Multi-level integrated landslide susceptibility evaluation method based on chimeric model
By constructing a multi-level ensemble method based on a chimeric model, the problems of insufficient data utilization and model accuracy in traditional landslide disaster assessment methods are solved, achieving higher accuracy in landslide susceptibility assessment and improving the model's generalization ability and predictive performance.
Patent Information
- Application Number
- CN202510856534.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-25
- Publication Date
- 2025-10-17
AI Technical Summary
Traditional landslide disaster assessment methods are inadequate in terms of data utilization, model accuracy, and predictive ability, making it difficult to meet the needs of high-precision disaster prevention. Single machine learning models suffer from overfitting and insufficient generalization ability when dealing with complex geological conditions.
A multi-level ensemble method based on a chimeric model is adopted. By combining data preprocessing, contribution factor calculation and extreme random tree layer and gradient boosting framework layer in the landslide disaster research area, a chimeric model is constructed. Feature selection and modeling are performed using contribution factor weighting and decision tree, the model training is optimized, and finally the landslide susceptibility is assessed.
This improves the accuracy and reliability of landslide susceptibility assessment. By enhancing model performance through multi-level integration, it scientifically and rationally selects features and weakens the influence of secondary factors, thereby achieving a more accurate landslide susceptibility assessment.
Smart Images
Figure CN120804907A_ABST
Abstract
Description
Technical Field
[0001] This specification relates to the technical field of landslide disaster prediction, and in particular to a multi-level integrated landslide susceptibility assessment method based on a chimeric model. Background Art
[0002] Landslides pose a serious threat to human life, property, and infrastructure. Traditional assessment methods rely on geological surveys, terrain analysis, and empirical models, but these methods are limited in data utilization, model accuracy, and predictive capabilities. Advances in machine learning have led to widespread application in geological hazard assessment. However, single models still suffer from overfitting and insufficient generalization when dealing with complex geological conditions, making them inadequate for high-precision disaster prevention. Summary of the Invention
[0003] In view of the above-mentioned deficiencies in the prior art, the present invention provides a multi-level integrated landslide susceptibility assessment method based on a mosaic model, which solves the problem of low accuracy in landslide susceptibility risk assessment.
[0004] In order to achieve the above-mentioned object, the technical solution adopted by the present invention is: a multi-level integrated landslide susceptibility assessment method based on a mosaic model, comprising: S1: Preprocess the evaluation factor data of the landslide hazard study area, unify the spatial resolution and coordinate system of multiple types of data, and obtain the original data table of evaluation factors; S2: Calculate the contribution factors of the original data table of the evaluation factors to obtain a training set and a validation set; S3: Construct a chimeric model using the extreme random tree layer and the gradient boosting framework layer; S4: using the chimeric model, processing the training set and the validation set, and training using a loss function to obtain a trained chimeric model; S5: Analyze the evaluation factor data using the trained chimeric model to obtain a multi-level integrated landslide susceptibility assessment result, thereby completing the assessment of landslide susceptibility.
[0005] The beneficial effects of the present invention are: a multi-level integrated landslide susceptibility assessment method based on a chimeric model. By integrating the contribution factor value into feature engineering, a unique chimeric model is constructed, giving full play to the advantages of both and improving model performance. The contribution factor weighting is used to highlight important factors and weaken the influence of secondary factors, making feature selection more scientific and reasonable. At the same time, in the model structure, the decision tree is used for feature selection and preliminary modeling, key features are extracted and a basic model is constructed, and the results are then input into the gradient boosting framework model for final prediction, and model training is optimized. Through multi-level integrated evaluation factors, more accurate landslide susceptibility assessment results are obtained.
[0006] Furthermore, the S1 includes: obtain topographic data, geological data and meteorological data of a landslide disaster research area; wherein the topographic data, the geological data and the meteorological data belong to evaluation factor data; unify and resample the topographic data, the geological data and the meteorological data in spatial resolution and coordinate system to obtain an evaluation factor original data table.
