Landslide susceptibility analysis method based on causal discovery and structural learning

Abstract

The present application provides a landslide susceptibility analysis method based on causal discovery and structural learning. First, multi-source geological data and landslide catalogs of the monitoring area are obtained to construct a set of hazard factors, which are divided into a causal discovery factor set and a susceptibility analysis factor set, wherein the susceptibility analysis factor set needs to be discretized; the monitoring area is divided into a causal discovery area and a susceptibility analysis area, and sample sets are constructed in sequence; an unsupervised causal discovery algorithm is used to construct a data-driven causal directed acyclic graph, and redundant data in the susceptibility analysis factor set and the sample set are eliminated based on the result; a Bayesian network model is constructed using a causal directed acyclic graph, and the susceptibility analysis sample set is used to perform training, prediction, and model evaluation in sequence, and finally a landslide susceptibility analysis is performed on the monitoring area. This method can fully explore the coupling relationship between landslides and hazard factors, construct a low-redundancy, highly interpretable model, and achieve more accurate landslide susceptibility analysis.

Description

Technical Field

[0001] The present application relates to the technical field of geological disaster monitoring, and in particular to a landslide susceptibility analysis method based on causal discovery and structural learning. Background Art

[0002] Landslide is one of the most common geological disasters, which can cause huge economic losses and casualties. Landslide susceptibility analysis based on machine learning algorithms is the basis and important decision-making support for data-driven intelligent disaster prevention and mitigation. Therefore, accurate and automated landslide susceptibility analysis has important academic and social significance.

[0003] Landslide susceptibility analysis refers to the use of multi-source hazard factors to evaluate the possibility of slope instability in a specific area, and visualize it in the form of a zoning map to intuitively understand the possibility of landslides in each evaluation unit in the area. Accurate landslide susceptibility analysis results can provide scientific decision-making references for disaster prevention and mitigation. Landslide susceptibility analysis can be divided into qualitative description and quantitative analysis. Qualitative description refers to the morphological description, kinematic path analysis, force analysis and disaster mechanism speculation of landslides based on geological dynamics through investigation and field surveys, providing susceptibility analysis for areas with the same terrain and disaster-prone environment. Quantitative analysis refers to the construction of models through geoscience data, geographic information science data processing methods and mathematical methods (mathematical formulas, models or algorithms), and the quantitative estimation of the possibility of slope instability for landslide susceptibility analysis. With the popularization of geological big data and machine learning algorithms, the research on geological disasters such as landslides has gradually shifted from the previous qualitative analysis to more applicable and objective quantitative analysis. According to the specific mathematical methods used, landslide susceptibility analysis can be divided into statistical learning methods and machine learning methods. Statistical methods, that is, the contribution of each disaster-prone environmental feature to the disaster is calculated through mathematical statistics and its corresponding formula, and the relative estimate of the high and low landslide susceptibility in all feature combinations is obtained through weighting, assignment, nonlinear transformation and other methods. The more commonly used methods are binary discriminant statistics, multivariate statistical discriminant methods, weighted certainty coefficient method, information method, and evidence weight method. Machine learning methods refer to the automatic learning of the causal laws and patterns of practice through models based on a large number of samples. The more commonly used methods are logistic regression models, naive Bayes models, decision tree models, random forest models, support vector machine models, artificial neural networks, convolutional neural networks, etc.

[0004] At present, there are problems in landslide susceptibility analysis, such as insufficient analysis of the coupling relationship between hazard factors, and most data-driven machine learning models ignore the actual disaster mechanism of landslides, resulting in low interpretability of landslide susceptibility analysis models, and structural learning models cannot fully utilize data coupling relationships. How to fully explore the coupling relationship between landslides and hazard factors, build a low-redundancy, highly interpretable landslide susceptibility analysis model, and achieve more accurate landslide susceptibility analysis is a problem that needs to be solved in this field. Summary of the invention

[0005] The purpose of the embodiments of the present application is to provide a landslide susceptibility analysis method based on causal discovery and structural learning, so as to fully explore the coupling relationship between landslides and disaster-causing factors, construct a landslide susceptibility analysis model with low redundancy and high interpretability, and achieve more accurate landslide susceptibility analysis.

[0006] In order to achieve the above purpose, the embodiments of the present application are implemented in the following manner:

[0007] In a first aspect, an embodiment of the present application provides a landslide susceptibility analysis method based on causal discovery and structural learning, comprising: obtaining multi-source geological data and a landslide catalog of a monitoring area, wherein the multi-source geological data is data containing geographic information from different sources, and the landslide catalog is data recording landslide information in the monitoring area, and the landslide information includes the time, location, scale, disaster type, development characteristics, and impact range of the landslide; based on the multi-source geological data, constructing a set of disaster factors, and further processing the disaster factor set to form a causal discovery factor set and a susceptibility analysis factor set, wherein the disaster factor set includes a variety of continuous disaster factors and a variety of discrete disaster factors; dividing the monitoring area into a causal discovery area and a susceptibility analysis area, and based on the causal discovery factor set and the causal discovery area A causal discovery sample set is constructed based on the landslide catalog of the susceptibility analysis factor set and the landslide catalog of the susceptibility analysis area, and a susceptibility analysis sample set is constructed, wherein the causal discovery sample set includes multiple positive samples and negative samples, and the susceptibility analysis sample set includes multiple positive samples and negative samples, the positive sample indicates that the sample label is a sample with a landslide, and the negative sample indicates that the sample label is a sample without a landslide; causal discovery is performed based on the causal discovery sample set and the causal discovery factor set to obtain a causal directed acyclic graph; a Bayesian network model is constructed using the causal directed acyclic graph, and the Bayesian network model is trained using the susceptibility analysis sample set to obtain a trained Bayesian network model; the trained Bayesian network model is used to perform landslide susceptibility analysis on the monitoring area to obtain landslide susceptibility analysis results.

