Method and device for evaluating landslide susceptibility constrained by disaster-gestational environment, equipment and medium
By constructing slope units and graph neural network models using the meandering watershed method, and combining environmental and node feature constraints, the problem of insufficient correlation between landslides and disaster-prone environments in traditional methods is solved, achieving a more reliable landslide susceptibility assessment and reducing false alarm and missed alarm rates.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- LANZHOU JIAOTONG UNIV
- Filing Date
- 2023-12-14
- Publication Date
- 2026-05-26
Smart Images
Figure CN122088560A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of geological exploration and information technology, and to methods, devices, equipment, and media for assessing landslide susceptibility under disaster-prone environmental constraints, but is not limited to such methods. Background Technology
[0002] In landslide susceptibility assessment, machine learning methods, when faced with classification tasks in complex scenarios, enhance their ability to generate end-to-output by increasing layer width and model depth, thereby improving the ability to generate more complex nonlinear mappings. Correspondingly, this requires more sample label data to accommodate the larger model capacity. The construction of landslide sample labels typically requires expert knowledge and field survey results. However, in complex environments, especially high-altitude, high-vegetation areas, obtaining a large number of recorded landslide samples through field surveys or landslide interpretation presents significant challenges. Furthermore, this approach can only establish landslide samples in localized areas, resulting in a limited and uneven sample size.
[0003] Landslide samples vary significantly across different environments. Subjectively selecting non-landslide samples randomly makes it difficult to guarantee their representativeness and balance. In complex environments, this can lead to imbalanced predictions, severely impacting the generalization ability and reliability of machine learning methods for assessing landslide susceptibility. Most supervised machine learning models are considered black-box models due to the difficulty in discerning their internal learning mechanisms, resulting in limited generalization ability. It is challenging to learn a general model from one dataset and then appropriately transfer it to other (unseen) datasets. Therefore, models trained with limited and imbalanced labeled samples inevitably lead to overfitting. When landslide samples contain noise, this can even result in training incorrect models. However, these models have numerous hyperparameters, requiring sufficient experience and prior knowledge to optimize, significantly increasing the workload.
[0004] It can be seen that many achievements have been made in machine learning methods for assessing landslide susceptibility, but there are still some key issues that scholars have not explored or have not solved well, at least in the following two aspects: the full correlation between landslides and disaster-prone environments needs to be explored; and reliable learning under insufficient and imbalanced landslide samples has not yet been solved. Summary of the Invention
[0005] This application provides a landslide susceptibility assessment method, apparatus, equipment, and medium constrained by a disaster-prone environment. It at least solves the problem that traditional machine learning landslide susceptibility assessment methods only consider the characteristics of landslides and local disaster-prone environments, without taking into account similar disaster-prone environments that are geographically distant but directly related. This makes it impossible to deeply explore the potential correlation between landslide characteristics, resulting in poor reliability of landslide susceptibility assessment.
[0006] The technical solution of this application embodiment is implemented as follows:
[0007] In a first aspect, embodiments of this application provide a method for assessing landslide susceptibility under disaster-prone environmental constraints, the method comprising:
[0008] Slope units are constructed and multiple geographic nodes are generated based on the meander rate watershed method. Attribute features of each geographic node are extracted using spatial statistics. A directed weighted graph relating landslides to the disaster-prone environment is constructed by combining environmental feature constraints and node feature constraints. The environmental feature constraints characterize the environmental consistency between any two geographic nodes, and the node feature constraints characterize the similarity of attribute features between any two geographic nodes. Based on the directed weighted graph, a trained graph neural network model is used to assess landslide susceptibility. The graph neural network model is trained using a balanced sample dataset.
[0009] Secondly, embodiments of this application provide a landslide susceptibility assessment device under disaster-prone environmental constraints, including a node generation module, a sign extraction module, a correlation graph construction module, and a landslide assessment module, wherein:
[0010] The node generation module is used to construct slope units and generate multiple geographic nodes based on the meander watershed method.
[0011] The feature extraction module is used to extract the attribute features of each geographic node through spatial statistics.
[0012] The association graph construction module is used to construct a directed weighted graph that associates landslides with disaster-prone environments by combining environmental feature constraints and node feature constraints; wherein, the environmental feature constraints represent the environmental consistency between any two geographic nodes, and the node feature constraints represent the attribute feature similarity between any two geographic nodes.
[0013] The landslide assessment module is used to assess landslide susceptibility based on the directed weighted graph and a trained graph neural network model; wherein the graph neural network model is trained using a balanced sample dataset.
[0014] Thirdly, embodiments of this application provide an electronic device, including a memory and a processor. The memory stores a computer program that can run on the processor. When the processor executes the program, it implements the steps in the above-described landslide susceptibility assessment method under disaster-prone environmental constraints.
[0015] Fourthly, embodiments of this application provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the landslide susceptibility assessment method under the aforementioned disaster-prone environmental constraints.
