A landslide susceptibility intelligent evaluation method based on feature stability guidance and cost self-adaptive tightening

Abstract

The application discloses a landslide susceptibility intelligent evaluation method based on feature stability guidance and cost adaptive tightening, which comprises the following steps: collecting and normalizing landslide sample data; calculating feature importance and sorting by random forest; traversing feature subset scale by gradient boosting tree, determining global optimal subset by cross-validation AUC; calculating feature stability coefficient, generating tightening priority queue; trying to remove features in turn by asymmetric cost-sensitive AUC measurement, and accepting if the performance decline is not more than the threshold; adaptively adjusting the threshold and re-tightening after the confrontation verification, recording the historical optimal subset; retraining the model to output the landslide probability. The application realizes feature stability guidance, asymmetric cost-sensitive tightening and self-evolution threshold closed loop, reduces redundant features while maintaining accuracy, and improves generalization and interpretability. The measured AUC reaches 0.897, which is 3.2% higher than that of the traditional GBDT, and the feature is reduced by 40%.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of geological disaster prevention and control and artificial intelligence. Specifically, it relates to an intelligent evaluation method for landslide susceptibility based on feature stability guidance and cost adaptive compression, which is applicable to scenarios such as regional landslide disaster risk zoning, early warning and land spatial planning. Background Technology

[0002] Landslides are among the most frequent and destructive geological hazards globally. Scientifically and accurately assessing regional landslide susceptibility is fundamental to geological hazard risk management and reducing casualties and economic losses. Traditional landslide susceptibility assessment methods mainly fall into two categories: one is qualitative-semi-quantitative methods based on expert experience (such as the analytic hierarchy process and fuzzy comprehensive evaluation method), whose assessment results are highly subjective and difficult to promote and apply on a large scale with high precision requirements; the other is deterministic models based on physical mechanics (such as the infinite slope model and SHALSTAB), which require detailed geotechnical parameters. In practical applications, these methods are limited by the difficulty of obtaining parameters and spatial heterogeneity, making rapid assessment at the regional scale difficult.

[0003] In recent years, with the deep integration of Geographic Information System (GIS) technology and machine learning algorithms, data-driven models have gradually become the mainstream paradigm for landslide susceptibility assessment. Domestic and international scholars have successively introduced algorithms such as logistic regression, support vector machines, random forests, gradient boosting trees, and artificial neural networks into this field, achieving higher prediction accuracy than traditional methods. However, existing machine learning-based landslide susceptibility assessment methods still face the following common technical bottlenecks in practical applications: (1) The feature selection mechanism has defects. Most current studies directly use all features or filtering methods such as correlation analysis and variance thresholding for coarse screening. On the one hand, they fail to fully consider the nonlinear interaction between features, and on the other hand, they cannot adaptively determine the size of the optimal feature subset. Some scholars use evolutionary algorithms such as genetic algorithms and particle swarm optimization for feature selection, but they have inherent defects such as premature convergence, high computational cost, and hyperparameter sensitivity, and it is difficult to guarantee global optimality.

[0004] (2) Feature redundancy and overfitting are prominent issues. Landslide susceptibility assessment usually involves multiple sources of features such as topography, geology, hydrology, and human activities, resulting in high feature dimensionality and a large amount of redundant information. Directly using high-dimensional features to train the model not only leads to low computational efficiency but also easily causes overfitting, reducing the model's generalization ability in areas not involved in the training. Existing dimensionality reduction methods, such as principal component analysis, can compress feature dimensionality, but they destroy the physical interpretability of the original features, making them difficult for experts in the field of geological hazards to adopt.

[0005] (3) The evaluation indicators are out of touch with business needs. Landslide disasters are characterized by typical sample imbalance (the number of landslide samples is far less than the number of non-landslide samples), and the consequences of "missed reports" are far more serious than "false reports". However, existing models generally use symmetric evaluation indicators (such as accuracy, AUC, F1 score) in the feature selection and model training stages, failing to incorporate the cost of asymmetric misclassification into the optimization objective. This leads to the model tending to misclassify landslide samples as non-landslide samples in actual early warning, which greatly reduces the practicality of the early warning system.

