Landslide susceptibility evaluation method combining model weight and decision path
By constructing a weighted-decision dual-path fusion model, and combining historical data and conditional factors of the landslide study area, the nonlinear coupling relationship and spatial heterogeneity problems of existing landslide susceptibility assessment models are solved, achieving more accurate landslide susceptibility assessment and disaster prevention decision support.
Patent Information
- Application Number
- CN202511650085.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-12
- Publication Date
- 2026-02-06
AI Technical Summary
Existing landslide susceptibility assessment models neglect the nonlinear coupling relationship between geological parameters, resulting in a lack of physical mechanism explanation and an inability to quantify the contribution of individual features, making it difficult to reflect spatial heterogeneity.
A landslide susceptibility assessment method combining model weights and decision paths is proposed. By obtaining historical landslide catalogs and conditional factors of the landslide study area, preprocessing and feature screening are performed to construct a weight-decision dual-path fusion model. Leaf node features are extracted using random forest and trained using a TabNet classifier to generate a landslide susceptibility assessment model.
It significantly improves the model's discriminative ability, can quantify feature contribution and provide reliable disaster prevention decision-making basis, and shows excellent performance indicators such as high AUC and F1 score, which is in line with the theoretical expectation of spatial differentiation of landslide susceptibility.
Smart Images
Figure CN121479569A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of geological disaster prevention and control, specifically to a landslide susceptibility assessment method that combines model weights and decision-making paths. Background Technology
[0002] Landslides cause over 4,000 deaths and billions of dollars in economic losses globally each year. Landslide susceptibility assessment, as a crucial foundation of disaster risk prevention and control systems, directly determines the scientific rigor and timeliness of disaster prevention and mitigation decisions. Traditional landslide susceptibility assessment models (such as logistic regression and single random forest) rely solely on end-to-end training of raw data features, neglecting the nonlinear coupling relationships between geological parameters. Furthermore, the decision-making process cannot quantify the contribution of individual features, hindering the precise deployment of prevention and control measures. Therefore, obtaining interactive information between features is of paramount importance. Summary of the Invention
[0003] The purpose of this invention is to provide a landslide susceptibility assessment method that combines model weights and decision paths. This method aims to address the problems in existing technologies, such as the black-box nature of decision-making in data-driven models (e.g., random forests, neural networks), which leads to poor interpretation of results due to a lack of physical mechanisms, and the difficulty in quantifying nonlinear interactions between factors and reflecting spatial heterogeneity in physical-driven models (e.g., logistic regression).
[0004] To achieve the above objectives, this invention provides a landslide susceptibility assessment method that combines model weights and decision paths, comprising the following steps:
[0005] Step 1: Obtain the historical landslide catalog and landslide condition factors for the landslide study area, and unify their spatial resolution and geographic coordinate system;
[0006] Step 2: Preprocess the landslide condition factors by screening key factors through correlation analysis (Pearson coefficient) and multicollinearity test (VIF value) to form a standardized landslide condition factor system;
[0007] Step 3: Construct a balanced sample set using the buffer control method. Generate a buffer with historical landslide points as the center, randomly select non-landslide sample points outside the buffer with an equal number of landslide samples, and merge positive and negative samples to construct the dataset.
[0008] Step 4: Input the dataset obtained in Step 3 into the weight-decision dual-path fusion model. The decision path extracts leaf node features through random forest, and the weight path generates weighted features based on feature importance. After fusing the dual-path features, input them into the TabNet classifier for training and testing to obtain the trained susceptibility evaluation model.
[0009] Step 5: Input the landslide condition factors of the study area in Step 2 into the evaluation model trained in Step 4 for susceptibility analysis, obtain the probability of landslide occurrence for each grid point in the entire study area, and use the natural breakpoint method to divide the probability values into five levels, generating a landslide susceptibility classification map on the GIS platform.
[0010] Furthermore, the landslide condition factors mentioned in step one include geographical and geomorphological factors, geological factors, environmental factors, and human activity factors.
[0011] Furthermore, the geomorphological factors include elevation, aspect, slope, plan curvature, profile curvature, and topographic humidity index;
[0012] The geological factors include distance from the fault, lithology, and soil.
[0013] The environmental factors include distance from the river and normalized vegetation coefficient;
[0014] The human activity factors include land use type and distance from roads.
[0015] Furthermore, the raster resolution of all data is 30m×30m.
