Explainable machine learning combined with class imbalance handling for landslide susceptibility evaluation
By combining SMOTE-Tomek hybrid sampling and soft voting ensemble learning with the SHAP interpreter, the class imbalance and interpretability issues in landslide susceptibility assessment were resolved, the model's classification boundary and prediction reliability were optimized, and an interpretable landslide susceptibility assessment map was generated.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- JILIN UNIVERSITY
- Filing Date
- 2026-05-18
- Publication Date
- 2026-06-16
Smart Images
Figure CN122221054A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of engineering geology, and in particular to a landslide susceptibility assessment method that combines interpretable machine learning with class-disequilibrium treatment. Background Technology
[0002] Landslides are among the most common geological hazards in the world, causing numerous casualties and economic losses every year. Landslide susceptibility assessment is a fundamental task in geological hazard prevention and control, aiming to predict the spatial distribution of landslide probability based on regional geological environmental conditions. Traditional landslide susceptibility assessment methods mainly rely on expert experience for qualitative or semi-quantitative analysis, such as the analytic hierarchy process (AHP) and fuzzy comprehensive evaluation method, which suffer from strong subjectivity and difficulty in quantifying complex nonlinear relationships.
[0003] In recent years, machine learning technology has been widely used in geological hazard assessment. Models such as logistic regression, support vector machine, random forest, and neural network have been used to predict landslide susceptibility and have achieved good results. However, existing methods still face class imbalance problems. As landslides are low-probability events, the number of historical disaster point samples is far less than that of non-landslide samples. Furthermore, the selection and weighting of influencing factors largely depend on expert experience, lacking objective quantitative screening methods. Moreover, multicollinearity may exist among factors, affecting model stability. In addition, existing machine learning models generally suffer from black box problems, and the prediction results lack interpretability, making it difficult to reveal the contribution mechanism of each influencing factor to landslide occurrence. This results in low model credibility and makes it difficult for geological hazard prevention and control departments to adopt them. Summary of the Invention
[0004] In view of the aforementioned existing problems, the present invention is proposed.
[0005] Therefore, this invention provides an interpretable machine learning method combined with class imbalance processing for landslide susceptibility assessment, which solves the problems of sample imbalance and lack of interpretability of machine learning models in the landslide susceptibility assessment process.
[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides an interpretable machine learning method combined with class-imbalanced processing for landslide susceptibility assessment. The method includes: collecting historical landslide disaster data and multi-source influencing factors in a target area; constructing landslide and non-landslide samples based on the historical landslide disaster data, and using the SMOTE-Tomek hybrid sampling method to perform class-imbalanced processing on the landslide and non-landslide samples to obtain a balanced training set; filtering multi-source influencing factors through a dual screening mechanism combining random forest feature scoring and collinearity diagnosis to obtain an optimal feature subset; constructing and training a soft-voting ensemble learning model based on the optimal feature subset and the balanced training set, and outputting a landslide susceptibility index; using the SHAP interpreter to perform interpretability analysis on the trained soft-voting ensemble learning model, generating contribution analysis results of each influencing factor to landslide susceptibility; dividing the target area into susceptibility zones according to the landslide susceptibility index, and interpreting the zoning results based on the contribution analysis results of each influencing factor to landslide susceptibility, generating a landslide susceptibility assessment map.
[0007] As a preferred embodiment of the landslide susceptibility assessment method combining interpretable machine learning with class-disequilibrium processing described in this invention, the collection of historical landslide disaster point data and multi-source influencing factors in the target area includes collecting historical landslide disaster point data and extracting topographic factors, geological factors, hydrological factors, vegetation factors, and human activity factors from digital elevation models, geological maps, remote sensing images, and vector map data to form multi-source influencing factors.
[0008] As a preferred embodiment of the landslide susceptibility assessment method combining interpretable machine learning with class-disequilibrium processing described in this invention, the construction of landslide samples and non-landslide samples refers to establishing a buffer zone with a preset range centered on the landslide point based on historical landslide disaster data, extracting grid cells within the buffer zone as landslide samples, and randomly sampling non-landslide samples in landslide-free areas.
[0009] As a preferred embodiment of the landslide susceptibility assessment method combining interpretable machine learning with class-imbalanced processing described in this invention, the step of using the SMOTE-Tomek hybrid sampling method to process landslide samples and non-landslide samples for class-imbalanced processing to obtain a balanced training set is as follows: After using SMOTE to expand the number of landslide samples to the same number as the number of non-landslide samples, a balanced dataset was obtained. Tomek Link pairs are identified in the balanced dataset, and non-slippage samples in the Tomek Link pairs are removed to obtain a balanced training set.
