Landslide disaster case library construction and intelligent landslide prediction method based on data driving

Abstract

The invention discloses a landslide disaster case library construction and landslide intelligent prediction method based on data driving. According to the method, statistical analysis is carried out on existing historical landslide disaster case data, and various landslide types are marked; performing data preprocessing on the historical landslide disaster data under each grade according to multiple types of influence factors, re-dividing the grade of the case library by using a clustering algorithm, establishing a landslide disaster case library with a risk grade label and a landslide disaster risk prediction model, and obtaining a landslide disaster risk prediction model by taking precision as an index and based on a D-S evidence theory. Constructing a landslide disaster grade prediction model; and screening the collected real-time landslide disaster data of the current to-be-predicted area, and inputting the screened data into the landslide disaster grade prediction model to carry out landslide disaster risk prediction so as to obtain a landslide disaster risk prediction result. According to the invention, landslide disaster risk prediction is carried out on the real-time landslide data of the to-be-predicted area through the multi-class model, and landslide risk disasters are avoided.

Description

Technical Field

[0001] This invention relates to the field of landslide disaster prevention technology, and in particular to a data-driven method for constructing a landslide disaster case library and an intelligent landslide prediction method. Background Technology

[0002] With the increasing prominence of global warming and extreme climate change, landslides have become a common geological problem in reservoirs and mountains, posing a serious threat to human communities and natural ecosystems. This natural phenomenon is complex and variable, potentially triggered by natural factors such as river erosion, groundwater activity, rainwater inundation, and earthquakes, as well as by human activities such as blasting and slope cutting. Therefore, establishing scientifically effective landslide risk prediction models is crucial for effectively mitigating landslide risks.

[0003] However, various factors influence landslides, and these factors control different types of landslides to varying degrees. Building predictive models around only a few factors cannot comprehensively reflect all the characteristics of landslides. Limited by the incompleteness of historical landslide sample data—specifically, the lack of corresponding data across multiple influencing factors—currently, there are no methods for constructing supplementary landslide databases or corresponding risk prediction models. Unlike traditional empirical qualitative models, data-driven models can rely on deterministic models to supplement and improve historical landslide databases, utilize weighting systems to simplify the complex relationships between different related influencing factors, and integrate multiple machine learning algorithms to construct comprehensive decision-making models to predict potential landslide disaster risks. Summary of the Invention

[0004] The purpose of this invention is to provide a data-driven method for constructing a landslide disaster case library and for intelligent landslide prediction, aiming to solve the problems of incomplete and missing real-world data collection, large workload of existing technologies, poor robustness of prediction results, and low practicality of landslide disaster evaluation factors.

[0005] The technical solution provided by this invention is as follows: A data-driven method for constructing a landslide disaster case library and for intelligent landslide prediction, the method comprising the following steps: S1. Collect historical landslide big data, label the landslide types, and obtain several large landslide datasets based on the landslide types; S2. Preprocess the landslide disaster case database, including outlier detection and replacement, missing data supplementation, dataset standardization, and dataset dimensionality reduction; S3. Use an adaptive constraint clustering algorithm to reclassify the case library levels and establish a landslide disaster case library with risk classification labels; S4. Construct a landslide disaster risk prediction model based on a data-driven landslide disaster case library, and based on the landslide disaster risk prediction model, construct a landslide disaster level prediction model with accuracy as the indicator and based on DS evidence theory. S5. Collect real-time landslide disaster data of the current area to be predicted, filter the real-time landslide disaster data of the current area to be predicted, and then input it into the landslide disaster level prediction model to predict the landslide disaster risk and obtain the landslide disaster risk prediction result.

[0006] Preferably, in step S1, the landslide type label includes soil landslide, rock landslide, shallow landslide, and deep landslide; The process of obtaining several large landslide datasets based on landslide types specifically involves: establishing a large historical landslide database based on the collected historical landslide data, according to qualitative and quantitative influencing factors.

[0007] Preferably, the qualitative influencing factors include slope, aspect, slope type, slope position, slope variability, aspect variability, profile curvature, planar curvature, land use type, topographic relief, topographic cutting depth, topographic humidity index, SPI index, TWI index, NDVI, and elevation.