[0008] In combination with the first aspect, in a first possible implementation of the first aspect, a disaster factor set is constructed based on multi-source geoscientific data, and the disaster factor set is further processed to form a causal discovery factor set and a susceptibility analysis factor set, including: preprocessing the multi-source geoscientific data to achieve the unification of data spatial benchmarks and spatial resolutions, and obtaining 22 disaster factors, wherein the disaster factors include the normalized vegetation index, average annual precipitation, water body distribution density, groundwater distribution, terrain humidity index, normalized water body index, and improved Normalized water body index, elevation, slope, terrain undulation index, surface curvature, horizontal curvature, profile curvature, fault distribution density, normalized building index, road distribution density and engineering geological rock groups, slope aspect, land use type, distance to roads, distance to faults, and distance to water bodies that belong to discrete hazard factors; all continuous hazard factors and discrete hazard factors are taken as causal discovery factor sets; all continuous hazard factors are discretized to obtain discrete hazard factors; the discrete hazard factors and discrete hazard factors are taken together as susceptibility analysis factor sets.

[0009] In combination with the first aspect, in a second possible implementation of the first aspect, a causal discovery sample set is constructed based on a causal discovery factor set and a landslide catalog of a causal discovery area, and a susceptibility analysis sample set is constructed based on a susceptibility analysis factor set and a landslide catalog of a susceptibility analysis area, including: using each disaster-causing factor in the causal discovery factor set, extracting attributes from the landslide catalog of the causal discovery area to form a causal discovery sample set, wherein the causal discovery sample set includes multiple positive samples and negative samples, the positive sample represents a sample with a sample label of a landslide, and the negative sample represents a sample with a sample label of no landslide; using each disaster-causing factor in the susceptibility analysis factor set, extracting attributes from the landslide catalog of the susceptibility analysis area to form a susceptibility analysis sample set, wherein the susceptibility analysis sample set includes multiple positive samples and negative samples, the positive sample represents a sample with a sample label of a landslide, and the negative sample represents a sample with a sample label of no landslide.

[0010] In combination with the first aspect, in a third possible implementation method of the first aspect, causal discovery is performed based on the causal discovery sample set and the causal discovery factor set to obtain a causal directed acyclic graph, including: based on the causal discovery sample set and the machine learning model, determining the pre-trained landslide susceptibility analysis result corresponding to each sample as the result variable corresponding to each sample; taking each disaster-causing factor in the causal discovery sample set as a cause variable; and performing causal discovery based on the result variable and the cause variable to determine the causal directed acyclic graph.

[0011] In combination with the third possible implementation method of the first aspect, in a fourth possible implementation method of the first aspect, causal discovery is performed based on the result variables and the cause variables to determine a causal directed acyclic graph, including: inputting the result variables and the cause variables into the NOTEARS algorithm for preliminary causal discovery, and determining a number of important disaster-causing factors from the cause variables; inputting the result variables and the important disaster-causing factors into the NOTEARS algorithm for secondary causal discovery, and constructing a causal directed acyclic graph, wherein, in the causal directed acyclic graph, important disaster-causing factors point to important disaster-causing factors, or, important disaster-causing factors point to the result variables.

[0012] In combination with the fourth possible implementation method of the first aspect, in the fifth possible implementation method of the first aspect, a Bayesian network model is constructed using a causal directed acyclic graph, and the Bayesian network model is trained using a susceptibility analysis sample set to obtain a trained Bayesian network model, including: reading the causal directed acyclic graph and defining a Bayesian network framework based on it; using the causal directed acyclic graph to perform factor screening on samples in the susceptibility analysis sample set, retaining the attributes extracted by the factors in each sample that are consistent with the causal directed acyclic graph; dividing the screened susceptibility analysis sample set into a training set and a test set; calculating the conditional probability of each node through the training set, thereby assigning conditional probability to each node in the Bayesian network framework to obtain a Bayesian network; using the test set to make predictions, and evaluating the model performance to obtain a trained Bayesian network model.

[0013] In combination with the fifth possible implementation manner of the first aspect, in a sixth possible implementation manner of the first aspect, calculating the conditional probability of each node through the training set includes: calculating the joint probability distribution of each node using the following formula:

[0014]

[0015] Among them, P(X 1 , X 2 , ..., X n ) represents the joint probability distribution, and the susceptibility analysis factor set X = {X 1 , X 2 , ..., X i , ..., X n}, representing all nodes in the Bayesian network framework, X i is the ith hazard factor, n is the total number of hazard factor types in the susceptibility analysis factor set, P a (X i ) is X i The parent node set of P(X i |P a (X i )) is X iThe conditional probability of the corresponding parent node in turn; the conditional probability of each node is calculated using the following formula:

[0016]

[0017] Among them, F(θ) is the likelihood function, which represents the joint probability of the model parameter E. The model parameter θ is a necessary parameter of the likelihood function. i ,θ) is the node X i The conditional probability density function of .

