A dynamic evaluation method for landslide susceptibility
By optimizing the index interval division and negative sample selection for landslide susceptibility assessment, and combining hyperparameter optimization of machine learning models with surface deformation information, the uncertainties and dynamic issues in landslide susceptibility assessment were resolved, achieving higher-precision landslide disaster prediction and dynamic assessment.
Patent Information
- Application Number
- CN202410468295.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-04-18
- Publication Date
- 2024-11-12
- Estimated Expiration
- 2044-04-18
AI Technical Summary
Existing landslide susceptibility assessments suffer from problems such as high uncertainty in the construction of the indicator system, selection bias of non-landslide negative samples, difficulty in selecting hyperparameters for machine learning modeling, and insufficient dynamic evaluation methods, resulting in insufficient accuracy in landslide disaster prediction.
The evaluation indicators were divided into intervals using the equal interval classification method, standard deviation method, quantile method, geometric interval method, natural discontinuity method, and ChiMerge discretization method. Reasonable intervals were selected in combination with the geographic detector model. Non-landslide negative samples were selected by the weighted information content model. The hyperparameters of the XGBoost, support vector machine, and random forest models were optimized using the dung beetle algorithm. Surface deformation information was extracted by combining SBAS-InSAR technology for dynamic evaluation.
It improves the accuracy and reliability of landslide susceptibility prediction models, provides a dynamic evaluation matrix, and can more accurately predict the likelihood of landslide disasters, supporting early identification and prevention.
Smart Images

Figure CN118296961B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of dynamic assessment of landslides, and in particular to a method for dynamic assessment of landslide susceptibility. Background Art
[0002] Landslide is a serious geological disaster with great destructiveness, strong harm and widespread occurrence, which poses a major threat to the safety of life and property. In recent years, human activities and global changes have led to a significant increase in the frequency of landslide disasters, causing great harm to social and economic development. Landslide is a deformation and destruction process and phenomenon of slope rock and soil under the action of gravity and external factors (such as rainfall, water level, earthquake, human engineering activities, etc.). Driven by gravity and external forces, the slope rock and soil will break, accompanied by the incubation and penetration of internal potential sliding surfaces, resulting in external macroscopic deformation and eventually instability and destruction to form landslides. To solve this problem, landslide prediction and risk assessment are the key to reducing the loss of life and property. Landslide Susceptibility Assessment (LSA) aims to predict the spatial distribution and possibility of landslides. Landslide Susceptibility Assessment points out the prone areas of landslides. It is a basic step in assessing landslide hazards and formulating mitigation strategies. Decision makers, scientists, engineers and the public can use the assessment results to avoid catastrophic landslides. Therefore, it is very important to carry out research on landslide susceptibility assessment.
[0003] Landslide susceptibility assessment research is crucial to avoid and mitigate regional landslide disasters. It is a comprehensive measure of the probability of landslides. In short, landslide susceptibility modeling is a landslide risk assessment technology that uses geographic information systems and computer technology to obtain detailed and accurate spatial information, that is, to study whether a region is likely to experience a landslide disaster. It is an important tool for avoiding losses related to landslide disasters. It can identify high-risk areas for landslides, which usually helps government decision-making and the formulation of disaster prevention and mitigation measures, and can indirectly reduce casualties and economic losses caused by landslides.
[0004] The existing landslide assessment still has some problems in theory and technology: (1) In the construction of the landslide susceptibility evaluation index system, the uncertainty of landslide susceptibility evaluation index processing is high, and the selection method of numerical evaluation index interval division is highly subjective and lacks broad representativeness. (2) Non-landslide negative samples are important data in landslide susceptibility modeling. Sample bias may occur in the process of selecting non-landslide negative samples. How to improve the rationality of negative sample selection still requires relevant research. (3) Machine learning modeling is a key part in the landslide susceptibility evaluation process. The selection of model hyperparameters is crucial. How to improve the accuracy of landslide susceptibility prediction results through model optimization still requires further research. (4) The occurrence of landslides is a dynamically changing geological phenomenon. How to combine the temporal dynamic characteristics and the spatial characteristics of landslide evaluation indicators to conduct dynamic evaluation of landslide susceptibility is still a problem to be solved. Summary of the invention
[0005] In order to overcome the shortcomings and deficiencies of the prior art, the present invention provides a method for dynamically evaluating landslide susceptibility.
[0006] The method includes:
[0007] Step S1, acquiring landslide information: acquiring landslide catalog information per unit area by applying technical means and statistical data, comprehensively considering the landslide disaster mechanism and disaster-pregnant environment per unit area, analyzing the landslide distribution characteristics per unit area from two aspects of disaster type and distribution law, and extracting 10 evaluation indicators of five categories;
[0008] Step S2, constructing an evaluation index system; using equal interval classification method, standard deviation method, quantile method, geometric interval method, natural break point method and ChiMerge discretization method to divide the data-based evaluation index into intervals, using geographic detectors to calculate the evaluation index q value obtained by different interval methods, and selecting the evaluation index interval division method with the largest q value to achieve reasonable interval division of the evaluation index; selecting the evaluation index that passes the multicollinearity test to make a preliminary prediction of regional landslide susceptibility based on the weighted information model, and randomly selecting non-landslide points that are consistent with the number of landslide points in the extremely low susceptibility area of the weighted information model as non-landslide negative samples, and the selection of negative samples meets the requirements of randomness and low susceptibility;
[0009] Step S3, machine learning model construction; integrating historical landslide positive samples and non-landslide negative samples to form a training sample set for unit area regional landslide susceptibility modeling, introducing the dung beetle algorithm to optimize the parameters of the XGBoost model, the support vector machine model, and the random forest model, selecting the optimal parameters to establish a machine learning landslide susceptibility prediction model, completing the landslide susceptibility evaluation of the unit area, drawing a landslide disaster susceptibility zoning map, performing zoning statistics and machine learning model accuracy evaluation, comparing the results of the XGBoost model, the support vector machine model, and the random forest model from three aspects: confusion matrix, statistical parameters, and receiver operating curve, and analyzing the accuracy of the unit area regional landslide susceptibility prediction model;
[0010] Step S4, dynamic evaluation of landslide susceptibility; based on the ascending orbit Sentinel-1A image data of unit area, the SBAS-InSAR technology is used to extract the surface deformation information, and the spatial distribution characteristics of landslide hazards in the unit area are analyzed in combination with the deformation information; the application and performance of the DBO-XGBoost model, the DBO-SVM model, and the DBO-RF model in the prediction results of landslide susceptibility in unit area are compared, and the landslide susceptibility result map obtained by the DBO-RF model and the SABS-InSAR deformation information are combined to establish a dynamic evaluation matrix for landslide susceptibility in unit area, and perform dynamic evaluation of landslide susceptibility in unit area.
[0011] Furthermore, the five categories are topography, geological structure, meteorology and hydrology, environmental conditions, and human activities; the ten evaluation indicators are elevation, slope, slope aspect, terrain undulation, lithology, distance to rivers, land use, distance to roads, POI kernel density, and human footprint index.
[0012] Furthermore, the equal-interval classification method uses equal intervals to divide the range of the evaluation index into several sub-ranges of equal size, and uses the reclassification tool in ArcGIS Pro software to perform the equal-interval classification interval division of the evaluation index raster layer, and the number of intervals is set to 6 based on the subjective experience method;
[0013] The natural break point method is used to automatically classify the intervals of the evaluation index in ArcGIS Pro software, and the natural break point classification method interval division of the evaluation index raster layer is performed using the reclassification tool in ArcGIS Pro software, and the number of intervals is set to 6 based on the subjective experience method;
[0014] The quantile method uses the reclassification tool in ArcGIS Pro software to perform quantile classification interval division of the evaluation index raster layer, and the number of intervals is set to 6 based on subjective experience;
[0015] The geometric interval method uses the reclassification tool in ArcGIS Pro software to perform geometric interval classification interval division of the evaluation index raster layer, and the number of intervals is set to 6 based on subjective experience;
[0016] The standard deviation method uses the reclassification tool in ArcGIS Pro software to divide the evaluation index raster layer into standard deviation classification intervals, and the interval is set to 1 times;
[0017] The ChiMerge discretization method divides the data-type evaluation indicators into intervals. In IBM SPSS Statistics 26.0 software, the card binning tool is used to perform ChiMerge discretization on the continuous evaluation indicators of elevation, slope, and terrain relief, and the threshold is set to 0.05.