[0008] Preferably, in the historical landslide database established in step S1, data that deviates significantly from the value range is removed according to the qualitative evaluation model, and blank spaces are reserved for data that cannot be obtained.

[0009] Preferably, in step S2, the preprocessing procedure for outlier detection and replacement is as follows: outlier detection is performed on the data using the isolated forest method, the number of trees is set, a spatial correlation weighting mechanism for landslide factors is introduced, and the segmentation depth of the trees is dynamically adjusted according to the local similarity of geographical features such as slope and aspect; for the detected outliers, continuous variables are replaced with the truncated mean of the dataset after removing outliers, and discrete variables are replaced with the mode; The preprocessing process for supplementing the missing dataset is as follows: for continuous variables, spatially constrained KNN is used, with topographic humidity index and NDVI as distance metrics, and samples with the same slope aspect are selected first for local mean calculation; for missing discrete variables, the random forest classifier incorporates prior knowledge of expert scoring method, and transforms expert qualitative evaluation into class weights through fuzzy membership function. The standardized preprocessing of the dataset is as follows: the influencing factors are normalized to [0,1] in the value range, and the mean of each factor is calculated as a supplementary benchmark value; the mean of all other relevant factors participating in the input of the BP neural network model is selected, and then the BP neural network pattern recognition model is constructed based on the Scikit-Learn library in Python. A 3-layer BP neural network is constructed, and landslide level sensitivity analysis is added during the validation of the BP neural network; gradient perturbation test is performed on the normalized factors. If the change in landslide risk level is inconsistent with the factor gradient, adaptive normalization interval adjustment is triggered. The dimensionality reduction preprocessing of the dataset is as follows: Principal component analysis is used to divide the 16 classes of data according to the degree of association between the subordinate categories. The parameters of each class are fitted with the fitness values ​​and the final function is established. The processed data are used as variables of each class. Principal component analysis is performed using the Scikit-Learn library of Python to calculate the contribution rate of each principal component. Finally, the principal components are selected as the data for dimensionality reduction of the original main influencing factors.

[0010] Preferably, in step S3, the clustering algorithm is as follows: an adaptive constraint clustering algorithm is used to refine the risk level prediction of the landslide case library. Based on minimizing the intra-class distance using the traditional K-means algorithm, a multi-source landslide dynamic monitoring index is introduced as a clustering constraint.

[0011] Preferably, in step S3, the clustering algorithm is: using an improved weighted Gaussian mixture model to reconstruct the risk level classification process.

[0012] Preferably, in step S4, the landslide disaster risk prediction model includes a decision tree landslide prediction model, a support vector machine landslide prediction model, and a random forest landslide disaster risk prediction model.

[0013] Preferably, in step S5, after obtaining the landslide prediction results, the results are compared with the landslide labels that have been verified in the field. At the same time, the prediction performance of the landslide disaster level prediction model is compared with that of the base model, and the prediction performance of the model is evaluated based on some accuracy indicators. The base model includes decision tree, support vector machine, and random forest model.

[0014] This invention overcomes the shortcomings of existing technologies and provides a data-driven method for constructing a landslide disaster case library and intelligent landslide prediction. Specifically, the method involves: statistically analyzing existing historical landslide disaster case data and labeling various landslide types; preprocessing historical landslide disaster data at each level according to multiple influencing factors, including cluster value detection and replacement, missing data supplementation, dataset standardization, and dimensionality reduction; reclassifying the case library using a clustering algorithm into four risk levels: weak risk, medium risk, strong risk, and extremely strong risk; establishing a landslide disaster case library with risk level labels; constructing decision tree landslide prediction models, support vector machine landslide prediction models, and random forest landslide disaster risk prediction models, using accuracy as an indicator, and based on DS evidence theory to construct a landslide disaster level prediction model; collecting real-time landslide disaster data for the area to be predicted, filtering the real-time landslide disaster data for the area to be predicted, and then inputting it into the landslide disaster risk prediction model to predict the landslide disaster risk, obtaining the landslide disaster risk prediction result.

