1 This is a flowchart of an intelligent prediction method for landslide susceptibility based on graph convolutional networks according to an embodiment of the present invention. Figure 2 This is a schematic diagram of a slope unit structure that has experienced a landslide, according to an embodiment of the present invention. Figure 3 This is a sub-diagram of a ramp element in an embodiment of the present invention; Figure 4 This is a schematic diagram of feature extraction according to an embodiment of the present invention; Figure 5This is a schematic diagram of the graph convolutional neural network structure according to an embodiment of the present invention; Figure 6 This is a schematic diagram of the receiver operating characteristic curves of different models in embodiments of the present invention; Figure 7 This is a schematic diagram of the precision-recall curves for different models in embodiments of the present invention. Figure 8 This is a framework diagram of a landslide susceptibility intelligent prediction system based on graph convolutional networks, according to an embodiment of the present invention. Detailed Implementation

[0018] Specific embodiments of the present invention will now be described in detail. It should be noted that the embodiments described herein are for illustrative purposes only and are not intended to limit the invention. In the following description, numerous specific details are set forth in order to provide a thorough understanding of the invention. However, it will be apparent to those skilled in the art that these specific details are not necessary to practice the invention. In other instances, well-known circuits, software, or methods have not been specifically described to avoid obscuring the invention.

[0019] Throughout this specification, references to "an embodiment," "an embodiment," "an example," or "an example" mean that a particular feature, structure, or characteristic described in connection with that embodiment or example is included in at least one embodiment of the invention. Therefore, the phrases "in an embodiment," "in an embodiment," "an example," or "an example" appearing in various places throughout the specification do not necessarily refer to the same embodiment or example. Furthermore, specific features, structures, or characteristics can be combined in one or more embodiments or examples in any suitable combination and / or sub-combination. Moreover, those skilled in the art will understand that the illustrations provided herein are for illustrative purposes and are not necessarily drawn to scale.

[0020] Please see Figure 1 One embodiment of the present invention provides an intelligent prediction method for landslide susceptibility based on graph convolutional networks, the method comprising the following steps: S1. Obtain landslide adjustment factors, evaluate and select the landslide susceptibility evaluation factors.

[0021] The occurrence of landslides is influenced by a combination of basic geological environmental conditions (such as topography and geological conditions) and triggering factors (such as earthquakes and human activities).

[0022] In this embodiment, drawing on relevant research findings both domestically and internationally, and considering the landslide situation in the study area, the following 11 landslide susceptibility evaluation factors were initially selected: topographic factors (elevation, slope, aspect, curvature, surface roughness, and surface relief); geological factors (lithology and distance from faults); human engineering activity factors (distance from roads); and triggering factors (distance from water systems and rainfall).

[0023] S2. Correlation analysis and feature importance ranking are performed on the landslide susceptibility evaluation factors to obtain optimized landslide susceptibility evaluation factors.

[0024] In this embodiment, correlation analysis is performed on the landslide susceptibility evaluation factors based on the Pearson correlation coefficient to obtain non-redundant features in the landslide susceptibility evaluation factors; the variable importance score of the landslide susceptibility evaluation factors is obtained based on the Gini index, and the feature importance is ranked based on the variable importance score to obtain the feature ranking result; the landslide susceptibility evaluation optimization factor is obtained by combining the non-redundant features and the feature ranking result.

[0025] Feature selection before training a model is a crucial step in improving model performance and accuracy. Pearson correlation analysis is used to remove redundant features, reducing noise interference in the training data and enhancing the model's ability to learn truly useful information. The Pearson correlation coefficient, by measuring the linear correlation between two variables, can effectively identify and eliminate highly correlated features, thus preventing them from excessively influencing the weight distribution during model learning. The Pearson correlation coefficient is a statistical method for measuring the degree of linear correlation between two variables, with values ​​ranging from -1 to 1. When the correlation coefficient is close to 1 or -1, it indicates a strong positive or negative correlation between the two variables; when it is close to 0, it indicates no significant linear relationship.

[0026] Pearson correlation analysis quantifies the linear relationship between evaluation factors and identifies redundant or highly correlated features. During feature selection, excessively high correlation may indicate that certain features provide redundant information in the model; therefore, these features need to be removed or merged to avoid performance degradation due to multicollinearity. By eliminating highly correlated redundant features based on the Pearson correlation coefficient, non-redundant features are obtained, retaining the core factors most influential on landslide susceptibility and ensuring the model's simplicity and effectiveness. This not only improves the model's training efficiency but also enhances its predictive performance and generalization ability.