[0010] As a preferred embodiment of the landslide susceptibility assessment method combining interpretable machine learning with class-imbalanced processing described in this invention, the optimal feature subset is obtained by combining random forest feature scoring with collinearity diagnosis. The feature importance of each influencing factor in the multi-source influencing factors is calculated through random forest, influencing factors with importance greater than the feature importance screening threshold are retained, and then the Pearson correlation coefficient between the retained factors is calculated and the highly correlated influencing factors are processed.
[0011] As a preferred embodiment of the landslide susceptibility assessment method combining interpretable machine learning with class imbalance processing described in this invention, the soft-voting ensemble learning model includes at least three base learners selected from logistic regression, support vector machine, random forest, XGBoost, and lightweight gradient boosting machine.
[0012] As a preferred embodiment of the landslide susceptibility assessment method combining interpretable machine learning and class-disequilibrium processing described in this invention, the specific steps for outputting the landslide susceptibility index are as follows: Based on the balanced sample set, the feature data corresponding to the optimal feature subset are selected to construct the training dataset, and the training dataset is divided into the training set and the validation set. The soft voting ensemble learning model is trained using the training set to obtain the trained soft voting ensemble learning model. The predicted probabilities and AUC values of each base learner are obtained on the validation set. Based on the AUC values of each base learner on the validation set, the predicted probabilities of each base learner are weighted and averaged to obtain the landslide susceptibility index.
[0013] As a preferred embodiment of the landslide susceptibility evaluation method combining interpretable machine learning with class imbalance processing described in this invention, the step of using the SHAP interpreter to perform interpretability analysis on the trained soft-voting ensemble learning model is based on the trained soft-voting ensemble learning model interpreting and analyzing the training dataset, and calculating the SHAP value corresponding to each sample in the training dataset.
[0014] As a preferred embodiment of the landslide susceptibility assessment method combining interpretable machine learning and class-disequilibrium processing described in this invention, the specific steps for generating the contribution analysis results of each influencing factor to landslide susceptibility are as follows: Based on the SHAP value, a SHAP bee diagram is generated to determine the dominant influencing factors affecting landslide development and the direction of their effects. SHAP waterfall plots are generated based on the SHAP values corresponding to individual historical landslide samples to determine the dominant influencing factors affecting landslide susceptibility and the direction of their influence in individual historical landslide samples. Based on the SHAP values corresponding to continuous impact factors, a SHAP dependency graph is generated to determine the variation law of the impact of continuous impact factors on landslide susceptibility and the impact threshold range. Based on the average SHAP value, a feature importance bar chart was generated to determine the ranking of the contribution of each influencing factor to landslide susceptibility, and the contribution analysis results of each influencing factor to landslide susceptibility were obtained.
[0015] As a preferred embodiment of the landslide susceptibility assessment method combining interpretable machine learning and class-imbalanced processing described in this invention, the specific steps of dividing the target area into susceptibility zones based on the landslide susceptibility index, interpreting the susceptibility zone results based on the contribution analysis results of each influencing factor to landslide susceptibility, and generating a landslide susceptibility assessment map are as follows: The target area is divided into grids, the landslide susceptibility index of each grid is calculated, and the natural breakpoint method is used to divide the landslide susceptibility index into high susceptibility areas, medium susceptibility areas and low susceptibility areas. Based on the analysis of the contribution of each influencing factor to landslide susceptibility, the results of each susceptibility zone are interpreted, and a landslide susceptibility evaluation map is generated.
[0016] The beneficial effects of this invention are as follows: it effectively solves the class imbalance problem and optimizes the classification boundary through SMOTE-Tomek hybrid sampling; it integrates the advantages of multiple models through soft voting ensemble learning, and can be widely applied to scenarios such as regional geological disaster investigation, land spatial planning, and site selection of major projects. At the same time, it uses the SHAP interpreter to perform global and local interpretability analysis, revealing the contribution mechanism of each influencing factor to landslide susceptibility, transforming the black box model into an interpretable decision support tool, enhancing the credibility of the model's prediction results, and providing technical support for geological disaster prevention and control. Attached Figure Description
[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0018] Figure 1 A flowchart for an interpretable machine learning-based method for assessing landslide susceptibility using class imbalance processing.