[0015] The beneficial effects of this invention after adopting the above technical solution are as follows: This invention publicly collects historical landslide data, selects various influencing factors to establish several large landslide databases, and employs methods including outlier detection and replacement, missing data supplementation, and dataset standardization and dimensionality reduction to supplement and integrate the databases, thereby improving the accuracy of historical landslide data and the prediction accuracy of machine learning. Furthermore, this invention establishes multiple models to predict landslide disaster risks using real-time landslide data of the area to be predicted, thus mitigating landslide risk disasters. Attached Figure Description

[0016] Figure 1 This is a flowchart of the steps of the method of the present invention; Figure 2 This is a graph showing the prediction results of the support vector machine model on the test set; Figure 3 This is a graph showing the prediction results of the decision tree model on the test set; Figure 4 This is a graph showing the prediction results of the random forest model on the test set. Detailed Implementation

[0017] The technical solution of the present invention will be further described in detail below with reference to specific embodiments, but this does not constitute any limitation on the present invention.

[0018] This invention discloses a data-driven method for constructing a landslide disaster case library and for intelligent landslide prediction. Figure 1 As shown, the procedure includes the following steps: S1. Collect historical landslide big data, label the landslide types, and obtain several large landslide datasets based on the landslide types; S2. Preprocess the landslide disaster case database, including outlier detection and replacement, missing data supplementation, dataset standardization, and dataset dimensionality reduction; S3. Use an adaptive constraint clustering algorithm to reclassify the case library levels and establish a landslide disaster case library with risk classification labels; S4. Construct a landslide disaster risk prediction model based on a data-driven landslide disaster case library, and based on the landslide disaster risk prediction model, construct a landslide disaster level prediction model with accuracy as the indicator and based on DS evidence theory. S5. Collect real-time landslide disaster data of the current area to be predicted, filter the real-time landslide disaster data of the current area to be predicted, and then input it into the landslide disaster level prediction model to predict the landslide disaster risk and obtain the landslide disaster risk prediction result.

[0019] This invention combines the isolated forest method, BP neural network theory, principal component analysis, adaptive constrained clustering algorithm, and improved weighted Gaussian mixture model (WGMM) to propose a data-driven method for constructing a landslide disaster case database and intelligent landslide prediction. This invention performs statistical analysis on historical landslide big data. Through data preprocessing (outlier detection and replacement using box plots), data standardization using BP neural network theory, and data division into four groups based on the correlation of influencing factors, dimensionality reduction using principal component analysis, and the establishment of databases with different labels, this invention uses clustering algorithms to train and test multiple machine learning models using data from the database. Once all indicators meet reliability requirements, landslide risk prediction for the area to be predicted can be performed.

[0020] In step S1, landslide-related data are collected through UAV topographic surveys, geological borehole sampling, meteorological station rainfall monitoring, satellite remote sensing (NDVI, elevation), and historical disaster archives. Landslide types are labeled as soil landslides, rock landslides, shallow landslides, and deep landslides. Based on the collected historical landslide data, a large database of historical landslides is established according to 16 qualitative and quantitative influencing factors, including slope, aspect, slope type, slope position, slope variability, aspect variability, profile curvature, plane curvature, land use type, topographic relief, topographic cutting depth, topographic humidity index, SPI index, TWI index, NDVI, and elevation. At the same time, according to the qualitative evaluation model, data that deviates significantly from the range, such as slope >75° and abnormally high topographic humidity index (>15), are removed, and blank spaces are reserved for data that cannot be obtained.

[0021] In step S2, since most landslide datasets lack completeness and gaps appear in most historical landslide databases as the number of statistical influencing factors increases, it is necessary to preprocess the landslide disaster case database, including outlier detection and replacement, missing dataset supplementation, dataset standardization, and dataset dimensionality reduction.

[0022] For outlier detection, the Isolation Forest method was used to detect outliers in the data. The number of trees was set to 100, and a spatial correlation weighting mechanism for landslide factors was introduced. The segmentation depth of the trees was dynamically adjusted according to the local similarity of geographical features such as slope and aspect, so as to avoid misjudgment of mountain areas caused by fixed parameters.