[0016] The beneficial effects of the technical solutions provided in this application include at least the following:
[0017] In this embodiment, firstly, slope units are divided based on the meandering watershed method to construct geographic nodes, and geographic node features are extracted using spatial statistical methods. Then, a directed weighted graph is constructed based on both environmental and node feature constraints. Finally, a graph neural network model is built to learn the feature relationships of the directed weighted graph, enabling susceptibility prediction for unknown geographic nodes. This approach visually represents landslides and their associated environments through node-based graph construction, and utilizes graph neural networks to effectively process and capture similar environmental geographic information across spatial boundaries for susceptibility assessment. Based on a graph neural network method that fully correlates landslides and their associated environments, the potential correlation features of landslides are deeply mined, enabling landslide susceptibility assessment of graph nodes, reducing false alarm and missed alarm rates, and improving the reliability of landslide susceptibility assessment. This provides a more stable and practical method for landslide disaster investigation and prevention in complex environments and sparse sample areas. Attached Figure Description
[0018] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein:
[0019] Figure 1 A flowchart illustrating the landslide susceptibility assessment method under disaster-prone environmental constraints provided in this application embodiment;
[0020] Figure 2 A flowchart illustrating the construction process of a geographic node provided in an embodiment of this application;
[0021] Figure 3 A schematic diagram illustrating the feature extraction process for geographic nodes provided in an embodiment of this application;
[0022] Figure 4 A schematic diagram illustrating the environment rule-based graph construction process provided in this application embodiment;
[0023] Figure 5 This is a schematic diagram of the key process of the GraphSAGE model provided in the embodiments of this application;
[0024] Figure 6 A system block diagram of the GraphSAGE model provided in the embodiments of this application;
[0025] Figure 7 This is a schematic diagram illustrating landslide susceptibility results under different sampling methods provided in the embodiments of this application;
[0026] Figure 8Schematic diagrams of landslide susceptibility results under different graphical structures provided in the embodiments of this application;
[0027] Figure 9 A schematic diagram illustrating the landslide susceptibility results of a specific watershed provided in an embodiment of this application;
[0028] Figure 10 A schematic diagram showing the landslide susceptibility evaluation results of the four models provided in the embodiments of this application;
[0029] Figure 11 A schematic diagram of the landslide sensitivity distribution of the four models provided in the embodiments of this application;
[0030] Figure 12 Google imagery, field photos, and landslide susceptibility maps of typical areas provided in this application's embodiments;
[0031] Figure 13 A schematic diagram of the composition structure of a landslide susceptibility assessment device under disaster-prone environmental constraints provided in this application embodiment;
[0032] Figure 14 This is a schematic diagram of the hardware entity of an electronic device provided in an embodiment of this application. Detailed Implementation
[0033] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. The following embodiments are used to illustrate this application, but are not intended to limit the scope of this application. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0034] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.
[0035] It should be noted that the terms "first, second, and third" used in the embodiments of this application are merely to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first, second, and third" can be interchanged in a specific order or sequence where permitted, so that the embodiments of this application described herein can be implemented in an order other than that illustrated or described herein.
[0036] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which the embodiments of this application pertain. It should also be understood that terms such as those defined in general dictionaries should be understood to have a meaning consistent with their meaning in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.
[0037] Landslide susceptibility assessment is the prerequisite and foundation for landslide disaster evaluation. It mainly refers to the discovery and identification of the location, extent, and probability of potential landslides at different scales, i.e., spatial prediction of landslides. Machine learning-based landslide susceptibility assessment modeling mainly includes many steps such as landslide cataloging, landslide factor extraction, model architecture construction, model training, and model evaluation. During model construction, input data, network structure, parameter settings, and algorithm optimization all constrain the accuracy of landslide susceptibility assessment to varying degrees. Machine learning methods, with their ability to express non-linear relationships, have been widely explored and applied in landslide susceptibility assessment. Examples include logistic regression, decision trees, random forests, support vector machines, artificial neural networks, and other neural networks, whose performance surpasses that of general statistical models in landslide susceptibility assessment.
[0038] While machine learning methods can automatically identify the main influencing factors of landslides and fit target classifications, learning the feature distribution of landslides using only independent sample features performs well locally but produces unbalanced predictions when faced with complex geographical environments. For example, landslide susceptibility assessment based on convolutional neural networks (CNNs) considers the local spatial information of landslide samples through convolutional operations. However, due to the complex and heterogeneous geographical background, landslide distribution is relatively discrete, and the features of adjacent landslides are uncertain or even unrelated, making it difficult for local convolutional operations to obtain effective geographical environmental information, thus limiting the accuracy of the assessment. Based on the geographical theory that "the more similar the geographical environment, the more similar the geographical features," similar environments (including similar geology, geomorphology, environmental factors, etc.) will produce similar landslide conditions regardless of distance. This correlation between sub-regions helps to better understand the characteristics of landslides from their essence.
[0039] The issue of fully linking landslides with their nucleus-forming environment, which has not yet been addressed in existing research, requires further exploration. While there are promising results in using machine learning to assess landslide susceptibility across different regions, there is limited research on machine learning methods for assessing landslide susceptibility by fully linking landslides with their nucleus-forming environment. In geography, the principle of "the more similar the geographical environment, the more similar the geographical features" suggests that points with similar geographical environments to landslide samples are highly likely to experience future landslides. It is necessary to find or derive methods for expressing and calculating the link between landslides and their nucleus-forming environment, which is essential for machine learning-based landslide susceptibility assessment. However, current applications of this link lack theoretical support.
[0040] Limited sample sizes and complex landslide-prone environments make it difficult for commonly used machine learning methods to learn effective landslide representations. The key to solving this problem lies in applying prior knowledge of landslide-environment associations to improve learning performance. Furthermore, in complex environments, negative samples cannot guarantee environmental equilibrium with positive samples under conventional sampling methods, necessitating the use of geographical environmental consistency constraints to determine sample selection for various environmental types. Currently, research on designing balanced sampling strategies and learning based on landslide-prone environments is relatively scarce.
[0041] A graph, composed of a set of nodes and edges, typically uses nodes to represent entities and edges to describe the relationships between entities. It is a typical non-linear data structure. It can flexibly and effectively represent unstructured, complex data, unconstrained by spatial distance. Considering the complexity of landslides and their nucleation environments, graph structures are more likely to effectively represent the complex non-linear relationships between landslides and their nucleation environments compared to single feature sequences or images. In graph structures, the relationships between nodes are no longer constrained by spatial distance, allowing for the association of landslide points with similar nucleation environments over a wider range, thereby achieving sample expansion and balanced sampling.
[0042] Therefore, this application proposes constructing a graph structure to display the relationship between landslides and the disaster-prone environment. Specifically, this involves constructing a graph with geographic environmental consistency constraints on disaster-prone landslide factors. For insufficient and unbalanced landslide samples, designing a balanced sampling strategy also requires relying on the disaster-prone environment association graph. Clearly, obtaining a landslide susceptibility assessment necessitates feature learning from the graph structure. Graph Neural Networks (GNNs) are an effective method for processing graphs, capable of handling and defining non-Euclidean irregular data in graphs. They possess the ability to aggregate and transform the neighborhood feature information of each graph node, and can effectively process and capture similar geographic environmental information across spaces through information transfer between graph nodes.
