Extreme rainfall group-occurring landslide susceptibility evaluation method fusing diffusion generation and interpretable learning

Abstract

The invention relates to the field of vulnerability evaluation of extreme rainfall induced mass landslide, in particular to an extreme rainfall mass landslide vulnerability evaluation method fusing diffusion generation and interpretable learning, which comprises the following steps: firstly, constructing a multi-dimensional factor data set, and realizing rasterization preprocessing; secondly, establishing a feature response analysis module, identifying main control factors of landslide occurrence and eliminating redundant variables; then, a denoising diffusion probability model is introduced, and an enhanced balance sample library is constructed; then, carrying out landslide probability prediction based on a multi-algorithm fusion integrated discrimination model, and outputting a grid-level landslide occurrence probability; and finally, minimizing intra-class variance and maximizing inter-class variance by using a natural breakpoint method, determining an optimal grading threshold set, and realizing landslide susceptibility grade division and spatial mapping output. Under the extreme rainfall condition, the method can effectively characterize the susceptibility zoning influence of the mass-occurring landslide flow slip catastrophe process, and achieves high-precision landslide susceptibility evaluation.

Description

Technical Field

[0001] This invention relates to the field of susceptibility assessment for landslides induced by extreme rainfall, specifically a hierarchical susceptibility assessment method for landslides induced by extreme rainfall that integrates diffusion generation and interpretable learning. Background Technology

[0002] In recent years, concentrated extreme rainfall events have significantly increased the triggering frequency of shallow landslides in Southeast China, especially in areas with fractured geological structures, undulating topography, and high-intensity, short-duration rainfall. Landslides have become a major disaster factor affecting traffic safety, the stability of residential settlements, and the operation of major engineering projects. Driven by heavy rainfall, such shallow landslides undergo rapid shear softening and an integrated evolutionary process of sliding and flowing along the slope, exhibiting typical "flow-slide" disaster characteristics, further exacerbating the scope of damage and the disaster effect. Existing landslide susceptibility assessment methods are mostly based on statistical or empirical models constructed from multiple sources of factors such as topography, geology, ecology, rainfall, and engineering disturbances. However, the following technical bottlenecks still exist: First, there are significant nonlinear coupling relationships among multiple factors, making it difficult for traditional methods to fully characterize the triggering mechanism; second, in extreme rainfall scenarios, landslide samples are scarce and unevenly distributed, resulting in insufficient model generalization ability and a significant decrease in prediction accuracy.

[0003] In recent years, interpretable machine learning methods have been introduced into landslide susceptibility research to improve predictive capabilities and reveal characteristic effects. However, existing methods mostly rely on fitting existing samples, failing to address the problems of "insufficient key samples" and "fuzzy class boundaries." Meanwhile, existing generative models focus on general data augmentation, lacking constraint mechanisms specific to landslide scenarios. They cannot generate highly reliable supplementary samples for extreme landslides or high-uncertainty areas, and still harbor the potential for error accumulation and overestimation or underestimation of risk.

[0004] Based on the above shortcomings, there is an urgent need for a technology that can simultaneously achieve multi-factor coupled cognition, key sample enhancement, and high-precision interpretable assessment of the susceptibility to cluster landslides, in order to support regional early warning, risk classification, and disaster prevention decision-making. Summary of the Invention

[0005] This invention addresses the problems in landslide disaster prediction under extreme rainfall backgrounds, such as sample scarcity, difficulty in identifying boundary samples, difficulty in interpreting multi-factor coupling relationships, and insufficient usability of spatial classification results. It proposes a hierarchical intelligent assessment method for landslide susceptibility that integrates diffusion-generative probabilistic modeling and interpretable learning mechanisms, thereby achieving high-precision prediction and hierarchical expression of the probability of landslide occurrence in the target area.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A method for assessing the susceptibility of landslides triggered by extreme rainfall clusters, which integrates diffusion generation and interpretable learning, includes the following steps: S1: Multi-source factor data construction and rasterization preprocessing; For multi-source data including topography, geology, hydrology, vegetation, engineering disturbance and rainfall structure indicators, spatial projection unification, resolution resampling and numerical standardization are performed to form a multi-factor raster input matrix for model training.