[0023] For detected outliers, continuous variables are replaced with the truncated mean of the dataset after removing outliers, and discrete variables are replaced with the mode to avoid distortion of the meaning of the median of categorical variables; For missing data in the dataset, appropriate values ​​are used to supplement them: for continuous variables, spatially constrained KNN is used, with topographic moisture index (TWI) and NDVI as distance metrics. Samples with the same slope aspect are preferentially selected for local mean calculation to ensure that the imputed values ​​retain the spatial continuity of landslide-prone factors. In handling missing discrete variables, a random forest classifier incorporates prior knowledge from expert scoring, and a fuzzy membership function is used to transform expert qualitative evaluations into category weights, improving the prediction accuracy of discrete data such as land use types.

[0024] For dataset standardization, the influencing factors are normalized to [0,1] within their value range, and the mean of each factor is calculated as a supplementary baseline value. The mean of all other relevant factors input to the BP neural network model is selected. Then, based on the Scikit-Learn library in Python, a BP neural network pattern recognition model is constructed with three layers (16 nodes in the input layer, 8 nodes in the hidden layer, and 1 node in the output layer). During BP neural network validation, a landslide risk sensitivity analysis is incorporated: gradient perturbation tests are performed on the normalized factors. If the change in landslide risk level is inconsistent with the factor gradient (e.g., increased slope leading to increased risk), adaptive normalization interval adjustment is triggered to avoid distortion of landslide features after standardization.

[0025] Principal component analysis was used to reduce the dimensionality of the dataset. The 16 data categories were divided into four classes based on their association with each other: A1 (slope, aspect, slope type, slope position), A2 (slope variability, aspect variability, profile curvature, planar curvature), A3 (land use type, topographic relief, topographic dissection depth, topographic humidity index), and A4 (SPI index, TWI index, NDVI, elevation). For each class, four sets of parameters A... ij , respectively with fitness value Y iPerform function fitting and establish the final function Y i =A i1 +A i2 +A i3 +A i4 The four sets of data, after processing, were used as four types of variables. Principal component analysis was performed using Python's Scikit-Learn library to calculate the contribution rates of the first, second, third, and fourth principal components. The first, second, and third principal components were ultimately selected as the data for dimensionality reduction of the original main influencing factors. After principal component analysis, nine main influencing factors were finally selected from the initial 16 influencing factors.

[0026] In step S3, a clustering algorithm is used to reclassify the case library levels. In this embodiment, two clustering algorithms are used to classify the case library levels. Both algorithms are implemented in the Scikit-Learn library of Python.

[0027] The first clustering algorithm employs an adaptive constraint clustering algorithm to refine the risk level prediction of the landslide case library. Based on the traditional K-means algorithm's minimization of intra-cluster distance, it introduces dynamic monitoring indicators of multi-source landslides as clustering constraints. The basic process of this algorithm is as follows: S31. The algorithm first determines the search range of K value (3~6 classes) through Canopy pre-clustering. S32. Subsequently, when calculating the sum of squared distances within a class (SSE), the standard deviation coefficient of landslide displacement rate (with a threshold set to ≤0.35) and the rainfall-induced response factor (based on the critical rainfall intensity curve retrieved from historical landslide events) are integrated to construct the objective function: (1) Where α and β are dynamic weighting coefficients, x is a landslide case in the database, μ represents the centroid of the class, and σ v R is the standard deviation of displacement rate within the class, and R is the overlap of rainfall response among samples within the class. hist The overlap of historical rainfall responses; S33. When the objective function simultaneously satisfies σ v When both K and R are minimized and maximized, the system automatically locks the optimal K value without manual intervention. S34. After clustering, a four-dimensional feature space is constructed based on the landslide stability index. By calculating the Mahalanobis distance between the cluster centers of areas where no landslides have occurred and those where landslides have occurred, and combining the Bayesian probability to determine the overlap, the risk level is divided into four risk levels. S35. Generate a landslide disaster case library with multi-dimensional labels, where the labels include cluster center feature vectors, risk probability values, and dynamic constraint parameters.