[0043] This application provides a landslide susceptibility assessment method under disaster-prone environmental constraints, applied to an electronic device. The electronic device includes, but is not limited to, mobile phones, laptops, tablets and handheld internet devices, multimedia devices, streaming media devices, mobile internet devices, wearable devices, or other types of electronic devices. The functions implemented by this method can be achieved by a processor in the electronic device calling program code. The program code can be stored in a computer storage medium; therefore, the electronic device includes at least a processor and a storage medium. The processor can be used to process the landslide susceptibility assessment, and the memory can be used to store the data required and generated during the landslide susceptibility assessment.
[0044] Figure 1 A flowchart illustrating a landslide susceptibility assessment method under disaster-prone environmental constraints provided in this application embodiment is shown below. Figure 1 As shown, the method includes at least the following steps:
[0045] Step S110: Construct slope units and generate multiple geographic nodes based on the meandering watershed method.
[0046] Here, creating appropriate geographic nodes is crucial for constructing a graph neural network model. Appropriate geographic nodes should maximize the uniformity of landslide characteristics and the heterogeneity of individual units. There are generally two methods for creating geographic nodes: grid cells and slope cells.
[0047] Grid cells are simple in structure and easy to calculate, but they are difficult to accurately reflect landforms and landslide-prone environmental characteristics, and therefore cannot be applied to large-scale studies. Slope cells, on the other hand, are topographic units defined by ridgelines and valley lines. They are related to the hydrological conditions and processes that shape the topography and can accurately reflect the topographic features of different slope sections. They are widely used in large-scale, refined assessments of landslide susceptibility, hazard, or risk.
[0048] Currently, there are two methods for dividing slope units: hydrological analysis and the curvature watershed method. The hydrological analysis method often results in numerous small, broken surfaces and unreasonable elongated surfaces, requiring extensive manual corrections later on—a tedious and inefficient process. The curvature watershed method, however, overcomes the limitation of the hydrological analysis method in creating slopes that conform to the topography in flat areas. Furthermore, the resulting slope units have uniform slope and aspect, satisfying the fundamental assumption of homogeneity in landslide mechanics analysis. Therefore, the curvature watershed method is used to construct slope units to generate geographic nodes.
[0049] The process of constructing the geographic nodes provided in this application is as follows: Figure 2As shown, firstly, the curvature of the elevation (distance from a point along the vertical line to the absolute datum) is calculated to obtain the average curvature, and a concave watershed is obtained through the catchment area. Then, the elevation is reversed to obtain the reverse elevation, and the reverse curvature is calculated again, resulting in a convex watershed through the catchment area. Finally, the concave and convex watersheds are merged to obtain slope elements, and the turning points are analyzed to obtain geographic nodes. This achieves disaster-prone environment clustering using slope elements to obtain disaster-prone environment geographic nodes.
[0050] Step S120: Extract the attribute features of each geographic node using spatial statistics.
[0051] Here, the attributes of slope units are important features of map nodes, such as elevation, slope, aspect, topographic humidity index, annual average cumulative rainfall, normalized difference vegetation index, distance from road, and land use type. Based on the ArcGIS platform, a zonal statistical approach is used to calculate the attribute characteristics of each node. Landslide influencing factors generally fall into two categories: continuous and discrete, requiring two different statistical methods for their analysis.
[0052] Step S130: Combining environmental feature constraints and node feature constraints, construct a directed weighted graph relating landslides to the disaster-prone environment.
[0053] Here, the environmental feature constraints characterize the environmental consistency between any two geographic nodes, and the node feature constraints characterize the attribute feature similarity between any two geographic nodes. The directed weighted graph is a correlation graph of geographic environmental similarity constraints, used to display the correlation between landslides and the disaster-prone environment.
[0054] Step S140: Based on the directed weighted graph, landslide susceptibility is assessed using a trained graph neural network model.
[0055] Here, the Graph Neural Networks (GNN) model is trained using a balanced sample dataset to learn the feature relationships of a directed weighted graph. To obtain a landslide susceptibility assessment, feature learning is necessary for the graph structure relating landslides to their geological environment. GNNs are an effective method for processing graphs, capable of handling and defining irregular non-Euclidean spatial data. They possess the ability to aggregate and transform neighborhood feature information of each graph node and can effectively process and capture similar environmental geographic information across spatial boundaries through information transfer between graph nodes. This provides a graph neural network-based method to deeply mine potential landslide association features, achieve landslide susceptibility assessment of graph nodes, reduce false alarm and missed alarm rates, and improve the reliability of landslide susceptibility assessments.
[0056] In this embodiment, firstly, slope units are divided based on the meandering watershed method to construct geographic nodes, and geographic node features are extracted using spatial statistical methods. Then, a directed weighted graph is constructed based on both environmental and node feature constraints. Finally, a graph neural network model is built to learn the feature relationships of the directed weighted graph, enabling susceptibility prediction for unknown geographic nodes. This approach visually represents landslides and their associated environments through node-based graph construction, and utilizes graph neural networks to effectively process and capture similar environmental geographic information across spatial boundaries for susceptibility assessment. Based on a graph neural network method that fully correlates landslides and their associated environments, the potential correlation features of landslides are deeply mined, enabling landslide susceptibility assessment of graph nodes, reducing false alarm and missed alarm rates, and improving the reliability of landslide susceptibility assessment. This application can provide a more stable and practical method for landslide disaster investigation and prevention in complex environments and sparse sample areas.
[0057] In some implementations, the extraction of attribute features for each geographic node using spatial statistical methods includes: for continuous landslide influencing factors, determining the average value of the influencing factors within the corresponding slope unit as the attribute feature of the geographic node; for discrete landslide influencing factors, first reclassifying them according to category, and then statistically analyzing the influencing factor with the highest frequency within the corresponding slope unit as the attribute feature of the geographic node. Here, landslide influencing factors are selected through multicollinearity analysis, such as... Figure 3 As shown, for continuous factors, the average value of the influencing factor within the slope unit is calculated, that is, the maximum value of the continuous factor is taken as the attribute of the geographic node; for discrete factors, they are first reclassified according to category, and then the category with the highest frequency of the original influencing factor within the slope unit is counted, that is, the mode of the discrete factor is taken as the attribute feature of the geographic node.