[0007] S2: Construct a feature response parsing module to identify the main control factors and eliminate redundant variables; S3: Construct a generative supplementation module to expand the distribution range of training samples and enhance the ability to identify boundary samples; S4: Establish an integrated discrimination module to output rasterized probability results and level expressions; S5: Outputs layered landslide susceptibility assessment results, expressing the assessment results as risk level maps for use in early warning linkage, risk management and engineering treatment plan formulation.

[0008] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention introduces a diffusion-generative supplementation mechanism, which effectively compensates for the lack of training samples in extreme rainfall scenarios by performing constrained generation on boundary samples and scarce category samples. This significantly reduces the bias risk of the classifier in a highly imbalanced sample space, thereby improving the stability and reliability of landslide probability prediction under complex working conditions.

[0009] 2. This invention uses an interpretable learning mechanism to construct a factor contribution response system, which can realize the causal relationship analysis between input factors and output results without destroying the model's discrimination performance, making the landslide susceptibility results traceable in engineering.

[0010] 3. This invention constructs a hierarchical landslide susceptibility level output system, which transforms continuous probability expression into executable risk classification results. It can be directly connected to spatial risk zoning, dynamic early warning scheduling and prevention and control resource allocation, realizing seamless transformation from model output to business execution, and improving the applicability of assessment results in emergency management and disaster prevention of transportation infrastructure.

[0011] 4. This invention overcomes the shortcomings of traditional empirical weighting methods, such as strong subjectivity, uninterpretable pure data-driven models, and difficulty in implementation due to the integrated process design of generation enhancement, interpretation modeling, and hierarchical expression. It has engineering transformation and promotion value. Attached Figure Description

[0012] Figure 1 This is a flowchart of the overall process of the present invention; Figure 2 This is a schematic diagram illustrating the construction of multi-source landslide disaster-causing factor data in an embodiment of the present invention; Figure 3 This is a schematic diagram of the generative replenishment module in an embodiment of the present invention; Figure 4 This is a schematic diagram of the integrated discrimination and classification module according to an embodiment of the present invention; Figure 5 This is a schematic diagram of a two-layer integrated strategy composite discrimination framework according to an embodiment of the present invention; Figure 6 This is a schematic diagram of the landslide susceptibility classification results according to an embodiment of the present invention. Detailed Implementation

[0013] The technical solution provided in this application will be further described below with reference to specific embodiments and accompanying drawings. The advantages and features of this application will become clearer from the following description.

[0014] like Figure 1 As shown, this invention proposes a method for assessing the susceptibility of landslides triggered by extreme rainfall clusters by fusing diffusion generation and interpretable learning, comprising five steps: S1: Multi-source factor data construction and rasterization preprocessing; For multi-source data such as topography, geology, hydrology, vegetation, engineering disturbance and rainfall structure indicators, spatial projection unification, resolution resampling and numerical standardization are performed to form a multi-factor raster input matrix for model training.

[0015] S2: Construct a feature response parsing module to identify the main control factors and eliminate redundant variables; S3: Construct a generative supplementation module to expand the distribution range of training samples and enhance the ability to identify boundary samples; S4: Establish an integrated discrimination module to output rasterized probability results and level expressions; S5: Outputs layered landslide susceptibility assessment results, expressing the assessment results as risk level maps for use in early warning linkage, risk management and engineering treatment plan formulation.

[0016] By combining multi-source spatial data fusion, generative model supplementation, and interpretable machine learning methods, the entire process of landslide susceptibility calculation, from probability prediction to grade stratification, is realized.