[0006] (4) Lack of quantification and evaluation mechanism for feature stability. The training data for landslide susceptibility assessment models are often limited by the number of historically recorded samples, and the importance of features fluctuates greatly under different training data divisions. Some "pseudo-important" features only show high importance on specific training folds, and their performance drops sharply after changing folds, seriously weakening the robustness and cross-regional transferability of the model. At present, there are no effective means to identify and remove such features.

[0007] (5) Key hyperparameters are fixed and lack adaptive capability. Hyperparameters such as tolerance threshold and penalty coefficient in existing feature selection methods are usually preset manually and cannot be dynamically adjusted according to the data distribution characteristics, resulting in drastic fluctuations in the generalization performance of the model in different regions and sample sizes.

[0008] In summary, there is an urgent need for an intelligent assessment method for landslide susceptibility that can globally search for the optimal feature subset, quantify feature stability, integrate asymmetric cost evaluation, and have adaptive threshold adjustment capabilities, in order to break through the existing technical bottlenecks and enhance the practical value of the model in complex geological environments. Summary of the Invention

[0009] I. Technical Issues

[0010] This invention aims to overcome the technical shortcomings of existing landslide susceptibility assessment methods in terms of global feature selection, feature redundancy removal, asymmetric cost adaptation, and hyperparameter adaptive adjustment. It provides an intelligent landslide susceptibility assessment method based on feature stability guidance and cost adaptive compression, which achieves global optimal search of high-dimensional feature subsets, stability-aware precise compression, asymmetric cost-driven assessment, and threshold self-evolution driven by adversarial verification. This significantly improves the model's generalization ability and interpretability while ensuring assessment accuracy.

[0011] II. Technical Solution

[0012] To achieve the above objectives, the present invention provides the following technical solution: S1. Data Acquisition and Preprocessing S101. Obtain historical landslide disaster data for the target area and determine positive landslide sample units; use an equal number of non-landslide units as negative samples to construct a balanced sample set.

[0013] S102. Extract multi-source geographic information feature factors that affect landslide occurrence, including but not limited to: elevation, slope, aspect, plane curvature, profile curvature, topographic humidity index, distance from fault, distance from river, distance from road, rainfall, normalized vegetation index, lithology, land use type, vertical vibration frequency, and curvature.

[0014] S103. Perform Min-Max normalization on continuous characteristic factors. The calculation formula is as follows:

[0015] One-hot encoding is used for categorical feature factors; the processed feature matrix is ​​combined with landslide labels (landslide = 1, non-landslide = 0) to form the training sample set D. train With test sample set D test .

[0016] S2. Ranking of global feature importance

[0017] S201. Use random forest as the feature importance evaluator, and set the number of decision trees n. trees =200, the random seed is fixed at 42, in the training set D train Training will be conducted on [the platform].

[0018] S202. Calculate each feature f i The Ginii importance values ​​are calculated and sorted in descending order to generate a global feature importance sequence: R={f (1) ,f (2) ,…,f (m)} Where m is the total number of features, f (1) This is the most important feature.

[0019] S3. Global Optimal Feature Subset Search

[0020] S301. Use gradient boosting decision trees as the base learners, with fixed hyperparameters: number of trees nestimators=400, maximum depth max_depth=7, learning rate=0.01, subsample ratio subsample=0.85, minimum number of leaf node samples min_samples_leaf=3, maximum feature ratio max_features=0.8, and random seed random_state=42.

