Rainfall type landslide susceptibility evaluation method, system, equipment and medium

Abstract

The invention provides a rainfall-type landslide susceptibility evaluation method, system and device and a medium, and belongs to the technical field of geological disaster risk evaluation and spatial information modeling. The method comprises the following steps: collecting and preprocessing multi-source environment and rainfall data of a target area; constructing a multi-time scale rainfall factor system, and reducing variable colinearity through intra-group screening and a multi-stage feature engineering method; modeling the static environment factor by using a random forest model, obtaining a landslide background susceptibility probability, and converting the landslide background susceptibility probability into a logarithmic probability offset item; under the constraint of an offset term, constructing a Bayesian spatial statistical model, introducing a nonlinear effect term and a spatial random effect term of a rainfall factor, and carrying out joint modeling and Bayesian inference on the occurrence probability of the landslide; and landslide susceptibility dynamic prediction and spatial mapping are realized. According to the method, the nonlinear prediction capability of machine learning and the physical constraint and spatial modeling capability of Bayesian statistics are fused, so that the stability, prediction precision and mechanism interpretation of an evaluation result are improved.

Description

Technical Field

[0001] This invention relates to the field of geological disaster risk assessment and spatial information modeling technology, and in particular to a method, system, equipment and medium for evaluating the susceptibility of rainfall-induced landslides. Background Technology

[0002] Landslides are a typical geological hazard influenced by multiple factors such as topography, geological conditions, and rainfall. Their occurrence process exhibits significant nonlinear characteristics and spatial heterogeneity. Conducting reliable landslide susceptibility assessments is crucial for regional disaster prevention and mitigation, infrastructure site selection, and land spatial planning. Existing landslide susceptibility assessment methods mainly include empirical statistical models and machine learning-based predictive models. Traditional statistical methods typically rely on linear or weakly nonlinear assumptions, making it difficult to fully characterize the landslide occurrence mechanism under the coupling of multiple factors. While machine learning methods such as random forests and support vector machines have certain advantages in predictive accuracy, their internal decision-making processes are complex, and the model results are difficult to interpret, limiting their application in engineering practice and risk management.

[0003] In existing technologies, most methods focus on improving prediction accuracy, often incorporating relatively static factors such as topography and geology with dynamic triggering factors such as rainfall into a single model for training. This fails to effectively distinguish the different mechanisms of landslide background susceptibility and rainfall triggering processes, resulting in a lack of physical consistency in the model's response to time-varying rainfall conditions. Furthermore, rainfall indicators at multiple time scales generally exhibit strong correlations in practical applications. Without a systematic screening and constraint mechanism, multicollinearity problems can easily be introduced, thereby reducing model stability and generalization ability.

[0004] Furthermore, existing landslide susceptibility assessment methods still have significant shortcomings in interpreting results. On the one hand, there is a lack of quantitative interpretation methods for the contributions of static factors such as topography and geology, making it difficult to clarify the relative roles of different factors in the formation of landslide backgrounds. On the other hand, for the influence of dynamic factors such as rainfall, most methods only provide a single weight or empirical threshold, making it difficult to reveal the nonlinear response characteristics of rainfall triggering effects as intensity or cumulative processes change, let alone characterize its spatial variability. This lack of interpretive capability makes it difficult for model results to support refined disaster prevention decision-making and dynamic risk assessment.

[0005] Therefore, there is an urgent need for a landslide susceptibility assessment method that can distinguish between the background susceptibility of landslides and the dynamic triggering mechanism of rainfall under physical and statistical constraints, while taking into account both prediction accuracy and interpretability, and quantitatively revealing the contribution of static factors, the nonlinear threshold effect of rainfall, and spatial heterogeneity, so as to make up for the shortcomings of existing technologies in terms of stability, interpretability, and practical application. Summary of the Invention

[0006] The purpose of this invention is to provide a method, system, device and medium for evaluating the susceptibility of rainfall-induced landslides, in order to solve the problems in the prior art, such as model instability caused by the high collinearity of rainfall factors, difficulty in explicitly characterizing the nonlinear threshold characteristics and spatial heterogeneity of rainfall triggering effects, and insufficient physical interpretability of model prediction results.