[0017] Example As an example, the input data sources of this invention are from publicly available remote sensing, surveying and mapping basic data, and measured geological data from natural resources departments: the DEM data source is the PALSAR DEM (12.5m) product from the ALOS satellite (used to calculate slope, aspect, curvature, etc.); landslide information, lithology maps, and fault distribution are taken from regional geological maps and geological survey databases; NDVI is taken from Sentinel satellite remote sensing vegetation products; road vectors are taken from publicly available map vector data; rainfall data is taken from measured rainfall at local rain gauge stations; and other hydrological derivative indicators (such as TWI) are calculated from the aforementioned data.

[0018] S1: Construction and Preprocessing of Multi-Source Factor Data like Figure 2 As shown, a multi-source factor data construction and preprocessing module is constructed to generate the multi-dimensional input factor matrix required for landslide susceptibility analysis. The input data includes topographic factors, geological factors, hydrological factors, ecological factors, geotechnical factors, and anthropogenic disturbance factors. Topographic factors are calculated using a digital elevation model (DEM) and include slope, aspect, curvature, and topographic relief. Geological factors include lithology and fault distance. Hydrological factors include the topographic moisture index (TWI). Ecological factors use the normalized difference vegetation index (NDVI). Geotechnical factors are softening coefficients reflecting the strength and permeability of soil and rock masses, obtained from exploration and testing data, assigned values ​​according to lithology, and then uniformly rasterized. Anthropogenic disturbance factors include distance from roads.

[0019] To ensure spatial data consistency, this invention employs a unified coordinate system and a unified resolution (12.5 meters) for spatial projection transformation, and achieves data registration through resampling, interpolation, and filtering operations. All input data undergoes standardization to form a feature matrix of uniform dimension. ,in The number of sample units. For the number of input factors, the first The feature vector corresponding to each sample unit is denoted as . The sample unit is a raster pixel in a unified coordinate system with a resolution of 12.5m. The feature matrix serves as the input basis for subsequent feature analysis and model training.

[0020] This module ensures that data from different sources remain consistent in spatial scale, coordinate system, and numerical domain, thereby avoiding modeling bias caused by data heterogeneity and providing accurate input for subsequent generative sampling and probabilistic discrimination.

[0021] S2: Feature Response Parsing Module The feature response analysis module is used to identify the main controlling factors of landslide occurrence and eliminate redundant variables, thereby achieving dimensionality reduction and optimization of input factors.

[0022] The module's processing procedure includes the following four sub-steps: S2.1: Calculate the discriminative contribution of each factor to the target variable (landslide / non-landslide) based on the information gain ratio (IGR) to obtain the factor importance ranking sequence; The definition and calculation method of Information Gain Ratio (IGR) are existing technologies. Information gain is used to suppress bias towards factors with a large number of values, making the contributions of different factors comparable. For continuous factors, binning discretization is first used to obtain a finite range of values, and then the IGR values ​​are calculated and ranked according to the standard definition.

[0023] S2.2: Employ interpretable learning mechanisms to quantify the causal response of the model output and use feature contribution explanatory vectors. To measure the impact of changes in input factors on prediction results The marginal impact; in , Indicates the first The contribution strength of each factor to the prediction result of this sample, the sign indicates the direction of action, and the absolute value indicates the degree of influence.

[0024] S2.3: Set contribution threshold According to information gain ratio and The combined result eliminates low-sensitivity factors, retaining only the main control variables; the combined result refers to simultaneously considering the factor information gain rate ranking and the global contribution ranking: when a factor is below the same contribution threshold in both rankings. If an input is deemed low-sensitivity or redundant, it will not be included in subsequent modeling inputs; otherwise, it will be retained.

[0025] Normalization and scaling are performed on the retained factors to form an optimized input set.

[0026] S2.4: The obtained optimized input set serves as the final input factor set, providing a reduced and efficient set of input variables for the S3 generative supplementation module. This serves as the basis for directional identification of boundary samples and scarce category samples, and as the generation condition variables. Simultaneously, it receives the sample consistency feedback signal returned by S3, and adjusts the contribution threshold accordingly. The set of retained factors is iteratively corrected until the feedback indicators meet the preset consistency requirements.