[0027] The above Pearson correlation coefficients satisfy the following relationship: in, The Pearson correlation coefficient is used. For covariance, As evaluation factors variance, Evaluation factors The variance.

[0028] It should be noted that, These represent different evaluation factors; for example, It can represent the slope. It can represent the distance from a water system; Covariance measures the degree to which two variables change together. Variance measures the degree of dispersion of a variable. That is, standard deviation.

[0029] In this embodiment, the Gini index of Random Forest (RF) is used to rank feature importance. Compared with other methods, the Gini index of RF has the advantages of high computational efficiency and good adaptability to various types of data. It also provides a clear ranking of feature importance, facilitating the understanding and interpretation of feature contributions in the model. By combining Pearson correlation analysis and the Gini index of RF, feature redundancy and noise are reduced while ensuring model performance, thereby effectively reducing computational complexity and the risk of model overfitting. The variable importance scores of the initially selected 11 features are calculated using the Gini index.

[0030] The Gini index above satisfies the following relationship: in, For the Gini index, The number of categories in the self-service sample set. For index variables, For nodes The sample belongs to the first The probability estimate of the class.

[0031] When the sample is binary classification data ( ),node The Gini index satisfies the following relationship: in, For the Gini index, For samples at nodes The probability estimate of belonging to any class.

[0032] Furthermore, for variable X, at node The importance of nodes The change in the Gini index before and after branching satisfies the following relationship: in, The change in the Gini index, For the Gini index, For nodes The Gini index of the left child node generated after the split. For nodes The Gini index of the right child node generated after the split.

[0033] If variables In the Appearing in the trees Next, then the variable In the The importance of each tree satisfies the following relationship: in, For variables In the Variable importance scores in a decision tree For variables In the The total number of times a splitting feature is used in a tree. For index variables, This represents the change in the Gini index.

[0034] variable The importance of the Gini index in RF satisfies the following relationship: in, For the first The average change in node splitting impurity for each variable across all trees in a randomized controlled trial (RF). The number of classification trees in the random forest. For index variables, For variables In the The importance score of variables in a decision tree.

[0035] S3. Combining the landslide susceptibility evaluation optimization factors, construct a slope unit structure diagram based on the contour line characteristics of the slope.

[0036] In this embodiment, contour lines are generated based on the digital elevation model, and nodes are generated on the contour lines at equal intervals to form an initial node set; the nodes in the initial node set are traversed one by one to obtain the two lower-level nodes that are closest to the node in the next layer of contour lines, so as to establish node connection edges; the node steepness of adjacent contour lines is obtained, and the lower-level node attributes are obtained based on the node steepness; the slope unit structure diagram is constructed by combining the initial node set, node connection edges, and node attributes.

[0037] Please see Figure 2The diagram illustrates the structural features of a slope unit where a landslide has occurred. Contour-based mapping methods can detect and identify ground conditions that may lead to landslides based on morphological characteristics. Identifying these areas allows decision-makers to focus on these geographic locations for further assessment or prediction. Contour-based mapping methods assign nodes at regular intervals along each contour line, then connect them to nodes on the next level of contour lines according to certain rules. A steepness threshold is set to identify key topographic features of landslides. Based on data collection from past landslides, combined with field investigations and analysis of existing literature, the steepness of the landslide's rear edge is often between 0.65 and 0.9. Therefore, the node steepness threshold is set to 0.7. When the node steepness is equal to or greater than the threshold, the lower-level node is classified as a sill node, identifying the sill area of ​​the slope. When the node steepness is less than the threshold, the lower-level node is classified as a flow node, identifying the downslope area of ​​the slope. The landslide's flow path lies in these areas, i.e., soil deposits or fans. Includes the following steps: The first step is to generate contour lines based on the digital elevation model, and then generate nodes on each contour line at equal intervals to form an initial node set; the spacing between contour lines and the spacing between nodes on the same contour line are both 30m.

[0038] The second step is to traverse the contour lines. layer( Each node on ) Search for its next level contour line layer( The two nearest nodes to the node) , To establish node connection edges.

[0039] The third step is to assess the contour lines. Layers and Calculate the steepness of nodes between layers and the contour lines. layer( Each node on the top layer and its corresponding contour line on the next layer. layer( There exists a node steepness between connected nodes on the ); when the node steepness is equal to or greater than a node steepness threshold, ( Nodes on ) are considered steep slope nodes, in Figure 2 The center is represented by a white hollow dot.