[0021] S302. For k=1,2,…,m, select the first k features of the global feature importance sequence to form a subset S.k ={f (1) ,f (2) ,…,f (k)}, 5-fold cross-validation is used to compute subset S on the training set. k The corresponding average ROC-AUC value:

[0022] S303. Determine the globally optimal number of features and the corresponding optimal subset :

[0023] S4. Characteristic Stability Self-Evaluation and Tightening Priority Construction

[0024] S401. For the optimal subset For each feature f in the dataset, within a 5-fold cross-validation framework, calculate the mean μ of the validation set AUC when that feature is included. f With standard deviation σ f Define the characteristic stability coefficient:

[0025] The closer the coefficient is to 1, the more stable the performance of the feature is under different training data partitions.

[0026] S402. [The following appears to be a separate, unrelated sentence:] Will... The features are sorted in ascending order of stability coefficient, generating a compact priority queue:

[0027] Prioritize removing features with the worst stability (i.e., the greatest volatility).

[0028] S5. Asymmetric Cost-Sensitive Contraction

[0029] S501. Set the initial tolerance threshold δ0 = 0.001, the cost weights for positive and negative samples are w1 = 0.8 and w0 = 0.2 respectively, and the current feature subset S = Current subset performance .

[0030] S502. For each feature f in queue Q, perform a compaction attempt in sequence: Constructing candidate subsets ; Five-fold cross-validation is used on the candidate subset S′, and the mean csAUC is calculated using cost-sensitive AUC as the evaluation metric. candThe cost-sensitive AUC is calculated as follows: In each training fold, landslide samples are assigned weight w1, and non-landslide samples are assigned weight w0 to train the base learner; during validation, the AUC is also calculated using a sample weighting method, and the weighted AUC is defined as:

[0031] Where P and N are the positive and negative sample sets in the validation set, respectively, p i Let be the probability that the model predicts sample i as positive, 1( ) is an indicator function.

[0032] S503. Determine whether to accept removal:

[0033] If removal is accepted, update S←S′, csAUC. cur ←csAUC cand Otherwise, retain the current feature and continue trying the next feature in the queue.

[0034] S6. Self-evolutionary threshold closed loop and historical optimum backtracking

[0035] S601. The feature subset S finally obtained in step S5 final Perform adversarial validation: Label training set samples as 0 and test set samples as 1, merge them to form the adversarial validation set; use a random forest as a binary classifier to train it to distinguish the source of samples, and calculate the adversarial AUC:

[0036] in For sample source labeling, Predict probabilities for the classifier.

[0037] S602. The preset adversarial threshold θ = 0.7. If AUCadv > θ, it is determined that the current feature subset has a significant distribution shift, and the compaction threshold needs to be adaptively adjusted.

[0038] Reset S= Then proceed to step S502 to re-execute the compression.

[0039] S603. Record the tolerance threshold δ, cost-sensitive AUC csAUC, and performance gain Δ=csAUC- for each iteration. Counteracting AUC adv When the iteration terminates (i.e., AUC...) adv When ≤θ and there are no more features to remove, from all features satisfying AUC advAmong the records ≤ θ, the feature subset corresponding to the record with the highest cost-sensitive AUC is selected as the historical optimal subset S. best .

[0040] S7. Final Model Training

[0041] S701. In the historical best feature subset S best The gradient boosting decision tree is retrained using the same hyperparameters as in step S301 to obtain the intelligent evaluation model M for landslide susceptibility. final .

[0042] S702. The model output is the probability of landslide occurrence p∈[0,1]. The larger the p value, the higher the landslide susceptibility of the unit.

[0043] S8. Regional landslide susceptibility assessment

[0044] S801. Divide the area to be evaluated into regular grid units, extract the multi-source geographic information feature factors corresponding to each unit, and perform normalization processing according to step S103.

[0045] S802. Input the preprocessed feature matrix into the trained model M final Calculate the landslide susceptibility probability index p for each unit.

[0046] S803. Based on the natural discontinuity method or the equal interval method, the probability values ​​are divided into five susceptibility levels: extremely low, low, medium, high, and extremely high. A landslide susceptibility zoning map is generated to complete the regional intelligent evaluation.