[0007] Further, the S2 comprises: processing the evaluation factor original data table to remove missing values and outliers to obtain an evaluation factor complete data table; analyzing each evaluation factor in the evaluation factor complete data table to obtain a corresponding contribution factor value; multiplying the contribution factor value and the original value of the corresponding evaluation factor to obtain a weighted data set; dividing the weighted data set by proportion to obtain a training set and a validation set.
[0008] By calculating the CF value of all factors and multiplying it with the original value of the factor to generate a new weighted data set. The CF value reflects the contribution degree of each evaluation factor to the occurrence of landslide, which is calculated based on the statistical distribution of landslide points and non-landslide points. Through CF weighting, factors with greater contribution to landslide occurrence are highlighted, and the influence of factors with smaller contribution is weakened, making feature selection more scientific and providing a better data basis for model training, which significantly improves the model performance.
[0009] Further, the chimeric model comprises: an extreme random tree layer for randomly selecting a feature subset from the weighted data set to construct each decision tree; for each selected feature subset, a split point is randomly selected in the value range thereof to split a node; based on the selected split point, the data set is recursively divided into sub-data sets until a preset stopping condition is reached, the contribution of each feature at the split node is recorded to obtain an evaluation feature vector; a gradient boosting framework layer for calculating a landslide prediction value based on the evaluation feature vector; using the current landslide prediction value, a residual between the prediction value and the true value is calculated; based on the residual, a new decision tree is constructed, and by evaluating all features and possible split points, a split point that can most reduce the residual is determined; based on the best split point, the data set is divided into sub-data sets; the above steps are recursively repeated until a preset stopping condition is reached to obtain a landslide susceptibility probability value.
[0010] By using Extratrees as the primary model for feature selection and initial modeling, the key features in the data are extracted and a basic model is constructed, and then the results are input into the LGB model for final prediction. The LGB model is based on the gradient boosting framework and has the ability to efficiently process large-scale data and avoid overfitting. This hybrid model structure fully leverages the strengths of both algorithms, enabling the model to have stronger feature extraction capabilities and prediction performance when dealing with complex geological conditions.
[0011] Further, the expression of the loss function of the trained hybrid model is: ; Wherein, represents the loss function result, represents the number of samples, represents the true label of the i-th sample, represents the probability of the i-th sample predicted by the model as a landslide point.
[0012] Further, the S5 comprises: Using the trained hybrid model, the evaluation factor data is analyzed to obtain a landslide susceptibility probability value corresponding to each grid cell; The landslide susceptibility probability values are arranged in ascending order, and arranged based on the level plan to obtain a multi-level integrated landslide susceptibility evaluation result, and the landslide susceptibility is evaluated.
[0013] One or more embodiments of the present specification provide a multi-level integrated landslide susceptibility evaluation device based on a hybrid model, comprising a processor configured to execute a multi-level integrated landslide susceptibility evaluation method based on a hybrid model.
[0014] One or more embodiments of the present specification provide a computer-readable storage medium storing computer instructions, wherein when a computer reads the computer instructions in the storage medium, the computer executes a multi-level integrated landslide susceptibility evaluation method based on a hybrid model. BRIEF DESCRIPTION OF DRAWINGS
[0015] The present specification will be further illustrated in the form of exemplary embodiments, which will be described in detail with reference to the accompanying drawings. These embodiments are not limiting, and in these embodiments, the same numbers represent the same structures, wherein: Figure 1 is an exemplary flowchart of a multi-level integrated landslide susceptibility evaluation method based on a hybrid model according to some embodiments of the present specification. DETAILED DESCRIPTION
[0016] The specific embodiments of the present application are described below to facilitate the understanding of the present application for those skilled in the art, but it should be clear that the present application is not limited to the scope of the specific embodiments, and for those skilled in the art, all the inventions utilizing the concept of the present application are within the scope of the present application as long as various changes are obvious within the spirit and scope of the present application defined and determined by the appended claims.
[0017] Embodiment one Figure 1 is an exemplary flowchart of a multi-level integrated landslide susceptibility assessment method based on a chimeric model according to some embodiments of the present specification. As shown in Figure 1 , the flow includes the following steps. In some embodiments, the flow can be executed by a processor.