[0007] To achieve the above objectives, the present invention provides a method for evaluating the susceptibility of rainfall-induced landslides, comprising the following steps: Step S1: Obtain topographic and geomorphological data, geological and engineering geological data, hydrological data, land use and cover data, historical landslide disaster cataloging data, and rainfall data for the target area; Step S2: Construct a multi-timescale rainfall factor system based on rainfall data, and perform factor screening through in-group screening and multi-stage feature engineering methods; Step S3: Construct a mixed landslide susceptibility assessment model: Step S31: Use a random forest model to model the selected topography, geological conditions and land use factors to obtain the probability of landslide background susceptibility for each spatial unit. Step S32: Convert the probability of landslide background susceptibility into log odds form and use it as a bias term; Step S33: Under the constraint of the bias term, construct a Bayesian spatial statistical model, introduce the nonlinear effect term and spatial random effect term of the rainfall factor, and jointly model the probability of landslide occurrence. Step S34: Obtain the posterior probability of landslide occurrence for each spatial unit through Bayesian inference, and perform interpretability analysis in conjunction with the SHAP method; Step S4: Using the trained hybrid model, predict the probability of landslide occurrence in the target area under different time scales or rainfall scenarios, and generate a spatial distribution map of landslide susceptibility.

[0008] Preferably, in step S2, the rainfall factor system includes the anterior cumulative rainfall (ARA) and the rainfall intensity index; the formula for calculating the anterior cumulative rainfall (ARA) is: ; in, Indicates the date before the landslide occurred. Daily rainfall of the day Indicates the length of the cumulative time window; Within-group screening involved dividing rainfall factors into two groups based on their physical meaning: Anterior Accumulated Rainfall (ARA) and Rainfall Intensity. Within each group, the explanatory power of the variables for landslide occurrence was assessed.

[0009] Preferably, in step S2, the multi-stage feature engineering method includes: Step S221: Use Pearson correlation coefficient to perform linear correlation analysis between variables and remove redundant variables whose absolute value of correlation coefficient exceeds the threshold. Step S222: Perform multicollinearity analysis using the variance inflation factor (VIF) and remove variables with VIF values ​​higher than a preset threshold. Step S223: Perform recursive feature elimination (RFE) based on the random forest model, gradually eliminating variables with low contribution to obtain the optimal feature subset.

[0010] Preferably, in step S32, the formula for calculating the bias term is: ; in, Indicates the bias term. This represents the background susceptibility probability output by the random forest model.

[0011] Preferably, in step S33, the Bayesian spatial statistical model takes the following form: ; in, Represents the logarithmic probability function. Represents the intercept term of the model. This represents the log-odds term of the background landslide susceptibility generated by the random forest model. The nonlinear effect function representing the rainfall factor, Indicates spatial random effects; A second-order random walk (RW2) model was used for modeling.

[0012] Preferably, in step S34, the Bayesian inference uses the Integrated Nested Laplace Approximation (INLA) method to calculate the posterior distribution, thereby obtaining the posterior mean, variance, and confidence interval of the landslide occurrence probability.

[0013] Preferably, in step S4, the spatial distribution map of landslide susceptibility is classified according to the predicted probability, including extremely low, low, medium, high and extremely high susceptibility levels, and compared and verified with the actual landslide disaster locations.

[0014] This invention also provides a rainfall-induced landslide susceptibility assessment system, comprising: The data collection and preprocessing module is used to acquire topographic and geomorphological data, geological and engineering geological data, hydrological data, land use and cover data, historical landslide disaster catalog data, and rainfall data of the target area, and to perform preprocessing. The rainfall factor construction and screening module is used to construct a multi-timescale rainfall factor system based on rainfall data, and to screen factors through in-group screening and multi-stage feature engineering methods to reduce variable collinearity. The hybrid model building module is used to model the selected topography, geological conditions and land use factors using the random forest model to obtain the probability of landslide background susceptibility for each spatial unit. The probability of landslide background susceptibility is converted into log odds as a bias term. Under the constraint of the bias term, a Bayesian spatial statistical model is built, introducing the nonlinear effect term of rainfall factor and the spatial random effect term. The posterior probability of landslide occurrence for each spatial unit is obtained through Bayesian inference, and interpretability analysis is performed in combination with the SHAP method. The landslide susceptibility simulation and mapping module is used to predict the probability of landslides in a target area under different time scales or rainfall scenarios using a trained hybrid model, and generate a spatial distribution map of landslide susceptibility.

[0015] The present invention also provides a computer device, including: a memory and a processor; the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described method for evaluating the susceptibility of rainfall-induced landslides.

[0016] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method for evaluating the susceptibility of rainfall-induced landslides.

[0017] Therefore, the present invention employs the above-mentioned method, system, equipment, and medium for assessing the susceptibility of rainfall-induced landslides, and the beneficial technical effects are as follows: (1) To address the problem of high correlation among rainfall indices at multiple time scales in existing methods, which can easily lead to multicollinearity when directly included in the model, this invention employs a systematic factor screening process that combines in-group screening with multi-stage feature engineering (including correlation analysis, VIF test, and recursive feature elimination). This process effectively reduces variable redundancy from both the physical meaning and statistical contribution perspectives. This technique alleviates the interference of collinearity on parameter estimation from the source, enabling the model to exhibit stronger numerical stability and generalization performance when dealing with different rainfall scenarios. It overcomes the shortcomings of traditional single-model models that are prone to overfitting or instability due to variable coupling.