[0040] Fourth step, assess the contour lines Layers and Steepness between nodes in different layers, and calculation of contour lines. layer( Each node on the top layer and its corresponding contour line on the next layer. layer( There exists a node steepness between connected nodes on the . When the node steepness is less than a node steepness threshold, ( Nodes on ) are considered to be flowing nodes. Figure 2 The center is represented by solid black dots.

[0041] In three-dimensional space, two adjacent points and The steepness of nodes between them satisfies the following relationship: in, For node steepness, The three-dimensional coordinates of the first node are: The coordinates of the second node are 3D coordinates.

[0042] S4. Establish a graph convolutional neural network structure, and combine the slope unit structure diagram and the graph convolutional neural network structure to realize intelligent prediction of landslide susceptibility.

[0043] In this embodiment, a slope unit dataset for the study area is established, and a slope unit sub-graph is obtained by combining the slope unit dataset and the slope unit structure diagram. Feature vectors are obtained by extracting node features from the slope unit sub-graph, and a feature matrix is ​​constructed based on the feature vectors. The landslide probability distribution is obtained by combining the graph convolutional neural network structure and the feature matrix, thereby realizing intelligent prediction of landslide susceptibility.

[0044] For the slope unit structure map, features are extracted for each node, and steepness settings are used to distinguish steep slope nodes in the steep slope area from flow nodes in the sediment area. A graph convolutional neural network is combined to learn the structural features of the landslide body, thereby further realizing landslide identification and susceptibility mapping.

[0045] Please see Figure 3 The diagram illustrates a slope unit subgraph. A graph convolutional neural network was used to build a slope unit dataset for the study area. Based on a limited number of landslides that have occurred and slope units with identified potential hazards, the landslide susceptibility of the entire study area was assessed. By constructing each slope as a slope unit subgraph, this graph structure effectively captures the spatial and semantic relationships between nodes, providing rich contextual information for the model. The slope unit subgraph satisfies the following relationship: in, This is a sub-graph of a slope unit. It is the size of The set of nodes assigned based on contour lines. It is the set of edges that represent the adjacency relationships between nodes.

[0046] Please see Figure 4 The diagram illustrates feature extraction. Each node in the slope unit subgraph possesses a feature vector containing multiple attribute information. These features include "elevation," "slope," "aspect," "curvature," "rainfall," "distance to fault," "distance to road," "distance to water system," and "lithology." These feature vectors constitute a feature matrix with dimensions N×D, where N is the number of nodes in each subgraph, and D is the feature dimension of each node.

[0047] For example, the "rainfall" feature can reflect the impact of water on soil stability, and nodes in areas with high rainfall may have a higher risk of landslides; the "distance from fault" feature indicates the proximity of a node to a geological fault, and nodes closer to faults are more susceptible to geological activity, increasing the likelihood of landslides.

[0048] Please see Figure 5 The diagram illustrates the structure of a graph convolutional neural network. The graph convolutional neural network structure includes an input layer, multiple convolutional layers, a max-pooling layer, and an output layer. The feature matrix is ​​input into the graph convolutional neural network structure through the input layer. The multiple convolutional layers perform convolution processing on the feature matrix to obtain the output feature matrix. The max-pooling layer performs global graph embedding on the output feature matrix, and a landslide-non-landslide binary classification prediction is performed based on the output feature matrix to obtain the landslide probability distribution. The output layer then outputs the landslide probability distribution.

[0049] Let the input feature matrix be The final output feature matrix is ​​obtained after processing through three GCN layers. For the first Layer GCN ( =1,2,3), the ReLU activation function is used in the first three layers, and its calculation formula satisfies the following relationship: in, For the first The feature matrix output after processing by the convolutional layer For activation function, for The degree matrix, This is the corrected adjacency matrix. For the first The feature matrix output after processing by the convolutional layer For the first The learnable weight matrix of a convolutional layer.

[0050] The corrected adjacency matrix satisfies the following relationship: in, This is the corrected adjacency matrix. It is an adjacency matrix. It is an identity matrix.

[0051] The elements on the diagonal of the degree matrix satisfy the following relationship: in, The elements on the diagonal of the degree matrix, For the node index of the matrix, This is the corrected adjacency matrix.