[0047] III. Beneficial Effects

[0048] Compared with the prior art, the present invention has the following significant advantages: (1) Global Optimal Feature Subset Deterministic Search: This invention abandons the traditional forward selection or early stopping strategy, and determines the global optimal number of features by traversing all feature subset sizes and based on cross-validation AUC, fundamentally avoiding the local optimum trap, and the computational complexity is only O. (m) Multiple cross-validations are perfectly acceptable in engineering.

[0049] (2) Feature stability quantification and precise elimination: The feature stability coefficient Stab is proposed for the first time. (f) This method quantifies the performance fluctuations of features across different training folds into an index ranging from 0 to 1, and uses this as the basis for prioritizing feature compression. This approach effectively identifies and eliminates "pseudo-important" features, significantly improving the robustness and interpretability of the model.

[0050] (3) Asymmetric cost-driven compression: In view of the actual need that the cost of underreporting is much higher than that of false alarms in the landslide susceptibility assessment, cost-sensitive AUC is introduced as the core evaluation index in the feature compression stage, and a minimum tolerance threshold δ0=0.001 is set so that the model can prioritize the retention of features that contribute greatly to the identification of landslide samples without losing performance, thus greatly improving the recall rate.

[0051] (4) Self-evolving threshold closure and historical best backtracking: Adversarial verification technology is introduced into the post-feature selection evaluation, and for the first time, the compaction threshold is dynamically and adaptively adjusted according to the data distribution characteristics. At the same time, the historical best backtracking mechanism ensures that the model performance does not degrade in multiple iterations, effectively balancing model complexity and generalization ability.

[0052] (5) Feature selection and model training are completely decoupled: The present invention uses the same base learner hyperparameters in the three stages of global search, compaction optimization and final training, ensuring that the performance improvement comes entirely from feature subset optimization rather than hyperparameter adjustment. The technical contribution is clear and highly reproducible.

[0053] (6) Excellent performance in actual tests: Experiments on a landslide dataset in a watershed in the southwestern mountainous area show that the AUC of the method of this invention on the test set reaches 0.897, which is 3.2% higher than that of the traditional GBDT. The number of features is reduced from 15 to 9 (a reduction of 40%), and the recall rate is increased by 6.8%. It is significantly better than the existing mainstream models and similar patented methods, and has extremely high engineering application value. Attached Figure Description

[0054] Figure 1 This is the overall flowchart of the present invention; Figure 2 This is a schematic diagram of the ROC curve. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0056] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0057] It should be noted that, unless otherwise specified, the embodiments and features described in this invention can be combined with each other.

[0058] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0059] In this invention, unless otherwise explicitly specified and limited, "above" or "below" the second feature can include direct contact between the first and second features, or contact between the first and second features through another feature between them. Furthermore, "above," "over," and "on top" of the second feature includes the first feature directly above or diagonally above the second feature, or simply indicates that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature includes the first feature directly below or diagonally below the second feature, or simply indicates that the first feature is at a lower horizontal level than the second feature.

[0060] It should be noted that, unless otherwise specified, the embodiments and features described in this invention can be combined with each other.

[0061] Example

[0062] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. Those skilled in the art should understand that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. S1. Data Acquisition and Preprocessing S101. This embodiment selects a county in Jiangxi Province as the research demonstration area. Located in the low-to-medium mountainous region of central and southern Jiangxi, the county has significant topographic relief, complex geological structures, abundant rainfall, and intense human engineering activities, making it a typical landslide-prone area. Through high-resolution remote sensing image interpretation, combined with geological hazard surveys and on-site verification, a total of 252 historical landslide disaster points were identified in the study area. To enhance sample representativeness and reduce the influence of spatial autocorrelation, a 30m radius buffer zone was established centered on each landslide point, and non-landslide raster units within the buffer zone were used as candidate positive samples. After deduplication and spatial filtering, a total of 756 positive sample units were obtained. Using a random sampling method, an equal number of non-landslide units were randomly selected from the non-landslide areas of the study area as negative samples, ultimately constructing a balanced dataset containing 1512 samples. These were randomly divided into a training set (1058 samples) and a test set (454 samples) at a 7:3 ratio, with the positive-to-negative sample ratio in the training set maintained at 1:1.