[0018] (2) To address the shortcomings of existing technologies in distinguishing between background susceptibility and dynamic triggering mechanisms, and in effectively expressing spatial dependencies, this invention proposes a two-stage modeling framework decoupling background susceptibility and rainfall triggering. First, random forests are used to capture the complex nonlinear background effects of static factors such as topography and geology, and these are converted into log-probability bias terms. Then, in a Bayesian spatial statistical model, nonlinear effect terms of rainfall factors (using an RW2 model to fit the threshold response) and spatial random effect terms are introduced to explicitly and jointly model the triggering effect and spatial autocorrelation structure of rainfall under physical constraints. This technical solution enables the model to not only quantitatively separate the contributions of static background and dynamic rainfall, but also to reveal the nonlinear threshold characteristics of rainfall triggering and their spatial differences, thereby more realistically reflecting the physical mechanisms and spatial patterns of landslide occurrence.

[0019] (3) To address the problem of insufficient mechanistic explanation caused by the "black box" nature of existing machine learning models, this invention incorporates the SHAP interpretability analysis method into a hybrid modeling framework to quantitatively decompose and visualize the contributions of each static environmental factor in the random forest. Simultaneously, the Bayesian model provides the posterior distribution of parameters and predicted probabilities, and can output uncertainty measures (such as confidence intervals). These techniques ensure that the model output is no longer just a single probability value, but a set of interpretable outputs containing information on factor contribution, influence direction, nonlinear response relationships, and spatial uncertainty, greatly improving the transparency and reliability of the model results. Attached Figure Description

[0020] Figure 1 This is a flowchart of a rainfall-induced landslide susceptibility assessment method according to the present invention; Figure 2 This is a topographic factor map of a rainfall-induced landslide susceptibility assessment method according to the present invention, wherein, Figure 2 (a) in the figure represents the elevation. Figure 2 (b) in the figure represents the slope. Figure 2 (c) in the text represents the slope aspect. Figure 2 In this context, (d) represents the profile curvature. Figure 2 In this context, (e) represents the curvature of the plane. Figure 2 (f) in the figure represents the topographic humidity index; Figure 3 This is a geological and soil factor map of a rainfall-induced landslide susceptibility assessment method according to the present invention, wherein, Figure 3 (a) in the text represents the lithology. Figure 3 (b) in the text represents the soil type. Figure 3 (c) in the figure represents the clay content. Figure 3 (d) represents the sand content; Figure 4 This is a map showing human engineering activities and land cover factors in a rainfall-induced landslide susceptibility assessment method according to the present invention. Figure 4 In the figure, (a) represents the normalized vegetation index. Figure 4 (b) in the text represents land cover. Figure 4 (c) in the figure represents the distance from the road. Figure 4 In this context, (d) represents the distance from the river; Figure 5 A graph showing the results of rainfall factor grouping and screening in the rainfall-based landslide susceptibility assessment method. Figure 6 Correlation analysis diagram of evaluation factors for rainfall-induced landslide susceptibility assessment method; Figure 7 The figure shows the results of factor multicollinearity analysis for the assessment method of landslide susceptibility caused by rainfall. Figure 8 The result of recursive elimination of factor features in the evaluation method for the susceptibility of rainfall-induced landslides is shown in the figure. Figure 9 A static susceptibility map of a randomized forest for evaluating the susceptibility of rainfall-induced landslides; Figure 10 The graph shows the rainfall factor response curves for the rainfall-induced landslide susceptibility assessment method. Figure 10 (a) in the figure is the response curve of 3-day cumulative rainfall. Figure 10 (b) in the figure is the response curve of the cumulative rainfall over 15 days; Figure 11 A Bayesian spatial statistical random field result for a rainfall-induced landslide susceptibility assessment method; Figure 12 ROC curves of the improved model and the basic model (random forest) for the assessment method of rainfall-induced landslide susceptibility; Figure 13 Summary diagram of SHAP interpretability analysis for rainfall-induced landslide susceptibility assessment method; Figure 14 The figure shows the prediction results of the improved Bayesian mixture model for assessing the susceptibility of rainfall-induced landslides. Figure 15 Probability density diagrams of different models for evaluating the susceptibility of rainfall-induced landslides; Figure 16 This is a comparison chart of the frequency-to-accuracy ratio between the improved model and the basic model of the rainfall-induced landslide susceptibility assessment method. Detailed Implementation

[0021] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0022] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0023] Example 1 like Figure 1 As shown, the flowchart of a rainfall-induced landslide susceptibility assessment method of the present invention specifically includes: Step S1: Data Acquisition and Preprocessing.