[0018] Furthermore, the geographic detector is used to reveal the driving force by calculating the spatial heterogeneity of the evaluation index, and the differentiation and factor detectors are selected to detect the degree of explanation of the landslide. The expression is:
[0019]
[0020] Where i represents the number of layers of landslide evaluation index, i = 1, 2, ..., m; N i represents the number of units in the i-th layer of the landslide evaluation index; N represents the number of units in the entire area of the evaluation index; represents the variance of the landslide Y value corresponding to the i-th layer of landslide evaluation index; σ 2 It represents the variance of the Y value of the landslide corresponding to the evaluation index in the whole area; SL represents the sum of the values within the layer; ST represents the total variance of the whole area.
[0021] Furthermore, the multicollinearity test is used to find the linear or nonlinear relationship between variables. When TOL>0.1 and VIF<10, there is no multicollinearity between the evaluation indicators, and the calculation expression is:
[0022]
[0023]
[0024] in, represents the coefficient of determination, Tolerances represents the tolerance, and VIF represents the variance inflation factor;
[0025] The collinearity diagnostic tool in the linear analysis of IBM SPSS Statistics 26.0 software was used to calculate the VIF and TOL of the landslide susceptibility evaluation indicators per unit area. The evaluation indicators with high collinearity were excluded, and three noun-type evaluation indicators and seven evaluation indicators with the optimal interval division method obtained by the geographic detector model were selected as input variables to build a landslide susceptibility prediction model.
[0026] Furthermore, the information volume model is calculated as follows:
[0027]
[0028]
[0029] Among them, I(y,x1,x2,…x n ) represents the evaluation index combination x1, x2,…, x n The amount of information provided on geological hazards; P(y|x1,x2,…,x n ) represents the evaluation index x1, x2,…, x n The probability of landslide disaster under combined conditions; P(y) represents the probability of landslide disaster; It indicates the amount of information provided by evaluation index x2 on the landslide susceptibility evaluation results when evaluation index x1 exists;
[0030] when I(y,x1,x2,…x n )>0, indicating that the factor combination is conducive to the occurrence of landslide disasters; on the contrary, I(y,x1,x2,…x n )<0, indicating that the factor combination is not conducive to the occurrence of landslide disasters;
[0031] The information value provided by each evaluation index at different levels on landslide disasters is calculated, and the expression is:
[0032]
[0033] Among them, P(x i ,D) indicates that in the evaluation index x i The probability of a certain state classification level; P(x i ) represents the probability of landslide hazard distribution;
[0034] The calculation expression of the information amount of each evaluation index is:
[0035]
[0036] Among them, N i Indicates that under a certain evaluation index and a certain grading state (xi ) represents the number of geological disasters; N represents the total number of geological disasters occurring in the entire unit area; S i Indicates a certain grading status of an evaluation factor (x i ) of the grid; S—the number of grid cells in the entire unit area;
[0037] The calculation expression for the total amount of information in each grid cell is:
[0038]
[0039] Where n represents the number of evaluation indicators involved in landslide susceptibility assessment.
[0040] Furthermore, the non-landslide negative samples are used to make a preliminary prediction of the landslide susceptibility per unit area using a weighted information model, and non-landslide sample points that are consistent with the number of landslide samples are randomly selected in the extremely low-susceptibility area in the prediction results. The landslide susceptibility evaluation per unit area is completed through a machine learning model based on raster units, and the landslide susceptibility zoning results of the weighted information model are used to randomly select non-landslide points that are consistent with the number of landslide points using a random point generation tool in ArcGIS Pro3.0 software.
[0041] Furthermore, the XGBoost model uses randomly selected training set data to establish a landslide susceptibility evaluation model based on the XGBoost model, uses test set data to verify the prediction accuracy of the model, and uses the dung beetle algorithm to optimize the n_estimators and max_depth parameters of the model to improve the accuracy of the model. n_estimators represents the number of trees, and max_depth represents the maximum depth of the tree. Each tree in the XGBoost model has a limited depth, and overfitting is prevented by limiting the depth of the tree. In the process of optimizing the XGBoost model by the dung beetle algorithm, the number of algorithm iterations is set to 50 times, and the model accuracy obtained by each iteration is calculated. After the iteration is completed, the XGBoost model under the optimal parameter combination is obtained.
[0042] Furthermore, the support vector machine model uses randomly selected training set data to establish a landslide susceptibility evaluation model based on the XGBoost model, uses test set data to verify the prediction accuracy of the model, and uses the dung beetle algorithm to optimize the C and gamma parameters of the model to improve the accuracy of the model. C represents the penalty coefficient, which determines the tolerance of the degree to which the sample data violates the interval. When the C value is very small, the model has a low penalty for misclassified points; when the C value is large, the model will reduce the number of misclassified samples, and the decision boundary has a smaller interval. The gamma parameter controls the distribution of the data after it is mapped to the high-dimensional space. The model's fit to the training data increases with the increase of the gamma parameter. If the gamma parameter is too large, the model will overfit.
[0043] Furthermore, the random forest model uses randomly selected training set data to establish a landslide susceptibility evaluation model based on the random forest model, uses test set data to verify the prediction accuracy of the model, uses the dung beetle algorithm to optimize the n_estimators and max_depth parameters of the model to improve the accuracy of the model, and in the process of optimizing the random forest model by the dung beetle algorithm, the number of algorithm iterations is set to 50 times, and the model accuracy obtained by each iteration is calculated; after the iteration is completed, the random forest model under the optimal parameter combination is obtained, and the model is rebuilt with the obtained optimal parameters. After the evaluation model training is completed, the overall data set of the unit area is input into the model, and the landslide susceptibility evaluation of the unit area is performed, and the prediction results are output, and finally the landslide occurrence probability of all grid cells in the entire unit area is obtained; the closer the prediction value corresponding to the grid cell is to 0, the more stable the grid cell is, and the smaller the probability of landslide; conversely, the closer the prediction value is to 1, the more unstable the grid cell is, and the greater the possibility of landslide.
[0044] Beneficial effects:
[0045] The present invention proposes a method for dynamic evaluation of landslide susceptibility, which combines machine learning and SBAS-InSAR technology. On the basis of considering the selection of evaluation index interval division method and non-landslide negative sample selection, the dung beetle algorithm is used to optimize the hyperparameters of the machine learning model, construct a landslide susceptibility evaluation model, complete the landslide susceptibility evaluation of the unit area, establish a dynamic evaluation matrix, combine SBAS-InSAR deformation information and landslide susceptibility evaluation zoning map, complete the dynamic evaluation of landslide susceptibility of the unit area, and combine different data discretization methods (equal interval classification method, standard deviation method, quantile method, geometric interval method, natural break point method and ChiMerge discretization method) and geographic detector model to achieve the optimal discretization of the evaluation index of the landslide susceptibility prediction process, and improve the regional applicability of the evaluation index interval division method. In view of the problem that the discretization of numerical evaluation indicators is highly subjective, based on the geographic detector model, the q value of the grouping result obtained by different interval division methods is calculated. The present invention proposes a new non-landslide negative sample selection method, which ensures the randomness and low susceptibility of non-landslide negative sample selection, thereby improving the accuracy and reliability of the landslide susceptibility prediction model. In view of the problem that the traditional negative sample sampling method lacks consideration of the randomness and low susceptibility of samples, the present invention proposes a negative sample selection method based on a weighted information volume model on the basis of a method for selecting an interval division of applicable evaluation indicators based on a geographic detector, combines the information volume method and the analytic hierarchy process, calculates the information volume weight of each indicator, establishes a unit area landslide susceptibility mapping based on the weighted information volume model, and selects non-landslide negative sample points that are consistent with the number of landslide samples from the extremely low susceptibility area in the susceptibility mapping. The present invention applies the dung beetle algorithm to the machine learning landslide susceptibility evaluation model, uses the dung beetle algorithm to perform hyperparameter optimization on the XGBoost model, the SVM model and the RF model, and thereby completes the construction of the unit area landslide susceptibility evaluation model. Based on the acquisition of ascending orbit Sentinel-1A image data per unit area, the present invention uses SBAS-InSAR technology to extract surface deformation information, divides the surface deformation information into five levels, combines the landslide susceptibility result map obtained by the best prediction model - DBO-RF model and SABS-InSAR deformation information, and completes the dynamic evaluation of landslide susceptibility per unit area based on the dynamic evaluation matrix, providing a theoretical reference for landslide disaster prediction and prevention. The method has high calculation accuracy and can effectively prevent the occurrence of landslide accidents. BRIEF DESCRIPTION OF THE DRAWINGS
[0046] Figure 1 It is a general step flow chart of the present invention;
[0047] Figure 2 It is the flowchart of the hierarchical analysis method of the present invention;
[0048] Figure 3It is a structural diagram of the hierarchical analysis method model of the present invention;
[0049] Figure 4 It is the XGBoost model flow chart of the present invention;
[0050] Figure 5 It is a flow chart of the support vector machine model of the present invention;
[0051] Figure 6 This is a flow chart of the random forest model of the present invention. DETAILED DESCRIPTION
[0052] It should be noted that, in the absence of conflict, the embodiments and features described in the embodiments of the present application may be combined with each other. The present application is further described in detail below in conjunction with the accompanying drawings and specific embodiments.