[0058] In some implementations, constructing a directed weighted graph relating landslides to the disaster-prone environment by combining environmental feature constraints and node feature constraints includes: determining a first similarity between the environmental factor features of any two geographic nodes and obtaining a candidate set of connection nodes for each geographic node; filtering neighboring nodes from the candidate set of connection nodes for each geographic node by calculating a second similarity between the attribute features of any two geographic nodes; connecting the corresponding geographic node and the neighboring node to form an edge and calculating the weight of the edge to obtain the directed weighted graph relating landslides to the disaster-prone environment.
[0059] It should be noted that the topographical conditions in which a landslide occurs are the basis for its formation, and unstable environmental conditions are the basic conditions for landslide occurrence. These conditions determine the basic probability of landslide formation, meaning that a consistent environment is the basic condition for the connection of nodes.
[0060] From the perspective of slope systems, landslides, as a form of slope failure, are characterized by their load-bearing structure composed of materials that remain relatively stable over a long period with minimal change. Within the same landslide body, these materials possess consistent lithology and soil types, and these materials directly determine or influence the long-term geological conditions of the landslide. Furthermore, the geological environment of landslides is also affected by faulting and river erosion, and historical landslides in the study area exhibit a distribution pattern along faults and rivers. Therefore, lithology, soil type, distance to faults, and distance to rivers are used as environmental constraints to classify unstable environments.
[0061] Here, to measure the consistency of the environments in which geographic nodes u and v are located, cosine similarity (Formula 1) is used to calculate the environmental factor feature vector x of nodes u and v. u and x v First similarity between Each geographic node u is only allowed to connect to the node with the most consistent environment category, that is, when equal When v is added to the candidate set N(u) of connected nodes of u, v is added to the set N(u).
[0062]
[0063]
[0064] After applying environmental consistency constraints, for each node u, there are still a large number of candidate nodes in N(u). Nodes with different features are meaningless for the overall landslide characteristics. Therefore, it is necessary to further calculate the similarity of node features to filter candidate nodes. Elevation, slope, aspect, topographic humidity index, annual average cumulative rainfall, normalized difference vegetation index, distance from road, and land use type are used as node features. Cosine similarity (Formula 3) is used to calculate the feature vector x of each geographic node u and v. u' and x v' Second similarity between To filter out redundant nodes, for each node, only the candidate node with the highest similarity is retained. That is, when... equal Only when v reaches the threshold η can it be added to the final neighbor set N'(u). This process continues until the size of N'(u) reaches the threshold η, at which point v is finally generated. Edges to u. In this invention, η = 10, meaning a node has 10 neighbors. To account for the inherent characteristics of nodes in the computation, a self-loop is added to each node, i.e., u∈N'(u).
[0065]
[0066]
[0067] In some implementations, the method further includes: for each edge formed by connection, using the product of the first similarity calculated by the corresponding geographic node and the second similarity calculated by the neighboring node as the weight of the edge.
[0068] It's worth noting that the original graph neural network model treats the feature information of each neighbor in the graph equally. However, in different disaster-prone environments, the influence of each neighbor node on the central node varies, and treating them equally will prevent the model from learning the most effective features. Therefore, a method is adopted to add weights to each neighbor during the model's aggregation of neighbor information. This helps the model better understand the relationships between nodes in the graph, generating more representative feature vectors and improving the model's accuracy. The environmental factor similarity between the central node u and neighbor node v is then considered. Similarity to node features The product of the two edges is used as the weight of the edge to comprehensively reflect the influence of the neighbor node v on the center node u. The calculation formula is shown in formula (5) below.
[0069]
[0070] Figure 4 The environment rule-based graph construction process provided in the embodiments of this application, such as Figure 4 As shown, different colored areas represent different environments. Within the same environment, black dots in each area represent candidate nodes for node u, and white dots represent excluded nodes. First, the consistency of the environment is calculated. Maintain the consistency of node u, and then calculate node similarity. Filter neighboring nodes for node u, and finally connect the edges between nodes and calculate the weights to obtain the association graph constrained by geographical environment similarity.
[0071] In some embodiments, the method further includes: classifying environments according to landslide impact factors, using cataloged landslide data as positive samples, and selecting negative samples in each environment in proportion to the positive samples to obtain the balanced sample dataset.
[0072] Here, the quality of the training sample set has a crucial impact on the accuracy of the landslide susceptibility map. Positive samples in the training sample set are based on historical landslide catalogs and are sampled in landslide-affected areas; their quality is high and can fully reflect the typical geographical environmental conditions for landslide occurrence. Negative samples, however, cannot be directly obtained and are usually randomly selected from the study area in equal numbers to the positive samples. These collected negative samples are not representative, lack typical significance, and are highly likely to contain potential landslide points. These false negative samples reduce the quality of the training samples, thus affecting the accuracy of the landslide susceptibility assessment map. Therefore, to improve the reliability of the landslide susceptibility assessment map, non-landslide samples are selected based on the similarity constraints of the aforementioned classification, and the balance between positive and negative samples is improved based on geographical environmental consistency. This ensures that the training dataset not only maintains a balance in the number of positive and negative landslide samples but also maintains a balance across different landslide features, ensuring that the model can learn more accurate, reliable, and richer features.
[0073] The specific steps are as follows: Based on similarity constraints, the environment is divided using cataloged landslide data as positive samples. In each environment, negative samples are selected in equal proportions to the positive samples to ensure that the resulting dataset includes landslide and non-landslide samples from all environment types, thus obtaining a sample dataset. The dataset is divided, with 70% used for training and 30% for validation. This application's embodiment constructs a graph structure to fully correlate landslides and disaster-prone environments, and designs a balanced sampling strategy for landslide and non-landslide points that considers geographical consistency constraints to ensure the reliability of the learning process.
[0074] In some implementations, the graph neural network model is a GraphSAGE network, including convolutional layers and fully connected layers. The landslide susceptibility assessment based on the directed weighted graph using the trained graph neural network model includes: for any node in the directed weighted graph, fusing the feature information of all neighboring nodes in the full neighbor set of the node layer by layer in the convolutional layer using an aggregation function; multiplying the fused feature information by the weight matrix and feeding it into the fully connected layer through a nonlinear activation function to generate a node representation for the next layer; and normalizing the node representation of the last layer to obtain the probability of whether the corresponding geographical node is a landslide, thus completing the landslide susceptibility assessment. Here, the aggregation function is the Aggregate function, and the normalization function is the softmax function.