[0063] S102. Extract multi-source geographic information characteristic factors that affect landslide occurrence. Based on the geological environmental conditions and data availability of the study area, select 15 landslide susceptibility evaluation characteristic factors, covering five categories: topography, geological structure, hydrological environment, human engineering activities, and vegetation cover. These include, but are not limited to: elevation, slope, aspect, plane curvature, profile curvature, topographic humidity index, distance from fault, distance from river, distance from road, rainfall, normalized difference vegetation index, lithology, land use type, vertical vibration frequency, and curvature.

[0064] S103. Perform Min-Max normalization on the 15 continuous characteristic factors. The calculation formula is as follows:

[0065] One-hot encoding is used for categorical feature factors; the processed feature matrix is ​​combined with landslide labels (landslide = 1, non-landslide = 0) to form the training sample set D. train With test sample set D test Finally, the feature matrix X∈R is constructed. 1512×15 The landslide label vector y∈{0,1} 1512 (1 indicates a landslide, 0 indicates a non-landslide).

[0066] S2. Ranking of global feature importance

[0067] S201. Use random forest as the feature importance evaluator, and set the number of decision trees n. trees =200, the random seed is fixed at 42, in the training set D train Training will be conducted on [the platform].

[0068] S202. Calculate each feature f i The Ginii importance values ​​are calculated and sorted in descending order to generate a global feature importance sequence: R={f (1) ,f (2) ,…,f (m)} Where m is the total number of features, f (1) This is the most important feature.

[0069] S3. Global Optimal Feature Subset Search

[0070] S301. Use gradient boosting decision trees as the base learners, with fixed hyperparameters: number of trees nestimators=400, maximum depth max_depth=7, learning rate=0.01, subsample ratio subsample=0.85, minimum number of leaf node samples min_samples_leaf=3, maximum feature ratio max_features=0.8, and random seed random_state=42.

[0071] S302. For k=1,2,…,15, select the top k features of the global feature importance sequence to form a subset S. k ={f (1) ,f (2) ,…,f (k)}, 5-fold cross-validation is used to compute subset S on the training set. k The corresponding average ROC-AUC value:

[0072] S303. Determine the globally optimal number of features and the corresponding optimal subset :

[0073] S4. Characteristic Stability Self-Evaluation and Tightening Priority Construction

[0074] S401. For the optimal subset For each feature f in the dataset, within a 5-fold cross-validation framework, calculate the mean μ of the validation set AUC when that feature is included. f With standard deviation σ f Define the characteristic stability coefficient:

[0075] The closer the coefficient is to 1, the more stable the performance of the feature is under different training data partitions.

[0076] S402. [The following appears to be a separate, unrelated sentence:] Will... The features are sorted in ascending order of stability coefficient, generating a compact priority queue:

[0077] Q=[NDVI(0.9486),VFT(0.9538),curvature(0.9562),lithology(0.9561),profilecurvature(0.9566),otherfeasors(0.9585)], prioritize removing features with the worst stability (i.e. the largest fluctuation).

[0078] S5. Asymmetric Cost-Sensitive Contraction

[0079] S501. Set the initial tolerance threshold δ0 = 0.001, the cost weights for positive and negative samples are w1 = 0.8 and w0 = 0.2 respectively, and the current feature subset S = Current subset performance .