[0053] A dynamic evaluation method for landslide susceptibility is developed by combining machine learning and SBAS-InSAR technology. The research focuses on the construction of an indicator system for landslide susceptibility prediction, the method of selecting negative samples, and the optimization of machine learning models. (1) In order to solve the problem that the index interval division in the landslide susceptibility evaluation per unit area is not widely representative, a geographic detector is combined to achieve a reasonable interval division of the evaluation index; (2) In order to solve the problem of unreasonable extraction of non-landslide negative samples, a negative sample selection method based on a weighted information volume model is proposed to provide important sample data that meets the needs of landslide susceptibility modeling. (3) In order to solve the problem of hyperparameter selection of machine learning models in landslide susceptibility evaluation, the dung beetle algorithm is used to automatically search for the optimal parameters to improve the prediction accuracy of the machine learning model in landslide susceptibility evaluation.
[0054] like Figure 1 As shown, a landslide susceptibility dynamic evaluation method comprises:
[0055] Step S1, landslide information acquisition: obtain the landslide catalog information per unit area by applying technical means, statistical data, etc., comprehensively consider the landslide disaster mechanism and disaster-pregnant environment per unit area, analyze the landslide distribution characteristics per unit area from two aspects of disaster type and distribution law, and extract 10 evaluation indicators of 5 categories (topography, geological structure, meteorology and hydrology, environmental conditions, and human activities), namely elevation, slope, slope aspect, terrain undulation, lithology, distance to rivers, land use, distance to roads, POI kernel density, and human footprint index.
[0056] Step S2, construction of evaluation index system; in order to explore the interval division method of landslide susceptibility evaluation index applicable to unit area, the equal interval classification method, standard deviation method, quantile method, geometric interval method, natural break point method and ChiMerge discretization method are used to divide the data type evaluation index (noun type evaluation index can be directly converted into discrete values and are not included in the discussion) into intervals, and the evaluation index q value obtained by different interval methods is calculated by using the geographic detector, and the evaluation index interval division method with the largest q value is selected to achieve reasonable interval division of the evaluation index; the evaluation index that passes the multicollinearity test is selected to make a preliminary prediction of regional landslide susceptibility based on the weighted information model, and non-landslide points that are consistent with the number of landslide points in the extremely low susceptibility area of the weighted information model are randomly selected as non-landslide negative samples, so that the selection of negative samples meets the requirements of randomness and low susceptibility.
[0057] Step S3, machine learning model construction; integrating "historical landslide" positive samples and non-landslide negative samples to form a training sample set for unit area landslide susceptibility modeling, introducing the dung beetle algorithm to optimize the parameters of the XGBoost model, support vector machine model, and random forest model, selecting the optimal parameters to establish a machine learning landslide susceptibility evaluation model, completing the landslide susceptibility evaluation of the unit area, drawing a landslide disaster susceptibility zoning map, performing zoning statistics and machine learning model accuracy evaluation, comparing the results of the XGBoost model, support vector machine model, and random forest model from three aspects: confusion matrix, statistical parameters, and receiver operating curve, and analyzing the accuracy of the unit area landslide susceptibility evaluation model.
[0058] Step S4, dynamic evaluation of landslide susceptibility; based on the ascending orbit Sentinel-1A image data of unit area from January 2021 to December 2023, use SBAS-InSAR technology to extract surface deformation information, and analyze the spatial distribution characteristics of landslide hazards per unit area in combination with the deformation information; compare the application and performance of DBO-XGBoost model, DBO-SVM model, and DBO-RF model in the evaluation results of landslide susceptibility per unit area, combine the landslide susceptibility result map obtained by the best evaluation model—DBO-RF model and SABS-InSAR deformation information, establish a dynamic evaluation matrix for landslide susceptibility per unit area, carry out dynamic evaluation of landslide susceptibility per unit area, and provide a theoretical reference for landslide disaster prediction and prevention.
[0059] The unit area is Fengjie County in the Three Gorges Reservoir area, which has a subtropical warm and humid monsoon climate. The region has abundant rainfall, with an average annual precipitation of 1,132 mm. It is located in the border area of the eastern mountains of the Sichuan Basin, the front edge of the Daba Mountains and the western Hubei mountains. The surface water system in the area is well developed. Years of water erosion and accumulation have formed multi-level river valley terraces and deep gullies. The stratigraphic types are mainly Quaternary, Triassic and Jurassic. The unit area is located in the heart of the Yangtze River Three Gorges Reservoir, adjacent to Wushan County in the east, Yunyang County in the south, Wanzhou District in the west, and Yichang City, Hubei Province in the north. It is 109°1′17″ to 109°45′58″ east longitude and 30°29′19″ to 31°22′33″ north latitude. It covers an area of about 4,087 square kilometers. It has a complex structure, with dense development of synclines and anticlines. The overall terrain is high in the southeast and northeast, while the central part is slightly flat in the west. The north is the southern foot of Daba Mountain, and the east and south are Wushan and Qiyao Mountain. The county seat is located in the center of the entire administrative area, among which Dufu Town and Yong'an Town are the seats of the county government.
[0060] The unit area mapping unit is selected. The landslide susceptibility evaluation unit is the smallest unit that cannot be divided in the susceptibility assessment. During the assessment process, each unit is abstracted into a feature vector and input into the machine learning model. The determination of the evaluation unit directly affects the model evaluation results. At present, the more mature landslide susceptibility evaluation units mainly include regional units, uniform condition units, slope units and grid units.
[0061] (1) Regional unit. Regional unit is an evaluation unit based on the associations and differences existing in the natural geographical environment. It is determined by the naturally formed regional boundaries that reflect the topography, landforms and geological properties. Regional units are usually used in land resource surveys and landform studies. The use of regional units in landslide susceptibility assessment can more clearly reflect the relationship between landforms and landslide occurrence. Regional units rely on a large amount of topographic and geological data and expert knowledge, and their setting is also subjective. They are more suitable for large-area macro-landslide susceptibility assessment.
[0062] (2) Uniform conditional unit. The uniform conditional unit does not rely on prior knowledge, but is directly based on the spatial overlap of each evaluation index, and the spatial overlapping grouping result is used as the evaluation unit. The division of the uniform conditional unit retains the independence of each evaluation index. The final evaluation unit is the maximum combination set of different values of each evaluation factor, which is suitable for the refined evaluation of small-area landslide susceptibility with high data accuracy. However, due to the division of the evaluation index itself, the uniform conditional unit has a certain degree of subjectivity in actual use.
[0063] (3) Slope cells can be created by extracting ridge lines and valley lines from high-precision DEM data or contour lines and dividing the area into individual slope surfaces accordingly. Slope cells are compatible with topographic and geological features and are suitable for deterministic modeling of small areas. However, slope cells require high-quality terrain data, require manual modification to represent, and have a low degree of automation.
[0064] (4) Grid unit: Grid unit is a regular grid, and the unit area is divided into regular grids of equal size. The advantages of grid unit are simple data structure, flexible sampling, fast calculation speed and high data availability. However, compared with the first three methods, the data generated by dividing grid units is more, which is not enough to intuitively reflect the overall situation of terrain and topography.
[0065] The present invention collects evaluation index data related to landslide occurrence, such as topography, geological structure, meteorology and hydrology, environmental conditions and human activities per unit area, to construct a spatial database of landslide evaluation indexes. Among them, topography and geomorphology data include elevation, slope, slope aspect and terrain relief index; geological structure includes lithology index; meteorology and hydrology include distance to river index; environmental conditions include land use index; human activities include distance to road, POI kernel density and human footprint index index.