[0019] Figure 2 A flowchart for handling class imbalances in SMOTE-Tomek hybrid sampling.
[0020] Figure 3 A flowchart for feature selection for random forest feature scoring and collinearity diagnosis.
[0021] Figure 4 This is a flowchart of the model interpretability analysis based on SHAP values. Detailed Implementation
[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.
[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.
[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.
[0025] Reference Figures 1-4 As one embodiment of the present invention, this embodiment provides a landslide susceptibility assessment method that combines interpretable machine learning with class-disequilibrium processing, comprising the following steps: S1. Collect historical landslide disaster data and multi-source influencing factors in the target area.
[0026] Historical landslide disaster data were collected, and topographic factors, geological factors, hydrological factors, vegetation factors, and human activity factors were extracted from digital elevation models, geological maps, remote sensing images, and vector map data to form multi-source influencing factors.
[0027] Specifically, the system collects historical landslide disaster data in a high mountain and canyon area of a watershed, obtains SRTM DEM data (30m resolution), and extracts topographic factors such as elevation, slope, aspect, topographic relief, and surface incision depth; it can obtain 1:200,000 geological maps and extract geoscientific factors such as stratigraphic lithology and distance from faults; it can obtain annual average rainfall data and extract precipitation factors; it can obtain Landsat-8 OLI imagery and calculate the NDVI vegetation index; it can extract river and road data, calculate factors such as distance from rivers and roads, and construct multi-source influence factors.
[0028] S2. Based on historical landslide disaster data, landslide samples and non-landslide samples are constructed. The SMOTE-Tomek hybrid sampling method is used to process the landslide samples and non-landslide samples to obtain a balanced training set.
[0029] Constructing landslide and non-landslide samples refers to establishing a buffer zone with a preset range centered on the landslide point based on historical landslide disaster data, extracting grid cells within the buffer zone as landslide samples, and randomly sampling non-landslide samples in landslide-free areas.
[0030] Specifically, a 500m buffer zone was established centered on the landslide point, and the grid within the buffer zone was extracted as landslide samples, resulting in a total of 1245 landslide samples. In the non-landslide area, 12450 non-landslide samples were randomly sampled, with an initial positive-to-negative sample ratio of 1:10. Landslide samples were considered positive samples, and non-landslide samples were considered negative samples.
[0031] After using SMOTE to expand the number of landslide samples to the same number of non-landslide samples, a balanced dataset is obtained. Tomek Link pairs are identified in the balanced dataset, and non-landslide samples in the Tomek Link pairs are removed to obtain a balanced training set.
[0032] Specifically, the SMOTE-Tomek hybrid sampling method was used for processing. SMOTE oversampling was performed to expand the number of landslide samples to 12,450, forming a balanced dataset. Tomek Link pairs were identified in the balanced dataset, and a total of 863 Tomek Link pairs were identified. 863 non-landslide samples were removed, resulting in 12,450 landslide samples and 11,587 non-landslide samples, with a positive to negative sample ratio of approximately 1:0.93.
[0033] It should be noted that the expression for generating new minority class samples through SMOTE oversampling is: ; in, For new samples, For the first One original landslide sample, Indicates from the first One original landslide sample The th nearest neighbor sample is randomly selected from the th nearest neighbor samples. Nearest neighbor samples, is a random number between [0, 1] (usually following a uniform distribution). Representing the nearest neighbor parameter (here) The value 5 indicates the number of nearest neighbor samples selected for the original landslide sample.
[0034] It should also be noted that the Tomek Links judgment mechanism is that if the sample pair (x i , x j A Tomek Link is formed, and non-slope samples x in the Tomek Link pair are removed. jThe condition for forming a Tomek Link pair is that x belongs to a landslide sample, x j This is a non-landslide sample.
[0035] S3. The optimal feature subset is obtained by screening multi-source influencing factors through a dual screening mechanism that combines random forest feature scoring and collinearity diagnosis.
[0036] A combination of random forest feature scoring and collinearity diagnosis is adopted. The feature importance of each influence factor in the multi-source influence factors is calculated by random forest. Influence factors with importance greater than the feature importance screening threshold are retained. Then, the Pearson correlation coefficient between the retained factors is calculated, and highly correlated influence factors are processed to obtain the optimal feature subset.