[0018] In combination with the sixth possible implementation manner of the first aspect, in a seventh possible implementation manner of the first aspect, using a test set for prediction includes: using each sample in the test set to test the Bayesian network, wherein the Bayesian network calculates the landslide susceptibility probability of the sample based on the following formula:

[0019]

[0020] Among them, P(Landslide|X 1 , X 2 , ..., X n ) indicates that the sample is in the susceptibility analysis factor set (X 1 , X 2 , ..., X n ) under the condition of the probability of landslide; P(X 1 , X 2 , ..., X n |Landslide) indicates that under the condition that a landslide occurs in a sample, the susceptibility analysis factor set (X 1 , X 2 , ..., X n ) occur simultaneously, P(Landslide) represents the probability that a sample is predicted to be a landslide, and P(X 1 , X 2 , ..., X n ) is the susceptibility analysis factor set (X 1 , X 2 , ..., X n ) occur simultaneously.

[0021] In combination with the fifth possible implementation method of the first aspect, in an eighth possible implementation method of the first aspect, a landslide susceptibility analysis is performed on the monitored area using a trained Bayesian network model to obtain a landslide susceptibility analysis result, including: processing the monitored area into multiple analysis units; for each analysis unit, extracting attributes of the analysis unit through a susceptibility analysis factor set to form input data corresponding to the analysis unit; inputting the input data into the trained Bayesian network model to obtain the landslide susceptibility analysis result output by the Bayesian network model.

[0022] Beneficial effects:

[0023] 1. This scheme obtains multi-source geological data and landslide catalogs in the monitoring area, constructs a disaster factor set based on the multi-source geological data, and further processes the disaster factor set to form a causal discovery factor set and a susceptibility analysis factor set, wherein the disaster factor set includes a variety of continuous disaster factors and a variety of discrete disaster factors; divides the monitoring area into a causal discovery area and a susceptibility analysis area, and constructs a causal discovery sample set based on the causal discovery factor set and the landslide catalogs in the causal discovery area, and constructs a susceptibility analysis sample set based on the susceptibility analysis factor set and the landslide catalogs in the susceptibility analysis area; then, causal discovery is performed based on the causal discovery sample set and the causal discovery factor set to obtain a causal directed acyclic graph; a Bayesian network model is constructed using the causal directed acyclic graph, and the Bayesian network model is trained using the susceptibility analysis sample set to obtain a trained Bayesian network model; the trained Bayesian network model is used to perform landslide susceptibility analysis on the monitoring area to obtain landslide susceptibility analysis results. This method fully realizes the mining of information in disaster geographic data, takes causal relationships as mapping, comprehensively considers the coupling relationship between landslides and hazard factors, and builds a connection between data and the landslide mechanism in the real world. Its application can provide new ideas for data-driven disaster prevention and mitigation. The model guided by causality can be studied on a larger scale, and the model structure has a strong migration ability (for quantitative research on similar landslides), so it can provide a stable landslide susceptibility analysis solution. At the same time, the directed acyclic graph constructed by causality can give full play to the advantages of Bayesian networks and build a model with interpretability, representativeness, and low feature space redundancy. Therefore, the unsupervised causal discovery combined with the structured learning Bayesian network model can fully explore the coupling relationship between landslides and hazard factors, build a low-redundancy and highly interpretable landslide susceptibility analysis model, and achieve more accurate landslide susceptibility analysis, providing application references for the industry and providing new ideas for data-driven intelligent disaster prevention and mitigation.

[0024] 2. The technical route of unsupervised causal discovery - structural learning - landslide susceptibility assessment adopted in this scheme is the first of its kind in this field. In the process of causal analysis, causality-based feature selection is used to constrain the directed acyclic graph (twice causal analysis) to optimize the model scale. For the probabilistic graphical model Bayesian network, the interpretability and stability of the model are enhanced by introducing causal relationships and causal directed acyclic graphs, so that it can be applied to larger-scale research areas.

[0025] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, preferred embodiments are specifically cited below and described in detail with reference to the attached drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments of the present application will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0027] Figure 1 A flow chart of a landslide susceptibility analysis method based on causal discovery and structural learning provided in an embodiment of the present application. DETAILED DESCRIPTION

[0028] The technical solutions in the embodiments of the present application will be described below in conjunction with the drawings in the embodiments of the present application.

[0029] To facilitate the understanding of this solution, some concepts involved in this solution are introduced here.

[0030] Landslide susceptibility analysis: Landslide susceptibility analysis refers to the use of multi-source hazard factors to evaluate the possibility of slope instability in a specific area, and visualize it in the form of a zoning map to intuitively determine the possibility of landslides in each evaluation unit in the area.

[0031] Machine learning: Machine learning refers to a research method that uses multidisciplinary methods such as probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory to simulate or implement human learning behavior by computer to acquire new knowledge or skills and perform automatic pattern recognition, classification, regression and other tasks.

[0032] Landslide catalog: Landslide catalog refers to a database that records information such as the time, location, scale, disaster type, development characteristics, and impact range of landslides.

[0033] Causal statistics: Causal statistics refers to causal inference based on statistical principles, which infers and proves how changes in one variable affect changes in another variable, thereby clarifying the cause-effect relationship in complex systems.

[0034] Causal discovery: Causal discovery in a broad sense refers to the discovery of causal relationships between variables. Data-driven causal discovery refers to the process of finding causal relationships between variables from observed data without experimental intervention. Its purpose is to infer causal relationships between variables based on statistical patterns and dependencies observed in the data, thereby revealing the inherent causal structure of the system.