[0066] The present invention uses ArcGIS Pro 3.0 software to extract and preprocess the data to obtain an evaluation index grid layer. In ArcGIS Pro 3.0 software, all layers are converted into a WGS-84-UTM-zone-49N projection coordinate system, and the evaluation index layer is superimposed on the landslide catalog grid layer to achieve two-way retrieval and query of historical landslide disasters and evaluation indicators.
[0067] To construct an index system for evaluating the susceptibility of landslides, it is necessary to classify the index factors. The classification of index factors refers to dividing a single evaluation index into multiple status categories according to certain standards. Evaluation indicators include noun-type evaluation indicators and numerical-type evaluation indicators. Noun-type evaluation indicators are discrete and discontinuous forms. For example, the values of land use types are: cultivated land, forest land, grassland, water area, construction land and unused land. The noun-type evaluation indicators of the present invention include "lithology, slope aspect, land use". The numerical-type evaluation indicators are continuous numerical representations, such as continuously changing elevation data. The numerical-type evaluation indicators of the present invention include "elevation, slope, terrain undulation, distance from the river, distance from the road, POI kernel density, and human footprint index".
[0068] Descriptive text cannot be directly applied to the landslide susceptibility assessment model, so it is necessary to convert the noun-type evaluation index into a discrete value. The present invention uses the model evaluation index regularization to process the noun-type evaluation index, and the processing method is to use numbers to represent the descriptive text.
[0069] The process of establishing the evaluation index system requires discretization (i.e., interval division) of numerical evaluation indicators such as elevation, slope, and terrain relief. Discretization of data-based evaluation indicators has the following advantages: ① Many machine learning models (such as XGBoost and random forest) cannot directly process continuous variables, and the discretization of continuous data can reduce the limitations of the modeling process; ② Discretization of continuous data can improve the discriminability of machine learning models, thereby improving the predictive ability of the model; ③ Discretization can simplify the data structure of evaluation indicators, reduce the impact of outliers, reduce dependence on data distribution assumptions, and improve computational efficiency.
[0070] Discretization is an important part in the construction of evaluation index body, and has a direct impact on the evaluation of landslide susceptibility by machine learning. At present, most studies on the discretization of continuous evaluation indicators adopt the interval division method of equal spacing method or natural break point method, which is simple and convenient, but not necessarily applicable to different unit area areas. In view of this problem, the present invention is based on the classification method (equal spacing classification method, standard deviation method, quantile method, geometric interval method, natural break point method) in ArcGIS Pro3.0 software and ChiMerge discretization method to divide 7 continuous evaluation indicators (due to the consistent categories of slope aspect, lithology and land use data, which are not discussed separately in different classification methods) into intervals, and the interval division results are loaded into the geographic detector model, and the q value results are counted, and the method with the largest q value in the classification method corresponding to each evaluation indicator is selected for interval division and landslide susceptibility evaluation modeling.
[0071] The equal-interval classification method uses equal intervals to divide the range of the evaluation index into several sub-ranges of equal size. The reclassification tool in ArcGIS Pro software is used to divide the evaluation index raster layer into equal-interval classification intervals, and the number of intervals is set to 6 based on subjective experience;
[0072] The natural break point method first sorts all the data within the evaluation index, combines statistical methods to find the "breakpoints" of the difference data between samples, and completes the classification based on the errors between and within the categories, so that the attribute differences between different classification intervals are maximized and the differences within the classification intervals are minimized. This method first calculates the variance of each classification interval, then calculates the sum of the variances, and finally obtains the optimal classification result with the smallest sum of variances. The intervals of the evaluation index are automatically classified in ArcGIS Pro software. The reclassification tool in ArcGIS Pro software is used to divide the natural break point classification method intervals of the evaluation index raster layer, and the number of intervals is set to 6 based on the subjective experience method;
[0073] The quantile method presents the classified objects in roughly the same number, and the classification results of each evaluation index are evenly distributed, that is, each classification interval contains an equal number of elements. The reclassification tool in ArcGIS Pro software is used to divide the quantile classification interval of the evaluation index raster layer, and the number of intervals is set to 6 based on subjective experience;
[0074] The geometric interval method establishes intervals based on geometric series and determines the classification intervals of the evaluation indicators. The principle is to minimize the sum of the squares of the number of elements in each classification interval as much as possible to ensure that the number of values in each class is as close as possible and to keep the changes between classification intervals consistent. The reclassification tool in ArcGIS Pro software is used to divide the geometric interval classification method intervals of the evaluation indicator raster layer, and the number of intervals is set to 6 based on subjective experience;
[0075] The standard deviation method calculates the mean and standard deviation of each. The discrete points are created by calculating the equal value range proportional to the standard deviation. This method can reflect the difference between the feature attribute value and the mean value. The reclassification tool in ArcGIS Pro software is used to divide the standard deviation classification interval of the evaluation index raster layer, and the interval is set to 1 times;
[0076] The ChiMerge discretization method uses the χ 2 (Chi-square statistic) discretizes continuous variables and determines whether the current breakpoint should be deleted and whether it should be merged with adjacent groups. The implementation steps of the ChiMerge method are: ① Sort the variable values; ② Define the initial interval; ③ Count and generate a frequency table; ④ Calculate the chi-square value of adjacent intervals, and determine whether the adjacent intervals should be merged based on the calculation results; ⑤ Based on the new interval distribution, repeat the work of step ④. The calculation expression is:
[0077]
[0078] E ij =R i ×C j / N
[0079]
[0080] Among them, χ 2 —Reflects the uniformity of the distribution of class j samples in two adjacent intervals of the node.
[0081] Chi-square binning stops when the number of bins is equal to the specified number of bins or the minimum chi-square value is greater than the chi-square threshold.
[0082] The present invention uses the ChiMerge discretization method to divide the data type evaluation indicators into intervals. In IBM SPSS Statistics 26.0 software, the card binning tool is used to implement ChiMerge discretization for continuous evaluation indicators such as elevation, slope, and terrain undulation, and the threshold is set to 0.05.
[0083] Geographic detectors can reveal the driving force of evaluation indicators by calculating their spatial heterogeneity, including differentiation and factor detectors, interaction detectors, risk zone detectors, and ecological detectors. The present invention selects differentiation and factor detectors that can detect the degree to which landslide evaluation indicators explain landslides. Let Y represent landslides, and X represent evaluation indicators that affect landslides. The impact of landslide evaluation indicators on landslides is measured by q (with a value range of 0-1), and the expression is:
[0084]
[0085] Where i represents the number of layers (classifications or partitions) of landslide evaluation indicators, i = 1, 2, ..., m; N i represents the number of units in the i-th layer of the landslide evaluation index; N represents the number of units in the entire area (all layers) of the evaluation index; represents the variance of the landslide Y value corresponding to the i-th layer of landslide evaluation index; σ 2 It represents the variance of the landslide Y value corresponding to the whole area (all layers) of the evaluation index; SL represents the sum of the values within the layer; ST represents the total variance of the whole area.
[0086] The numerical evaluation index interval division method and the number of classifications are different, and the q value obtained based on the geographic detector model will also be different. At present, the methods for discretizing numerical evaluation indicators in landslide evaluation indicators include equal spacing classification method, natural break point classification method, subjective experience method, etc. (due to the excessive uncertainty of subjective experience method in the process of evaluation index interval division, it is not included in the research of the present invention). The present invention is based on the reclassification tool in ArcGIS Pro3.0 software (classification method: equal spacing classification method, natural break point classification method, quantile classification method, geometric interval classification method and standard deviation classification method) and IBM SPSS Statistics 26.0 software (classification method: ChiMerge discretization method) to divide the 7 numerical evaluation indicators (slope aspect, lithology, and land use are also included in the geographic detector model to sort the importance of evaluation indicators) into intervals, and the q value results are counted, and the method with the largest q value in the classification method corresponding to each evaluation indicator is selected to divide the intervals.