[0024] Step S1.1: Obtain basic data for the target study area, including: historical landslide catalog data; rainfall data at multiple landslide sites; topographic factor data (elevation, slope, aspect, plan curvature, profile curvature, topographic humidity index, etc.). Figure 2 As shown in (a)-(f); geological and soil factor data (lithology, soil type, sand content, clay content) are as follows: Figure 3 As shown in (a)-(d); Human engineering activities and land cover factors (distance from road, distance from river, normalized difference vegetation index, land use type) are as follows: Figure 4 As shown in (a)-(d) in the figure.

[0025] Step S1.2: Construct time-series features for the rainfall data and calculate rainfall indicators at different time scales. These indicators include multi-time-window cumulative rainfall (Antecedent Rainfall, AR) (cumulative rainfall on the day of the disaster, 3 days prior, 7 days prior, 15 days prior, 30 days prior, 90 days prior, and 365 days prior). Maximum rainfall intensity (Rainfall Intensity, I) includes the maximum rainfall intensity 3 days prior to the disaster and the maximum rainfall intensity 7 days prior to the disaster.

[0026] Step S1.3: Spatially unify all environmental factors to ensure they have the same spatial resolution and spatial reference system, and standardize them to eliminate the impact of dimensional differences on model training.

[0027] Step S2: Construction and screening of rainfall factors.

[0028] Step S2.1: Based on rainfall time series data, construct multi-scale rainfall factors. The preferred rainfall factors include: the previous cumulative rainfall in different time windows; the average rainfall intensity and the maximum rainfall intensity at different time scales; and a combined rainfall index that characterizes the changes in the rainfall process.

[0029] This invention constructs a rainfall factor system from two complementary dimensions: antecedent rainfall accumulation (ARA) and rainfall intensity. Antecedent rainfall accumulation... Used to characterize the water memory effect of a slope in the period leading up to a landslide, it is defined as the period before the landslide occurs. Total rainfall within the day : ; in, Indicates the date before the landslide occurred. Daily rainfall of the day This indicates the length of the cumulative time window.

[0030] This invention comprehensively considers short-term, medium-term and long-term hydrological response processes, and constructs ARA indicators (cumulative rainfall on the day of disaster, 3 days before, 7 days before, 15 days before, 30 days before, 90 days before, and 365 days before) at multiple time scales to reflect the evolution characteristics of slope stability under different rainfall memory lengths.

[0031] To characterize the instantaneous triggering effect of rainfall events, this invention further introduces various rainfall intensity indicators, including the maximum daily rainfall and average rainfall intensity over 3 and 7 days. These indicators mainly reflect the rapid infiltration and sudden increase in pore water pressure caused by short-duration heavy rainfall, complementing the previous accumulated rainfall.

[0032] Since multi-timescale ARA and various intensity indices are highly correlated in their physical definitions, directly incorporating them all into the model can lead to significant multicollinearity, thus affecting the stability of parameter estimation and the interpretability of the model. Therefore, this invention proposes and employs an in-group screening strategy to preprocess rainfall variables. Specifically, rainfall factors are first divided into three groups according to their physical meaning: (1) long-term rainfall group (cumulative rainfall in the 15, 30, 90, and 365 days before the disaster); (2) intensity group (maximum rainfall intensity in the 3 days before and 7 days before); (3) short-term rainfall group (cumulative rainfall on the day of the disaster, the 3 days before, and the 7 days before). Subsequently, the explanatory power of individual variables for landslide occurrence is evaluated within each group (based on univariate model performance or importance ranking), and the screening results are as follows: Figure 5 As shown, the three groups with the highest correlation coefficients are the cumulative rainfall in the previous 15 days, the maximum rainfall intensity in the previous three days, and the cumulative rainfall in the previous three days.

[0033] Step S2.2: Use correlation analysis to calculate the degree of correlation between each environmental factor and the occurrence of landslides, and preliminarily eliminate factors with extremely weak correlation or high redundancy.

[0034] The Pearson correlation coefficient is used to assess the linear relationship between two variables, with values ​​ranging from -1 to 1. A coefficient close to 1 or -1 indicates a strong linear correlation, while a value close to 0 indicates a weak or no correlation. In the context of landslide susceptibility assessment, Pearson correlation analysis is used to quantify the relationship between causal factors. Identifying and eliminating highly correlated variables reduces data redundancy, thereby improving the stability and accuracy of the predictive model. The calculation process is as follows: ; in, Represents the correlation coefficient; Represents a single value of a variable ; Represents a single value of a variable ; This represents the mean of the variable x; Representing variables Mean; This represents the sample number; n represents the total sample size. The correlation analysis results are as follows: Figure 6 As shown, the correlation between the cumulative rainfall in the first three days and the maximum rainfall intensity in the first three days was as high as 0.97, and the correlation between sand content and clay content was 0.91. Based on the magnitude of the correlation between the factors and landslides, the maximum rainfall intensity in the first three days and the clay content were removed respectively. The correlations of the remaining factors were all below 0.8.