[0018] S1: Preprocess the evaluation factor data of the landslide disaster research area, unify the spatial resolution and coordinate system of multiple types of data, and obtain an evaluation factor original data table.
[0019] The evaluation factor data is multi-source data used to evaluate the landslide risk of the landslide disaster research area. For example, the evaluation factor data can include terrain data, geological data, and meteorological data; wherein the terrain data can include elevation, slope, slope direction, surface cutting coefficient, profile curvature, elevation variation coefficient, ground roughness, curvature, terrain wetness index (TWI), etc., the geological data can include river distance, distance from road, distance from fault, normalized difference vegetation index (NDVI), land use type, and lithology, etc., and the meteorological data can include rainfall, etc.
[0020] The evaluation factor original data table is a table with unified resolution and coordinate system including all evaluation factors.
[0021] In some embodiments, the processor can process the evaluation factor data of the landslide disaster research area, unify the spatial resolution and coordinate system, resample the layers to have the same grid size, and ensure that the number of rows and columns and the number of grids are consistent, to obtain the evaluation factor original data table.
[0022] In some embodiments, the processor can randomly generate non-landslide points outside the diameter of 1 kilometer from the landslide points to avoid sample bias caused by the close spatial relationship between landslide points and non-landslide points, and improve the model generalization ability.
[0023] In some embodiments, the processor can obtain terrain data, geological data, and meteorological data of the landslide disaster research area; unify and resample the spatial resolution and coordinate system of the terrain data, the geological data, and the meteorological data to obtain the evaluation factor original data table.
[0024] In some embodiments, the processor can utilize corresponding software to unify and resample the spatial resolution and coordinate system of the terrain data, geological data, and meteorological data to complete the construction of the original data table of the evaluation factors.
[0025] S2: Calculate the contribution factors of the original data table of the evaluation factors to obtain a training set and a validation set.
[0026] The contribution factor is a parameter used to reflect the impact of the evaluation factors on the landslide susceptibility risk weight.
[0027] The training set is the dataset used to train the chimeric model.
[0028] The validation set is a dataset used to evaluate the performance of the trained chimeric model.
[0029] In some embodiments, the processor can process the original data table of the evaluation factors, remove missing values and outliers, and obtain a complete data table of the evaluation factors; analyze each evaluation factor in the complete data table of the evaluation factors to obtain the corresponding contribution factor value; multiply the contribution factor value and the original value of the corresponding evaluation factor to obtain a weighted data set; divide the weighted data set proportionally to obtain a training set and a validation set.
[0030] The complete evaluation factor data table is an evaluation factor data table with missing values and outliers removed.
[0031] In some embodiments, the processor may use a table to process the original data table of the evaluation factors, remove missing values and outliers, ensure the integrity and accuracy of the data, and obtain a complete data table of the evaluation factors.
[0032] The contribution factor value is a numerical value that reflects the impact of the evaluation factor on the landslide susceptibility risk weight.
[0033] In some embodiments, the processor may obtain a contribution factor value (CF value) through calculation based on the statistical distribution of landslide points and non-landslide points.
[0034] In some embodiments, the contribution factor value may be expressed as: ; in, Represents the contribution factor value, represents the frequency ratio, k represents the kth evaluation factor, and n represents the number of evaluation factors.
[0035] The weighted dataset is a complete data set of evaluation factors that highlights the factors that contribute most to landslide occurrence.
[0036] In some embodiments, the processor may multiply the CF value of each evaluation factor by its original value to obtain a weighted data set.
[0037] In some embodiments, the processor can divide the weighted dataset in a 7:3 ratio, use a random sampling method to ensure consistent distribution of the training set and the validation set, avoid model performance bias caused by uneven data distribution, and obtain the training set and the validation set.
[0038] S3: Constructing a chimeric model using an extreme random tree layer and a gradient boosting framework layer.
[0039] The chimeric model is a neural network model for predicting landslide susceptibility probability, which is chimeric using an extreme random tree layer and a gradient boosting framework layer.
[0040] In some embodiments, the chimeric model can include an extreme random tree layer and a gradient boosting framework layer.