[0087] When establishing a landslide susceptibility evaluation model, the comprehensive influence between the evaluation indicators must be considered and analyzed. Theoretically, the accuracy of the landslide susceptibility evaluation model and the complexity of the model are positively correlated with the number of evaluation indicators. As the number of landslide susceptibility evaluation indicators increases, the probability of multicollinearity between the evaluation indicators will also increase. If there is a linear correlation between the evaluation indicators, it will have a negative impact on the stability and reliability of the landslide susceptibility prediction model. The multicollinearity between the evaluation indicators is tested using the variance inflation factor (VIF) and tolerance (TOL). The main function of the multicollinearity test is to discover the linear or nonlinear relationship between variables. When TOL>0.1 and VIF<10, there is no multicollinearity between the evaluation indicators, and the calculation expression is:
[0088]
[0089]
[0090] in, represents the coefficient of determination.
[0091] The present invention uses the collinearity diagnostic tool in the linear analysis of IBM SPSS Statistics 26.0 software to calculate the VIF and TOL of the landslide susceptibility evaluation index per unit area, excludes the evaluation index with high collinearity, and the TOL values of all evaluation indexes are greater than 0.3, and the VIF values are less than 4, meeting the critical values of TOL>0.1 and VIF<10. Among them, the TOL of the evaluation index elevation is the smallest and the VIF is the largest, which are 0.313.27 respectively, which does not exceed the collinearity diagnostic threshold, and there is no collinear correlation with other evaluation indexes. Therefore, the three noun-type evaluation indexes selected by the present invention and the seven evaluation indexes of the optimal interval division method obtained by the geographic detector model do not show strong collinearity, and can be used as input variables to construct a landslide susceptibility evaluation model.
[0092] The calculation formula of the information volume model is as follows:
[0093]
[0094]
[0095] Among them, I(y,x1,x2,…x n ) represents the evaluation index combination x1, x2,…, x n The amount of information provided on geological hazards; P(y|x1,x2,…,x n) represents the probability of landslide disaster under the combination of evaluation indicators x1, x2, …, xn; P(y) represents the probability of landslide disaster; It indicates the amount of information provided by evaluation index x2 on the landslide susceptibility evaluation results when evaluation index x1 exists.
[0096] when I(y,x1,x2,…x n )>0, indicating that the factor combination is conducive to the occurrence of landslide disasters; on the contrary, I(y,x1,x2,…x n )<0, indicating that the combination of factors is not conducive to the occurrence of landslide disasters.
[0097] The information provided by each evaluation index at different levels on landslide disasters is calculated as follows:
[0098]
[0099] Among them, P(x i ,D) indicates that in the evaluation index x i The probability of a state classification level appearing; P(x i ) represents the landslide hazard distribution probability.
[0100] For the calculation of the information amount of each evaluation index, the formula can be used:
[0101]
[0102] Among them, N i Indicates that under a certain evaluation index and a certain grading state (x i ) represents the number of geological disasters; N represents the total number of geological disasters occurring in the entire unit area; S i Indicates a certain grading status of an evaluation factor (x i ) is the number of grid cells; S represents the number of grid cells in the entire unit area.
[0103] To calculate the total amount of information in each grid cell, the formula can be used:
[0104]
[0105] Where n represents the number of evaluation indicators involved in landslide susceptibility assessment.
[0106] The susceptibility or probability of landslide disasters in the evaluation unit increases with the increase of information value. After the information value is determined, the landslide susceptibility area can be divided into unit area.
[0107] The analytic hierarchy process (AHP) is a structure-based decision-making procedure that uses simple mathematical models to solve complex decision-making problems, assign optimal weights to input factors and obtain appropriate results (such as landslide susceptibility evaluation). The analytic hierarchy process relies on the relative importance of each factor defined by importance, and due to its unique consistency test, the method can assign reasonable weights to the evaluation.
[0108] like Figure 2 The process of hierarchical analysis is shown in the figure. The steps include: establishing a hierarchical structure, constructing a judgment matrix for each factor according to the importance scale of the target, hierarchical single sorting and consistency test, hierarchical total sorting and consistency test, determining whether the relative importance of each factor is appropriate, and finally making a decision.
[0109] Combining the analytic hierarchy process with the information method, a weighted information model is constructed. Based on the information value of the existing evaluation indicators, the analytic hierarchy process is used to calculate the weight of each evaluation indicator. The information value of each evaluation indicator interval is multiplied by the corresponding weight, and the sum is obtained to obtain the information weight, as shown in formula (3-14):
[0110]
[0111] Where I represents the information volume of a unit in the evaluation area; W i Represents the weight value of the i-th evaluation index; I i represents the information content of the i-th evaluation index; S represents the total number of evaluation units in the unit area; S i Indicates that the unit area contains evaluation index X i The number of units; N represents the total number of units with disaster distribution per unit area; N i Indicates the distribution of evaluation index X i The number of disaster units within a particular category.
[0112] like Figure 3 As shown in the figure, the specific steps of using the weighted information model to make a preliminary prediction of landslide susceptibility per unit area are as follows:
[0113] (1) According to the relationship between the elements in the evaluation system, a hierarchical structure is established. The present invention divides the hierarchical structure into three layers: landslide susceptibility evaluation (A) is the target layer; evaluation index type (B) is the criterion layer; and evaluation index (C) is the solution layer.
[0114] (2) Construct a pairwise judgment matrix at each level. The order of the constructed recursive hierarchy is from the criterion level to the solution level. Use the 1-9 scaling method to construct a pairwise judgment matrix for different indicators at the same level. Based on the constructed pairwise judgment matrix, use the square root method to find the normalized relative importance weight vector W of each factor relative to the upper level factor, and complete the calculation in an Excel table. B2 and B3 only contain one influencing factor, which does not meet the standard for constructing a judgment matrix.
[0115] (3) Consistency test of judgment matrix: check whether all judgment matrices have satisfactory consistency.
[0116] (4) Calculate the combined weight of each evaluation index in layer C. Multiply the weight of each factor in layer B corresponding to each factor in layer C by the weight of each factor in layer C. Finally, the combined weight of each factor in layer C of the susceptibility assessment scheme is obtained.
[0117] (5) Calculate the information weight of each evaluation index. The calculation formula for the information weight of a single evaluation is:
[0118]
[0119] Among them, W i Represents the combined weight of each evaluation index; N i Indicates the distribution of evaluation index X i The number of landslide disasters in a unit area; N represents the number of landslide disasters per unit area; S i Represents the evaluation index X per unit area i The area occupied; S represents the total area per unit area.
[0120] Landslide susceptibility assessment is a process of calculating the landslide susceptibility index by constructing a relationship between landslides and evaluation indicators. In order to reduce the impact of the uncertainty of non-landslide sample selection on the landslide susceptibility evaluation results, the present invention proposes a new non-landslide negative sample selection method, which makes a preliminary prediction of the landslide susceptibility per unit area based on the weighted information model, randomly selects non-landslide sample points that are consistent with the number of landslide samples in the extremely low-prone areas in the prediction results, and completes the landslide susceptibility evaluation per unit area through a machine learning model based on grid units. The landslide susceptibility zoning results based on the weighted information model are used to randomly select 1437 non-landslide points that are consistent with the number of landslide points in ArcGIS Pro3.0 software using the random point generation tool.
[0121] Dung beetle optimizer (DBO) is a new swarm intelligence optimization algorithm that simulates the habits of dung beetles in nature and establishes a search framework based on the "rolling, dancing, foraging, stealing and breeding" model. Since the dung beetle algorithm takes into account both global exploration and local development, and has the characteristics of strong optimization ability and fast convergence, the present invention uses this algorithm to optimize the parameters of the landslide susceptibility evaluation model.
[0122] The XGBoost (Extreme Gradient Boosting, XGBoost) model uses a gradient boosting algorithm to iteratively train a series of decision trees and combine them to form a powerful integrated model. It was proposed by Chen and Guestrin in 2016. It is an integrated model guided by the Boosting integration idea. The structure and decision path of each tree are composed of multiple decision nodes and leaf nodes. Compared with the GBDT model, the XGBoost model uses Taylor's second-order expansion to optimize the loss function, supports multi-threaded parallel computing of the CPU, and adds a regular term to the loss function. The prediction effect is better, the processing speed for large-scale data is faster, and it supports custom loss functions. Therefore, the present invention selects this algorithm to evaluate the landslide susceptibility per unit area.
[0123] XGBoost modeling process: XGBoost learns a decision tree in each iteration, and uses the gradient boosting algorithm to fit the residual of the decision tree prediction result and the true value of the data, so as to establish the final prediction model. The sample prediction result expression of XGBoost is:
[0124]
[0125] Among them, f k represents the kth decision tree; x i represents the feature vector corresponding to sample i; K represents the total number of decision trees; Represents sample x i prediction results.