[0027] After processing through the steps described above in this module, the input dimensions can be significantly reduced, the model computation efficiency can be improved, and the interpretability and traceability of subsequent models can be enhanced.

[0028] S3: Generative replenishment module: A generative supplementation module is constructed to expand the distribution range of training samples and enhance the ability to identify boundary samples based on the sample data optimized by S2 features.

[0029] The generative supplementation module includes a denoising diffusion probability model, a positive sample generator, and a negative sample generator. The denoising diffusion probability model is used to perform the positive perturbation and reverse denoising generation processes. The positive sample generator generates samples in a targeted manner for landslide-type samples, while the negative sample generator generates samples in a targeted manner for non-landslide-type samples. Both take the factor set output by S2 as input. The module output is a set of generated samples, which is used for training the subsequent integrated discrimination module.

[0030] S3.1: First, extract boundary samples and scarce category samples based on the sample distribution characteristics to form a sample set to be enhanced. The methods for extracting boundary samples and scarce category samples are existing technologies.

[0031] S3.2: Introduce a denoising diffusion probability model to perform forward perturbation and reverse generation processes. The forward process involves adding Gaussian noise to the original samples to form a noise sequence. The reverse process uses a neural network model to gradually denoise and generate samples. This makes it approximate the true distribution.

[0032] S3.3: Set geological consistency constraints The generated samples undergo anomaly removal and attribute consistency verification. Geological consistency constraints include rule verification of the upper and lower limits of the value range of each factor in the generated samples, the legality of geological category codes, and the physical relationships between key factors, in order to remove unreasonable samples and ensure the physical rationality and geological interpretability of the generated samples.

[0033] If a systematic deviation occurs during the consistency check of the generated samples, a consistency evaluation index is output as a sample consistency feedback signal and sent back to S2 to trigger adaptive adjustment of the retained factor set and threshold τ. Then, supplementary sample generation is re-executed on the updated factor set. The adaptive adjustment refers to prioritizing the adjustment of the threshold when the consistency evaluation index fails to meet the preset threshold. To change the number of retention factors and synchronously update the set of retention factors, including: increasing... To eliminate more redundant factors, or reduce To introduce more constraint factors.

[0034] S3.4: Proportionately generate samples Mixed with the original samples to form an enhanced sample library The training objective is to minimize the Wasserstein distance between generated samples and real samples. in, Represents the distribution of the real samples. This represents the distribution of the generated samples; This is the set of all feasible joint distributions of the two. These are sample pairs obtained from the real sample distribution and the generated sample distribution, respectively; It is a distance metric between samples, used to characterize the differences between two distributions.

[0035] The generative supplementation module is trained independently with positive and negative sample generators, and performs supplementation at multiple candidate time steps. The next step is to generate samples, resulting in the corresponding generated sample set. For each... Calculate the generated sample distribution Compared with the true sample distribution Wasserstein distance between Select to make The optimal generation time is the time step that minimizes the time and whose change tends to stabilize within adjacent time steps. The optimal generation result is achieved when the Wasserstein distance simultaneously reaches a stable minimum.

[0036] Generative supplementation significantly improves the spatial coverage of the sample, providing balanced and complete data support for subsequent integrated discriminative models.

[0037] like Figure 3 As shown, the results of the distribution difference test between the real samples and the generated samples during the supplementation process are presented. The horizontal axis (0–9) represents the time step number of the diffusion model, and the bar proportions represent the comparison between the distances of the real samples and the generated samples, used to evaluate the degree to which the generated samples approximate the real distribution at different time steps. Optimal generation time. This refers to the time step number (0-9) that minimizes the difference between the distribution of the real samples and the generated samples. Figure 3 That is, time step 7.

[0038] S4: Integrated discrimination module: like Figure 4 and Figure 5 As shown, the integrated discrimination module of this invention is used to achieve quantitative prediction of landslide occurrence probability. The core of the model consists of a multi-algorithm fusion system, including two processing branches: a random forest sub-model and an XGBoost sub-model, and combines a Bagging / Boosting two-layer fusion strategy to form a composite discrimination framework.