[0016] Furthermore, the landslide condition factor preprocessing in step two specifically involves: reclassifying and grading continuous factors such as elevation, aspect, slope, plane curvature, profile curvature, topographic humidity index, fault distance, river distance, normalized vegetation coefficient, and distance from road using the natural breakpoint method, and assigning standardized numerical codes to each grading interval.
[0017] For discrete factors such as land use type, lithology, and soil, unique numerical values are directly assigned to each of their subcategories.
[0018] Furthermore, the dataset constructed in step three is divided into a training subset and a test subset proportionally using a random stratified sampling strategy. The training subset accounts for 70% ± 5% of the total sample size and is used for model parameter optimization; the test subset accounts for 30% ± 5% and is used for verifying the model's generalization ability.
[0019] Furthermore, the weight-decision dual-path fusion model in step four is executed according to the following path:
[0020] Weighting path: Input the preprocessed features into the random forest model, calculate the importance weight of each feature, and perform weighted calculation on the original feature values based on the weight to generate a weighted feature matrix;
[0021] Decision path: Input the features into the random forest model, extract the leaf index features of each decision tree terminal node, and generate a decision path feature matrix after one-hot encoding and standardization.
[0022] The weighted feature matrix is concatenated with the decision path feature matrix to form a fused feature matrix, which is then input into the TabNet classifier for training and testing.
[0023] Furthermore, the five levels divided using the natural breakpoint method in step five are extremely high susceptibility zone, high susceptibility zone, medium susceptibility zone, low susceptibility zone, and extremely low susceptibility zone.
[0024] The beneficial effects of this invention are as follows:
[0025] 1. By constructing a weighted-decision dual-path fusion model, which utilizes feature importance (weighted path) and the nonlinear decision rules of the decision tree (decision path), both quantitative information after importance weighting is retained, and structured information representing complex rules is incorporated. This fusion strategy can capture richer interactive information of geological features, thereby significantly improving the model's discriminative ability.
[0026] 2. The weighted path assigns weights to features based on their importance, quantifying the contribution of each feature to the final prediction. This not only helps in understanding the model's decision-making mechanism but also provides a more reliable theoretical basis for disaster prevention decisions.
[0027] 3. The method demonstrates excellent performance metrics (high AUC and F1 score) in a case study in Cenxi City, proving its effectiveness and superiority in real-world scenarios. Attached Figure Description
[0028] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.
[0029] Figure 1 This is a flowchart of a landslide susceptibility assessment method that combines model weights and decision paths according to the present invention.
[0030] Figure 2 This is a schematic diagram of the ROC curve in an embodiment of the present invention.
[0031] Figure 3 These are landslide susceptibility partitioning maps in various embodiments of the present invention, where a is the susceptibility partitioning map output by the logistic regression (LR) model, b is the susceptibility partitioning map output by the random forest (RF) model, c is the susceptibility partitioning map output by the TabNet model, and d is the susceptibility partitioning map output by the weight-decision dual-path fusion model (RF-TabNet). Detailed Implementation
[0032] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.
[0033] The embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.
[0034] Combination Figure 1 The flowchart shown illustrates specific embodiments of the present invention in detail:
[0035] Step 1: Obtain the historical landslide catalog and landslide condition factor data for the study area, preprocess them, extract all condition factor data to historical landslide points, export the extracted historical landslide point data, and use the multicollinearity method and Pearson correlation coefficient to analyze and screen the condition factors, and establish a landslide condition factor system.
[0036] Furthermore, step one specifically includes the following steps:
[0037] Step 1.1: In this embodiment, Cenxi City, Wuzhou City, is selected as the study area. The area has well-developed folds and fault structures. The regional tectonic trend is mainly northeast, followed by east-west, northwest, and near north-south. The tectonic faults are intertwined, and the superposition and interference of old and new structures are significant. The geological activity is relatively active, making it a high-risk area for landslide geological disasters. This embodiment obtained 281 landslide points in Cenxi City, the study area, from landslide logging data of the Wuzhou City Geological Environment Monitoring Station; 30m digital elevation model (DEM) data of the study area came from the Geospatial Data Cloud (https: / / www.gscloud.cn / ); road data came from the National Geographic Information Resource Catalog Service System (https: / / www.webmap.cn / main.do?method=index); NDVI data came from the National Ecological Science Data Center (https: / / www.nesdc.org.cn / ); land use data of the study area came from the 2020 GlobeLand30 dataset (http: / / www.globallandcover.com / ); 1:200,000 topographic and geological map of the study area came from the National Geological Archives (https: / / www.ngac.cn / 125cms / c / qggnew / index.htm); soil data came from the Nanjing Institute of Soil Science, Chinese Academy of Sciences (https: / / soil.geodata.cn / data / dataresource.html).
