A multi-view fusion landslide susceptibility robust evaluation method

Abstract

The present application relates to the technical field of geological disaster prediction, and more particularly to a multi-view fusion landslide susceptibility robust evaluation method. The steps are as follows: obtaining landslide influencing factor data and preprocessing; constructing landslide susceptibility prediction models from global, local and interactive perspectives; the global perspective model captures the overall pattern using all features, the local perspective model focuses on the top-ranked features to enhance robustness, and the interactive perspective model captures the synergistic effect between factors by constructing feature interaction terms; the fusion model is used to adaptively weight and integrate the prediction results of the three perspectives to generate the final susceptibility evaluation result. Through multi-view complementary analysis, the present application effectively overcomes the limitations of single model perspective and sensitivity to noise, significantly improves the accuracy, robustness and interpretability of landslide susceptibility evaluation, and provides reliable technical support for geological disaster risk prevention and control.

Description

Technical Field

[0001] This invention relates to the field of geological disaster prediction technology, and in particular to a robust evaluation method for landslide susceptibility based on multi-perspective fusion. Background Technology

[0002] Landslide susceptibility assessment is a core component of geological hazard risk assessment, aiming to identify potential future landslide locations at a regional scale. Traditional assessment methods are mostly based on deterministic models (such as infinite slope models) or statistical models (such as logistic regression and information content methods). These methods heavily rely on experts' prior knowledge and simplified assumptions about physical mechanisms, and have significant limitations when dealing with complex and nonlinear geological environmental systems.

[0003] In recent years, with the development of Geographic Information Systems (GIS) and machine learning technologies, various data-driven intelligent models have been widely used in landslide susceptibility modeling, such as support vector machines, random forests, gradient boosting decision trees, and neural networks. These methods can automatically learn the complex nonlinear relationships between a large number of environmental factors (such as elevation, slope, lithology, rainfall, and human activities) and landslide occurrence, and have shown superior predictive performance compared to traditional methods in many cases.

[0004] However, existing evaluation methods based on single machine learning models still face several challenges: First, insufficient model robustness. The fixed structure and assumptions of a single model make it susceptible to noise, outliers, or sample distribution biases in the training data, leading to unstable generalization ability of the evaluation results under different regions or data conditions. Second, limited perspective and information utilization. Most models indiscriminately input all features, failing to consciously separate and utilize the different levels of information contained within the features, such as global macroscopic laws, the local dominance of key driving factors, and the nonlinear interactive synergistic effects between different factors. Third, the balance between model interpretability and reliability. Simple models offer strong interpretability but may have limited performance, while complex ensemble models offer superior performance but their decision-making process is like a "black box," making it difficult for experts in the field of geological hazards to fully trust their results, thus hindering guidance for practical disaster prevention work.

[0005] Therefore, there is an urgent need for a new technical solution that can systematically integrate different analytical perspectives to improve the accuracy of landslide susceptibility prediction while enhancing the robustness of the model and the interpretability of the results, thereby providing more reliable and credible decision support for geological disaster risk management. Summary of the Invention

[0006] Therefore, it is necessary to provide a multi-perspective fusion robust assessment method for landslide susceptibility to address the aforementioned technical problems, thereby solving the shortcomings of existing landslide susceptibility assessment methods, such as single model perspective, insufficient robustness, and incomplete characterization of complex feature relationships.

[0007] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0008] A robust assessment method for landslide susceptibility based on multi-perspective fusion, comprising the following steps:

[0009] S1: Obtain geographical, geological, environmental, and historical landslide data for the target area; construct a feature set of landslide influencing factors and landslide occurrence labels; process missing values ​​and standardize the data; and divide the data into training and testing sets according to a preset ratio.

[0010] S2: Using all features in the training set, train the first machine learning model to learn the overall mapping relationship between landslide influencing factors and landslide occurrence from a global perspective, and obtain global prediction results.

[0011] S3: Calculate and rank the importance of each feature in the training set, and select the top K features by importance to form a local feature subset; train a second machine learning model based on the local feature subset to learn the relationship between key factors and landslide occurrence from a local focus perspective, and obtain local perspective prediction results;

[0012] S4: Based on the features of the training set, automatically construct interaction terms between features to form an interaction feature set; train a third machine learning model based on the interaction feature set to learn the relationship between the synergistic effect between factors and the occurrence of landslides from the perspective of feature interaction, and obtain the prediction results from the interaction perspective.