[0028] The second clustering algorithm uses an improved weighted Gaussian mixture model (WGMM) to reconstruct the risk level classification process. The basic process of this algorithm is as follows: S301. First, based on prior knowledge of the frequency of disasters occurring in slope type and slope location, initialize the Gaussian distribution parameters; S302. Calculate the posterior probability of each data point using the E-step algorithm after applying the Gaussian distribution. In the E-step algorithm, the posterior probability calculation includes an additional terrain relief constraint term to correct the latent membership degree. (2) Where, π j , π k x is the mixing coefficient. n Let μ be the feature vector of the nth sample point. k Let Σk be the mean vector of the k-th Gaussian distribution, Σk be the covariance matrix of the k-th Gaussian distribution, and ω(d) be the mean vector of the k-th Gaussian distribution. n ) represents the terrain relief weighting function, d n is the topographic relief index of the nth sample point.

[0029] Mixing coefficient π k Determined by the statistical frequency of slope-type disasters, the mean μ k Based on the adaptive allocation of the slope stability index, the covariance Σk is introduced into the topographic relief weight matrix (the weight is calculated by coupling the slope curvature and the elevation standard deviation).

[0030] S303. The algorithm converges by recalculating the covariance matrix using the M-step approach based on the posterior probability. During the M-step update, the covariance matrix is ​​estimated using weighted maximum likelihood. This optimization makes the Gaussian distribution better reflect the terrain-driven characteristics of landslide data.

[0031] S304. Monitor the convergence of the model using KL divergence (threshold < 0.001) and verify the rationality of the number of classifications based on the Bayesian Information Criterion (BIC).

[0032] S305. Finally, output the probability distribution diagrams of the four risk levels and the confidence intervals of each category to achieve uncertainty quantification.

[0033] After clustering the constructed database using two clustering algorithms—adaptive constrained clustering algorithm and improved weighted Gaussian mixture model (WGMM)—the results of the two clustering methods were analyzed. Ultimately, the landslide levels were classified into four categories: weak risk, medium risk, strong risk, and extremely strong risk. A landslide disaster case library with risk classification labels was established.

[0034] In step S4, a landslide disaster risk prediction model based on a data-driven landslide disaster case database is constructed. The database is divided into training and testing sets in an 8:2 ratio. To verify the rationality of the landslide disaster risk prediction model through multiple methods, three landslide disaster risk prediction models—a decision tree model in weak learners, a vector machine model in high-level learners, and a random forest model in ensemble learning—are used as inputs to train the model and obtain the optimal hyperparameters. All data is imported into the Scikit-Learn library in Python, and a model is built based on this library. The relevant model usage is as follows: The decision tree model in the weak learner includes: using feature splitting algorithms (information entropy and Gini coefficient) and pruning parameters as hyperparameters; limiting the maximum depth of the model to prevent overfitting; and setting hyperparameters such as random mode as default to ensure stability when processing high-dimensional landslide data. Based on landslide factor contribution analysis (e.g., the contribution of slope variability to landslides is greater than 30%), the model introduces an adaptive depth control mechanism: simultaneously calculating the weighted index of information gain rate and landslide risk entropy when splitting nodes; triggering pre-pruning when the factor correlation is below 10%; and dynamically adjusting the maximum depth threshold according to feature importance. Addressing the common occurrence of missing data in geological surveys, the node splitting criterion integrates the dual constraints of Gini coefficient and landslide risk entropy, effectively avoiding overfitting caused by missing data and improving the model's generalization ability in complex geological environments.

[0035] Support vector machine models in advanced learners include those employing discriminant formulas. (3) x represents the input vector to be classified, b represents the bias, and α is a non-negative parameter. o,i The `sign()` function is used to define the support vectors in the input vector. The data space is divided into m regions and m categories. If all data in the m-th category are "positive," then the remaining m-1 data are "negative." The support vector machine model employs a kernel function switching strategy: after dimensionality reduction through principal component analysis, the Pearson correlation coefficient between the factors and the landslide level is calculated. High-correlation factors (|r|>0.5) use a Gaussian kernel function to capture local nonlinear features, while low-correlation factors (|r|<0.2) switch to a multinomial kernel function. To address the class imbalance problem caused by the scarcity of historical landslide samples, the penalty coefficient C is dynamically weighted according to the proportion of landslide samples (positive sample weight = 1 + ln(total number of samples / number of landslide samples)). The sign function is used to optimize the spatial partitioning in multi-class scenarios, making the hyperplane closer to the high-risk sample cluster.