[0075] The core idea of the GraphSAGE model is to generate the embedding vector of the target node by learning a function that aggregates the representations of neighboring nodes. Using the information obtained from the aggregation of nodes, it aims to predict the node label, such as... Figure 5As shown, firstly, node sampling is performed, that is, uniform sampling of node u with a fixed neighborhood size (N(u)); then, information aggregation is performed, that is, the feature information of neighboring nodes is fused through the aggregation function Aggregate, that is, the representation of node u's neighboring nodes N(u) in layer l+1. (Formula 6); Finally, feature embedding is performed, that is, the concat operator is used to combine it with the current representation of the node. The concatenation is performed and multiplied with the weight matrix W, then fed into the fully connected layer using a non-linear activation function σ to generate... The representation used for the next step (Formula 7).
[0076]
[0077]
[0078] In some implementations, the method further includes: setting a weight for each neighbor node during the aggregation of neighbor information, using the attribute value of the edge as the weight.
[0079] The original GraphSAGE model treats each neighbor's information equally, failing to reflect the varying impacts of different neighbors on the embedding of the central node. Therefore, this application assigns a weight to each neighbor during the aggregation of neighbor information, assigning a weight to the edge's attribute value e. uv As weights (Formula 8), this enables the model to more effectively process and capture spatially similar environmental information during information transfer between graph nodes, allowing for training of geographically directed weighted graphs that better meet environmental consistency constraints. Since the neighbors of each node are selected during graph construction, the full neighbor set N(u) of node u is chosen for neighbor sampling, rather than performing a fixed-size random sampling of neighbors. Furthermore, self-loops are added to each node during graph construction, so... It contains the characteristic information of u itself, so there is no need to perform a concat operation to obtain the representation of node u in the next layer, which can be used for subsequent information transmission (Equation 9).
[0080]
[0081]
[0082] The GraphSAGE model designed in this application includes three convolutional layers (layer1, layer1-1, layer1-2) and two fully connected layers (fully connected layer 1, fully connected layer 2), as follows: Figure 6As shown, the GraphSAGE network employs supervised learning, trained using the Adam optimizer with binary cross-entropy loss as the loss function. The learning rate is dynamically adjusted to ensure the model learns more comprehensive features while accelerating convergence. To prevent overfitting, a dropout layer is placed between fully connected layer 1 and fully connected layer 2. Additionally, the ReLU activation function is used between convolutional layers to enhance the network's non-linear fitting capability. The final node u is represented... By applying the softmax function, the probability of whether a node is a landslide can be obtained, with a probability of 1 indicating a landslide and a probability of 0 indicating no landslide. Using the trained model to predict the landslide susceptibility of all nodes, the final landslide susceptibility assessment map of the study area can be generated.
[0083] The landslide susceptibility assessment method under disaster-prone environmental constraints described above will be explained below with reference to a specific embodiment. However, it should be noted that this specific embodiment is only for better illustration of this application and does not constitute an improper limitation of this application.
[0084] This application provides a graph neural network-based landslide susceptibility assessment method based on geographical environment consistency constraints. The main steps of this method are as follows: First, landslide influencing factors are selected through multicollinearity analysis, slope units are created based on the meander rate watershed method, geographical nodes are generated, and geographical node features are extracted using spatial statistical methods. Second, a correlation graph constrained by geographical environment similarity is constructed using the cosine similarity calculation method of environmental factors and unit features. The product of the environmental factors and the cosine similarity values of the unit features is used as the weight on the edge connecting two nodes to fully correlate landslides and disaster-prone environments. Then, an equal number of landslide and non-landslide samples are selected under similar geographical environment subdomains to achieve a balance of positive and negative samples under similar disaster-prone geographical environments. Finally, a graph neural network model is constructed to learn the feature relationships of the directed weighted correlation graph, enabling susceptibility prediction of unknown geographical nodes, and the accuracy of the results is analyzed.
[0085] The advantages of this application's embodiments lie in its use of node-based graph construction to visually represent landslides and their associated disaster-prone environments; its improvement of the balance between positive and negative samples based on geographical environmental consistency; and its utilization of graph neural networks to effectively process and capture similar geographical information across spatial boundaries for susceptibility assessment. This addresses the problem that traditional machine learning methods for landslide susceptibility assessment only consider the characteristics of landslides and their local disaster-prone environments, neglecting the characteristics of similar disaster-prone environments that are geographically distant but directly related, thus failing to deeply explore the potential correlations between landslide features and resulting in poor reliability of landslide susceptibility assessments. To verify the advantages of the sampling strategy and model proposed in this application, the following explanation focuses on sample sampling and model construction.
[0086] First, considering the advantages of this application's sampling strategy based on geographical similarity, a comparison is made with the conventional method of random sampling. The landslide susceptibility results under different sampling methods are as follows: Figure 7 As shown, the left column displays the results of random sampling, and the right column shows the structure of geographically similar sampling. From... Figure 7 It was found that landslide susceptibility assessment results obtained based on similar environment sampling are superior to traditional sampling. Without environmental constraints, all four models inevitably exhibited both underestimation and overestimation of landslides. Without the proposed sampling strategy, the RF model showed a wide distribution of moderately sensitive areas within the study area, failing to highlight typical problems; the SVM model showed a concentrated distribution of extremely and highly sensitive areas, failing to accurately and effectively identify landslide locations; the MLP model showed a wide distribution of highly sensitive areas, failing to reflect the actual landslide occurrence; and the model proposed in this application showed a wide concentration of extremely highly sensitive areas, failing to provide reference landslide location information. However, under environmental constraints, the landslide susceptibility assessment results of all four models were better, demonstrating that the proposed sampling strategy enables the model to comprehensively grasp the characteristics of the landslide-prone environment, reduces the impact of environmental differences on the model, and suppresses the overestimation of landslide susceptibility. This indicates that by defining environmental categories and using geographical similarity constraints to select samples for various environmental types, the model can learn the landslide-prone characteristics of different environments. This enables the model to deeply explore the potential correlations of landslide characteristics, which helps to summarize and convey landslide characteristics in different environments and improves prediction reliability.