[0080] S502. For each feature f in queue Q, perform a compaction attempt in sequence: Constructing candidate subsets ; Five-fold cross-validation is used on the candidate subset S′, and the mean csAUC is calculated using cost-sensitive AUC as the evaluation metric. cand The cost-sensitive AUC is calculated as follows: In each training fold, landslide samples are assigned weight w1, and non-landslide samples are assigned weight w0 to train the base learner; during validation, the AUC is also calculated using a sample weighting method, and the weighted AUC is defined as:

[0081] Where P and N are the positive and negative sample sets in the validation set, respectively, p i Let be the probability that the model predicts sample i as positive, 1( ) is an indicator function.

[0082] S503. Determine whether to accept removal:

[0083] If removal is accepted, update S←S′, csAUC. cur ←csAUC cand Otherwise, retain the current feature and continue trying the next feature in the queue. Taking the removal of NDVI as an example: the calculated csAUC is... cand =0.8333, a decrease of 0.0026 from the current value >0.001, therefore removal is rejected.

[0084] S6. Self-evolutionary threshold closed loop and historical optimum backtracking

[0085] S601. The feature subset S finally obtained in step S5 final Perform adversarial validation: Label training set samples as 0 and test set samples as 1, merge them to form the adversarial validation set; use a random forest as a binary classifier to train it to distinguish the source of samples, and calculate the adversarial AUC:

[0086] in For sample source labeling, Predict probabilities for the classifier.

[0087] S602. The preset adversarial threshold θ = 0.7. If AUCadv > θ, it is determined that the current feature subset has a significant distribution shift, and the compaction threshold needs to be adaptively adjusted.

[0088] Reset S= Then proceed to step S502 to re-execute the compression.

[0089] S603. Record the tolerance threshold δ, cost-sensitive AUC csAUC, and performance gain Δ=csAUC- for each iteration. Counteracting AUC adv When the iteration terminates (i.e., AUC...) adv When ≤θ and there are no more features to remove, from all features satisfying AUC adv Among the records ≤ θ, the feature subset corresponding to the record with the highest cost-sensitive AUC is selected as the historical optimal subset S. best Record the information for this iteration: δ=0.001, csAUC=0.8671, gain Δ=+0.0312, AUC adv =0.5005. Since the termination condition is met in only one iteration, this feature subset is directly used as the historical best subset S. best

[0090] S7. Final Model Training

[0091] S701. In the historical best feature subset S best The above (9 features) are used to retrain the gradient boosting decision tree with the exact same hyperparameters as in step S301, resulting in the intelligent landslide susceptibility assessment model M. final .

[0092] S702. The model predicts the landslide occurrence probability p∈[0,1] for 454 samples in the test set. The larger the p value, the higher the landslide susceptibility of the unit.

[0093] S8. Regional landslide susceptibility assessment

[0094] S801. Divide the area to be evaluated into regular grid units, extract the multi-source geographic information feature factors corresponding to each unit, and perform normalization processing according to step S103.

[0095] S802. Input the preprocessed feature matrix into the trained model M final Calculate the landslide susceptibility probability index p for each unit.

[0096] S803. Based on the natural discontinuity method or the equal interval method, the probability values ​​are divided into five susceptibility levels: extremely low, low, medium, high, and extremely high, generating a landslide susceptibility zoning map and completing the regional intelligent assessment. To verify the advancement of the method of this invention, three benchmark models—traditional GBDT, XGBoost, and LightGBM—are compared on the same test set. The hyperparameters of each model are independently optimized to their optimal state. ROC curve comparison is shown below. Figure 2 As shown.