[0036] The random forest model in ensemble learning includes: based on bagged ensemble theory, multiple training subsets are generated through bootstrap sampling, each decision tree independently learns the nonlinear relationship between landslide factors and disasters, and a voting mechanism is used to comprehensively output the risk level; during bootstrap sampling, the local mean of K-nearest neighbor interpolation is used as the sample weight factor (the voting weight of non-outlier samples is increased by 20%) to enhance the collaborative judgment ability of spatially adjacent samples; to address the differences in landslide causes in different slope aspects, the model implements slope aspect partitioning training: first, the data is divided into four subsets according to the slope aspect quadrant, each subset independently trains a decision tree with the geological unit as the smallest splitting unit, and finally the partitioning results are integrated with the global prediction in the voting stage; during the training process, a landslide risk-specific loss function is added, and 5-fold cross-validation is performed with F1-score as the core indicator. When the accuracy fluctuation of the test set exceeds 5%, grid search re-optimization is automatically triggered to ensure the robustness of the model under complex terrain.

[0037] Based on the above, a landslide disaster risk prediction model is constructed. Using accuracy as an indicator, a landslide disaster level prediction model is built based on the DS evidence theory. The DS evidence theory includes: an identification framework, a basic probability assignment function (BPA), and evidence combination rules. The fusion process based on the DS evidence theory is as follows: S41. Constructing the Initial BPA: Introducing a fuzzy risk level interval based on historical landslide data, using a composite hypothesis to characterize the uncertainty between categories, and generating the initial BPA by calculating the evidence distance between the output probabilities of each model, including DT, SVM, and RF, and the composite hypothesis; S42. Dynamically calculate the reliability weight of evidence sources: dynamically update the weights by combining the local consistency between model predictions and field monitoring data in the real-time data stream and the information entropy of the model's historical performance; apply a trust discount factor to evidence sources with low reliability and adjust their BPA. S43. Weighted Hybrid Evidence Combination: Based on the Dempster rule, input the weighted discounted BPA; monitor the global conflict coefficient K in real time. If K exceeds the adaptive threshold (dynamically adjusted according to the landslide warning level), start the secondary allocation algorithm based on the intersection complement weight, and redistribute the conflict quality according to the relative support of the proposition. S44. Before the final decision, the probability distribution of all composite hypotheses is transformed into the probability distribution of a single hypothesis through Pignistic probability transformation, providing a basis for risk level determination. S45. The entire fusion process is backtested through a historical landslide event database. The Bayesian update mechanism is used to continuously optimize the BPA generation and combination parameters to ensure stable decision-making performance even in extremely high conflict scenarios, and significantly reduce false positives and false negatives.

[0038] In step S5, real-time landslide disaster data of the current area to be predicted is collected, the real-time landslide disaster data of the current area to be predicted is filtered, and then input into the landslide disaster level prediction model to predict the landslide disaster risk and obtain the landslide disaster risk prediction result.

[0039] After obtaining the landslide prediction results, the results are compared with the landslide labels that have been verified in the field. At the same time, the prediction performance of the fusion model is compared with that of the base model. The prediction performance of the model is evaluated based on some accuracy indicators (AUC, G-mean, F1 score, Kappa coefficient, Matthews correlation coefficient).

[0040] To verify the predictive effectiveness of the fusion model, another set of data from the region was selected as the model input. As can be seen from the data in Table 1, different machine learning algorithms have different accuracies for landslide data of different levels. The DS fusion model can select more accurate prediction results and solve some of the erroneous predictions produced by the algorithms.

[0041] Table 1 Prediction Results

[0042] Based on the same test set data, the prediction accuracy of the three base models (SVM, DT, RF) is as follows: Figures 2-4 As shown, the prediction accuracy of the fused model algorithm is 81.66%, which verifies the effectiveness of the fused model algorithm.

[0043] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A data-driven method for constructing a landslide disaster case library and for intelligent landslide prediction, characterized in that, The method includes the following steps: S1. Collect historical landslide big data, label the landslide types, and obtain several large landslide datasets based on the landslide types; S2. Preprocess the landslide disaster case database, including outlier detection and replacement, missing data supplementation, dataset standardization, and dataset dimensionality reduction; S3. Use an adaptive constraint clustering algorithm to reclassify the case library levels and establish a landslide disaster case library with risk classification labels; S4. Construct a landslide disaster risk prediction model based on a data-driven landslide disaster case library, and based on the landslide disaster risk prediction model, construct a landslide disaster level prediction model with accuracy as the indicator and based on DS evidence theory. S5. Collect real-time landslide disaster data of the current area to be predicted, filter the real-time landslide disaster data of the current area to be predicted, and then input it into the landslide disaster level prediction model to predict the landslide disaster risk and obtain the landslide disaster risk prediction result.