[0035] Step S2.3: VIF is used to assess the degree of multicollinearity among variables. Multicollinearity can lead to unstable regression coefficients, thus affecting the predictive accuracy of the model. A VIF value exceeding 10 usually indicates a serious multicollinearity problem. Removing variables with high VIF values ​​can reduce the impact of multicollinearity on the model, thereby improving the model's interpretability and predictive performance. Therefore, this invention uses the variance inflation factor (VIF) analysis method to detect and eliminate multicollinearity among environmental factors. The calculation process of the VIF value is as follows: ; in, Indicates the first The coefficient of determination of the regression model established with one variable as the dependent variable and the remaining variables as independent variables. The collinearity analysis results in this embodiment are as follows: Figure 7 As shown, the expansion coefficients of the selected factors are all less than 5, indicating a weak collinearity among the factors.

[0036] Step S2.4: Employ the Recursive Feature Elimination (RFE) method, combining landslide sample data with a random forest model to iteratively filter the remaining factors, obtaining the optimal subset of factors that contribute significantly to landslide occurrence. In each iteration, the model is trained based on the current feature subset and, according to feature importance, removes one or more features with the lowest contribution. Let the original feature set be... for: ; in, Indicates the first feature in the feature set One characteristic variable, This represents the total number of original features.

[0037] In the In the next iteration, the current feature subset is used. Training base learning model And obtain the relative importance score for each feature. : ; in, This represents the contribution of a feature to the current model. Based on the importance ranking, an elimination strategy is defined: ; In each round, the feature with the lowest importance or the lowest set of features is removed. This process is repeated recursively until a termination condition is met, and the final result is as follows: Figure 8 As shown.

[0038] Step S3: Construct a mixed landslide susceptibility evaluation model.

[0039] Step S3.1: Background susceptibility modeling based on random forest.

[0040] In landslide susceptibility assessment, the occurrence of landslides is influenced by a variety of environmental and rainfall factors, with significant nonlinear relationships and higher-order interaction effects among the variables. The random forest model, through dual random sampling of the sample space and feature space, can effectively characterize this complex nonlinear mapping relationship without relying on any prior distribution assumptions.

[0041] Using the selected environmental factors and rainfall factors as input features, and whether a landslide occurs as the output variable, let the original training sample set be... for: ; in, Indicates the first The environmental factor and rainfall factor vector of each spatial unit Indicate whether a landslide has occurred. This indicates the total number of spatial units.

[0042] A random forest landslide susceptibility assessment model is constructed to learn the nonlinear mapping relationship between landslide occurrence and multiple factors, outputting the background landslide susceptibility probability for each spatial unit. For classification problems, each decision tree in the random forest selects the optimal splitting feature and splitting threshold by maximizing the node purity gain. Commonly used node purity metrics include the Gini index or information entropy. Taking the Gini index as an example, node... Impurity is defined as: ; in, Indicates that the node belongs to the first The proportion of class samples; This represents the total number of categories in a classification problem. By selecting the splitting method that maximizes the decrease in the Gini index, the decision tree can progressively approach the optimal discriminant boundary between landslide and non-landslide samples.

[0043] For any spatial unit x, the b-th decision tree outputs the predicted probability that it belongs to the landslide category. The final output of the random forest model is the average of the predictions from all decision trees: ; in, This represents the total number of decision trees participating in the ensemble in the random forest.

[0044] This probability value is a comprehensive probability estimate of the occurrence of a landslide in this spatial unit under given environmental conditions. Combined with static factors, the susceptibility results obtained from random forests are as follows: Figure 9 As shown.

[0045] Step S3.2: Construction of Bayesian spatial statistical model.

[0046] In regional-scale landslide susceptibility assessments, adjacent spatial units often share similar topographic, geological, and hydrological conditions, leading to significant spatial clustering of landslides. Traditional machine learning models typically assume independence between samples, making it difficult to explicitly characterize this spatial dependency, potentially introducing spatial autocorrelation bias and unexplained spatial heterogeneity. To overcome these issues, this invention introduces a Bayesian spatial statistical model based on the output of a random forest model. By explicitly setting a spatial random effects term in the regression structure, it models the spatial dependency, thereby spatially constraining and correcting the prediction results of the random forest model.

[0047] Suppose the study area is divided into several spatial units, for the first... Whether a landslide occurs in a spatial unit can be considered as a binary random variable: ; in, Indicates the first The probability of a landslide occurring in a spatial unit. Using the logit function as the connection function, a generalized additive model is constructed, whose linear predictor can be expressed as: ; in, This is the intercept term of the model; The background landslide susceptibility log odds term generated for the random forest model; A nonlinear effect function representing the rainfall factor; It represents a spatial random effect, used to characterize the spatial autocorrelation structure.