[0041] The extreme random tree layer is used to randomly select a feature subset from the weighted dataset to construct each decision tree; for each selected feature subset, a split point is randomly selected within its value range to split the node; based on the selected split point, the dataset is recursively divided into sub-datasets until the preset stopping condition is reached, the contribution of each feature at the split node is recorded, and an evaluation feature vector is obtained.
[0042] The evaluation feature vector is a feature vector related to landslide susceptibility in the weighted dataset.
[0043] Extratrees is a decision tree-based ensemble learning algorithm. For example, the parameters of extratrees can include max_depth (maximum depth): None, indicating that the depth of the tree is not limited and can grow indefinitely; min_samples_leaf (minimum number of samples in leaf nodes): 1, i.e., each leaf node must contain at least 1 sample; min_samples_split (minimum number of samples required for a split node): 2, meaning that a node must contain at least 2 samples to be split; n_estimators (number of trees): 200, indicating that 200 trees are used for ensemble, which helps to improve the stability and prediction performance of the model; random_state (random seed): 42, used to ensure the reproducibility of the results.
[0044] In some embodiments, the preset stopping condition of the extreme random tree layer can include the maximum depth of the tree, the minimum number of samples of the node, etc.
[0045] The gradient boosting framework layer is used to calculate the landslide prediction value based on the evaluation feature vector; use the current landslide prediction value to calculate the residual between it and the true value; construct a new decision tree based on the residual, and determine the split point that can minimize the residual by evaluating all features and possible split points; split the node based on the optimal split point to divide the dataset into sub-datasets; and recursively repeat the above steps until the preset stopping condition is reached to obtain the landslide susceptibility probability value.
[0046] The Gradient Boosting framework (LGB) is an efficient neural network model based on the gradient boosting framework. For example, parameters for the gradient boosting framework include: learning_rate (learning rate): 0.2, which controls the step size of model learning; max_depth (maximum depth): 20, indicating that each tree can have a maximum depth of 20 layers; n_estimators (number of trees): 200, using 200 trees for ensemble; num_leaves (number of leaf nodes): 40, which controls the complexity of the tree; more leaf nodes can lead to a more complex model (but increases the risk of overfitting); and random_state (random seed): 42, which ensures the reproducibility of results.
[0047] In some embodiments, the preset stopping conditions of the gradient boosting framework layer may include, for example, the maximum depth of the tree, the minimum number of samples of the leaf nodes, etc.
[0048] The predictions of the newly constructed decision trees are weighted (usually multiplied by a learning rate) and then updated to the model's total predictions. The learning rate controls the influence of each decision tree on the total prediction. A smaller learning rate can make the model more stable, but more decision trees are needed to achieve good predictions.
[0049] S4: Using the chimeric model, the training set and the validation set are processed, and training is performed using a loss function to obtain a trained chimeric model.
[0050] In some embodiments, a chimeric model can be trained using a labeled training set. For example, the labeled training set can be input into an initial chimeric model, and a loss function can be constructed using the labels and the results of the initial chimeric model. Based on the loss function, the parameters of the initial chimeric model are iteratively updated using methods such as grid search and cross-validation. When preset conditions are met, model training is completed, resulting in a trained chimeric model. The preset conditions may include convergence of the loss function, a threshold number of iterations, and the like.
[0051] In some embodiments, the label may be the actual landslide situation corresponding to each grid cell (positive sample is 1, negative sample is 0). The label may be manually annotated.
[0052] In some embodiments, the expression of the loss function can be: ; wherein, represents the loss function result, represents the number of samples, represents the true label of the i-th sample, represents the probability of the i-th sample predicted by the model as a landslide point.
[0053] In some embodiments, the processor can input the verification set into the embedded model, evaluate the prediction performance of the model by using the area under the curve (AUC), accuracy (ACC), F1 score, and precision indicators, and obtain an evaluation result.
[0054] In some embodiments, the expression of the area under the curve can be: ; wherein, represents the area under the curve, represents the ROC curve.
[0055] The ROC curve is a curve reflecting the points of the true positive rate (TPR) and the false positive rate (FPR) at different thresholds.