[0087] Secondly, considering the advantages of the weighted graph structure constructed in this application, under the same data conditions, compared with the model using the unweighted graph structure, the obtained model indices and landslide susceptibility results are shown in Table 1. Figure 8 As shown in Table 1, the weighted graph structure model outperforms the unweighted graph structure model by approximately 1.8% in all metrics. This indicates that the weighted graph structure model more accurately captures the strength and influence of relationships between nodes, allowing for a more refined consideration of the connections between different nodes and thus better learning of node representations. Furthermore, from... Figure 8 Looking at the landslide susceptibility maps of the two maps, the prediction results of the weighted graph structure model on the left (a) show clear classification of each category. Most of the extremely high and high sensitivity areas coincide with the distribution of historical landslides, and the predicted extremely high sensitivity areas are consistent with the direction of the Bailong River basin, which is consistent with the actual situation. In contrast, the prediction results of the unweighted graph structure model on the right (b) show that the high sensitivity areas are widely distributed and only overlap with some historical landslide locations, providing no valuable reference.
[0088] The above demonstrates that the weighted graph structure model designed in this application is more reasonable, better taking into account the impact of different disaster-prone environments on landslides and the interaction between geographical nodes. The feature vectors obtained by aggregating neighbor information are more typical, and the model prediction results can accurately reflect the susceptibility of each node. Its results can provide an effective approach and key technical support for landslide disaster investigation and prevention.
[0089] Table 1. Model index results under different graph structures
[0090]
[0091] like Figure 9 As shown, this application selects the Bailongjiang River Basin as the study area. Located in southern Gansu Province, the Bailongjiang River Basin covers an area of over 20,000 square kilometers. The average annual precipitation in the southern climate zone is 450-800 mm, while in the northern zone it is 400-700 mm. The study area is a typical mountainous region with significant topographic relief and a complex geographical environment. Influenced by tectonic movements, concentrated rainfall, and human activities, landslides are frequent and widespread, making it one of the four major landslide and debris flow disaster-prone areas in my country. Landslides are concentrated in clusters in the tectonically active Zhouqu-Wudu section, mostly loess landslides, sedimentary landslides, and fractured rock landslides. The 2008 Wenchuan earthquake exacerbated the damage to the mountain and soil structures, triggering a series of new landslide and debris flow disasters, creating numerous potential geological hazards. The 2013 Wenxian earthquake (magnitude 4.5), the Minxian and Zhangxian earthquakes (magnitude 6.6), and the 2017 Jiuzhaigou earthquake (magnitude 7.0) deformed and fractured the soil and rock, providing favorable channels for surface seepage and groundwater activity, and creating favorable material and structural conditions for the formation, development, and occurrence of landslides. The landslide hazard in the Bailongjiang River basin not only poses a serious threat to the lives and property of local people and critical infrastructure, but also threatens the ecological security and sustainable development of the region and even the vast upper reaches of the Yangtze River.
[0092] Under environmentally constrained data conditions, this application uses four pre-trained models—RF, SVM, MLP, and GraphSAGE—to predict landslide susceptibility in the Bailongjiang River basin. The predicted values are assigned to each slope unit to generate a landslide susceptibility map. The landslide susceptibility map is a visual means of intuitively reflecting the model's prediction results. Figure 10 The landslide susceptibility assessment results of four models are presented. Figure (a) shows the prediction results of the RF model, (b) shows the prediction results of the SVM model, (c) shows the prediction results of the MLP model, and (d) shows the model prediction results in GraphSAGE. Based on the natural discontinuity method, the results are categorized into five levels: extremely low sensitivity, low sensitivity, moderate sensitivity, high sensitivity, and extremely high sensitivity. To quantify the model prediction results, the proportions of the four models at different sensitivities are statistically analyzed, such as... Figure 11 As shown.
[0093] from Figure 10 and Figure 11 It was found that the RF model predicted the most moderately sensitive areas among the four models, and these areas were widely distributed, accounting for approximately 23.19% of the study area. The SVM model predicted slightly better susceptibility classification results than the RF model, but the extremely high and high susceptibility areas were widely distributed, accounting for 11.33% and 18.27% of the study area, respectively. In the MLP prediction results, the classification of each category was improved compared to RF and SVM. In the prediction results of the method proposed in this application, the extremely high and extremely low sensitivity areas accounted for the highest proportion, accounting for 13.21% and 57.61% of the study area, respectively, and these areas coincided with the locations of most historical landslides.
[0094] To further verify the reliability of the model's predictions, field investigations can be conducted to confirm whether the highly sensitive areas predicted by the model actually experience landslides. To better demonstrate the prediction results, a landslide area not listed in the landslide inventory can be selected to assess the model's accuracy. Two typical areas were selected through field investigations. Figure 12 These are landslide susceptibility maps, Google imagery, and drone field survey images for regions A and B respectively. (a) and (b) are Google imagery of region A, (c) is a field survey photograph of region A, (d) to (g) are landslide susceptibility maps for region A, (h) to (j) are Google imagery of region B, (k) is a field survey photograph of region B, and (l) to (o) are landslide susceptibility maps for region B. Field surveys revealed that both landslide events occurred in June 2023. For region A, the proposed model predicts it to be an extremely sensitive region, while the SVR model predicts it to be a highly sensitive region, and the MLP and RF models predict it to be low and moderately sensitive. For region B, the proposed model predicts it to be an extremely sensitive region, while the SVR and MLP models predict it to be a highly sensitive region, and the RF model predicts it to be moderately sensitive. The proposed model can efficiently predict the location of landslides and is of great significance for landslide prevention and management.
[0095] This application first divides slope units based on the meander rate watershed method, constructs geographical nodes, and builds a directed weighted graph constrained by both environmental and node feature constraints. Second, it divides the environment according to environmental factors and creates a balanced sample dataset. Then, it constructs a graph neural network model to assess landslide susceptibility and analyzes the accuracy of the results. Finally, it compares the proposed method with existing classical machine learning methods to demonstrate its advantages.