[0037] Specifically, a random forest model is used to calculate the feature importance scores of multi-source influencing factors. Factors with an importance greater than the feature importance screening threshold of 0.05 (example) are retained. The Pearson correlation coefficient between the retained factors is calculated, and the factor with the relatively low feature importance among highly correlated factor pairs with an absolute correlation coefficient greater than 0.7 (example) is removed. The correlation coefficient between slope and topographic relief is calculated to be 0.82 (example), so topographic relief is removed (slope is usually a more direct factor for landslide evaluation, while topographic relief is a derived indicator of slope, so removing the latter is reasonable). The correlation coefficient between elevation and surface incision depth is 0.76 (example), so surface incision depth is removed (elevation itself is a basic topographic factor, while surface incision depth is a local elevation difference; when the two are highly correlated, retaining elevation is more common). The optimal factors, including elevation, slope, aspect, stratigraphic lithology, distance from fault, distance from river, and NDVI, are obtained and integrated to obtain the optimal feature subset.
[0038] It should be noted that the expression for calculating the characteristic importance of each impact factor in the multi-source impact factors is as follows: ; in, Indicates the characteristic importance of the influencing factor. Indicates the number of decision trees. Indicates the first Using features in a tree The set of nodes to be split. This indicates the reduction in Gini impurity before splitting. This indicates the reduction in Gini impurity after splitting. Represents the decision tree index variable. Represents a set of nodes For any node in, retain Factors.
[0039] The expression for calculating the Pearson correlation coefficient among retained factors is as follows: ; in, Indicating the impact factor With impact factor The Pearson correlation coefficient between them This represents a retention factor index. Indicates and Different retention factor indices, Indicates the number of samples. Indicates the sample index. Indicates the first Influence factors of each sample The value of , Indicates the first Influence factors of each sample The average value across all samples. Indicates the first Influence factors of each sample The value of , Indicates the first Influence factors of each sample The average value across all samples. If If so, one of the factors is removed (usually the more important one is retained).
[0040] S4. Based on the optimal feature subset and balanced training set, construct and train a soft-voting ensemble learning model to output a landslide susceptibility index.
[0041] Soft voting ensemble learning models include at least three base learners from logistic regression, support vector machine, random forest, XGBoost, and lightweight gradient booster.
[0042] Based on a balanced sample set, feature data corresponding to the optimal feature subset are selected to construct a training dataset, which is then divided into a training set and a validation set.
[0043] Specifically, the study area was divided into a 30m×30m grid, generating a total of 856,432 grid cells. The seven optimal factor values of each grid cell were extracted, merged with the balanced sample set, and a training dataset was constructed. The training set and validation set were then divided in a 7:3 ratio.
[0044] The soft voting ensemble learning model is trained using the training set to obtain the trained soft voting ensemble learning model.
[0045] Specifically, the optimal feature subset corresponding to each sample in the training set is input into the soft-voting ensemble learning model, which includes logistic regression, support vector machine, random forest, extreme gradient boosting, and lightweight gradient boosting. Logistic regression learns a linear discriminant boundary based on the correspondence between the optimal feature subset in the training set and the sample class. Support vector machine learns a class separating hyperplane based on the distribution relationship between the optimal feature subset in the training set and the sample class. Random forest constructs multiple decision trees based on the optimal feature subset in the training set and the sample class and completes classification learning. Extreme gradient boosting performs round-robin classification based on the optimal feature subset in the training set and the sample class. After completing the boosting learning, the lightweight gradient booster performs gradient boosting learning based on the optimal feature subset and sample class in the training set. After training the logistic regression, support vector machine, random forest, extreme gradient boosting, and lightweight gradient booster, the logistic regression, support vector machine, random forest, extreme gradient boosting, and lightweight gradient booster output the classification results for the training set samples and the predicted probabilities for the validation set samples, respectively. The soft voting ensemble learning model summarizes the training results of the logistic regression, support vector machine, random forest, extreme gradient boosting, and lightweight gradient booster to obtain the trained soft voting ensemble learning model.
[0046] The predicted probabilities and AUC values of each base learner are obtained on the validation set. Based on the AUC values of each base learner on the validation set, the predicted probabilities of each base learner are weighted and averaged to obtain the landslide susceptibility index.