[0048] Because the triggering effect of rainfall on landslides typically exhibits nonlinear, threshold-based, and significant cumulative effects: during periods of low rainfall, the probability of landslides changes slowly; however, once rainfall reaches a certain threshold, pore water pressure rises rapidly, slope stability decreases significantly, and the probability of landslides increases sharply. Therefore, this invention introduces a second-order random walk (RW2) model to model the nonlinear effects of rainfall factors. In the RW2 model, it is assumed that the second difference between three adjacent points follows a zero-mean Gaussian distribution: ; in, This represents the smoothing parameter, which controls the smoothness of the function curve. In this embodiment, the random walk results of the cumulative rainfall over the previous 3 days and the cumulative rainfall over the previous 15 days are as follows: Figure 10 As shown.

[0049] Step S3.3: Uncertainty quantification.

[0050] After completing the Bayesian spatial statistical model construction in step S3.2, this invention further employs Bayesian inference to estimate the unknown parameters in the model, thereby obtaining the posterior distribution of the landslide susceptibility prediction results and achieving a quantitative characterization of the prediction uncertainty. Specifically, this includes the following steps: (1) In the Bayesian inference framework, the unknown parameters in the model are treated as random variables and a prior distribution is assigned to them. The unknown parameters include at least: the model intercept parameter; the smoothing parameter corresponding to the nonlinear effect function of rainfall; and the precision or variance parameter corresponding to the spatial random effect.

[0051] (2) Based on historical landslide disaster cataloging data, whether a landslide has occurred is taken as a binary observation variable, assuming that it follows a Bernoulli distribution, and the Bayesian spatial logistic regression model constructed in step S3.2 is used as the probability structure to form the likelihood function of the observation data. By combining the prior distribution and the likelihood function of the observation data, a complete Bayesian probability model is established.

[0052] (3) The Integrated Nested Laplace Approximation (INLA) method is used to approximate the posterior distribution of the model parameters. This method uses numerical approximation to quickly calculate the posterior distribution, avoiding the problem of low computational efficiency of the traditional Markov chain Monte Carlo method in high-dimensional space models. Through Bayesian inference, the posterior mean, posterior variance, and confidence interval of the model parameters and the probability of landslide occurrence are obtained. The spatial random field obtained by uncertainty quantification is as follows: Figure 11As shown, red indicates landslide clustering and a strong spatial effect, indicating an increased susceptibility to landslides in this area. Blue indicates sparse landslides and a weaker spatial effect, indicating a decreased susceptibility to landslides in this area. To further verify the improved model's effectiveness in enhancing prediction accuracy, this embodiment plotted ROC curves based on existing landslide inventory data, as shown below. Figure 12 As shown in the figure. The analysis results indicate that the improved model has a significant improvement over the traditional model, demonstrating its superior overall discriminative ability.

[0053] Step S3.4, interpretability analysis.

[0054] After completing the landslide susceptibility modeling and inference in steps S3.1 to S3.3, this invention further introduces an interpretability analysis method to quantitatively analyze the contribution of each input feature in the random forest model to the landslide susceptibility prediction results, thereby improving the interpretability and transparency of the model results.

[0055] The SHAP (Shapley Additive Explanations) method originates from the Shapley value theory in cooperative game theory. Its core idea is to treat the model's prediction result as a "payoff" resulting from the combined effect of multiple features, and quantify the marginal contribution of each feature to the prediction result by reasonably allocating this payoff. In its implementation, the SHAP method measures the contribution of a feature to the model output by comparing the degree of influence of a feature on the magnitude of change in the prediction result when it participates in model prediction versus when it does not.

[0056] For the random forest landslide susceptibility assessment model constructed in step S3.1, this invention employs the SHAP calculation method suitable for tree model structures to decompose the feature contribution of the prediction results for each spatial unit. Specifically, for any spatial unit, its landslide susceptibility prediction value can be expressed as the sum of the SHAP values ​​of each input feature and the model baseline prediction value, where: The baseline predictions of the model represent the average probability of a landslide occurring without introducing any feature information; The SHAP value corresponding to each feature represents the positive or negative impact of that feature on the probability of landslide occurrence under the current spatial unit conditions.

[0057] This method decomposes the originally difficult-to-interpret random forest prediction results into multiple quantifiable and comparable feature contributions.

[0058] Therefore, based on the positive and negative directions of SHAP values ​​and their distribution characteristics as a function of feature values, this invention can further identify the direction of influence and nonlinear impact characteristics of different environmental factors and rainfall factors on landslide susceptibility. This embodiment performs SHAP interpretation analysis on the model results, and the results are as follows: Figure 13 As shown.

[0059] S4. Landslide susceptibility simulation and mapping.