Claims

1. A landslide susceptibility intelligent evaluation method based on feature stability guidance and cost adaptive contraction, characterized in that, Includes the following steps: S1. Obtain historical sample data of landslide disasters in the target area, extract multi-source geographic information feature factors, and construct labeled training sample set and test sample set after normalization processing; S2. Calculate the Gini importance of each feature factor using the random forest algorithm, and generate a global feature importance sequence R = {f1, f2, …, f} in descending order. m }, where m is the total number of features; S3. Using the gradient boosting decision tree as the base learner, traverse the subsets consisting of the first k features of the global feature importance sequence, k = 1, 2, ..., m. Calculate the mean ROC-AUC of each subset through v-fold cross-validation. Record the number of features corresponding to the maximum AUC value as the global optimal feature number K, and the corresponding feature subset as S. S4. For the optimal feature subset S For each feature f in the model, calculate its stability coefficient Stab(f) = 1 - σ. f / μ f , where μ f σ f These are the mean and standard deviation of the cross-validation AUC when the feature is included; the subset S The features in the sequence are arranged in ascending order of stability coefficient, forming a compact priority queue Q; S5. Set an initial tolerance threshold δ0 = 0.001, and attempt to remove features from the priority queue Q sequentially; For each candidate subset after removal, cross-validation is performed using cost-sensitive AUC as the performance metric. The cost-sensitive AUC is calculated by weighting the samples, with positive sample weight w1 = 0.8 and negative sample weight w0 = 0.

2. If the decrease in cost-sensitive AUC after removal is less than or equal to the current tolerance threshold δ, then removal is accepted and the current subset is updated; otherwise, removal is rejected. S6. Perform adversarial verification on the feature subset finally obtained in step S5: Construct a random forest binary classifier to distinguish between training set and test set samples, and calculate the adversarial AUC; if the adversarial AUC > the preset threshold θ = 0.7, then increase the current tolerance threshold δ by a factor of 1.2, and use the globally optimal subset S obtained in step S3. Starting from the point of view, jump to step S5 and re-execute the compaction process. At the same time, record the tolerance threshold, cost-sensitive AUC, performance gain and adversarial AUC for each iteration. When the iteration ends, select the feature subset corresponding to the highest cost-sensitive AUC from all records with adversarial AUC≤θ as the historical best subset. S7. On the historical optimal feature subset, retrain the gradient boosting decision tree using the same hyperparameters as in step S3 to obtain the landslide susceptibility intelligent evaluation model; S8. Input the corresponding characteristic factors of the area to be evaluated into the evaluation model, output the landslide susceptibility probability index, and complete the intelligent evaluation of the area's landslide susceptibility.

2. The intelligent assessment method for landslide susceptibility according to claim 1, characterized in that, The multi-source geographic information feature factors mentioned in step S1 include one or more of the following: elevation, slope, aspect, curvature, stratigraphic lithology, fault distance, water system distance, road distance, land use type, normalized vegetation index, and rainfall intensity.

3. The intelligent assessment method for landslide susceptibility according to claim 1, characterized in that, The hyperparameters of the gradient boosting decision tree described in steps S3 and S7 are completely identical, specifically including: number of trees n_estimators = 400, maximum depth max_depth = 7, learning rate learning_rate = 0.01, subsample ratio subsample = 0.85, minimum number of leaf node samples min_samples_leaf = 3, maximum feature ratio max_features = 0.8, and random seed random_state = 42.

4. The intelligent assessment method for landslide susceptibility according to claim 1, characterized in that, The stability coefficient Stab(f) mentioned in step S4 has a value range of [0,1]. The closer its value is to 1, the more stable the performance of the feature is under different training data partitions.

5. The intelligent evaluation method for landslide susceptibility according to claim 1, characterized in that, The initial tolerance threshold δ0 = 0.001 mentioned in step S5 is determined by grid search to be the optimal value for maximizing feature compression efficiency without significantly degrading model performance.

6. The intelligent evaluation method for landslide susceptibility according to claim 1, characterized in that, In step S6, the adversarial AUC threshold θ = 0.

7. When the adversarial AUC ≤ 0.7, it is considered that the distribution offset between the training set and the test set on the current feature subset is within an acceptable range. The 1.2-fold increase ratio has been experimentally verified to quickly converge to a stable feature subset within 2 to 3 iterations.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the landslide susceptibility intelligent evaluation method according to any one of claims 1 to 6.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the landslide susceptibility intelligent evaluation method according to any one of claims 1 to 6.

Images