[0047] Specifically, the soft-voting ensemble learning model includes five base learners: Logistic Regression (LR), Support Vector Machine (SVM), Random Forest (RF), XGBoost (XGB), and Lightweight Gradient Boosting Machine (LGBM). For example, the AUC values of each base learner on the validation set are LR-0.812, SVM-0.856, RF-0.913, XGB-0.925, and LGBM-0.921, respectively. Using the AUC value as weight, the predicted probabilities of each base learner are weighted and averaged to obtain the landslide susceptibility index. When the AUC value of the soft-voting model reaches 0.938, it outperforms any single model.
[0048] It should be noted that the set of base learners is . Each base learner for samples The landslide probability is predicted to be The AUC value on the validation set is Then calculate the landslide susceptibility index. The expression is: ; Among them, the landslide susceptibility index The range of values is , This represents the index variable of the base learner.
[0049] S5. The SHAP interpreter is used to perform interpretability analysis on the trained soft-voting ensemble learning model to generate the contribution of each influencing factor to landslide susceptibility.
[0050] The interpretability analysis of the trained soft-voting ensemble learning model using the SHAP interpreter is based on the interpretation analysis of the training dataset by the trained soft-voting ensemble learning model, and the calculation of the SHAP value corresponding to each sample in the training dataset.
[0051] It should be noted that the expression for calculating the SHAP value is: ; in, Representation of features SHAP; Indicates the impact factor index variable; Represents the set consisting of all features; Indicates that it does not contain features A subset of features; Indicates using only a subset of features The predicted value of the time model.
[0052] Based on the SHAP value, a SHAP bee diagram is generated to determine the dominant influencing factors affecting landslide development and the direction of their effects.
[0053] Specifically, the SHAP values of each influencing factor in all training samples are arranged in correspondence with the original values of each influencing factor to generate a SHAP bee diagram. Based on the SHAP bee diagram, the distribution range and concentration of the SHAP values of each influencing factor are compared to obtain the overall strength of the influence of each influencing factor on landslide susceptibility. Influencing factors that consistently show a contribution in absolute SHAP value are regarded as the dominant influencing factors affecting landslide development. Combining the distribution position of the original values of each influencing factor in the SHAP bee diagram with the positive and negative distribution of SHAP values, the direction of the dominant influencing factor is obtained. When the original values are mainly distributed in the positive SHAP value area, the corresponding dominant influencing factor has a positive effect on landslide susceptibility. When the original values are mainly distributed in the negative SHAP value area, the corresponding dominant influencing factor has a negative effect on landslide susceptibility.
[0054] A SHAP waterfall plot is generated based on the SHAP value corresponding to a single historical landslide sample to determine the dominant influencing factors affecting landslide susceptibility and the direction of their effects in a single historical landslide sample. A SHAP dependency plot is generated based on the SHAP value corresponding to continuous influencing factors to determine the variation law of the influence of continuous influencing factors on landslide susceptibility and the range of influence thresholds.
[0055] Specifically, after interpreting and analyzing the training dataset based on the trained soft-voting ensemble learning model and obtaining the SHAP values corresponding to each sample, the SHAP values of each influencing factor corresponding to a single historical landslide sample are selected. The SHAP values are then sorted according to their positive and negative contributions to the landslide susceptibility prediction results. Starting from the prediction baseline, the SHAP values corresponding to each influencing factor are stacked one by one to generate a SHAP waterfall plot. By comparing the absolute value and positive / negative direction of the SHAP values corresponding to each influencing factor, the dominant influencing factor affecting landslide susceptibility in a single historical landslide sample and the direction of their influence are obtained.
[0056] For continuous impact factors, the original values of continuous impact factors and their corresponding SHAP values are extracted from the training dataset. A mapping is performed between the original values of continuous impact factors and their corresponding SHAP values to generate a SHAP dependency graph. By observing the changing trend, inflection point, and range of SHAP values during the change of the original values of continuous impact factors, the changing pattern of the impact of continuous impact factors on landslide susceptibility and the range of impact thresholds are obtained.
[0057] Based on the average SHAP value, a feature importance bar chart was generated to determine the ranking of the contribution of each influencing factor to landslide susceptibility, and the contribution analysis results of each influencing factor to landslide susceptibility were obtained.
[0058] Specifically, for each influencing factor, the absolute value of the corresponding SHAP value is extracted from all training samples, and the average of these absolute values is calculated to obtain the average SHAP value for each influencing factor. The names of each influencing factor are then matched with their average SHAP values, and a feature importance bar chart is generated in descending order of average SHAP value. Based on the arrangement order and corresponding bar length of the influencing factors in the feature importance bar chart, the contribution ranking of each influencing factor to landslide susceptibility is obtained. The order of contribution of each influencing factor to landslide susceptibility is then output, yielding the contribution analysis results of each influencing factor to landslide susceptibility.