[0038] Step 1.2: After obtaining all data, unify its spatial resolution and geographic coordinate system. In ArcGIS, the spatial resolution is unified to 30m×30m through resampling. All data are projected and transformed to the WGS_1984_UTM_Zone_49N coordinate system. Then, ArcGIS's multi-value extraction to point tool is used to extract the disaster-causing factor attribute values corresponding to all landslide locations. Finally, the extraction results are exported to form an original attribute table containing disaster-causing factor information for all historical landslide points.
[0039] Step Two: Statistical evaluation of the conditional factor attribute dataset of landslide sites was conducted based on correlation analysis (Pearson coefficient) and multicollinearity test (VIF value). Through the above analysis, significantly correlated conditional factors with low collinearity were selected and determined as the final set of landslide conditional factors. The final set of landslide conditional factors includes thirteen conditional factors: elevation, aspect, slope, plane curvature, profile curvature, topographic moisture index, fault distance, river distance, normalized vegetation coefficient, distance from road, land use type, lithology, and soil.
[0040] Step 3: Construct a balanced sample set using the buffer control method. Generate a buffer with historical landslide points as the center, randomly select non-landslide sample points outside the buffer with an equal number of landslide samples, and merge positive and negative samples to construct the dataset.
[0041] Step 4: The dataset is input into the weight-decision dual-path fusion model. The decision path extracts leaf nodes through random forest and performs one-hot encoding to obtain leaf features. ,in These are the leaf nodes obtained from training the random forest.
[0042] In the weighted path, the weighted feature is obtained by multiplying the feature importance by the feature value, and the calculation formula is:
[0043]
[0044] in, As a weighted feature, For standardized features, The weights represent the importance of the features.
[0045] Feature importance weights It is obtained from the following formula,
[0046]
[0047] in, It is the number of trees in a random forest. It is the splitting characteristic of node t. It is the reduction in impurity caused by node splitting.
[0048] After fusing the dual-path features, the model is input into the TabNet classifier for training and testing. The model performance metrics are then validated and evaluated to obtain a trained susceptibility evaluation model.
[0049] Step 5: Input the landslide condition factor data of the study area in Step 2 into the evaluation model trained in Step 4 for susceptibility analysis, obtain the landslide occurrence probability of each grid point in the entire study area, and use the natural breakpoint method to divide the probability value into five levels, and generate a landslide susceptibility classification map on the GIS platform.
[0050] Furthermore, the main performance metrics of the model include the area under the curve (AUC), accuracy, precision, recall, and F1 score.
[0051] Furthermore, AUC is the area under the receiver operating characteristic (ROC) curve. The ROC curve is a curve that shows the overall discriminative ability of the model by plotting the relationship between the True Positive Rate (TPR) and the False Positive Rate (FPR) at different classification thresholds. Its y-axis represents the True Positive Rate, which indicates the proportion of all actual positive classes that are correctly predicted as positive, and the x-axis represents the False Positive Rate, which indicates the proportion of all actual negative classes that are incorrectly predicted as positive.
[0052] Furthermore, the formulas for calculating each performance indicator are as follows:
[0053]
[0054] Wherein, TP is the number of classes that are actually positive and are predicted as positive by the model; TN is the number of classes that are actually negative and are correctly predicted as negative; FP is the number of classes that are actually negative but are incorrectly predicted as positive; and FN is the number of classes that are actually positive but are incorrectly predicted as negative by the model.
[0055] Furthermore, such as Figure 2 As shown in Table 1, in the comparison of key prediction accuracy indicators (accuracy, precision, recall, F1 score, and AUC), the weighted-decision dual-path fusion model (RF-TabNet) significantly outperforms the LR, RF, and TabNet models. This fully verifies the effectiveness and positive promoting effect of the proposed weighted-decision dual-path fusion strategy in improving the prediction accuracy of landslide susceptibility models. Combined with the statistical data in Table 2 and... Figure 3A comprehensive analysis of the spatial distribution results revealed that, in the susceptibility zoning statistics, the output of the weighted-decision dual-path fusion model (RF-TabNet) exhibited a gradient characteristic where the area / sample ratio of susceptibility zones increased sequentially from extremely low to extremely high susceptibility. This distribution pattern aligns with the theoretical expectation of spatial differentiation in landslide susceptibility, further validating the effectiveness of the proposed weighted-decision dual-path fusion strategy.