[0035] Correlation: Correlation refers to the statistical relationship between two variables. It measures how closely changes in one variable are associated with changes in another variable.

[0036] Causality: Causality refers to the relationship between cause and effect, where a change in one variable directly leads to a change in another variable.

[0037] Directed Acyclic Graph (DAG): In digital graph theory and computer science, a directed acyclic graph refers to a directed graph without loops, which is used to represent the relationship between variables in a system, where the edges (arrows) between nodes (variables) represent directional relationships. The graph has direction but no cycles (i.e., it is impossible to return to the starting node along the path on the edge). A causal directed acyclic graph refers to a directed acyclic graph that reflects causal dependencies.

[0038] Unsupervised learning / Unsupervised methods: Unsupervised is a type of machine learning where a model is trained on input data without explicit supervision or labeled outputs. The goal is to learn the underlying structure, patterns, or relationships in the data without being given a specific target or outcome to predict.

[0039] Probabilistic graphical models: Probabilistic graphical models are graphical representations of probability distributions that capture the probabilistic relationships between a set of random variables and provide a way to visualize and reason about complex probabilistic dependencies in a structured way.

[0040] Causal weight score: Causal discovery algorithms can be divided into score-based algorithms and constraint-based algorithms. Score-based algorithms calculate the quantitative results of the causal effect of each factor relative to other factors and present them as scores.

[0041] Bayesian network: A Bayesian network is a probabilistic graphical model based on Bayesian probability theory and graph theory, also known as a Bayesian belief network. It completes specific tasks by utilizing the graphical probabilistic relationship between a set of variables.

[0042] Coupling relationship: Coupling relationship refers to the interaction or relationship between different components or systems in a complex system, in which the change of one component or system will affect the behavior or state of another component or system.

[0043] Multi-source geoscientific data refers to data containing geographic information (i.e. spatial location and attribute information) with different sources, formats, and acquisition methods.

[0044] Engineering geological rock group: The engineering geological rock group is a rock group divided according to the lithology classification standards commonly used in engineering geology. The rock group is a formal rock stratigraphic unit equivalent to the first level of normal rock stratigraphic unit group, which is further divided according to the characteristics of rock (or rock combination) in the metamorphic rock group.

[0045] Normalized Difference Vegetation Index: Normalized Difference Vegetation Index is one of the important parameters reflecting the growth of crops in remote sensing image processing. By calculating this index, the characteristics of vegetation in the image are highlighted, which can more conveniently and accurately extract vegetation categories or estimate green biomass.

[0046] Normalized Water Index / Improved Normalized Water Index: Both are important parameters that reflect water areas in remote sensing image processing. By calculating this index, the characteristics of water bodies in the image are highlighted. The difference between the Normalized Water Index and the Improved Normalized Water Index is that the remote sensing image bands involved in the calculation are different.

[0047] Normalized Building Index: Normalized Building Index is one of the important parameters in remote sensing image processing that reflects the distribution of buildings and impervious surfaces. By calculating this index, the features of buildings and impervious surfaces in the image are highlighted, making it easier to extract the corresponding objects.

[0048] The above is an introduction to some relevant concepts that are helpful for understanding this solution. The following is an introduction to this solution.

[0049] See also Figure 1 , Figure 1 A flow chart of a landslide susceptibility analysis method based on causal discovery and structural learning provided in an embodiment of the present application. In this embodiment, the landslide susceptibility analysis method based on causal discovery and structural learning is run by an electronic device (such as a personal computer, server, etc.), and may include step S10, step S20, step S30, step S40, and step S50.

[0050] First, the electronic device may execute step S10.

[0051] Step S10: Obtain multi-source geological data and a landslide catalog of the monitoring area, wherein the multi-source geological data is data containing geographic information from different sources, and the landslide catalog is data recording landslide information in the monitoring area, and the landslide information includes the time, location, scale, disaster type, development characteristics, and impact range of the landslide.

[0052] In this embodiment, the electronic device can obtain multi-source geological data and landslide catalogs in the monitoring area. Multi-source geological data is data containing geographical information from different sources, and landslide catalogs are data recording landslide information in the monitoring area, including the time, location, scale, disaster type, development characteristics, and impact range of the landslide.

[0053] After obtaining the multi-source geoscientific data and the landslide catalog of the monitoring area, the electronic device may execute step S20.

[0054] Step S20: Based on multi-source geological data, a disaster factor set is constructed, and the disaster factor set is further processed to form a causal discovery factor set and a susceptibility analysis factor set, wherein the disaster factor set includes a variety of continuous disaster factors and a variety of discrete disaster factors.

[0055] In this embodiment, the electronic device can construct a set of disaster factors based on multi-source geological data.

[0056] Electronic equipment can pre-process multi-source geological data, unify the data spatial benchmark and spatial resolution, and obtain 22 disaster-causing factors.

[0057] Specifically, Arcmap and Google Earth Engine platforms were used to process all data into 22 disaster factors in .tiff format, and the spatial resolution was unified to 30m and the same data spatial reference system. Among them, the disaster factors include the normalized vegetation index, average annual precipitation, water distribution density, groundwater distribution, terrain humidity index, normalized water index, improved normalized water index, elevation, slope, terrain relief index, surface curvature, horizontal curvature, profile curvature, fault distribution density, normalized building index, road distribution density, and engineering geological rock group, slope aspect, land use type, distance to road, distance to fault, and distance to water body, which are continuous disaster factors. The disaster factors are shown in Table 1 below:

[0058] Table 1. Disaster factors

[0059]

[0060]

[0061] After constructing the disaster-causing factor set, the electronic device can further process the disaster-causing factor set to form a causal discovery factor set and a susceptibility analysis factor set.