[0059] S6. Based on the landslide susceptibility index, the target area is divided into susceptibility zones. Based on the analysis results of the contribution of each influencing factor to landslide susceptibility, the susceptibility zone results are interpreted, and a landslide susceptibility evaluation map is generated.
[0060] The target area is divided into grids, the landslide susceptibility index of each grid is calculated, and the natural breakpoint method is used to divide the landslide susceptibility index into high-susceptibility, medium-susceptibility and low-susceptibility areas.
[0061] Specifically, the seven optimal factor values of all grids in the study area are input into the soft voting model to calculate the landslide susceptibility index of each unit. The susceptibility index is divided into three levels: high susceptibility zone (1.0-0.6), medium susceptibility zone (0.6-0.3), and low susceptibility zone (0.3-0) using the natural breakpoint method.
[0062] Based on the analysis of the contribution of each influencing factor to landslide susceptibility, the results of each susceptibility zone are interpreted, and a landslide susceptibility evaluation map is generated.
[0063] Specifically, the landslide susceptibility index, the susceptibility zone to which each grid unit belongs, and the contribution of each influencing factor to landslide susceptibility for each grid unit are extracted from the target area. The contribution of each influencing factor to landslide susceptibility for all grid units within the same susceptibility zone is summarized, and the contribution direction and magnitude of each influencing factor within the same susceptibility zone are statistically analyzed to obtain the dominant influencing factors and their directions of action for high-susceptibility zones (1.0-0.6), medium-susceptibility zones (0.6-0.3), and low-susceptibility zones (0.3-0). Influencing factors with higher contribution and positive contributions within the high-susceptibility zone are identified. The main disaster-causing factors in high-risk areas were identified, the moderately contributing factors in medium-risk areas were identified as the main control factors in medium-risk areas, and the negative or low-contribution factors in low-risk areas were identified as the limiting factors in low-risk areas. The spatial distribution results of high-risk, medium-risk, and low-risk areas were correspondingly labeled with the dominant influencing factors and their directions of action for each area. The spatial location, risk level, and contribution of each influencing factor to landslide susceptibility of each risk zone were expressed in the same landslide susceptibility evaluation map. The interpretation of each risk zone was completed and a landslide susceptibility evaluation map was generated.
[0064] In summary, this invention effectively addresses class imbalance and optimizes classification boundaries through SMOTE-Tomek hybrid sampling, and integrates the advantages of multiple models through soft voting ensemble learning. It can be widely applied to scenarios such as regional geological disaster investigation, land spatial planning, and site selection for major projects. Furthermore, by using the SHAP interpreter for local and global interpretability analysis, it reveals the contribution mechanism of various influencing factors to landslide susceptibility, transforming the black-box model into an interpretable decision support tool, enhancing the credibility of model prediction results, and providing technical support for geological disaster prevention and control.
[0065] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A landslide susceptibility assessment method that can be interpreted by combining machine learning with class-disequilibrium processing, characterized in that, include: Collect historical landslide disaster data and multi-source influencing factors in the target area; Landslide samples and non-landslide samples were constructed based on historical landslide disaster data. The SMOTE-Tomek hybrid sampling method was used to process the landslide samples and non-landslide samples to obtain a balanced training set. The optimal feature subset is obtained by screening multi-source influencing factors through a dual screening mechanism that combines random forest feature scoring and collinearity diagnosis. Based on the optimal feature subset and balanced training set, a soft-voting ensemble learning model is constructed and trained to output a landslide susceptibility index. The SHAP interpreter was used to perform interpretability analysis on the trained soft-voting ensemble learning model, generating the contribution analysis results of each influencing factor to landslide susceptibility. The target area is divided into susceptibility zones based on the landslide susceptibility index. The susceptibility zone results are then interpreted based on the contribution analysis of each influencing factor to landslide susceptibility, and a landslide susceptibility evaluation map is generated.
2. The landslide susceptibility assessment method combining interpretable machine learning with class-disequilibrium processing as described in claim 1, characterized in that, The collection of historical landslide disaster data and multi-source influencing factors in the target area includes collecting historical landslide disaster data and extracting topographic factors, geological factors, hydrological factors, vegetation factors, and human activity factors from digital elevation models, geological maps, remote sensing images, and vector map data to form multi-source influencing factors.