[0013] S5: Using the global perspective prediction results, local perspective prediction results, and interactive perspective prediction results as input features, train the fusion model, and let the fusion model learn the optimal combination weights of the prediction results from each perspective to generate the final comprehensive prediction model for landslide susceptibility.

[0014] S6: Input the feature data of the area to be evaluated into the trained comprehensive prediction model, and output the landslide susceptibility probability value or risk level of the area; based on the probability value or risk level, generate a landslide susceptibility zoning map.

[0015] As a preferred embodiment of the multi-perspective fusion robust evaluation method for landslide susceptibility provided by the present invention, in step S2, the first machine learning model is a gradient boosting decision tree model.

[0016] As a preferred embodiment of the multi-perspective fusion robust evaluation method for landslide susceptibility provided by the present invention, in step S3, the second machine learning model is an efficient and lightweight model based on the gradient boosting framework.

[0017] As a preferred embodiment of the multi-perspective fusion robust evaluation method for landslide susceptibility provided by the present invention, in step S3, the feature importance is calculated by a random forest model.

[0018] As a preferred embodiment of the multi-perspective fusion robust evaluation method for landslide susceptibility provided by the present invention, in step S4, the method for constructing feature interaction terms includes, but is not limited to, multiplication, addition, or combination operations between features.

[0019] As a preferred embodiment of the multi-perspective fusion robust evaluation method for landslide susceptibility provided by the present invention, in step S4, the third machine learning model is an ensemble learning model.

[0020] As a preferred embodiment of the multi-perspective fusion robust evaluation method for landslide susceptibility provided by the present invention, in step S5, the fusion model is a logistic regression model.

[0021] A multi-perspective integrated robust assessment system for landslide susceptibility includes:

[0022] The data preprocessing module is used to acquire and process landslide impact factor data, including missing value imputation, standardization, and dataset partitioning.

[0023] The global perspective modeling module is used to train and apply a global perspective prediction model based on all features;

[0024] The local perspective modeling module is used to train and apply a local perspective prediction model based on a feature subset selected by importance.

[0025] The interactive perspective modeling module is used to train and apply an interactive perspective prediction model based on the constructed feature interaction items.

[0026] The multi-view fusion module receives the prediction outputs from the global view modeling module, the local view modeling module, and the interactive view modeling module, and generates the final susceptibility evaluation result through the fusion model.

[0027] The results output and visualization module is used to output the probability of landslide susceptibility, risk level, and generate susceptibility zoning maps.

[0028] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the methods described herein.

[0029] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, performs the steps of any of the methods described.

[0030] It is clear without a doubt that the technical solution described above in this application can solve the technical problem that this application aims to address.

[0031] Meanwhile, through the above technical solutions, the present invention has at least the following beneficial effects:

[0032] 1. The present invention provides a multi-perspective fusion robust evaluation method for landslide susceptibility. Through complementary analysis from multiple perspectives, it effectively overcomes the shortcomings of single models, such as limited perspective and sensitivity to noise, and significantly improves the accuracy, robustness and interpretability of landslide susceptibility evaluation, providing reliable technical support for geological disaster risk prevention and control.

[0033] 2. This invention is a robust evaluation method that integrates three complementary perspectives: global, local, and interactive. By integrating the advantages of different perspective models, it achieves a more comprehensive simulation of landslide occurrence mechanisms, and ultimately outputs susceptibility evaluation results with higher accuracy, stronger stability, and greater interpretability, providing a reliable tool for the refined management of geological disasters.

[0034] 3. The present invention focuses on key features from a local perspective, reducing redundancy and noise interference; the multi-view fusion mechanism reduces the risk of unstable results due to overfitting of a single model or sensitivity to specific data distributions, and improves the generalization ability in different geographical environments.

[0035] 4. This invention automates the entire process from feature processing and interaction construction to model training and fusion, reducing excessive reliance on human experience and improving evaluation efficiency.

[0036] 5. The model structure of this invention has clear physical and statistical significance (overall laws, key factors, factor interactions). The output feature importance and perspective fusion weights can provide geological experts with intuitive basis for understanding model decision-making, thus enhancing the credibility of the results. Attached Figure Description

[0037] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0038] Figure 1 This is a flowchart of the present invention;

[0039] Figure 2 This is a schematic diagram of the ROC curve of the present invention. Detailed Implementation

[0040] 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. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0041] To enable those skilled in the art to better understand the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0042] It should be noted that, unless otherwise specified, the embodiments and features and technical solutions in the present invention can be combined with each other.