[0126] The objective function expression of XGBoost is:
[0127]
[0128] Where i represents the i-th sample in the data set; m represents the number of samples; T represents the number of leaf nodes; γ represents the penalty term that will be pruned for each additional leaf, which serves as the regularization parameter of T; ω j represents the leaf weight on the jth leaf node; λ is the regularization parameter of the L2 norm; represents the loss function; Indicates the complexity of the model.
[0129] From this, is the L2 normal form of the leaf node weight vector. The essence of model building is a step-by-step iterative process based on the tree model. The iterative results of the first k-1 trees can be expressed as So for the kth iteration, Can be replaced by The objective function expression after conversion is:
[0130]
[0131] Using Taylor binomial for approximate expansion, the resulting expression is:
[0132]
[0133] in, —Constant term.
[0134] Since the constant term is irrelevant to the kth iteration, it is removed here, and the new objective function expression is:
[0135]
[0136] In the solution, ω and T get the best tree structure, which is defined as: the set I containing samples on the leaf with index j j , the expression is:
[0137]
[0138] For the convenience of description, we define G j and H j Substituting into the objective function,
[0139]
[0140] Among them, G j H represents the cumulative sum of the first-order derivatives of the samples contained in leaf node j; j Represents the cumulative sum of the second-order derivatives of the samples contained in the leaf node j;
[0141] In order to minimize the calculated Obj value, j Find the first-order derivative and set it to 0, and you get the result:
[0142]
[0143]
[0144] The objective function calculates the Obj value based on the corresponding structure of each tree, and obtains the optimal XGBoost model by minimizing the Obj value.
[0145] Support vector machine is a supervised learning model that can classify linear data and nonlinear data. The algorithm uses nonlinear mapping to map training data to a high-dimensional feature space according to certain rules, finds the optimal classification hyperplane in the new feature space, and constructs a linear classifier model with the largest interval in the feature space, so that the classification interval of the training data is maximized. The present invention constructs a support vector machine model, uses nonlinear mapping to map landslide evaluation indicators to a high-dimensional feature space, and searches for the optimal separation hyperplane that distinguishes landslide and non-landslide sample data in the new space.
[0146] For nonlinearly separable data (x i ,y i )(x i (x i ∈R d ) is the evaluation index; y i Indicates whether a landslide occurs or not; i is the number of samples), through nonlinear mapping Map the data into a new high-dimensional feature space. is a hyperplane equation. In this case, the classification interval is equal to 2 / ||ω||. To maximize 2 / ||ω||, we can make ||ω|| 2 Minimum, classification line constraints:
[0147] y i (ω·x i +b)≥1-ε i ,(ε i ≥0)
[0148] Among them, ε i represents the slack variable.
[0149] While finding the classification hyperplane, ε i The smaller the value of is, the better. Therefore, the original problem is transformed into a problem of finding A quadratic programming problem of the minimum value, where C is the penalty factor, is solved to obtain the judgment function:
[0150]
[0151] in, is the kernel function.
[0152] The commonly used kernel functions of support vector machine models are: linear kernel function, polynomial kernel function, radial basis kernel function, neural network kernel function and hyperbolic tangent kernel function.
[0153] Since the Gaussian radial basis kernel function can capture the complex relationship of data and has good effects on both large and small samples, the present invention selects the Gaussian radial basis kernel function to perform the processing and analysis steps.
[0154] Random Forest (RF) is an ensemble algorithm. Its core principle comes from ensemble learning and bagging. It integrates multiple decision trees through the idea of ensemble learning (integrating more models to avoid the model limitations of a single model or a set of parameters), thereby improving the accuracy and stability of the model, reducing the sensitivity to noise and outliers, and avoiding overfitting. The basic unit is generally a decision tree. A decision tree is a tree structure in which each node represents a feature and each leaf node represents a category or a value. The learning process is recursive, dividing the data into subsets according to the selected features until the stopping condition is reached. In the evaluation of landslide susceptibility, random forest uses the high variance between individual decision trees to make the decision trees vote one by one to determine the category membership, and then assign each category according to the majority vote.
[0155] In landslide susceptibility modeling, it is essential to evaluate the model performance. Confusion Matrix (CM) is the main evaluation indicator in the field of machine learning. The confusion matrix measures the prediction effect of the classification problem model and is the most effective way to show the performance of the algorithm. Compared with other methods, the confusion matrix can more easily distinguish whether the model is confused between two categories, that is, one category is wrongly classified into another category.
[0156] The confusion matrix is applied to the evaluation of the landslide susceptibility model of the present invention: True Positive (TP): the number of samples correctly classified as "landslide" (the true category of the sample is landslide, and the result of model recognition is also landslide); False Negative (FN): the number of samples incorrectly classified as non-landslide (the true category of the sample is landslide, but the model recognizes it as non-landslide); False Positive (FP): the number of samples incorrectly classified as "landslide" (the true category of the sample is non-landslide, but the model recognizes it as landslide); True Negative (TN): the number of samples correctly classified as non-landslide (the true category of the sample is landslide, and the model recognizes it as non-landslide).
[0157] The present invention uses the accuracy, recall, precision and F1-Score based on the confusion matrix as auxiliary evaluation indicators of model performance. The calculation expressions of each indicator are as follows:
[0158]
[0159]
[0160]
[0161]
[0162] The value range of accuracy, precision, recall and FI-Score is [0,1]. The closer the accuracy of the model is to l, the higher the overall accuracy of the landslide susceptibility assessment model is; the closer the precision of the model is to l, the lower the probability of misjudgment in the landslide prediction result of the landslide susceptibility assessment model is; the closer the recall of the model is to l, the stronger the prediction ability of the landslide susceptibility assessment model is for landslides; F1-Score combines precision and recall, and the larger its value is, the better the prediction performance of the landslide susceptibility assessment model is.
[0163] In order to better evaluate the accuracy of the landslide susceptibility assessment model, the present invention adopts the receiver operating characteristic (ROC) curve widely used in current research to evaluate the performance of the model. The ROC curve is used to describe the trade-off between the sensitivity and specificity of the classifier, and can intuitively find out the classification model's recognition ability for samples at a certain threshold. The ROC curve uses "sensitivity as the Y-axis" and "specificity as the X-axis." The area enclosed by the ROC curve, the X-axis and X=1 is called AUC. The AUC calculation expression is:
[0164]
[0165]
[0166]
[0167] Among them, P represents all landslide samples; N represents all non-landslide samples.
[0168] The overall prediction accuracy (OPA) is calculated by superimposing the landslide susceptibility zoning layer and the landslide catalog data, comparing the number of landslides in the susceptibility zone and the degree of overlap of the corresponding susceptibility zone. That is, the number of landslides increases with the increase of susceptibility level. The specific expression is:
[0169]
[0170] Among them, P L represents the landslide density in each sub-area, that is, the proportion of landslides in the sub-area to the total number of landslides; P A It indicates the ratio of the zone area to the total unit area area.
[0171] Applying machine learning to landslide susceptibility modeling is a binary classification problem, which includes positive samples and negative samples, i.e., landslides and non-landslides. Considering that machine learning models are more sensitive to data in the interval [0-1], the landslide attributes are set as follows before landslide susceptibility modeling: landslide samples are assigned the label "1" and non-landslide samples are assigned the label "0". In order to avoid the imbalance of model training due to too much or too little sample data, 1437 non-landslide negative sample points with low susceptibility are randomly selected in a 1:1 ratio from the extremely low susceptibility area of the landslide susceptibility results initially predicted based on the weighted information model, which are consistent with the number of landslide points. The Spatial Analyst tool in ArcGIS Pro software is used to obtain the evaluation index attribute values of landslide points and non-landslide points, and a sample database is constructed. The evaluation index obtained by the interval partitioning method based on the geographic detector is used as the input data of the model, and the susceptibility index (landslide: 1, non-landslide: 0) is used as the output item of model prediction for susceptibility modeling.
[0172] The model training data and test data set are divided according to the most widely used 7:3 ratio. 70% of the landslide and non-landslide data are randomly selected from the established sample database to form the model training data as the training set for building the global model; the remaining 30% are used as the model test data and the test set for verifying the model. In order to verify the accuracy of the model, all the training data and test data are summarized as the validation set to verify the accuracy of the model. The unit area is divided into 30m×30m grid units using the Create Fishnet tool in ArcGIS Pro software, and the attribute values of each grid point are extracted to obtain the overall data.