3. The landslide susceptibility assessment method combining interpretable machine learning with class-disequilibrium processing as described in claim 1, characterized in that, The construction of landslide and non-landslide samples refers to establishing a buffer zone with a preset range centered on the landslide point based on historical landslide disaster data, extracting grid cells within the buffer zone as landslide samples, and randomly sampling non-landslide samples in landslide-free areas.
4. The landslide susceptibility assessment method combining interpretable machine learning with class-disequilibrium processing as described in claim 3, characterized in that, The SMOTE-Tomek hybrid sampling method is used to process the landslide samples and non-landslide samples to achieve a balanced training set. The specific steps are as follows: After using SMOTE to expand the number of landslide samples to the same number as the number of non-landslide samples, a balanced dataset was obtained. Tomek Link pairs are identified in the balanced dataset, and non-slippage samples in the Tomek Link pairs are removed to obtain a balanced training set.
5. The landslide susceptibility assessment method combining interpretable machine learning with class-disequilibrium processing as described in claim 1, characterized in that, The optimal feature subset is obtained by combining random forest feature scoring with collinearity diagnosis. The random forest is used to calculate the feature importance of each influence factor in the multi-source influence factors, retain influence factors whose importance is greater than the feature importance screening threshold, calculate the Pearson correlation coefficient between the retained factors, and process the highly correlated influence factors.
6. The landslide susceptibility assessment method combining interpretable machine learning with class-disequilibrium processing as described in claim 1, characterized in that, The soft-voting ensemble learning model includes at least three base learners selected from logistic regression, support vector machine, random forest, XGBoost, and lightweight gradient boosting machine.
7. The landslide susceptibility assessment method combining interpretable machine learning with class-disequilibrium processing as described in claim 6, characterized in that, The specific steps for outputting the landslide susceptibility index are as follows: Based on the balanced sample set, the feature data corresponding to the optimal feature subset are selected to construct the training dataset, and the training dataset is divided into the training set and the validation set. The soft voting ensemble learning model is trained using the training set to obtain the trained soft voting ensemble learning model. The predicted probabilities and AUC values of each base learner are obtained on the validation set. Based on the AUC values of each base learner on the validation set, the predicted probabilities of each base learner are weighted and averaged to obtain the landslide susceptibility index.
8. The landslide susceptibility assessment method combining interpretable machine learning with class-disequilibrium processing as described in claim 1, characterized in that, The interpretation analysis of the trained soft-voting ensemble learning model using the SHAP interpreter is based on the interpretation analysis of the training dataset by the trained soft-voting ensemble learning model, and the calculation of the SHAP value corresponding to each sample in the training dataset.
9. The landslide susceptibility assessment method combining interpretable machine learning with class-disequilibrium processing as described in claim 1, characterized in that, The specific steps for generating the contribution analysis results of each influencing factor to landslide susceptibility are as follows: Based on the SHAP value, a SHAP bee diagram is generated to determine the dominant influencing factors affecting landslide development and the direction of their effects. SHAP waterfall plots are generated based on the SHAP values corresponding to individual historical landslide samples to determine the dominant influencing factors affecting landslide susceptibility and the direction of their influence in individual historical landslide samples. Based on the SHAP values corresponding to continuous impact factors, a SHAP dependency graph is generated to determine the variation law of the impact of continuous impact factors on landslide susceptibility and the impact threshold range. Based on the average SHAP value, a feature importance bar chart was generated to determine the ranking of the contribution of each influencing factor to landslide susceptibility, and the contribution analysis results of each influencing factor to landslide susceptibility were obtained.
10. The landslide susceptibility assessment method combining interpretable machine learning with class-disequilibrium processing as described in claim 9, characterized in that, The process involves dividing the target area into susceptibility zones based on the landslide susceptibility index, interpreting the susceptibility results based on the contribution analysis of each influencing factor to landslide susceptibility, and generating a landslide susceptibility evaluation map. The specific steps are as follows: The target area is divided into grids, the landslide susceptibility index of each grid is calculated, and the natural breakpoint method is used to divide the landslide susceptibility index into high susceptibility areas, medium susceptibility areas and low susceptibility areas. Based on the analysis of the contribution of each influencing factor to landslide susceptibility, the results of each susceptibility zone are interpreted, and a landslide susceptibility evaluation map is generated.