[0056] Table 1 Comparison of prediction accuracy of different evaluation models
[0057]
[0058] Table 2 Comparison of Susceptibility Results
[0059]
[0060] In summary, compared with existing technical solutions, the advantages of this invention are:
[0061] 1. By constructing a weight-decision dual-path fusion model, and utilizing feature importance (weight path) and the nonlinear decision rule of decision tree (decision path), a more profound characterization of complex geological patterns is achieved, and the discrimination ability significantly surpasses that of a single model.
[0062] 2. The weighted path assigns weights to features based on their importance, quantifying the contribution of each feature to the final prediction. This not only helps in understanding the model's decision-making mechanism but also provides a more reliable theoretical basis for disaster prevention decisions.
[0063] 3. The method demonstrates excellent performance metrics (high AUC and F1 score) in a case study in Cenxi City, proving its effectiveness.
Claims
1. A landslide susceptibility assessment method combining model weights and decision paths, characterized in that, Includes the following steps: Step 1: Obtain the historical landslide catalog and landslide condition factors for the landslide study area, and unify their spatial resolution and geographic coordinate system; Step 2: Preprocess the landslide condition factors, screen key factors through correlation analysis and multicollinearity test, and form a standardized landslide condition factor system. Step 3: Construct a balanced sample set using the buffer control method. Generate a buffer with historical landslide points as the center, randomly select non-landslide sample points outside the buffer with an equal number of landslide sample points, and merge positive and negative samples to construct the dataset. Step 4: Input the dataset obtained in Step 3 into the weight-decision dual-path fusion model. The decision path extracts leaf node features through random forest, and the weight path generates weighted features based on feature importance. After fusing the dual-path features, input them into the TabNet classifier for training and testing to obtain the trained susceptibility evaluation model. Step 5: Input the environmental factor data of all grid cells in the entire study area after the preprocessing in Step 2 into the evaluation model trained in Step 4 for susceptibility analysis, obtain the landslide occurrence probability of each grid cell in the entire study area, and use the natural breakpoint method to divide the probability value into five levels, and generate a landslide susceptibility classification map on the GIS platform.
2. The landslide susceptibility assessment method combining model weights and decision paths as described in claim 1, characterized in that, The landslide condition factors mentioned in Step 1 include geographical and geomorphological factors, geological factors, environmental factors, and human activity factors. The geographical and geomorphological factors include elevation, aspect, slope, plane curvature, profile curvature, and topographic humidity index. The geological factors include distance from faults, lithology, and soil. The environmental factors include distance from rivers and normalized vegetation coefficient. The human activity factors include land use type and distance from roads.
3. The landslide susceptibility assessment method combining model weights and decision paths as described in claim 1, characterized in that, The landslide condition factor preprocessing in step two specifically involves: reclassifying and grading continuous factors such as elevation, aspect, slope, plane curvature, profile curvature, topographic humidity index, fault distance, river distance, normalized vegetation coefficient, and distance from road using the natural breakpoint method, and assigning standardized numerical codes to each grading interval; and directly assigning unique numerical values to each subcategory of discrete factors such as land use type, lithology, and soil.
4. The landslide susceptibility assessment method combining model weights and decision paths as described in claim 1, characterized in that, The dataset constructed in step three is divided into a training subset and a test subset according to a random stratified sampling strategy. The training subset accounts for 70% ± 5% of the total sample size and is used for model parameter optimization; the test subset accounts for 30% ± 5% and is used for model generalization ability verification.
5. The landslide susceptibility assessment method combining model weights and decision paths as described in claim 1, characterized in that, The weight-decision dual-path fusion model described in step four is executed along the following path: Weighting path: Input the preprocessed features into the random forest model, calculate the importance weight of each feature, and perform weighted calculation on the original feature values based on the weight to generate a weighted feature matrix; Decision path: Input the features into the random forest model, extract the leaf index features of each decision tree terminal node, and generate a decision path feature matrix after one-hot encoding and standardization. The weighted feature matrix is concatenated with the decision path feature matrix to form a fused feature matrix, which is then input into the TabNet classifier for training and testing.
6. The landslide susceptibility assessment method combining model weights and decision paths as described in claim 1, characterized in that, In step five, the five levels divided using the natural breakpoint method are extremely high susceptibility zone, high susceptibility zone, medium susceptibility zone, low susceptibility zone, and extremely low susceptibility zone.