[0039] Specifically as follows: S4.1: Read the feature matrix of the training samples output from S3. and label vectors and construct Cross-validation partitioning. Based on the landslide and non-landslide labels of the aforementioned sample units, the following values ​​are obtained: if the sample unit is a landslide sample, the value is 1; otherwise, the value is 0.

[0040] Output: Folded data subset ， For cross-validation folds; the first fold. compromise Indicates a training subset. This indicates a subset to be verified.

[0041] S4.2: Random Forest Sub-model Training (Bagging Branch): Train the random forest sub-model based on the Bagging mechanism to obtain the samples. probability output .

[0042] Output: Probability output of Random Forest .

[0043] S4.3: XGBoost Sub-model Training (Boosting Branch): Train the XGBoost sub-model with cross-entropy as the objective function: in, For the sample size, For the first Each sample label For XGBoost, the first The output is the probability that each sample is predicted to be a landslide. This is the set of internal parameters for the XGBoost model.

[0044] Introducing a complexity penalty term to control model size: in, The number of leaf nodes in the tree. The leaf node weight vector, for Regularization coefficient, This is the penalty coefficient for the number of leaf nodes, used to control model complexity.

[0045] The final optimization goal is: Gradient boosting iterative updates (calculation of optimal weights for leaf nodes) are employed: in, Indicates the first During the first round of gradient boosting, the... The output of each tree for a sample, i.e. the optimal leaf weight when the sample falls into the corresponding leaf node, is used as the incremental update of the model logit for that round. and The loss function is respectively paired with the first... The first and second gradients of the predicted values ​​for each sample. for The regularization coefficient is summed over the set of samples contained in the current leaf node.

[0046] Output: XGBoost's logit output and probability output .

[0047] S4.4: Perform a two-level fusion of the outputs from the Bagging and Boosting branches to obtain the comprehensive logit: Output the probability of a landslide occurring: Output: Overall probability .

[0048] S4.5: Parameter Optimization and Model Determination. Using the cross-entropy loss output from S4.3. Using the average value under K-fold cross-validation as the objective function, Bayesian optimization is employed in the parameter space. Inner iterative search, To improve the learning rate of XGBoost iterations, leaf node weights Regularization coefficient, The penalty coefficient for the number of leaf nodes. The maximum depth of a single tree. To increase the number of trees generated in each iteration, each set of candidate parameters is generated. That is, calling S4.4 to complete model training and validation and obtain the corresponding loss and evaluation metrics, until the optimal parameters are obtained. Based on this, the final integrated discriminant model is solidified. .

[0049] S4.6: In the final integrated discriminant model The SHAP mechanism is introduced to calculate the feature contribution vector of the sample. Quantify the impact of each factor The marginal impact.

[0050] Output: Single-sample contribution vector And rasterized contribution layers for each factor.

[0051] The module's final output: a continuous landslide probability layer for each raster cell. and feature contribution results .in As an input to the S5 hierarchical susceptibility level output module, it is used for susceptibility level zoning.

[0052] Specifically, feature contribution vector The calculation method is as follows: Based on the Shapley value concept of game theory, the model output is decomposed into the baseline value and the sum of the contributions of each feature. For a single sample... The model output changes are calculated when each feature is added or omitted under different feature combinations. The expected value of all combined contributions is then calculated to obtain the contribution value of each factor. .

[0053] S5: Layered Susceptibility Level Output: like Figure 4 and Figure 6 As shown, the hierarchical susceptibility level output module of the present invention is used to output the final integrated discriminant model in S4. The continuous probability output is transformed into a discrete hierarchical distribution, realizing a spatially hierarchical representation.

[0054] The optimal grading threshold is calculated using the natural breakpoint method, and its core objective function is: in, Indicates the first The first of the levels The probability value of landslide occurrence corresponding to each spatial unit , For the first Class mean The population mean For the first Number of samples per class This represents the number of levels.