[0060] After completing the construction, parameter estimation, and interpretability analysis of the mixed landslide susceptibility evaluation model in step S3, this invention further conducts landslide susceptibility simulation and mapping to verify the predictive ability of the method under real rainfall events. Taking a typical area in the Marche region of Italy as the research object, the landslide disaster induced by heavy rainfall on September 15, 2022, was selected as a verification case. Specifically, based on the daily rainfall observation data before and after this time point, the rainfall factor input and evaluation factors at the corresponding time scale were constructed according to the method in step S2.

[0061] The constructed feature input data is then input into the hybrid landslide susceptibility assessment model trained in step S3, and the following calculation process is executed sequentially: The random forest model is used to predict the background landslide susceptibility of each spatial unit, and the initial landslide occurrence probability is obtained without considering the spatial dependence effect. The output of the random forest is introduced as a bias term into the Bayesian spatial statistical model. By combining spatial random effects and rainfall nonlinear functions, the probability of landslide occurrence is spatially constrained and corrected. Based on Bayesian inference results, the posterior distribution of the landslide occurrence probability for each spatial unit under a given rainfall scenario is calculated.

[0062] The final output is the predicted probability of landslide susceptibility for each spatial unit under the rainfall conditions on September 15, 2022. Based on the calculated landslide probability results, a spatial map of landslide susceptibility is created for the study area. In practice, landslide susceptibility can be divided into extremely low susceptibility, low susceptibility, moderate susceptibility, high susceptibility, and extremely high susceptibility levels according to preset probability thresholds or classification rules, and corresponding landslide susceptibility distribution maps can be generated in the form of raster or spatial units, such as... Figure 14 As shown.

[0063] To verify the effectiveness of the method of this invention, the landslide susceptibility prediction results were compared and analyzed with the actual distribution of landslide disaster sites in the Marche area on September 15, 2022. Furthermore, landslide susceptibility density distribution maps of the improved model and the traditional random forest model were plotted, as shown below. Figure 15 As shown in the figure. Simultaneously, the frequency ratio (FR) accuracy of the two models at different susceptibility levels was statistically analyzed and compared, and the results are as follows. Figure 16 The analysis results show that, at medium, high, and very high susceptibility levels, the improved model's frequency-to-accuracy ratio is significantly better than that of the traditional random forest model, indicating that this method has higher reliability and discriminative ability in landslide susceptibility identification.

[0064] Example 2 A rainfall-induced landslide susceptibility assessment system includes: The data collection and preprocessing module is used to acquire topographic and geomorphological data, geological and engineering geological data, hydrological data, land use and cover data, historical landslide disaster catalog data, and rainfall data of the target area, and to perform preprocessing. The rainfall factor construction and screening module is used to construct a multi-timescale rainfall factor system based on rainfall data, and to screen factors through in-group screening and multi-stage feature engineering methods to reduce variable collinearity. The hybrid model building module is used to model the selected topography, geological conditions and land use factors using the random forest model to obtain the probability of landslide background susceptibility for each spatial unit. The probability of landslide background susceptibility is converted into log odds as a bias term. Under the constraint of the bias term, a Bayesian spatial statistical model is built, introducing the nonlinear effect term of rainfall factor and the spatial random effect term. The posterior probability of landslide occurrence for each spatial unit is obtained through Bayesian inference, and interpretability analysis is performed in combination with the SHAP method. The landslide susceptibility simulation and mapping module is used to predict the probability of landslides in a target area under different time scales or rainfall scenarios using a trained hybrid model, and generate a spatial distribution map of landslide susceptibility.

[0065] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, 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 a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0066] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0067] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0068] It is worth noting that all contents not described in detail in this invention are existing technologies and are well known to those skilled in the art.

[0069] Therefore, this invention adopts the above-mentioned method, system, equipment and medium for evaluating rainfall-induced landslide susceptibility. By constructing a two-stage hybrid modeling framework that decouples background susceptibility and rainfall triggering effect, it combines the nonlinear fitting ability of the random forest model with the physical constraints and spatial modeling ability of the Bayesian spatial statistical model. Furthermore, it uses within-group screening and multi-stage feature engineering methods to reduce collinearity of rainfall factors. This effectively solves the problems of insufficient stability due to redundant dependent variables in existing technologies, difficulty in explicitly characterizing the nonlinear effects and spatial heterogeneity of rainfall triggering, and weak interpretability of prediction results. It significantly improves the prediction accuracy and interpretability of landslide susceptibility evaluation under variable rainfall conditions.