[0052] Furthermore, at the last layer GCN (i.e. After outputting the data, the Softmax function is used to perform binary classification prediction of landslide / non-landslide for the sample points. The Softmax function transforms the model's feature matrix into a landslide probability distribution, such that the sum of the probabilities of each sample point being assigned to the "landslide" or "non-landslide" category is 1.

[0053] In this embodiment, since the original data of the factors influencing landslides have relatively small differences, the learning rate of the mean squared error loss function decreases as the gradient descent occurs. Therefore, a cross-entropy loss function is introduced during training. The cross-entropy loss function helps to increase the gap between correct and incorrect predictions, reducing computational errors. When the cross-entropy loss function is used as the loss function for the last layer of the neural network, the gradient of the weights in the last layer will only be related to the difference between the output value and the true value, thereby accelerating convergence. The cross-entropy loss function satisfies the following relationship: in, This represents the cross-entropy loss value. For the total sample size, For index variables, For the first The true label of each sample For the graph convolutional neural network structure, the first... The probability value of a sample being predicted as the "landslide" category.

[0054] Based on the above methods, structural information within the slope can be integrated to determine whether a landslide has occurred and to predict the likelihood of a landslide disaster.

[0055] S5. Perform performance testing on the graph convolutional neural network structure.

[0056] In this embodiment, a GCN model based on slope units and their contour terrain features (abbreviated as CTFGCN) is constructed. This model extracts slope units and combines them with contour lines to generate nodes and edges to construct a graph structure, thus integrating contour morphology knowledge with graph neural network technology.

[0057] The CTFGCN model was compared and analyzed with a simple graph convolutional neural network model based on slope center point (SCPGCN), as well as traditional machine learning methods commonly used in landslide susceptibility assessment, such as RF and Support Vector Machine (SVM).

[0058] Feature extraction was performed using different methods on the same source data. In the CTFGCN model, features were extracted from multiple nodes within each slope unit. That is, different nodes within a slope unit have different feature attributes. For example, most nodes within a slope have a south slope, while some nodes have a southwest slope and some nodes have a southeast slope. In the other three models, features were extracted only from the center point of each slope unit. For example, the slope of a slope is south.

[0059] The model performance was tested using the following metrics: accuracy, precision, recall, and F1 score. The area under the receiver operating characteristic curve (ROC) and the area under the precision-recall curve (PR) of different models were compared.

[0060] The test results and model prediction accuracy are shown in Table 1: Table 1 Please see Figure 6The figure shows the receiver operating characteristic (ROC) curves for different models and the ROC curves for different algorithm models. The horizontal axis represents the false positive rate, and the vertical axis represents the true positive rate. The closer the curve is to the top left corner, the better the model performance. The figure contains four curves, corresponding to the CTFGCN, SCPGCN, RF, and SVM algorithm models. The CTFGCN model's curve is closest to the top left corner, with an area under the curve (AUC) of 0.93, indicating that this model performs best in balancing true and false positive rates. The SCPGCN model has an AUC of 0.90, showing slightly better performance. The RF model has an AUC of 0.84, and the SVM model has an AUC of 0.81, indicating relatively weaker performance. Overall, CTFGCN demonstrates superior performance in classification tasks.

[0061] Please see Figure 7 The figure shows the precision-recall curves for different models, illustrating the PR curves of different algorithm models. The horizontal axis represents recall, and the vertical axis represents precision. The larger the area enclosed by the curve and the coordinate axis, the better the model performance. The figure contains four curves, corresponding to four algorithm models: CTFGCN, SCPGCN, RF, and SVM. The area enclosed by the curve for the CTFGCN model is 0.94, indicating that this model performs best in balancing precision and recall; the area for the SCPGCN model is 0.91, indicating slightly better performance; the area for the RF model is 0.83, and the area for the SVM model is 0.82, indicating relatively weaker performance. Overall, CTFGCN demonstrates superior performance in classification tasks.

[0062] Combined with Table 1, Figure 6 and Figure 7 As can be seen, the CTFGCN model performs best across all metrics. The CTFGCN model achieves accuracy, precision, recall, F1 score, AUROC, and AUPR of 0.86, 0.94, 0.79, 0.86, 0.93, and 0.94 respectively, significantly outperforming other models and demonstrating its superior ability to capture complex spatial relationships and structural features. In comparison, while the SCPGCN model performs slightly worse than CTFGCN, its landslide susceptibility prediction accuracy is still superior to RF and SVM because it aggregates and transfers features (i.e., contextual information) from topologically connected slope units to the current slope unit.