[0055] The optimal set of hierarchical thresholds is obtained through a dual-objective optimization that minimizes intra-class variance and maximizes inter-class variance. . The first one obtained by the natural break method A probability grading threshold is used to divide the probability interval into several risk levels.

[0056] This embodiment takes =5, the probability interval is divided into five risk levels: extremely low ( ),Low( ),middle( ),high( ) and extremely high ( ).

[0057] To ensure the continuity of spatial results, this invention further introduces a smoothing and spatial consistency verification mechanism to correct for neighborhood probability differences using a weighted average: in, For the sample The neighborhood set, Spatial weights, For neighborhood probability, As the normalization factor, This represents the smoothed probability.

[0058] The spatial weight It is a function that decays with distance based on spatial distance. ,in For the sample With neighboring units distance, This is the attenuation coefficient.

[0059] Finally, a layered landslide susceptibility level map is generated (e.g.) Figure 6 As shown in the figure, different risk levels are represented by five color gradients, from green (very low) to red (very high). This result can be directly imported into a geographic information system platform to achieve spatial visualization and risk zoning management.

[0060] This embodiment combines generative data augmentation with interpretable ensemble learning to improve sample utilization efficiency in extreme rainfall-induced landslide scenarios. The output five-level landslide susceptibility map can provide quantitative support for regional disaster prevention and mitigation, realizing an integrated technical path from "data fusion - generative modeling - risk classification - results application," and possesses engineering promotion value and scientific guidance significance.

[0061] The above description is merely a description of preferred embodiments of this application and is not intended to limit the scope of this application in any way. Any changes or modifications made by those skilled in the art based on the above-disclosed technical content should be considered as equivalent and valid embodiments and fall within the scope of protection of the technical solution of this application.

Claims

1. A method for assessing the susceptibility to landslides caused by extreme rainfall clusters, integrating diffusion generation and interpretable learning, characterized in that... Includes the following steps: S1: Multi-source factor data construction and rasterization preprocessing; For multi-source data of topography, geology, hydrology, vegetation, engineering disturbance and rainfall structure indicators, spatial projection unification, resolution resampling and numerical standardization are performed to form a multi-factor raster input matrix for model training; S2: Construct a feature response parsing module to identify the main control factors and eliminate redundant variables; S3: Construct a generative supplementation module to expand the distribution range of training samples and enhance the ability to identify boundary samples; S4: Establish an integrated discrimination module to output rasterized probability results and level expressions; S5: Outputs layered landslide susceptibility assessment results, expressing the assessment results as risk level maps for use in early warning linkage, risk management and engineering treatment plan formulation.

2. The evaluation method according to claim 1, characterized in that, Step S1 specifically involves: A multi-source factor data construction and preprocessing module is constructed to generate the multi-dimensional input factor matrix required for landslide susceptibility analysis. The input data includes topographic factors, geological factors, hydrological factors, ecological factors, geotechnical factors, and anthropogenic disturbance factors. Topographic factors are calculated using a digital elevation model and include slope, aspect, curvature, and topographic relief. Geological factors include lithology and fault distance. Hydrological factors include the topographic humidity index. Ecological factors use the normalized vegetation index. Geotechnical factors are softening coefficients reflecting the strength and permeability of soil and rock masses, obtained from exploration and testing data, assigned values ​​according to lithology, and then uniformly rasterized. Anthropogenic disturbance factors include distance from roads. To ensure spatial data consistency, a unified coordinate system and resolution are used for spatial projection transformation, and data registration is achieved through resampling, interpolation, and filtering operations. All input data are standardized to form a feature matrix of uniform dimension. ,in The number of sample units. For the number of input factors, the first The feature vector corresponding to each sample unit is denoted as . The sample unit is a raster pixel in a unified coordinate system and at a resolution of 12.5m.