[0043] 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.

[0044] Example 1

[0045] Reference Figures 1-2 A robust assessment method for landslide susceptibility based on multi-perspective fusion is proposed, with the following steps:

[0046] S1: Data Acquisition and Standardization Processing

[0047] S101: Landslide logging and sample point generation.

[0048] Through high-resolution remote sensing image interpretation (such as GF-2 satellite imagery) and verification through three years of field investigation, 203 historical landslide disaster sites were accurately identified and delineated within Cili County. To address the requirement of combining landslide areal features with point-based model input, and to avoid the problem of insufficient representativeness of single pixels, a buffer circular area with a radius of 30 meters was generated, centered on the geometric center of each landslide body. This buffer effectively covers the main landslide area, and sampling points were systematically deployed within this buffer, ultimately generating 609 positive sample points representing "landslide occurrence". To ensure sample balance, in areas where no landslides had occurred and where geological and geographical conditions were representative (such as flat, intact bedrock areas and stable floodplains), a spatially constrained random sampling method was used to select an equal number of 609 negative sample points. Thus, a model training and validation dataset (X, Y) with a total of 1218 sample points was constructed, with each sample point associated with a binary label (Y: landslide = 1, non-landslide = 0).

[0049] In this embodiment, multi-source geospatial data of the target area are collected, including digital elevation model (DEM), geological map, land use map, rainfall data, seismic activity data, and historical landslide logging map.

[0050] S102: Landslide Influencing Factor Extraction and Database Construction. Based on multi-source data including digital elevation model (DEM, 12.5m resolution), geological map (1:200,000), land use dataset, and annual precipitation interpolation map, 16 environmental factors closely related to landslide development were extracted in ArcGIS and Python environments, forming an initial feature set. Mainly includes:

[0051] Topographic factors: elevation, slope, aspect, plane curvature, profile curvature, topographic relief, and topographic humidity index.

[0052] Geological condition factors: stratigraphic lithology (coded as a categorical variable), distance from fault.

[0053] Hydrological factors: distance from the river, multi-year average rainfall.

[0054] Land cover and human activity factors: Normalized Difference Vegetation Index (NDVI), land use type, and distance from roads.

[0055] right Missing values ​​are checked, and numerical features are imputed using the median. All numerical features are normalized and scaled to the [0,1] interval, as shown in the following formula:

[0056] ;

[0057] in, The value of the j-th feature of the i-th sample after normalization. Let jj be the feature value of the i-th sample. and Let X be the minimum and maximum values ​​of feature j, respectively. Let X be the processed feature set.

[0058] S103: Data preprocessing and partitioning.

[0059] For continuous factors, missing values ​​were checked (no missing values ​​were found in this case), and normalization was performed using Min-Max standardization. Categorical variables (such as lithology) were uniquely coded. The processed standardized feature set is denoted as X. Stratified sampling was used to randomly divide the 1218 samples into a training set of 852 samples and an independent test set of 366 samples in a 7:3 ratio. , This ensures that the ratio of positive to negative samples remains consistent in the training and test sets (approximately 1:1).

[0060] S2: Global Perspective Model Construction and Prediction

[0061] S201: Model Selection and Training. Configure the XGBoost model using all 16 features of the training set. The parameters are:

[0062] n_estimators=150,max_depth=6,learning_rate=0.05.

[0063] The model is trained on a training set to capture the complex, non-linear global correlation patterns between all factors and landslides. Extreme Gradient Boosting (XGBoost) is employed as the global perspective model. The model minimizes the following regularized objective function through additive training:

[0064] ;

[0065] In the formula, Let be the objective function value at iteration t; n be the total number of training samples; i be the sample index; l(⋅,⋅) be the loss function used to measure the difference between the true value and the predicted value (such as log loss); The true label for the i-th sample is (landslide occurred = 1, no landslide occurred = 0). For the front The predicted probability of the i-th sample by the round ensemble model; For the first The newly added decision tree for the samples The predicted contribution value; This is a regularization term used to control the... The complexity of the tree structure is reduced to prevent overfitting. For the first The feature vector of each sample.

[0066] in, Complexity control terms, where T is the number of leaf nodes, w is the leaf weight, and γ and λ are hyperparameters. In the training set ( , ) on Train it and optimize its parameters.