[0063] The graph construction method proposed in this invention forms a graph structure by constructing nodes and edges based on contour lines within slope units. Combined with a graph convolutional neural network, it effectively captures the complexity of the terrain and its interactions. Compared with the method of constructing a graph based on the center point of the slope unit (SCPGCN), it fully captures the micro-topographic features that affect landslide occurrence, such as slope, aspect, and curvature, and better reflects the refined topographic features within the slope. Features are extracted from multiple nodes within each slope unit. Different nodes have different feature attributes, covering various topographic and geological factors such as elevation, slope, and aspect. Compared with traditional models that only extract features from the center point of the unit, it can obtain richer information. This effectively compensates for the shortcomings of traditional machine learning methods in accurately extracting key topographic features of slope units.

[0064] The graph construction method proposed in this invention can effectively capture the complexity of terrain and its interactions, breaking through the traditional method of feature extraction for slope units using statistical methods. Node allocation within slope units based on contour lines helps to comprehensively capture micro-topographic features affecting landslide occurrence, such as slope, aspect, and curvature, and identify terrain features such as steep slopes and flow zones within the slope. The CTFGCN model effectively captures non-Euclidean spatial correlations, effectively capturing local correlations while maintaining translation invariance, thus overcoming the shortcomings of traditional machine learning methods that neglect the spatial correlations and mutual influences between adjacent grids or slope units. The dual-drive mode, which integrates contour morphology knowledge with graph neural network technology, reduces dependence on sample size and lowers the requirements for geological field surveys.

[0065] Please see Figure 8 In an optional embodiment, the present invention provides an intelligent landslide susceptibility prediction system based on graph convolutional networks. The system includes an input device, an output device, a processor, and a memory, all interconnected. The memory stores a computer program comprising program instructions, and the processor is configured to invoke the program instructions to execute specific steps as described in the relevant embodiments of the intelligent landslide susceptibility prediction method based on graph convolutional networks provided by the present invention. The intelligent landslide susceptibility prediction system based on graph convolutional networks provided by the present invention has a complete structure, is objective and stable, and enhances the overall applicability and practical application capability of the present invention.

[0066] In summary, the present invention provides an intelligent prediction method and system for landslide susceptibility based on graph convolutional networks. It constructs a graph convolutional neural network structure based on slope units and their contour topographic features. By extracting slope units and combining them with contour lines to generate nodes and edges, a graph structure is built. This dual-drive mode integrates contour morphology knowledge with graph neural network technology, reducing dependence on landslide sample size. By calculating the steepness between nodes, key topographic features such as landslide embankments and flow zones are identified. Based on the graph convolutional neural network, landslide identification and susceptibility prediction are performed on slope units, generating a landslide susceptibility zoning map. Tests show that this method can significantly improve the accuracy of landslide susceptibility assessment. The method is easy to understand, computationally simple, requires less workload, and is convenient for engineering applications, providing a theoretical foundation and technical support for the further development of the fields of geological disaster and geographic information technology.

[0067] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. A method for intelligent prediction of landslide susceptibility based on graph convolutional networks, characterized in that, Includes the following steps: Landslide adjustment factors are obtained, and the landslide adjustment factors are evaluated and selected to obtain landslide susceptibility evaluation factors; Correlation analysis and feature importance ranking were performed on the landslide susceptibility evaluation factors to obtain optimized landslide susceptibility evaluation factors; Based on the landslide susceptibility evaluation optimization factors, a slope unit structure diagram is constructed based on the contour line characteristics of the slope. A graph convolutional neural network structure is established, and the slope unit structure diagram and the graph convolutional neural network structure are combined to realize intelligent prediction of landslide susceptibility.

2. The intelligent prediction method for landslide susceptibility based on graph convolutional networks according to claim 1, characterized in that, The process of obtaining landslide adjustment factors, evaluating and selecting landslide susceptibility evaluation factors includes: The landslide susceptibility evaluation factors include topographic factors, geological factors, human engineering activity factors, and triggering factors. The topographic and geomorphological factors include elevation, slope, aspect, curvature, surface roughness, and surface relief. The geological factors include lithology and distance from faults; The human engineering activity factor includes distance from the road; The triggering factors include distance from the water system and rainfall.