[0096] This application fully considers the topographic and geological characteristics of the landslide-prone environment, introducing geographical environmental feature constraints to fully associate landslides with the landslide-prone environment. This significantly improves the transferability of machine learning methods and enhances the reliability and accuracy of large-scale landslide susceptibility assessment. On one hand, it proposes using a graph structure to associate landslides with the landslide-prone environment, using slope unit clustering of the landslide-prone environment to obtain geographical nodes of the landslide-prone environment, clearly expressing environmental characteristics, and constructing a directed weighted graph with geographical consistency constraints to optimize the distribution of landslide features. On the other hand, it introduces a graph neural network model, utilizing the structural information of the directed weighted graph with geographical consistency constraints to construct a graph neural network model for aggregating environmental information, generating a discriminative representation, and using this representation for supervised learning to achieve reliable landslide susceptibility assessment, effectively improving the generalization ability and reliability of machine learning-based landslide susceptibility assessment methods.
[0097] Based on the foregoing embodiments, this application further provides a landslide susceptibility assessment device under disaster-prone environmental constraints. The device includes various modules and units included in each module, which can be implemented by a processor in an electronic device; of course, it can also be implemented by specific logic circuits. In the implementation process, the processor can be a central processing unit (CPU), a microprocessor (MPU), a digital signal processor (DSP), or a field programmable gate array (FPGA), etc.
[0098] Figure 13 This application provides a schematic diagram of the composition of a landslide susceptibility assessment device under environmental constraints, as illustrated in the embodiments of this application. Figure 13 As shown, the device 1300 includes a node generation module 1310, a vital sign extraction module 1320, a correlation graph construction module 1330, and a landslide assessment module 1340, wherein:
[0099] The node generation module 1310 is used to construct slope units and generate multiple geographic nodes based on the meander watershed method.
[0100] The feature extraction module 1320 is used to extract the attribute features of each geographic node through spatial statistics.
[0101] The association graph construction module 1330 is used to construct a directed weighted graph that associates landslides with disaster-prone environments by combining environmental feature constraints and node feature constraints; wherein, the environmental feature constraints represent the environmental consistency between any two geographic nodes, and the node feature constraints represent the attribute feature similarity between any two geographic nodes.
[0102] The landslide assessment module 1340 is used to assess landslide susceptibility based on the directed weighted graph and a trained graph neural network model; wherein the graph neural network model is trained using a balanced sample dataset.
[0103] In some possible embodiments, the association graph construction module 1330 includes: a first determining unit, configured to determine a first similarity between the environmental factor features of any two geographic nodes and obtain a candidate set of connection nodes for each geographic node; a filtering unit, configured to filter neighbor nodes from the candidate set of connection nodes for each geographic node by calculating a second similarity between the attribute features of any two geographic nodes; and a connecting unit, configured to connect the corresponding geographic node and the neighbor node to form an edge and calculate the weight of the edge to obtain a directed weighted graph of the landslide and the disaster-prone environment association.
[0104] In some possible embodiments, the association graph construction module 1330 further includes a weight calculation unit, which is used to take the product of the first similarity calculated by the corresponding geographic node and the second similarity calculated by the neighboring node as the weight of the edge for each edge formed by the connection.
[0105] In some possible embodiments, the feature extraction module 1320 includes: a second determining unit, used to determine the average value of the influencing factors within the corresponding slope unit as the attribute feature of the geographic node for continuous landslide influencing factors; and a third determining unit, used to first reclassify the discrete landslide influencing factors according to categories, and then statistically analyze the influencing factors that appear most frequently within the corresponding slope unit as the attribute feature of the geographic node.
[0106] In some possible embodiments, the graph neural network model is a GraphSAGE network, including convolutional layers and fully connected layers. The landslide assessment module 1340 includes: a feature fusion unit, used to fuse the feature information of all neighboring nodes in the full neighbor set of any node in the directed weighted graph layer by layer using an aggregation function in the convolutional layer; a node representation unit, used to multiply the fused feature information with the weight matrix and then feed it to the fully connected layer through a nonlinear activation function to generate a node representation for the next layer; and a probability prediction unit, used to normalize the node representation of the last layer to obtain the probability of whether the corresponding geological node is a landslide, thereby completing the landslide susceptibility assessment.
[0107] In some possible embodiments, the landslide assessment module 1340 further includes a weight setting unit, used to set a weight for each neighbor node during the process of aggregating neighbor information, using the attribute value of the edge as the weight.
[0108] In some possible embodiments, the apparatus further includes a sample production module for classifying environments according to landslide impact factors, using cataloged landslide data as positive samples, and selecting negative samples in each environment in proportion to the positive samples to obtain the balanced sample dataset.
[0109] It should be noted that the descriptions of the above device embodiments are similar to those of the above method embodiments, and have similar beneficial effects. For technical details not disclosed in the device embodiments of this application, please refer to the descriptions of the method embodiments of this application for understanding.
[0110] It should be noted that, in the embodiments of this application, if the above-mentioned landslide susceptibility assessment method under disaster-prone environmental constraints is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiments of this application, or the part that contributes to the related technology, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause an electronic device to execute all or part of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), magnetic disks, or optical disks. Thus, the embodiments of this application are not limited to any specific hardware and software combination.
[0111] Correspondingly, embodiments of this application provide a computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements the steps in any of the landslide susceptibility assessment methods for disaster-prone environmental constraints described in the above embodiments. Correspondingly, embodiments of this application also provide a computer program product, which, when executed by a processor of an electronic device, is used to implement the steps in any of the landslide susceptibility assessment methods for disaster-prone environmental constraints described in the above embodiments.
[0112] Based on the same technical concept, this application provides an electronic device for implementing the landslide susceptibility assessment method with disaster-prone environmental constraints as described in the above method embodiments. Figure 14 This is a hardware entity diagram of an electronic device provided in an embodiment of this application, such as... Figure 14 As shown, the electronic device 1400 includes a memory 1410 and a processor 1420. The memory 1410 stores a computer program that can run on the processor 1420. When the processor 1420 executes the program, it implements the steps in the landslide susceptibility assessment method with disaster-prone environmental constraints according to any of the embodiments of this application.
[0113] The memory 1410 is configured to store instructions and applications executable by the processor 1420, and can also cache data to be processed or already processed by the processor 1420 and various modules in the electronic device (e.g., image data, audio data, voice communication data and video communication data), which can be implemented by flash memory or random access memory (RAM).
[0114] When processor 1420 executes a program, it implements the steps of the landslide susceptibility assessment method under the disaster-prone environmental constraints described above. Processor 1420 typically controls the overall operation of electronic equipment 1400.