[0067] S202: Global perspective prediction. After training, use... Predictions are made on both the training and test sets to obtain the probability of landslide occurrence from a global perspective:

[0068] ;

[0069] in, For global perspective models (such as XGBoost).

[0070] S3: Local Perspective Model Construction and Prediction

[0071] S301: Feature Importance Evaluation and Selection. A random forest model is trained on the training set, and its built-in feature importance (Gini importance or average impurity reduction) evaluation mechanism is used to calculate the importance of each feature. Importance score Based on the scores, ranked from highest to lowest, the top 5 features (slope, lithology, elevation, distance from river, and rainfall) were selected to form a local feature subset. .

[0072] S302: Local Model Training. A lightweight gradient boosting machine (LightGBM, n_estimators=120, max_depth=4) algorithm is used as the local viewpoint model (ML). LightGBM employs gradient-based one-sided sampling (GOSS) and mutually exclusive feature binding (EFB) to improve efficiency. Its objective function is similar to XGBoost but with optimized implementation. This is achieved in the local feature subset of the training set. ( ) on Conduct training.

[0073] S303: Local viewpoint prediction. Using pre-trained... Predicting the corresponding local feature subset yields the probability of landslide occurrence from a local perspective:

[0074] ;

[0075] in, This is a local perspective model (such as Lightweight Gradient Boosting Machine LightGBM).

[0076] S4: Interactive Perspective Model Construction and Prediction

[0077] S401: Automatically construct interactive features. The system automatically combines the top M important features (e.g., the top 5) from the original feature set XX in pairs to generate new interactive features. Combination methods include:

[0078] Multiplication interaction term: ;

[0079] Addition interaction item: ;

[0080] The system automatically combines the above five key factors in pairs. For example, it generates interactive features such as "slope × rainfall" and "elevation + distance from river". This case study generates a total of [number missing] features. A new interaction feature is added, which is combined with the original 5 key factors to form a 25-dimensional interaction feature set. .

[0081] S402: Interactive Model Training. The Random Forest algorithm is used as the interactive perspective model. Random forests, by constructing multiple decision trees and integrating their outputs, have a natural ability to capture complex interactions and nonlinear relationships between features. In the interaction feature set of the training set... Top Conduct training.

[0082] S403: Interactive viewpoint prediction. Using pre-trained... Predicting the probability of landslide occurrence from the interaction perspective by analyzing the corresponding interaction feature set:

[0083] ;

[0084] in, For interactive perspective models (such as random forests).

[0085] S5: Multi-view Adaptive Fusion

[0086] S501: Construct a fusion training set. Combine the probability prediction results from the three perspectives obtained on the training set. As a new feature, a three-column fusion feature matrix is ​​constructed. =[ The corresponding tag is still []. .

[0087] S502: Training the fusion model. Logistic Regression is used as the fusion model. Logistic regression achieves adaptive fusion by learning the weights of prediction results from different perspectives. Its model form is as follows:

[0088] ;

[0089] Where p is the final fusion prediction probability, The intercept, , , These are the weighting coefficients assigned to the prediction results from the global, local, and interactive perspectives, respectively. , Training The optimal weights are obtained. The final learned weight coefficients are: =0.45, =0.30, =0.25. This weight indicates that global information contributes the most, followed by key local factors, and factor interactions also provide important supplementation, forming an effective complementarity among the three.

[0090] S503: Generate a comprehensive prediction model. This involves combining all models trained in steps S2, S3, S4, and S502. , , , This is integrated into a complete multi-perspective fusion comprehensive prediction model for landslide susceptibility.

[0091] S6: Susceptibility Assessment and Outcome Output

[0092] S601: Regional Susceptibility Calculation. The feature data (Xunit) of each evaluation unit (e.g., raster pixel) in the target area is sequentially input into the comprehensive prediction model: first, through... , , get Then input these three probability values. The final landslide susceptibility probability value of this unit was calculated. .

[0093] S602: Risk Level Classification. Based on probability values. The study area was divided into different risk levels using the natural breakpoint method: extremely low risk area, low risk area, medium risk area, high risk area, and extremely high risk area.

[0094] S603: Results Visualization and Output. Based on a GIS platform, the calculated probability value or risk level of each unit is spatially rendered to generate a landslide susceptibility zoning map. The MVRL (Multi-View Robust Landslide susceptibility evaluation) model proposed in this invention outperforms any single-view model on all key indicators, with an AUC (Area Under the Curve) reaching 0.872 (e.g., ...). Figure 2 This improves performance by approximately 1.6% compared to the suboptimal global perspective model, demonstrating the effectiveness of multi-view fusion.