3. The evaluation method according to claim 2, characterized in that, In step S2, The feature response analysis module is used to identify the main controlling factors of landslide occurrence and eliminate redundant variables, thereby achieving dimensionality reduction and optimization of input factors; The processing procedure includes the following steps: S2.1: Calculate the discriminative contribution of each factor to the target variable based on the information gain ratio (IGR) to obtain the factor importance ranking sequence; By suppressing the preference for factors with a large number of values ​​through information gain, the contributions of different factors are made comparable; for continuous factors, binning discretization is first used to obtain a finite range of values, and then the IGR value is calculated and sorted according to the standard definition. S2.2: Employ interpretable learning mechanisms to quantify the causal response of the model output and use feature contribution explanatory vectors. To measure the impact of changes in input factors on prediction results The marginal impact; in , Indicates the first The contribution strength of each factor to the prediction result of this sample, the sign indicates the direction of action, and the absolute value indicates the degree of influence; S2.3: Set contribution threshold According to information gain ratio and The combined result eliminates low-sensitivity factors, retaining only the main control variables; the combined result refers to simultaneously considering the factor information gain rate ranking and the global contribution ranking: when a factor is below the same contribution threshold in both rankings. If the input is deemed low-sensitivity or redundant, it will not be included in subsequent modeling input; otherwise, it will be retained. Normalization and scaling are performed on the retained factors to form an optimized input set; S2.4: The obtained optimized input set serves as the final input factor set, providing a reduced and efficient set of input variables for the S3 generative supplementation module. This serves as the basis for directional identification of boundary samples and scarce category samples, and as the conditional variable for generation. Simultaneously, it receives the sample consistency feedback signal returned by S3, and adjusts the contribution threshold accordingly. The set of retained factors is iteratively corrected until the feedback indicators meet the preset consistency requirements.

4. The evaluation method according to claim 3, characterized in that, In step S3, The generative supplementation module includes: a denoising diffusion probability model, a positive sample generator, and a negative sample generator. The denoising diffusion probability model is used to perform the positive perturbation and reverse denoising generation process. The positive sample generator generates samples in a targeted manner for landslide-type samples, and the negative sample generator generates samples in a targeted manner for non-landslide-type samples. Both of them take the factor set output by S2 as input. The module output is a set of generated samples for use in the training of the subsequent integrated discrimination module. The processing procedure is as follows: S3.1: First, extract boundary samples and scarce category samples based on the sample distribution characteristics to form a sample set to be enhanced. ; S3.2: Introduce a denoising diffusion probability model to perform forward perturbation and reverse generation processes; the forward process generates a noise sequence by adding Gaussian noise to the original samples. The reverse process uses a neural network model to gradually denoise and generate samples. This makes it approximate the true distribution; S3.3: Set geological consistency constraints The generated samples are subjected to anomaly removal and attribute consistency verification. Geological consistency constraints include rule verification of the upper and lower limits of the value range of each factor in the generated samples, the legality of the geological category code, and the physical relationship between key factors, in order to remove unreasonable samples and ensure the physical rationality and geological interpretability of the generated samples. If a systematic deviation occurs in the consistency check of the generated samples, the consistency evaluation index is output as a sample consistency feedback signal and sent back to S2 to trigger the adjustment of the retention factor set and threshold. The adaptive adjustment is performed, and supplementary sampling is re-executed on the updated factor set; the adaptive adjustment refers to prioritizing the adjustment of the threshold when the consistency evaluation index does not meet the preset threshold. To change the number of retention factors and synchronously update the set of retention factors, including: increasing... To eliminate more redundant factors, or reduce To introduce more constraint factors; S3.4: Proportionately generate samples Mixed with the original samples to form an enhanced sample library The training objective is to minimize the Wasserstein distance between generated samples and real samples. in, Represents the distribution of the real samples. This represents the distribution of the generated samples; This is the set of all feasible joint distributions of the two. These are sample pairs obtained from the real sample distribution and the generated sample distribution, respectively. It is a distance metric between samples, used to characterize the difference between two distributions; The generative supplementation module is trained independently with positive and negative sample generators, and performs supplementation at multiple candidate time steps. The following steps generate samples, resulting in a corresponding set of generated samples; for each Calculate the generated sample distribution Compared with the true sample distribution Wasserstein distance between Select to make The optimal generation time is the time step that minimizes the time and whose change tends to stabilize within adjacent time steps. The optimal generation result is achieved when the Wasserstein distance simultaneously reaches a stable minimum.