2. The method as described in claim 1, characterized in that, In step S1, the landslide type marking includes soil landslide, rock landslide, shallow landslide, and deep landslide; The process of obtaining several large landslide datasets based on landslide types specifically involves: establishing a large historical landslide database based on the collected historical landslide data, according to qualitative and quantitative influencing factors.

3. The method as described in claim 2, characterized in that, The qualitative influencing factors include slope, aspect, slope type, slope position, slope variability, aspect variability, profile curvature, planar curvature, land use type, topographic relief, topographic cutting depth, topographic humidity index, SPI index, TWI index, NDVI, and elevation.

4. The method as described in claim 2, characterized in that, In the historical landslide database established in step S1, data that deviates significantly from the value range is removed based on the qualitative evaluation model, and blank spaces are reserved for data that cannot be obtained.

5. The method as described in claim 1, characterized in that, In step S2 The preprocessing process for outlier detection and replacement is as follows: outlier detection is performed on the data using the isolated forest method, the number of trees is set, a spatial correlation weighting mechanism for landslide factors is introduced, and the segmentation depth of the trees is dynamically adjusted based on the local similarity of geographical features such as slope and aspect; for the detected outliers, continuous variables are replaced with the truncated mean of the dataset after removing outliers, and discrete variables are replaced with the mode. The preprocessing process for supplementing the missing dataset is as follows: continuous variables are based on spatially constrained KNN, with topographic humidity index and NDVI as distance metrics, and samples with the same slope aspect are preferentially selected for local mean calculation. In handling missing discrete variables, the random forest classifier incorporates prior knowledge from expert scoring and transforms expert qualitative evaluations into class weights through a fuzzy membership function. The standardized preprocessing of the dataset is as follows: the influencing factors are normalized to [0,1] in the value range, and the mean of each factor is calculated as a supplementary benchmark value; the mean of all other relevant factors participating in the input of the BP neural network model is selected, and then the BP neural network pattern recognition model is constructed based on the Scikit-Learn library in Python. A 3-layer BP neural network is constructed, and landslide level sensitivity analysis is added during the validation of the BP neural network; gradient perturbation test is performed on the normalized factors. If the change in landslide risk level is inconsistent with the factor gradient, adaptive normalization interval adjustment is triggered. The dimensionality reduction preprocessing of the dataset is as follows: Principal component analysis is used to divide the 16 classes of data according to the degree of association between the subordinate categories. The parameters of each class are fitted with the fitness values ​​and the final function is established. The processed data are used as variables of each class. Principal component analysis is performed using the Scikit-Learn library of Python to calculate the contribution rate of each principal component. Finally, the principal components are selected as the data for dimensionality reduction of the original main influencing factors.

6. The method as described in claim 1, characterized in that, In step S3, the clustering algorithm is as follows: an adaptive constraint clustering algorithm is used to refine the risk level prediction of the landslide case library. Based on the traditional K-means algorithm to minimize the intra-class distance, a multi-source landslide dynamic monitoring index is introduced as a clustering constraint.

7. The method as described in claim 1, characterized in that, In step S3, the clustering algorithm is: to reconstruct the risk level classification process using an improved weighted Gaussian mixture model.

8. The method as described in claim 1, characterized in that, In step S4, the landslide disaster risk prediction model includes a decision tree landslide prediction model, a support vector machine landslide prediction model, and a random forest landslide disaster risk prediction model.

9. The method as described in claim 1, characterized in that, In step S5, after obtaining the landslide prediction results, the results are compared with the landslide labels that have been verified in the field. At the same time, the prediction performance of the landslide disaster level prediction model is compared with that of the base model, and the prediction performance of the model is evaluated based on some accuracy indicators. The base model includes decision tree, support vector machine, and random forest model.

Images