3. The intelligent prediction method for landslide susceptibility based on graph convolutional networks according to claim 1, characterized in that, The process of performing correlation analysis and feature importance ranking on the landslide susceptibility evaluation factors to obtain optimized landslide susceptibility evaluation factors includes: Correlation analysis was performed on the landslide susceptibility evaluation factors based on the Pearson correlation coefficient to obtain non-redundant features in the landslide susceptibility evaluation factors; The landslide susceptibility assessment factors are evaluated based on the Gini index, and the feature importance is ranked based on these scores to obtain the feature ranking results. The landslide susceptibility evaluation optimization factor is obtained by combining the non-redundant features and the feature ranking results.

4. The intelligent prediction method for landslide susceptibility based on graph convolutional networks according to claim 1, characterized in that, The process of constructing a slope unit structure map based on the contour line characteristics of the slope, in conjunction with the landslide susceptibility evaluation optimization factors, includes: The contour lines are generated based on the digital elevation model, and nodes are generated on the contour lines at equal intervals to form an initial node set. The nodes in the initial node set are traversed one by one to obtain the two lower-level nodes that are closest to each node in the contour lines of the next layer, so as to establish node connection edges; Obtain the node steepness of adjacent contour lines, and obtain the lower-level node attributes based on the node steepness; The slope unit structure diagram is constructed by combining the initial node set, the node connection edges, and the node attributes.

5. The intelligent prediction method for landslide susceptibility based on graph convolutional networks according to claim 4, characterized in that, The step of obtaining the node steepness of adjacent contour lines and obtaining the lower-level node attributes based on the node steepness includes: In three-dimensional space, the kurtosis of the nodes satisfies the following relationship: in, For node steepness, The three-dimensional coordinates of the first node are: The three-dimensional coordinates of the second node; A node steepness threshold is set. When the node steepness is equal to or greater than the node steepness threshold, the lower-level node is classified as a steep ridge node. When the node steepness is less than the node steepness threshold, the lower-level node is classified as a flowing node.

6. The intelligent prediction method for landslide susceptibility based on graph convolutional networks according to claim 1, characterized in that, The establishment of the graph convolutional neural network structure, combined with the slope unit structure diagram and the graph convolutional neural network structure, enables intelligent prediction of landslide susceptibility, including: A slope unit dataset for the study area is established, and a slope unit sub-map is obtained by combining the slope unit dataset and the slope unit structure diagram. Node features are extracted from the slope unit subgraph to obtain feature vectors, and a feature matrix is ​​constructed based on the feature vectors; By combining the graph convolutional neural network structure and the feature matrix, a landslide probability distribution is obtained, enabling intelligent prediction of landslide susceptibility.

7. The intelligent prediction method for landslide susceptibility based on graph convolutional networks according to claim 6, characterized in that, The process of obtaining the landslide probability distribution by combining the graph convolutional neural network structure and the feature matrix includes: The graph convolutional neural network structure includes an input layer, multiple convolutional layers, a max pooling layer, and an output layer; The feature matrix is ​​input into the graph convolutional neural network structure through the input layer; The feature matrix is ​​obtained by performing convolution processing on the feature matrix using the multiple convolutional layers; The output feature matrix is ​​subjected to global graph embedding based on the max pooling layer, and the landslide probability distribution is obtained by performing binary classification prediction of landslide-non-landslide based on the output feature matrix. The landslide probability distribution is output based on the output layer.

8. The intelligent prediction method for landslide susceptibility based on graph convolutional networks according to claim 7, characterized in that, The step of using the multiple convolutional layers to perform convolution processing on the feature matrix to obtain the output feature matrix includes: in, For the first The feature matrix output after processing by the convolutional layer For activation function, for The degree matrix, This is the corrected adjacency matrix. For the first The feature matrix output after processing by the convolutional layer For the first The learnable weight matrix of a convolutional layer.

9. The intelligent prediction method for landslide susceptibility based on graph convolutional networks according to claim 7, characterized in that, The graph convolutional neural network structure includes: The graph convolutional neural network structure is trained using a cross-entropy loss function, which satisfies the following relationship: in, This represents the cross-entropy loss value. For the total sample size, For index variables, For the first The true label of each sample For the graph convolutional neural network structure, the first... The probability value of a sample being predicted as the "landslide" category.

10. A landslide susceptibility intelligent prediction system based on graph convolutional networks, characterized in that, The system includes an input device, an output device, a processor, and a memory, which are interconnected. The memory stores a computer program, which includes program instructions. The processor is configured to invoke the program instructions to execute the landslide susceptibility intelligent prediction method based on graph convolutional networks as described in any one of claims 1-9.

Images