[0070] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for evaluating the susceptibility of rainfall-induced landslides, characterized in that, Includes the following steps: Step S1: Obtain topographic and geomorphological data, geological and engineering geological data, hydrological data, land use and cover data, historical landslide disaster cataloging data, and rainfall data for the target area; Step S2: Construct a multi-timescale rainfall factor system based on rainfall data, and perform factor screening through in-group screening and multi-stage feature engineering methods; Step S3: Construct a mixed landslide susceptibility assessment model: Step S31: Use a random forest model to model the selected topography, geological conditions and land use factors to obtain the probability of landslide background susceptibility for each spatial unit. Step S32: Convert the probability of landslide background susceptibility into log odds form and use it as a bias term; Step S33: Under the constraint of the bias term, construct a Bayesian spatial statistical model, introduce the nonlinear effect term and spatial random effect term of the rainfall factor, and jointly model the probability of landslide occurrence. Step S34: Obtain the posterior probability of landslide occurrence for each spatial unit through Bayesian inference, and perform interpretability analysis in conjunction with the SHAP method; Step S4: Using the trained hybrid model, predict the probability of landslide occurrence in the target area under different time scales or rainfall scenarios, and generate a spatial distribution map of landslide susceptibility.

2. The method for evaluating the susceptibility of rainfall-induced landslides according to claim 1, characterized in that, In step S2, the rainfall factor system includes the antecedent cumulative rainfall (ARA) and the rainfall intensity index; the formula for calculating the antecedent cumulative rainfall (ARA) is: ； in, Indicates the date before the landslide occurred. Daily rainfall of the day Indicates the length of the cumulative time window; Within-group screening involved dividing rainfall factors into two groups based on their physical meaning: Anterior Accumulated Rainfall (ARA) and Rainfall Intensity. Within each group, the explanatory power of the variables for landslide occurrence was assessed.

3. The method for evaluating the susceptibility of rainfall-induced landslides according to claim 1, characterized in that, In step S2, the multi-stage feature engineering method includes: Step S221: Use Pearson correlation coefficient to perform linear correlation analysis between variables and remove redundant variables whose absolute value of correlation coefficient exceeds the threshold. Step S222: Perform multicollinearity analysis using the variance inflation factor (VIF) and remove variables with VIF values ​​higher than a preset threshold. Step S223: Perform recursive feature elimination (RFE) based on the random forest model, gradually eliminating variables with low contribution to obtain the optimal feature subset.

4. The method for evaluating the susceptibility of rainfall-induced landslides according to claim 1, characterized in that, In step S32, the formula for calculating the bias term is: ； in, Indicates the bias term. This represents the background susceptibility probability output by the random forest model.

5. The method for evaluating the susceptibility of rainfall-induced landslides according to claim 1, characterized in that, In step S33, the Bayesian spatial statistical model takes the following form: ； in, Represents the logarithmic probability function. Represents the intercept term of the model. This represents the log-odds term of the background landslide susceptibility generated by the random forest model. The nonlinear effect function representing the rainfall factor, This indicates a spatial random effect.

6. The method for evaluating the susceptibility of rainfall-induced landslides according to claim 1, characterized in that, In step S34, Bayesian inference uses the Integrated Nested Laplace Approximation (INLA) method to calculate the posterior distribution, obtaining the posterior mean, variance, and confidence interval of the landslide occurrence probability.

7. The method for evaluating the susceptibility of rainfall-induced landslides according to claim 1, characterized in that, In step S4, the spatial distribution map of landslide susceptibility is classified according to the predicted probability, including extremely low, low, medium, high and extremely high susceptibility levels, and compared and verified with the actual landslide disaster locations.

8. A rainfall-induced landslide susceptibility assessment system, characterized in that, include: The data collection and preprocessing module is used to acquire topographic and geomorphological data, geological and engineering geological data, hydrological data, land use and cover data, historical landslide disaster catalog data, and rainfall data of the target area, and to perform preprocessing. The rainfall factor construction and screening module is used to construct a multi-timescale rainfall factor system based on rainfall data, and to screen factors through in-group screening and multi-stage feature engineering methods to reduce variable collinearity. The hybrid model building module is used to model the selected topography, geological conditions and land use factors using the random forest model to obtain the probability of landslide background susceptibility for each spatial unit. The probability of landslide background susceptibility is converted into log odds as a bias term. Under the constraint of the bias term, a Bayesian spatial statistical model is built, introducing the nonlinear effect term of rainfall factor and the spatial random effect term. The posterior probability of landslide occurrence for each spatial unit is obtained through Bayesian inference, and interpretability analysis is performed in combination with the SHAP method. The landslide susceptibility simulation and mapping module is used to predict the probability of landslides in a target area under different time scales or rainfall scenarios using a trained hybrid model, and generate a spatial distribution map of landslide susceptibility.

9. A computer device, comprising: Memory and processor; The memory stores a computer program, characterized in that when the processor executes the computer program, it implements the steps of the rainfall-induced landslide susceptibility assessment method according to any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the rainfall-induced landslide susceptibility assessment method according to any one of claims 1-7.

Images