[0056] In some embodiments, the expressions of the true positive rate and the false positive rate can be: ; ; wherein, represents the true positive rate, represents the number of true positives, represents the number of true negatives, represents the false positive rate, represents the number of false positives, represents the number of false negatives.
[0057] The accuracy is an indicator representing the proportion of the number of correctly classified samples to the total number of samples.
[0058] In some embodiments, the expression of the accuracy can be: ; wherein, represents the accuracy.
[0059] The F1 score is a harmonic mean of precision and recall, which is an index for measuring the comprehensive performance of the model.
[0060] In some embodiments, the expression of the F1 score can be: ; ; ; wherein, denotes the F1 score, denotes the precision, denotes the recall.
[0061] S5: Using the trained hybrid model, analyze the evaluation factor data to obtain a multi-level integrated landslide susceptibility assessment result, and complete the assessment of landslide susceptibility.
[0062] The multi-level integrated landslide susceptibility assessment result is a landslide susceptibility probability result corresponding to the landslide disaster research area. For example, the multi-level integrated landslide susceptibility assessment result can include low, low, medium, high, and high levels.
[0063] Multi-level integration is mainly reflected in multiple links such as data processing, model construction, and result division. First, in the data processing stage, the evaluation factors of landslide points and non-landslide points are multiplied by CF to form a weighted data set, realizing data-level integration. Then in the model construction stage, Extratrees is used for feature selection and preliminary modeling, key features are extracted, and the results are input into the LightGBM model for final prediction. This hybrid model structure realizes multi-level integration at the model level, and fully utilizes the advantages of the two algorithms. Finally, in the result division stage, the landslide susceptibility probability values obtained by prediction are arranged in ascending order and divided into levels to obtain the multi-level integrated landslide susceptibility assessment result, providing decision support for landslide disaster prevention and management. Through integration at different levels, the accuracy and reliability of landslide susceptibility assessment are improved.
[0064] In some embodiments, the processor can use the trained hybrid model to analyze the evaluation factor data to obtain a landslide susceptibility probability value corresponding to each grid cell; arrange the landslide susceptibility probability values in ascending order based on the level planning to obtain a multi-level integrated landslide susceptibility assessment result, and complete the assessment of landslide susceptibility.
[0065] The landslide susceptibility probability value is a value reflecting the landslide susceptibility probability of a landslide disaster research area. After multiple iterations and construction of the decision tree, the final prediction value of the trained embedded model is the landslide susceptibility probability value of each sample (grid cell). These probability values reflect the likelihood of landslide occurrence in each grid cell.
[0066] In some embodiments, the processor can perform contribution factor calculation on the evaluation factor original data table, multiply the contribution factor value and the corresponding evaluation factor original value to obtain a weighted data set, perform feature selection on the weighted data set by using an extreme random tree layer to obtain an evaluation feature vector; perform gradient boosting processing and optimization on the evaluation feature vector by using a gradient boosting framework layer to obtain a multi-level integrated landslide susceptibility evaluation result, and complete the evaluation of landslide susceptibility.
[0067] In some embodiments of the present specification, a multi-level integrated landslide susceptibility evaluation method based on an embedded model is provided. By integrating the contribution factor value into feature engineering, a unique embedded model is constructed to fully leverage the advantages of both and improve model performance. The contribution factor is used to highlight important factors and weaken the influence of secondary factors, making feature selection more scientific and reasonable. At the same time, in the model structure, decision trees are used for feature selection and preliminary modeling to extract key features and build a basic model, and then the results are input into a gradient boosting framework model for final prediction, optimizing model training, and obtaining more accurate landslide susceptibility evaluation results through multi-level integrated evaluation factors.
[0068] Embodiment Two In some embodiments, a multi-level integrated landslide susceptibility evaluation device based on an embedded model includes a processor that can be used to execute a multi-level integrated landslide susceptibility evaluation method based on an embedded model.
[0069] Embodiment Three In some embodiments, a computer-readable storage medium stores computer instructions, and when a computer reads the computer instructions in the storage medium, the computer can execute a multi-level integrated landslide susceptibility evaluation method based on an embedded model.