5. The evaluation method according to claim 4, characterized in that, In step S4, The integrated discrimination module is used to realize the quantitative prediction of landslide occurrence probability. It consists of a multi-algorithm fusion system, including two processing branches: a random forest sub-model and an XGBoost sub-model. It also combines a Bagging / Boosting two-layer fusion strategy to form a composite discrimination framework. Specifically as follows: S4.1: Read the feature matrix of the training samples output from S3. and label vectors and construct Cross-validation partitioning; The values ​​are obtained from the landslide and non-landslide labels of the sample units: if the sample unit is a landslide sample, the value is 1; otherwise, the value is 0. Output: Folded data subset , For cross-validation folds; the first fold compromise Indicates a training subset. Indicates a subset for verification; S4.2: Random Forest Sub-model Training, also known as Bagging Branch: This involves training a random forest sub-model based on the Bagging mechanism to obtain a subset of samples. probability output ; Output: Probability output of Random Forest ; S4.3: XGBoost sub-model training, i.e., the Boosting branch: Training the XGBoos sub-model with cross-entropy as the objective function: in, For the sample size, For the first Each sample label For XGBoost to the first The output is the probability that each sample is predicted to be a landslide. This refers to the set of internal parameters of the XGBoost model. Introducing a complexity penalty term to control model size: in, The number of leaf nodes in the tree. The leaf node weight vector, for Regularization coefficient, This is the penalty coefficient for the number of leaf nodes, used to control model complexity; The final optimization goal is: Gradient boosting is used to iteratively update and calculate the optimal weights of the leaf nodes: in, Indicates the first During the first round of gradient boosting, the... The output of each tree for a sample, i.e., the optimal leaf weight when the sample falls into the corresponding leaf node, is used as the incremental update of the model's logit for that round. and The loss function is respectively paired with the first... The first and second gradients of the predicted values ​​for each sample. for Regularization coefficients, summed over the set of samples contained in the current leaf node; Output: XGBoost's logit output and probability output ; S4.4: Perform a two-level fusion of the outputs from the Bagging and Boosting branches to obtain the comprehensive logit: Output the probability of a landslide occurring: Output: Overall probability ; S4.5: Parameter optimization and model determination; Cross-entropy loss output by S4.3 As the objective function, Bayesian optimization is used in the parameter space. Inner iterative search, where To improve the learning rate of XGBoost iterations, leaf node weights Regularization coefficient, The penalty coefficient for the number of leaf nodes. The maximum depth of a single tree. To increase the number of trees generated in each iteration; each time a set of candidate parameters is generated That is, calling S4.4 to complete model training and validation and obtain the corresponding loss and evaluation metrics, until the optimal parameters are obtained. Based on this, the final integrated discriminant model is solidified. ; S4.6: In the final integrated discriminant model The SHAP mechanism is introduced to calculate the feature contribution vector of the sample. Quantify the impact of each factor The marginal impact; Output: Single-sample contribution vector and rasterized contribution layers for each factor; The module's final output: a continuous landslide probability layer for each raster cell. and feature contribution results ;in As an input to the S5 hierarchical susceptibility level output module, it is used for susceptibility level zoning.

6. The evaluation method according to claim 5, characterized in that, Feature contribution vector The calculation method is as follows: decompose the model output into the sum of the baseline value and the contribution of each feature, for a single sample. The model output changes are calculated when each feature is added or omitted under different feature combinations. The expected value of all combined contributions is then calculated to obtain the contribution value of each factor. .

Images