[0062] Exemplarily, the electronic device can use all continuous disaster factors and discrete disaster factors as a causal discovery factor set, that is, the causal discovery factor set includes normalized vegetation index, average annual precipitation, water body distribution density, groundwater distribution, terrain moisture index, normalized water body index, improved normalized water body index, elevation, slope, terrain undulation index, surface curvature, horizontal curvature, profile curvature, fault distribution density, normalized building index, road distribution density, engineering geological rock group, slope aspect, land use type, distance to road, distance to fault, and distance to water body.

[0063] Furthermore, the electronic device can discretize all continuous disaster factors to obtain discrete disaster factors, and then use the discretized disaster factors and discrete disaster factors together as a susceptibility analysis factor set. Then, the susceptibility analysis factor set includes engineering geological rock formations, slope aspect, land use type, distance to roads, distance to faults, distance to water bodies, as well as discretized normalized vegetation index, average annual precipitation, water body distribution density, groundwater distribution, terrain humidity index, normalized water body index, improved normalized water body index, elevation, slope, terrain relief index, surface curvature, horizontal curvature, profile curvature, fault distribution density, normalized building index, and road distribution density.

[0064] After constructing the causal discovery factor set and the susceptibility analysis factor set, the electronic device may execute step S30.

[0065] Step S30: Divide the monitoring area into a causal discovery area and a susceptibility analysis area, and construct a causal discovery sample set based on the causal discovery factor set and the landslide catalog of the causal discovery area, and construct a susceptibility analysis sample set based on the susceptibility analysis factor set and the landslide catalog of the susceptibility analysis area, wherein the causal discovery sample set includes multiple positive samples and negative samples, and the susceptibility analysis sample set includes multiple positive samples and negative samples, the positive sample indicates a sample with a sample label of a landslide, and the negative sample indicates a sample with a sample label of a landslide not occurring.

[0066] In this embodiment, the electronic device may divide the monitoring area into a causal discovery area and a susceptibility analysis area. Then, the electronic device may construct a causal discovery sample set based on the causal discovery factor set and the landslide catalog of the causal discovery area.

[0067] Exemplarily, the electronic device can use each disaster factor in the causal discovery factor set to extract attributes of the landslide catalog in the causal discovery area to form a causal discovery sample set, wherein the causal discovery sample set includes multiple positive samples and negative samples, and the positive sample indicates a sample with a sample label of a landslide, and the negative sample indicates a sample with a sample label of no landslide. In this embodiment, the causal discovery sample set is constructed with a ratio of positive samples to negative samples of 1:3. The number of positive samples in the causal discovery area can be determined by the landslide catalog, and then the number of negative samples can be determined to construct the causal discovery sample set. In other embodiments, other sample ratios can also be used to construct the causal discovery sample set, which is not limited here.

[0068] And, the electronic device can construct a susceptibility analysis sample set based on the susceptibility analysis factor set and the landslide catalog of the susceptibility analysis area.

[0069] Exemplarily, the electronic device can use each disaster-causing factor in the susceptibility analysis factor set to extract attributes of the landslide catalog in the susceptibility analysis area to form a susceptibility analysis sample set, wherein the susceptibility analysis sample set includes multiple positive samples and negative samples, and the positive sample represents a sample with a sample label of a landslide, and the negative sample represents a sample with a sample label of no landslide. In this embodiment, the causal discovery sample set is constructed with a ratio of positive samples to negative samples of 1:3. The number of positive samples in the susceptibility analysis area can be determined by the landslide catalog, and then the number of negative samples can be determined to construct the susceptibility analysis sample set. In other embodiments, other sample ratios can also be used to construct the susceptibility analysis sample set, which is not limited here.

[0070] After constructing the causal discovery sample set and the susceptibility analysis sample set, the electronic device may execute step S40.

[0071] Step S40: Perform causal discovery based on the causal discovery sample set and the causal discovery factor set to obtain a causal directed acyclic graph.

[0072] In this embodiment, the electronic device may perform causal discovery based on the causal discovery sample set and the causal discovery factor set.

[0073] Exemplarily, the electronic device can determine the pre-trained landslide susceptibility analysis result corresponding to each sample based on the causal discovery sample set and the machine learning model as the result variable corresponding to each sample. For example, a part of the causal discovery sample set is used to train the machine learning model (such as random forest, decision tree, support vector machine, etc.), and the pre-trained landslide susceptibility analysis is performed through the machine learning model to obtain the pre-trained landslide susceptibility analysis result as the result variable.

[0074] Afterwards, the electronic device can use each disaster-causing factor in the causal discovery sample set as a cause variable, perform causal discovery based on the result variable and the cause variable, and determine a causal directed acyclic graph.

[0075] For example, the electronic device can input the result variable and the cause variable into the NOTEARS algorithm for preliminary causal discovery, and determine several important disaster-causing factors from the cause variable. For example, preliminary causal discovery analyzes the important disaster-causing factors (i.e., the cause variable that has a more obvious impact on the result variable).

[0076] Then, the electronic device can input the result variables and important disaster factors into the NOTEARS algorithm for secondary causal discovery and construct a causal directed acyclic graph, wherein in the causal directed acyclic graph, important disaster factors point to important disaster factors, or, important disaster factors point to the result variables.