[0173] A landslide susceptibility assessment model based on the XGBoost model was established using randomly selected training set data, and the prediction accuracy of the model was verified using the test set data. The dung beetle algorithm was used to optimize the n_estimators and max_depth parameters of the model to improve the accuracy of the model. n_estimators is the number of trees. Too many trees may make the model too complex and overfit the noise of the training data. Too few trees may not capture the complex relationships in the data. Explore the balance between model overfitting and model performance by selecting the appropriate number of trees. max_depth is the maximum depth of the tree. Each tree in the XGBoost model has a limited depth. Overfitting can be effectively prevented by limiting the depth of the tree. The XGBoost model construction process is as follows: Figure 4 shown.
[0174] In the process of optimizing the XGBoost model with the dung beetle algorithm, the number of algorithm iterations is set to 50, and the model accuracy obtained in each iteration is calculated. After the iteration is completed, the XGBoost model with the optimal parameter combination is obtained.
[0175] The model is rebuilt with the optimal parameters obtained. After the evaluation model training is completed, the overall data set of the unit area is input into the model to evaluate the landslide susceptibility of the unit area, output the prediction results, and finally obtain the landslide probability of all grid cells in the entire unit area. The closer the prediction value corresponding to the grid cell is to 0, the more stable the grid cell is and the smaller the probability of landslide; conversely, the closer the prediction value is to l, the more unstable the grid cell is and the greater the possibility of landslide.
[0176] The DBO-SVM model construction process is as follows Figure 5 As shown in the figure, a landslide susceptibility evaluation model based on the XGBoost model is established using randomly selected training set data, and the prediction accuracy of the model is verified using the test set data. The dung beetle algorithm is used to optimize the C and gamma parameters of the model to improve the accuracy of the model. C is the penalty coefficient, which determines the tolerance of the degree to which the sample data violates the interval. If the C value is very small, the model will have a low penalty for misclassified points; when the C value is large, the model will minimize the number of misclassified samples, resulting in a smaller interval for the decision boundary. The gamma parameter controls the distribution of the data after it is mapped to the high-dimensional space. The model's fit to the training data increases with the increase of the gamma parameter. If the gamma parameter is too large, the model will overfit.
[0177] In the process of optimizing the support vector machine model with the dung beetle algorithm, the number of algorithm iterations is set to 50, and the model accuracy obtained in each iteration is calculated. After the iteration is completed, the support vector machine model under the optimal parameter combination is obtained.
[0178] The model is rebuilt with the optimal parameters obtained. After the evaluation model training is completed, the overall data set of the unit area is input into the model to evaluate the landslide susceptibility of the unit area, output the prediction results, and finally obtain the landslide probability of all grid cells in the entire unit area. The closer the prediction value corresponding to the grid cell is to 0, the more stable the grid cell is and the smaller the probability of landslide; conversely, the closer the prediction value is to l, the more unstable the grid cell is and the greater the possibility of landslide.
[0179] The DBO-RF model construction process is as follows Figure 6As shown in the figure, a landslide susceptibility assessment model based on the RF model is established using randomly selected training set data, and the prediction accuracy of the model is verified using the test set data. The dung beetle algorithm is used to optimize the n_estimators and max_depth parameters of the model to improve the accuracy of the model. n_estimators is the number of trees, and the balance between model overfitting and model performance is explored by selecting the appropriate number of trees. max_depth is the maximum depth of the tree. A tree that is too deep may overfit, while a tree that is too shallow may not fit the data well enough. The deeper the depth, the more complex the model is, and it can better capture the details of the training data, but it is also more prone to overfitting.
[0180] In the process of optimizing the random forest model with the dung beetle algorithm, the number of algorithm iterations is set to 50, and the model accuracy obtained in each iteration is calculated. After the iteration is completed, the random forest model with the optimal parameter combination is obtained.
[0181] The model is rebuilt with the optimal parameters obtained. After the evaluation model training is completed, the overall data set of the unit area is input into the model to evaluate the landslide susceptibility of the unit area, output the prediction results, and finally obtain the landslide probability of all grid cells in the entire unit area. The closer the prediction value corresponding to the grid cell is to 0, the more stable the grid cell is and the smaller the probability of landslide; conversely, the closer the prediction value is to l, the more unstable the grid cell is and the greater the possibility of landslide.
[0182] The model is rebuilt with the optimal parameters obtained. After the evaluation model training is completed, the overall data set of the unit area is input into the model to evaluate the landslide susceptibility of the unit area, output the prediction results, and finally obtain the landslide probability of all grid cells in the entire unit area. The closer the prediction value corresponding to the grid cell is to 0, the more stable the grid cell is and the smaller the probability of landslide; conversely, the closer the prediction value is to l, the more unstable the grid cell is and the greater the possibility of landslide.
[0183] The progress of InSAR technology has promoted the assessment of landslide hazards. In order to better assess the landslide hazards of the research, the present invention combines the deformation results of the unit area synthetic aperture radar data and the landslide susceptibility zoning results based on the DBO-RF model to carry out landslide hazard assessment on the unit area, and quotes the landslide susceptibility dynamic evaluation matrix, and combines the deformation information with the landslide susceptibility zoning map to draw a landslide susceptibility dynamic evaluation map that takes into account the deformation factors in the unit area. Landslide is a dynamic process, and its degree of hazard changes over time. Incorporating SBAS-InSAR deformation information as a dynamic indicator into the landslide susceptibility assessment can effectively reduce the interference of dynamic changes on the assessment results. The method that only considers dynamic evaluation indicators will cause the regional landslide hazard assessment results to be outdated and inaccurate. By combining the deformation information obtained by SBAS-InSAR technology with the landslide susceptibility assessment, the landslide hazard can be dynamically evaluated.
[0184] Landslide susceptibility assessment is to evaluate the possibility of landslide disasters occurring in a unit area based on sample data, and provide a scientific basis for early identification and prevention of landslide disasters. The theoretical essence of landslide susceptibility assessment is to predict the probability of future landslides by comparing the actual situation of previous landslides with the current landslides and the similarity between the current and past landslide situations.
[0185] Landslide evolution is a dynamic process, and the landslide risks in different regions are also different. SBAS-InSAR deformation information can provide important data for dynamic landslide hazard assessment. Based on Sentinel-1A radar images, displacement monitoring data per unit area from January 2023 to December 2023 were obtained. The surface deformation per unit area was obtained through interpolation analysis and resampling processing. A grading experiment was conducted based on the subjective experience method. The best results were selected to divide the deformation into 5 levels: -1 to 1mm / y (stable), -3 to -lmm / y and 1 to 3mm / y (basically stable), -5 to -3mm / y and 3 to 5mm / y (relatively unstable), -7 to -5mm / y and 5 to 7mm / y (unstable), and <-7mm / y and >7mm / y (extremely unstable).
[0186] Combining the SABS-InSAR deformation information and the landslide susceptibility zoning map obtained based on the DBO-RF model, a dynamic evaluation matrix of landslide susceptibility in the study area was established as shown in Table 1.
[0187] Table 1 Dynamic evaluation matrix of landslide susceptibility
[0188]
[0189]
[0190] According to the dynamic evaluation matrix, combined with the landslide susceptibility zoning map and surface deformation information, the dynamic evaluation results of landslide susceptibility in the final study area are obtained.
[0191] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "set", "install", "connect", "connect", and "fix" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood by specific circumstances.
[0192] Although embodiments of the present invention have been shown and described, it will be understood by those skilled in the art that various equivalent changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.