[0115] The aforementioned processor can be at least one of the following: Application Specific Integrated Circuit (ASIC), Digital Signal Processor (DSP), Digital Signal Processing Device (DSPD), Programmable Logic Device (PLD), Field Programmable Gate Array (FPGA), Central Processing Unit (CPU), Controller, Microcontroller, and Microprocessor. It is understood that other electronic devices can also implement the functions of the aforementioned processor, and this application does not specifically limit the specific implementation.
[0116] The aforementioned computer storage media / memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), magnetic random access memory (FRAM), flash memory, magnetic surface memory, optical disc, or compact disc read-only memory (CD-ROM), etc.; or it can be various electronic devices that include one or any combination of the above-mentioned memories, such as mobile phones, computers, tablet devices, personal digital assistants, etc.
[0117] It should be noted that the descriptions of the storage medium and device embodiments above are similar to the descriptions of the method embodiments above, and have similar beneficial effects. For technical details not disclosed in the storage medium and device embodiments of this application, please refer to the descriptions of the method embodiments of this application for understanding.
[0118] It should be understood that the phrase "one embodiment" or "an embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "in one embodiment" or "in an embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. It should be understood that in the various embodiments of this application, the sequence numbers of the above-described processes do not imply a sequential order of execution; the execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application. The sequence numbers of the above-described embodiments are merely descriptive and do not represent the superiority or inferiority of the embodiments.
[0119] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.
[0120] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.
[0121] The units described above as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units; some or all of the units may be selected to achieve the purpose of the embodiments of this application according to actual needs. Furthermore, each functional unit in the embodiments of this application may be integrated into one processing unit, or each unit may be a separate unit, or two or more units may be integrated into one unit; the integrated unit may be implemented in hardware or in a combination of hardware and software functional units. Alternatively, if the integrated unit of this application is implemented as a software functional module and sold or used as an independent product, it may be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiments of this application, or the part that contributes to the related technology, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause the device automatic test line to execute all or part of the methods described in the embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROMs, magnetic disks, or optical disks.
[0122] The methods disclosed in the several method embodiments provided in this application can be arbitrarily combined to obtain new method embodiments without conflict. The features disclosed in the several method or device embodiments provided in this application can be arbitrarily combined to obtain new method embodiments or device embodiments without conflict.
[0123] The above description is merely an embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A landslide susceptibility assessment method under disaster-prone environmental constraints, characterized in that, The method includes: Slope units were constructed and multiple geographic nodes were generated based on the meander watershed method. The attribute features of each geographic node are extracted using spatial statistical methods; By combining environmental and nodal characteristic constraints, a directed weighted graph relating landslides to their disaster-prone environment is constructed. ； The environmental feature constraint represents the environmental consistency between any two geographic nodes, and the node feature constraint represents the attribute feature similarity between any two geographic nodes. Based on the directed weighted graph, a trained graph neural network model is used to assess landslide susceptibility; wherein the graph neural network model is trained using a balanced sample dataset.
2. The method according to claim 1, characterized in that, The method combines environmental feature constraints and node feature constraints to construct a directed weighted graph relating landslides to the disaster-prone environment, including: Determine the first similarity between the environmental factor features of any two geographic nodes, and obtain a candidate set of connection nodes for each geographic node; By calculating the second similarity between the attribute features of any two geographic nodes, neighboring nodes are selected from the candidate set of connected nodes of each geographic node. By connecting the corresponding geographical nodes and the neighboring nodes to form edges and calculating the weights of the edges, a directed weighted graph relating the landslide to the disaster-prone environment is obtained.
3. The method according to claim 2, characterized in that, The method further includes: For each edge formed by the connection, the product of the first similarity and the second similarity calculated by the corresponding geographical node and the neighboring node is used as the weight of the edge.
4. The method according to claim 1, characterized in that, The extraction of attribute features for each geographic node using spatial statistical methods includes: For continuous landslide influencing factors, the average value of the influencing factors within the corresponding slope unit is determined as the attribute characteristic of the geographic node; For discrete landslide influencing factors, they are first reclassified according to category, and then the influencing factor with the highest frequency of occurrence in the corresponding slope unit is counted as the attribute feature of the geographic node.
5. The method according to any one of claims 1 to 4, characterized in that, The graph neural network model is a GraphSAGE network, including convolutional layers and fully connected layers. The landslide susceptibility assessment based on the directed weighted graph, using the trained graph neural network model, includes: For any node in the directed weighted graph, the feature information of all neighboring nodes in the full neighbor set of the node is fused layer by layer in the convolutional layer using an aggregation function; The fused feature information is multiplied by the weight matrix and then fed into the fully connected layer through a nonlinear activation function to generate node representations for the next layer. By normalizing the node representations of the last layer, the probability of whether the corresponding geological node is a landslide is obtained, thus completing the landslide susceptibility assessment.
6. The method according to claim 5, characterized in that, The method further includes: In the process of aggregating neighbor information, a weight is set for each neighbor node, and the attribute value of the edge is used as the weight.
7. The method according to any one of claims 1 to 4, characterized in that, The method further includes: The environment is divided according to the landslide impact factors. The cataloged landslide data is used as a positive sample. In each environment, a negative sample with an equal proportion to the positive sample is selected to obtain the balanced sample dataset.
8. A landslide susceptibility assessment device under environmental constraints, characterized in that, The device includes a node generation module, a vital sign extraction module, a correlation graph construction module, and a landslide assessment module, wherein: The node generation module is used to construct slope units and generate multiple geographic nodes based on the meander watershed method. The feature extraction module is used to extract the attribute features of each geographic node through spatial statistics. The association graph construction module is used to construct a directed weighted graph that associates landslides with disaster-prone environments by combining environmental feature constraints and node feature constraints; wherein, the environmental feature constraints represent the environmental consistency between any two geographic nodes, and the node feature constraints represent the attribute feature similarity between any two geographic nodes. The landslide assessment module is used to assess landslide susceptibility based on the directed weighted graph and a trained graph neural network model; wherein the graph neural network model is trained using a balanced sample dataset.
9. An electronic device comprising a memory and a processor, the memory storing a computer program executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method according to any one of claims 1 to 7.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the method according to any one of claims 1 to 7.