[0077] When using the NOTEARS algorithm for causal discovery, the learning process of the traditional causal directed acyclic graph is shown in formula (1):

[0078]

[0079] in, is the scoring function, d is the total number of nodes, F(W) is the constraint equation for causal discovery, G(W) is the causal directed acyclic graph constrained by the score matrix W under d nodes, and DAGs is a directed acyclic graph representing the causal relationship between variables.

[0080] In this embodiment, in order to process continuous variables in causal discovery, the search process of the causal directed acyclic graph is transformed into a pure continuous non-combinatorial constraint optimization of a real matrix. First, l 1 Regularization (||W|| 1 =||vecW|| 1 )Solve Therefore, the construction of the causal graph model is transformed into formula (2):

[0081]

[0082] Among them, ||X-XW|| F represents the Frobenius criterion, also known as L 2 paradigm, n represents the number of samples, limited by the graph structure of the directed acyclic graph, the solution hour, It is difficult to use continuous data for calculation. Therefore, a smooth function h(W) is used here to calculate G(W). For the function h, when h(W) = 0, there is Minimum, that is, the causal directed acyclic graph search effect is the best. Function h(W) satisfies:

[0083]

[0084] in, is the Hadamard transformation matrix, is the relevant matrix index, and tr represents the trace of the matrix, that is, the sum of the elements on the diagonal of the matrix. Based on the quadratic penalty term, the enhanced Lagrangian method is used to optimize the calculation to minimize the result. The quadratic penalty term is:

[0085]

[0086] Among them, ρ is the regularization parameter and α is the Lagrange multiplier.

[0087] Therefore, this embodiment first uses the sample landslide susceptibility analysis results (i.e., pre-trained landslide susceptibility analysis results) to train the NOTEARS algorithm. Since the causal directed acyclic graph structure finally obtained should not be too complex, this embodiment uses a two-round training method to successively eliminate the hazard factors with low causal correlation with the landslide. In the first round of training, 22 hazard factors were used for training to obtain 9 factors that were directly causally related to the landslide; in the second round of training, these factors were used for further modeling, and finally a causal directed acyclic graph was constructed according to the causal weight score and stored as a .dot format file.

[0088] After obtaining the causal directed acyclic graph, the electronic device may execute step S50.

[0089] Step S50: constructing a Bayesian network model using the causal directed acyclic graph, and training the Bayesian network model using the susceptibility analysis sample set to obtain a trained Bayesian network model.

[0090] In this embodiment, the electronic device can read the causal directed acyclic graph and define the Bayesian network framework accordingly. Then, the causal directed acyclic graph is used to screen the samples in the susceptibility analysis sample set, and the attributes extracted by the factors consistent with the causal directed acyclic graph in each sample are retained; the screened susceptibility analysis sample set is divided into a training set and a test set (for example, a 7:3 division), and the conditional probability of each node is calculated through the training set, so as to assign a conditional probability to each node in the Bayesian network framework and obtain a Bayesian network.

[0091] Exemplarily, the electronic device may calculate the conditional probability of each node in the following manner. Specifically, the joint probability distribution of each node may be calculated using the following formula:

[0092]

[0093] Among them, P(X 1 , X 2 , ..., X n) represents the joint probability distribution, and the susceptibility analysis factor set X = {X 1 , X 2 , ..., X i , ..., X n}, representing all nodes in the Bayesian network framework, X i is the ith hazard factor, n is the total number of hazard factor types in the susceptibility analysis factor set, P a (X i ) is X i The parent node set of P(X i |P a (X i )) is X i The conditional probabilities of the corresponding parent nodes are calculated in turn. And the conditional probability of each node is calculated using the following formula:

[0094]

[0095] Among them, F(θ) is the likelihood function, which represents the joint probability of the model parameter θ. The model parameter θ is a necessary parameter of the likelihood function. i ,θ) is the node X i The conditional probability density function of .

[0096] Thus, the electronic device can assign conditional probabilities to each node in the Bayesian network framework to obtain a Bayesian network. Then, the test set can be used for prediction, specifically:

[0097] The electronic device can use each sample in the test set to test the Bayesian network, where the Bayesian network calculates the landslide susceptibility probability of the sample based on the following formula:

[0098]

[0099] Among them, P(Landslide|X 1 , X 2 , ..., X n ) indicates that the sample is in the susceptibility analysis factor set (X 1 , X 2 , ..., X n ) under the condition of the probability of landslide; P(X 1 , X 2 , ..., X n |Landslide) indicates that under the condition that a landslide occurs in a sample, the susceptibility analysis factor set (X 1 , X 2 , ..., X n ) occur simultaneously, P(Landslide) represents the probability that a sample is predicted to be a landslide, and P(X 1, X 2 , ..., X n ) is the susceptibility analysis factor set (X 1 , X 2 , ..., X n ) occur simultaneously.

[0100] In this way, the landslide susceptibility probability of each sample in the test set is predicted, and then the model can be evaluated based on the test situation, and finally a trained Bayesian network model is obtained.

[0101] After obtaining the Bayesian network model, the electronic device may execute step S60.

[0102] Step S60: using the trained Bayesian network model to perform landslide susceptibility analysis on the monitoring area to obtain landslide susceptibility analysis results.

[0103] In this embodiment, the electronic device can process the monitoring area into multiple analysis units. For each analysis unit, the attributes of the analysis unit are extracted through the susceptibility analysis factor set to form input data corresponding to the analysis unit, and then the input data is input into the trained Bayesian network model to obtain the landslide susceptibility analysis results output by the Bayesian network model.