[0070] It should be noted that different embodiments can have different beneficial effects, and in different embodiments, the beneficial effects that can be produced can be any one or a combination of the above, or any other beneficial effects that can be obtained.
Claims
1. A multi-level integrated landslide susceptibility assessment method based on a mosaic model, characterized by: include: S1: Preprocess the evaluation factor data of the landslide hazard study area, unify the spatial resolution and coordinate system of multiple types of data, and obtain the original data table of evaluation factors; S2: Calculate the contribution factors of the original data table of the evaluation factors to obtain a training set and a validation set; S3: Construct a chimeric model using the extreme random tree layer and the gradient boosting framework layer; S4: using the training set and the validation set to train the chimeric model to obtain a trained chimeric model; S5: Analyze the evaluation factor data using the trained chimeric model to obtain a multi-level integrated landslide susceptibility assessment result, thereby completing the assessment of landslide susceptibility.
2. The multi-level integrated landslide susceptibility assessment method based on the mosaic model according to claim 1 is characterized in that: Said S1 comprises: Acquiring topographic data, geological data, and meteorological data of a landslide hazard study area; wherein the topographic data, the geological data, and the meteorological data are evaluation factor data; The topographic data, the geological data and the meteorological data are subjected to spatial resolution and coordinate system unification and resampling processing to obtain an original data table of evaluation factors.
3. The multi-level integrated landslide susceptibility assessment method based on the mosaic model according to claim 1 is characterized in that: The S2 includes: Processing the original data table of the evaluation factors to remove missing values and outliers to obtain a complete data table of the evaluation factors; Analyze each evaluation factor in the complete data table of evaluation factors to obtain the corresponding contribution factor value; Multiplying the contribution factor value and the original value of the corresponding evaluation factor to obtain a weighted data set; The weighted data set is divided into a training set and a validation set according to a certain ratio.
4. The multi-level integrated landslide susceptibility assessment method based on a mosaic model according to claim 1 is characterized in that: The chimeric model includes: The extreme random tree layer randomly selects a subset of features from the weighted dataset to construct each decision tree. For each selected feature subset, a split point is randomly selected within its value range to split the node. The node is split based on the selected split point, recursively partitioning the dataset into sub-datasets until a preset stopping condition is reached. The contribution of each feature at the split node is recorded to obtain an evaluation feature vector. The gradient boosting framework layer is used to calculate the landslide prediction value based on the evaluation feature vector; use the current landslide prediction value to calculate the residual between it and the true value; construct a new decision tree based on the residual, and determine the split point that can minimize the residual by evaluating all features and possible split points; split the node based on the optimal split point to divide the dataset into sub-datasets; and recursively repeat the above steps until the preset stopping condition is reached to obtain the landslide susceptibility probability value.
5. The multi-level integrated landslide susceptibility assessment method based on a mosaic model according to claim 1 is characterized in that: The loss function of the trained chimeric model is expressed as: ； in, Represents the loss function result, represents the number of samples, represents the true label of the i-th sample, It represents the probability that the i-th sample predicted by the model is a landslide point.
6. The multi-level integrated landslide susceptibility assessment method based on a mosaic model according to claim 1, characterized in that: The S5 includes: Analyzing the evaluation factor data using the trained chimeric model to obtain a landslide susceptibility probability value corresponding to each grid unit; The landslide susceptibility probability values are arranged in ascending order and arranged based on a grade planning to obtain a multi-level integrated landslide susceptibility assessment result, thereby completing the assessment of landslide susceptibility.
7. A multi-level integrated landslide susceptibility assessment device based on a mosaic model, characterized in that: The method comprises a processor configured to execute the multi-level integrated landslide susceptibility assessment method based on a mosaic model according to any one of claims 1 to 6.
8. A computer-readable storage medium, characterized in that The storage medium stores computer instructions. When the computer reads the computer instructions in the storage medium, the computer executes the multi-level integrated landslide susceptibility assessment method based on a chimeric model as described in any one of claims 1 to 6.