[0095] Example 2

[0096] Based on the above-described embodiment one, a multi-perspective fusion robust assessment system for landslide susceptibility is disclosed, comprising:

[0097] The data preprocessing module is used to acquire and process landslide impact factor data, including missing value imputation, standardization, and dataset partitioning.

[0098] The global perspective modeling module is used to train and apply a global perspective prediction model based on all features;

[0099] The local perspective modeling module is used to train and apply a local perspective prediction model based on a feature subset selected by importance.

[0100] The interactive perspective modeling module is used to train and apply an interactive perspective prediction model based on the constructed feature interaction items.

[0101] The multi-view fusion module is used to receive the prediction outputs of the above three view modules and generate the final susceptibility evaluation result through the fusion model.

[0102] The results output and visualization module is used to output the probability of landslide susceptibility, risk level, and generate susceptibility zoning maps.

[0103] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A robust evaluation method for landslide susceptibility based on multi-perspective fusion, characterized in that, The steps are as follows: S1: Obtain geographical, geological, environmental, and historical landslide data for the target area; construct a feature set of landslide influencing factors and landslide occurrence labels; process missing values ​​and standardize the data; and divide the data into training and testing sets according to a preset ratio. S2: Using all features in the training set, train the first machine learning model to learn the overall mapping relationship between landslide influencing factors and landslide occurrence from a global perspective, and obtain global prediction results. S3: Calculate and rank the importance of each feature in the training set, and select the top K features by importance to form a local feature subset; train a second machine learning model based on the local feature subset to learn the relationship between key factors and landslide occurrence from a local focus perspective, and obtain local perspective prediction results; S4: Based on the features of the training set, automatically construct interaction terms between features to form an interaction feature set; train a third machine learning model based on the interaction feature set to learn the relationship between the synergistic effect between factors and the occurrence of landslides from the perspective of feature interaction, and obtain the prediction results from the interaction perspective. S5: Using the global perspective prediction results, local perspective prediction results, and interactive perspective prediction results as input features, train the fusion model, and let the fusion model learn the optimal combination weights of the prediction results from each perspective to generate the final comprehensive prediction model for landslide susceptibility. S6: Input the feature data of the area to be evaluated into the trained comprehensive prediction model, and output the landslide susceptibility probability value or risk level of the area; Based on the probability value or risk level, a landslide susceptibility zoning map is generated.

2. The multi-perspective fusion robust assessment method for landslide susceptibility as described in claim 1, characterized in that, In step S2, the first machine learning model is a gradient boosting decision tree model.

3. The multi-perspective fusion robust assessment method for landslide susceptibility according to claim 1, characterized in that, In step S3, the second machine learning model is an efficient and lightweight model based on the gradient boosting framework.

4. The multi-perspective fusion robust assessment method for landslide susceptibility as described in claim 1, characterized in that, In step S3, feature importance is calculated using a random forest model.

5. The multi-perspective fusion robust assessment method for landslide susceptibility according to claim 1, characterized in that, In step S4, the methods for constructing feature interaction terms include, but are not limited to, multiplication, addition, or combination operations between features.

6. The multi-perspective fusion robust assessment method for landslide susceptibility according to claim 1, characterized in that, In step S4, the third machine learning model is an ensemble learning model.

7. The multi-perspective fusion robust evaluation method for landslide susceptibility according to claim 1, characterized in that, In step S5, the fusion model is a logistic regression model.

8. A multi-perspective fusion robust assessment system for landslide susceptibility, used to implement the method according to any one of claims 1-7, characterized in that, include: The data preprocessing module is used to acquire and process landslide impact factor data, including missing value imputation, standardization, and dataset partitioning. The global perspective modeling module is used to train and apply a global perspective prediction model based on all features; The local perspective modeling module is used to train and apply a local perspective prediction model based on a feature subset selected by importance. The interactive perspective modeling module is used to train and apply an interactive perspective prediction model based on the constructed feature interaction items. The multi-view fusion module receives the prediction outputs from the global view modeling module, the local view modeling module, and the interactive view modeling module, and generates the final susceptibility evaluation result through the fusion model. The results output and visualization module is used to output the probability of landslide susceptibility, risk level, and generate susceptibility zoning maps.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1-7.

10. 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 computer program, it implements the steps of the method as described in any one of claims 1-7.

Images