[0104] Landslide susceptibility analysis can be regarded as a binary classification process (landslide / non-landslide). The final output of landslide susceptibility indicates the possibility of a sample being classified as "landslide" in the range of [0, 1]. After constructing the Bayesian network, landslide susceptibility analysis is performed. The final result is that the landslide susceptibility level can be divided into five intervals with an interval of 0.2 from high to low, which are extremely low (0-0.2), low (0.2-0.4), medium (0.4-0.6), high (0.6-0.8) and extremely high (0.8-1), and finally the landslide susceptibility mapping result is obtained.

[0105] In summary, the embodiment of the present application provides a landslide susceptibility analysis method based on causal discovery and structural learning, by acquiring multi-source geological data and a landslide catalog in a monitoring area, constructing a disaster factor set based on the multi-source geological data, and further processing the disaster factor set to form a causal discovery factor set and a susceptibility analysis factor set, wherein the disaster factor set includes a variety of continuous disaster factors and a variety of discrete disaster factors; the monitoring area is divided into a causal discovery area and a susceptibility analysis area, and a causal discovery sample set is constructed based on the causal discovery factor set and the landslide catalog in the causal discovery area, and a susceptibility analysis sample set is constructed based on the susceptibility analysis factor set and the landslide catalog in the susceptibility analysis area; then, causal discovery is performed based on the causal discovery sample set and the causal discovery factor set to obtain a causal directed acyclic graph; a Bayesian network model is constructed using the causal directed acyclic graph, and the Bayesian network model is trained using the susceptibility analysis sample set to obtain a trained Bayesian network model; and a landslide susceptibility analysis result is obtained by performing a landslide susceptibility analysis on the monitoring area using the trained Bayesian network model. This method fully realizes the mining of information in disaster geographic data, takes causal relationships as mapping, comprehensively considers the coupling relationship between landslides and hazard factors, and builds a connection between data and the landslide mechanism in the real world. Its application can provide new ideas for data-driven disaster prevention and mitigation. The model guided by causality can be studied on a larger scale, and the model structure has a strong migration ability (for quantitative research on similar landslides), so it can provide a stable landslide susceptibility analysis solution. At the same time, the directed acyclic graph constructed by causality can give full play to the advantages of Bayesian networks and build a model with interpretability, representativeness, and low feature space redundancy. Therefore, the unsupervised causal discovery combined with the structured learning Bayesian network model can fully explore the coupling relationship between landslides and hazard factors, build a low-redundancy and highly interpretable landslide susceptibility analysis model, and achieve more accurate landslide susceptibility analysis, providing application references for the industry and providing new ideas for data-driven intelligent disaster prevention and mitigation.

[0106] The technical route of unsupervised causal discovery - structural learning - landslide susceptibility assessment adopted in this scheme is the first of its kind in this field. In the process of causal analysis, causality-based feature selection is used to constrain the directed acyclic graph (twice causal analysis) to optimize the model scale. For the probabilistic graphical model Bayesian network, the interpretability and stability of the model are enhanced by introducing causal relationships and causal directed acyclic graphs, so that it can be applied to larger-scale research areas.

[0107] In this document, relational terms such as first and second, etc. are used merely to distinguish one entity or operation from another entity or operation, but do not necessarily require or imply any such actual relationship or order between these entities or operations.

[0108] The above description is only an embodiment of the present application and is not intended to limit the protection scope of the present application. For those skilled in the art, the present application may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A landslide susceptibility analysis method based on causal discovery and structural learning, characterized in that: include: Obtain multi-source geoscientific data and landslide catalogs in the monitoring area, where multi-source geoscientific data refers to data from different sources containing geographic information, and landslide catalogs refer to data recording landslide information in the monitoring area, including landslide occurrence time, location, scale, disaster type, development characteristics, and impact range; Based on multi-source geological data, a disaster factor set is constructed, and the disaster factor set is further processed to form a causal discovery factor set and a susceptibility analysis factor set. The disaster factor set includes a variety of continuous disaster factors and a variety of discrete disaster factors. The monitoring area is divided into a causal discovery area and a susceptibility analysis area, and a causal discovery sample set is constructed based on the causal discovery factor set and the landslide catalog of the causal discovery area, and a susceptibility analysis sample set is constructed based on the susceptibility analysis factor set and the landslide catalog of the susceptibility analysis area, wherein the causal discovery sample set includes multiple positive samples and negative samples, and the susceptibility analysis sample set includes multiple positive samples and negative samples, where the positive sample indicates that the sample label is a sample with a landslide, and the negative sample indicates that the sample label is a sample with no landslide; Causal discovery is performed based on the causal discovery sample set and the causal discovery factor set to obtain a causal directed acyclic graph; A Bayesian network model is constructed using a causal directed acyclic graph, and the Bayesian network model is trained using a susceptibility analysis sample set to obtain a trained Bayesian network model; The trained Bayesian network model is used to analyze the landslide susceptibility of the monitoring area and obtain the landslide susceptibility analysis results; Causal discovery is performed based on the causal discovery sample set and the causal discovery factor set to obtain a causal directed acyclic graph, including: Based on the causal discovery sample set and the machine learning model, the pre-trained landslide susceptibility analysis results corresponding to each sample are determined as the result variables corresponding to each sample; Each hazard factor in the causal discovery sample set is taken as a causal variable; Based on the result variables and cause variables, causal discovery is performed to determine the causal directed acyclic graph; Based on the result variables and cause variables, causal discovery is performed to determine the causal directed acyclic graph, including: The outcome variables and cause variables were input into the NOTEARS algorithm for preliminary causal discovery, and several important disaster-causing factors were identified from the cause variables; The outcome variables and important disaster factors are input into the NOTEARS algorithm for secondary causal discovery, and a causal directed acyclic graph is constructed, in which important disaster factors point to important disaster factors, or important disaster factors point to outcome variables.