Claims
1. A dynamic evaluation method for landslide susceptibility, Features , the method comprises: Step S1, acquiring landslide information: acquiring landslide catalog information per unit area by applying technical means and statistical data, comprehensively considering the landslide disaster mechanism and disaster-pregnant environment per unit area, analyzing the landslide distribution characteristics per unit area from two aspects of disaster type and distribution law, and extracting 10 evaluation indicators of five categories; Step S2, constructing an evaluation index system; using equal interval classification method, standard deviation method, quantile method, geometric interval method, natural break point method and ChiMerge discretization method to divide the data-based evaluation index into intervals, using geographic detectors to calculate the evaluation index q value obtained by different interval division methods, and selecting the evaluation index interval division method with the largest q value to achieve reasonable interval division of the evaluation index; selecting the evaluation index that passes the multicollinearity test to make a preliminary prediction of regional landslide susceptibility based on the weighted information volume model, and randomly selecting non-landslide points that are consistent with the number of landslide points in the extremely low susceptibility area of the weighted information volume model as non-landslide negative samples, and the selection of negative samples meets the requirements of randomness and low susceptibility; Step S3, machine learning model construction; integrating historical landslide positive samples and non-landslide negative samples to form a training sample set for unit area regional landslide susceptibility modeling, introducing the dung beetle algorithm to optimize the parameters of the XGBoost model, the support vector machine model, and the random forest model, selecting the optimal parameters to establish a machine learning landslide susceptibility prediction model, completing the landslide susceptibility evaluation of the unit area, drawing a landslide disaster susceptibility zoning map, performing zoning statistics and machine learning model accuracy evaluation, comparing the results of the XGBoost model, the support vector machine model, and the random forest model from three aspects: confusion matrix, statistical parameters, and receiver operating curve, and analyzing the accuracy of the unit area regional landslide susceptibility prediction model; Step S4, dynamic evaluation of landslide susceptibility; based on the ascending orbit Sentinel-1A image data of the unit area, the SBAS-InSAR technology is used to extract the surface deformation information, and the spatial distribution characteristics of the landslide disasters in the unit area are analyzed in combination with the deformation information; the application and performance of the DBO-XGBoost model, the DBO-SVM model, and the DBO-RF model in the prediction results of landslide susceptibility in the unit area are compared, and the landslide susceptibility result map obtained by the DBO-RF model and the SABS-InSAR deformation information are combined to establish a dynamic evaluation matrix of landslide susceptibility in the unit area, and perform dynamic evaluation of landslide susceptibility in the unit area; The equal-interval classification method uses equal intervals to divide the range of the evaluation index into several sub-ranges of equal size, and uses the reclassification tool in ArcGIS Pro software to perform the equal-interval classification interval division of the evaluation index raster layer, and the number of intervals is set to 6 based on the subjective experience method; The natural break point method is used to automatically classify the intervals of the evaluation index in ArcGIS Pro software, and the natural break point classification method interval division of the evaluation index raster layer is performed using the reclassification tool in ArcGIS Pro software, and the number of intervals is set to 6 based on the subjective experience method; The quantile method uses the reclassification tool in ArcGIS Pro software to perform quantile classification interval division of the evaluation index raster layer, and the number of intervals is set to 6 based on subjective experience; The geometric interval method uses the reclassification tool in ArcGIS Pro software to perform geometric interval classification interval division of the evaluation index raster layer, and the number of intervals is set to 6 based on subjective experience; The standard deviation method uses the reclassification tool in ArcGIS Pro software to divide the evaluation index raster layer into standard deviation classification intervals, and the interval is set to 1 times; The ChiMerge discretization method divides the data-type evaluation indicators into intervals. In IBM SPSS Statistics 26.0 software, the Chi-square binning tool is used to perform ChiMerge discretization on the continuous evaluation indicators of elevation, slope, and terrain relief, and the threshold is set to 0.05; The geographic detector is used to reveal the driving force by calculating the spatial heterogeneity of the evaluation index, and the differentiation and factor detectors are selected to detect the degree of explanation of the landslide. The expression is: Where i represents the number of layers of landslide evaluation index, i = 1, 2, ..., m; N i represents the number of units in the i-th layer of the landslide evaluation index; N represents the number of units in the entire area of the evaluation index; represents the variance of the landslide Y value corresponding to the i-th layer of landslide evaluation index; σ 2 It represents the variance of the Y value of the landslide corresponding to the evaluation index in the whole area; SL represents the sum of the divisions within the layer; ST represents the total variance of the whole area; The multicollinearity test is used to find the linear or nonlinear relationship between variables. When TOL>0.1 and VIF<10, there is no multicollinearity between the evaluation indicators. The calculation expression is: in, represents the coefficient of determination, Tolerances represents the tolerance, and VIF represents the variance inflation factor; The collinearity diagnostic tool in the linear analysis of IBM SPSS Statistics 26.0 software was used to calculate the VIF and TOL of the landslide susceptibility evaluation indicators per unit area. The evaluation indicators with high collinearity were excluded. The three nominal evaluation indicators and the seven numerical evaluation indicators with the optimal interval division method obtained by the geographic detector model did not show strong collinearity and were used as input variables to build a landslide susceptibility prediction model. The information volume model is calculated as follows: Among them, I(y,x 1 ,x 2 ,…x n ) represents the evaluation index combination x 1 ,x 2 ,…,xn provides information on geological hazards; P(y|x 1 ,x 2 ,…,x n ) represents the evaluation index x 1 ,x 2 ,…,xn combination conditions, the probability of landslide disasters; P(y) represents the probability of landslide disasters; Represents the evaluation index x 1 In the presence of 2 The amount of information provided on the landslide susceptibility assessment results; when I(y,x 1 ,x 2 ,…x n )>0, indicating that the factor combination is conducive to the occurrence of landslide disasters; otherwise, I(y,x 1 ,x 2 ,…x n )<0, indicating that the factor combination is not conducive to the occurrence of landslide disasters; The information value provided by each evaluation index at different levels on landslide disasters is calculated, and the expression is: Among them, P(x i ,D) represents the probability of appearing in a certain state classification level of the evaluation index xi; P(x i ) represents the probability of landslide hazard distribution; The calculation expression of the information amount of each evaluation index is: Among them, N i Indicates that under a certain evaluation index and a certain grading state (x i ) represents the number of geological disasters; N represents the total number of geological disasters occurring in the entire unit area; S i Indicates a certain grading status of an evaluation factor (x i ) of the grid; S—the number of grid cells in the entire unit area; The calculation expression for the total amount of information in each grid cell is: Where n represents the number of evaluation indicators involved in landslide susceptibility assessment; The non-landslide negative samples are used to make a preliminary prediction of the landslide susceptibility per unit area by using a weighted information model, and non-landslide sample points that are consistent with the number of landslide samples are randomly selected in the extremely low-susceptibility area in the prediction results, and the landslide susceptibility evaluation per unit area is completed by a machine learning model based on grid units, and non-landslide points that are consistent with the number of landslide points are randomly selected by using a random point generation tool in ArcGIS Pro3.0 software using the landslide susceptibility zoning result of the weighted information model; The XGBoost model uses randomly selected training set data to establish a landslide susceptibility evaluation model based on the XGBoost model, uses test set data to verify the prediction accuracy of the model, and uses the dung beetle algorithm to optimize the n_estimators and max_depth parameters of the model to improve the accuracy of the model. n_estimators represents the number of trees, and max_depth represents the maximum depth of the tree. Each tree in the XGBoost model has a limited depth, and overfitting is prevented by limiting the depth of the tree. In the process of optimizing the XGBoost model by the dung beetle algorithm, the number of algorithm iterations is set to 50 times, and the model accuracy obtained by each iteration is calculated. After the iteration is completed, the XGBoost model under the optimal parameter combination is obtained; The random forest model uses randomly selected training set data to establish a landslide susceptibility evaluation model based on the random forest model, uses test set data to verify the prediction accuracy of the model, uses the dung beetle algorithm to optimize the n_estimators and max_depth parameters of the model to improve the accuracy of the model, and in the process of optimizing the random forest model by the dung beetle algorithm, the number of algorithm iterations is set to 50 times, and the model accuracy obtained by each iteration is calculated; after the iteration is completed, the random forest model under the optimal parameter combination is obtained, and the model is rebuilt with the obtained optimal parameters. After the evaluation model training is completed, the overall data set of the unit area is input into the model, the landslide susceptibility evaluation of the unit area is performed, and the prediction results are output, and finally the landslide occurrence probability of all grid cells in the entire unit area is obtained; the closer the prediction value corresponding to the grid cell is to 0, the more stable the grid cell is, and the smaller the probability of landslide; on the contrary, the closer the prediction value is to 1, the more unstable the grid cell is, and the greater the possibility of landslide.
2. A method for dynamic evaluation of landslide susceptibility as claimed in claim 1, It is characterized in that The five categories are topography, geological structure, meteorology and hydrology, environmental conditions, and human activities; the 10 evaluation indicators are elevation, slope, slope aspect, terrain undulation, lithology, distance from rivers, land use, distance from roads, POI kernel density, and human footprint index.