2. The landslide susceptibility analysis method based on causal discovery and structural learning according to claim 1 is characterized in that: Based on multi-source geoscience data, a disaster factor set is constructed, and the disaster factor set is further processed to form a causal discovery factor set and a susceptibility analysis factor set, including: Preprocess multi-source geoscience data to unify the spatial benchmark and spatial resolution of the data, and obtain 22 disaster factors, among which the disaster factors include the normalized vegetation index, average annual precipitation, water distribution density, groundwater distribution, terrain humidity index, normalized water index, improved normalized water index, elevation, slope, terrain relief index, surface curvature, horizontal curvature, profile curvature, fault distribution density, normalized building index, road distribution density, and engineering geological rock group, slope aspect, land use type, distance to road, distance to fault, and distance to water body, which belong to the continuous disaster factors; All continuous disaster factors and discrete disaster factors are regarded as causal discovery factor sets; Discretize all continuous disaster factors to obtain discrete disaster factors; Discrete disaster factors and discrete disaster factors are taken together as the susceptibility analysis factor set.

3. The landslide susceptibility analysis method based on causal discovery and structural learning according to claim 1 is characterized in that: Based on the causal discovery factor set and the landslide catalog of the causal discovery area, a causal discovery sample set is constructed. Based on the susceptibility analysis factor set and the landslide catalog of the susceptibility analysis area, a susceptibility analysis sample set is constructed, including: Using each disaster factor in the causal discovery factor set, the attributes of the landslides in the causal discovery area are cataloged and extracted to form a causal discovery sample set, where the causal discovery sample set contains multiple positive samples and negative samples. The positive sample indicates that the sample label is a sample with a landslide, and the negative sample indicates that the sample label is a sample without a landslide. Using each disaster factor in the susceptibility analysis factor set, the attributes of the landslide catalog in the susceptibility analysis area are extracted to form a susceptibility analysis sample set, where the susceptibility analysis sample set contains multiple positive samples and negative samples. Positive samples represent samples with a sample label of landslide, and negative samples represent samples with a sample label of no landslide.

4. The landslide susceptibility analysis method based on causal discovery and structural learning according to claim 1 is characterized in that: A Bayesian network model is constructed using a causal directed acyclic graph, and the Bayesian network model is trained using a susceptibility analysis sample set to obtain a trained Bayesian network model, including: Read the causal directed acyclic graph and define the Bayesian network framework based on it; The causal directed acyclic graph is used to screen the samples in the susceptibility analysis sample set, and the attributes extracted by the factors consistent with the causal directed acyclic graph in each sample are retained; Divide the screened susceptibility analysis sample set into a training set and a test set; The conditional probability of each node is calculated through the training set, so as to assign conditional probability to each node in the Bayesian network framework and obtain the Bayesian network; The test set is used to make predictions and evaluate the model performance to obtain the trained Bayesian network model.

5. The landslide susceptibility analysis method based on causal discovery and structural learning according to claim 4 is characterized in that: The conditional probability of each node is calculated through the training set, including: The joint probability distribution of each node is calculated using the following formula: Among them, P(X1,X2,…,X n ) represents the joint probability distribution, and the susceptibility analysis factor set X = {X1, X2, …, X i ,…,X n }, representing all nodes in the Bayesian network framework, X i is the ith hazard factor, n is the total number of hazard factor types in the susceptibility analysis factor set, P a (X i ) is X i The parent node set of P(X i |P a (X i )) is X i The conditional probabilities corresponding to the parent nodes in turn; The conditional probability of each node is calculated using the following formula: Among them, F(θ) is the likelihood function, which represents the joint probability of the model parameter θ. The model parameter θ is a necessary parameter of the likelihood function. i ,θ) is the node X i The conditional probability density function of .

6. The landslide susceptibility analysis method based on causal discovery and structural learning according to claim 5 is characterized in that: The test set is used for prediction, including: The Bayesian network is tested using each sample in the test set, where the Bayesian network calculates the landslide susceptibility probability of the sample based on the following formula: Among them, P(Landslide|X1,X2,…,X n ) indicates that the sample is in the susceptibility analysis factor set (X1, X2, …, X n ), the probability of landslide under the condition of n |Landslide) indicates that under the condition that a landslide occurs in a sample, the susceptibility analysis factor set (X1, X2, …, X n ) occur simultaneously, P(Landslide) represents the probability that a sample is predicted to be a landslide, P(X1,X2,…,X n ) is the set of susceptibility analysis factors (X1, X2, …, X n ) occur simultaneously.

7. The landslide susceptibility analysis method based on causal discovery and structural learning according to claim 4 is characterized in that: The trained Bayesian network model is used to analyze the landslide susceptibility of the monitoring area, and the landslide susceptibility analysis results are obtained, including: Processing the monitoring area into multiple analysis units; For each analysis unit, extract attributes of the analysis unit through the susceptibility analysis factor set to form the input data corresponding to the analysis unit; The input data is input into the trained Bayesian network model to obtain the landslide susceptibility analysis results output by the Bayesian network model.

Images