A landslide susceptibility evaluation method based on a comprehensive weighting coupling model

Abstract

The present invention discloses a landslide susceptibility evaluation method based on a comprehensive weighting coupling model, which relates to the field of landslide disaster prediction and specifically includes the following steps: selecting a research area to be studied, determining a sample training set and a sample validation set; obtaining landslide influencing factors through multi-source data of the research area to be studied; based on the sample training set, performing collinearity analysis and factor importance analysis on the landslide influencing factors to obtain target landslide influencing factors; based on the target landslide influencing factors, using the comprehensive weighting coupling model to predict landslide susceptibility and obtaining a landslide susceptibility index; classifying the risk level of the landslide susceptibility index, and using the ROC curve and the sample validation set to evaluate the accuracy of the landslide susceptibility prediction result, thereby completing the landslide susceptibility evaluation based on the comprehensive weighting coupling model.

Description

Technical Field

[0001] The present invention belongs to the field of landslide disaster prediction, and particularly relates to a landslide susceptibility evaluation method based on a comprehensive weighting coupling model. Background Art

[0002] Landslides are one of the most frequent natural disasters in the world. China is also one of the countries most severely affected by landslide disasters. Among them, the landslide disasters triggered by major earthquakes are the most serious, with the number often reaching thousands. The spatial distribution and the area affected by disturbances reach hundreds of square kilometers, and the disaster impact is significant, which is the primary problem faced by the safety of mountainous towns at present. In areas with a high risk of landslides, drawing a landslide susceptibility map is an important means to estimate the probability of landslide occurrence, providing important scientific basis for geological disaster prediction and assessment, the formulation of disaster prevention and mitigation strategies, and land use planning, etc.

[0003] In the current evaluation model methods, in order to overcome the subjectivity of traditional regional landslide spatial prediction methods and improve the prediction accuracy of LSM, different non-deterministic models such as statistical methods and machine learning methods have been continuously applied to landslide susceptibility evaluation. However, as the complexity of the model increases, the overall performance has not been significantly improved. Different types of prediction methods have their own advantages and disadvantages, but no method has been proven to have higher accuracy and better universality in the existing relevant research. In addition, although there is a certain trend in the integrated model based on a single model combination, the complexity of the evaluation method model cannot be blindly pursued, and the combination of different methods often does not consider the occurrence of landslides itself.

[0004] Therefore, there is an urgent need for a landslide susceptibility evaluation method based on a comprehensive weighting coupling model to solve the deficiencies in the prior art. Summary of the Invention

[0005] In order to improve the prediction accuracy and usability of landslides in the current single evaluation model and ordinary coupling model, the present invention proposes a landslide susceptibility evaluation method based on a comprehensive weighting coupling model to improve the interpretability, objectivity and accuracy of regional landslide prediction, making the landslide susceptibility evaluation method more in line with the background conditions of landslide occurrence and more scientific and reasonable.

[0006] To achieve the above object, the present invention provides a landslide susceptibility evaluation method based on a comprehensive weighting coupling model, specifically including the following steps:

[0007] Select the area to be studied, and determine the sample training set and the sample validation set;

[0008] Obtain landslide influencing factors through multi-source data of the area to be studied;

[0009] Based on the sample training set, perform collinearity analysis and factor importance analysis on the landslide influencing factors to obtain the target landslide influencing factors;

[0010] Based on the target landslide influencing factors, use the comprehensive weighting coupling model to predict the landslide susceptibility and obtain the landslide susceptibility index;

[0011] Classify the landslide susceptibility index according to the degree of danger, and use the ROC curve and the sample validation set to evaluate the accuracy of the landslide susceptibility prediction result, and complete the landslide susceptibility evaluation based on the comprehensive weighting coupling model.

[0012] According to a landslide susceptibility evaluation method based on a comprehensive weighting coupling model provided by the present invention, it is determined that the sample training set and the sample validation set include:

[0013] Based on Google Earth images and high-resolution remote sensing images, combined with on-site investigations, obtain a landslide catalog map;

[0014] Resample the landslide catalog map to obtain landslide samples and non-landslide samples and divide them into the sample training set and the sample validation set.

[0015] According to a landslide susceptibility evaluation method based on a comprehensive weighting coupling model provided by the present invention, through the multi-source data of the area to be studied, the landslide influencing factors obtained include:

[0016] Obtain the multi-source data of the area to be studied, wherein the multi-source data includes topographic and geological data, seismic data, meteorological data and land use data;

[0017] Rasterize the multi-source data to obtain the landslide influencing factors, wherein the landslide influencing factors include slope, terrain relief, aspect, terrain curvature, improved hydrological index, stratigraphic rock group, fracture zone density, land use, annual average rainfall, river incision density, distance from highway, seismic intensity and distance from epicenter.

[0018] According to a landslide susceptibility evaluation method based on a comprehensive weighting coupling model provided by the present invention, based on the sample training set, perform collinearity analysis and factor importance analysis on the landslide influencing factors to obtain the target landslide influencing factors including:

[0019] Based on the sample training set, use the correlation matrix to analyze the correlation between the landslide influencing factors and remove the data redundancy between the landslide influencing factors;

[0020] Analyze the landslide influencing factors after removing data redundancy through the variance inflation factor and tolerance to obtain the target landslide influencing factors.

[0021] A landslide susceptibility evaluation method based on a comprehensive weighting coupling model provided by the present invention, the correlation of the landslide influencing factors is as follows:

[0022]

[0023] where r xy is a normalized parameter, n is the number of pixels of the factor, x i is the i-th pixel of factor x, y i is the i-th pixel of factor y, is the average of all pixels of factor x, is the average of all pixels of factor y.

[0024] A landslide susceptibility evaluation method based on a comprehensive weighting coupling model provided by the present invention, based on the target landslide influencing factors, using the comprehensive weighting coupling model to predict landslide susceptibility, and obtaining the landslide susceptibility index includes:

[0025] Based on the target landslide influencing factors, using the frequency ratio method and the statistical index method for reclassification to obtain the final landslide influencing factors;

[0026] Construct the comprehensive weighting coupling model, where the comprehensive weighting coupling model includes an evidence weight model, a BPNN-GWO model, and a fusion model;

[0027] Input the target landslide influencing factors into the comprehensive weighting coupling model and output the landslide susceptibility index.

[0028] A landslide susceptibility evaluation method based on a comprehensive weighting coupling model provided by the present invention, based on the target landslide influencing factors, using the frequency ratio method and the statistical index method for reclassification to obtain the final landslide influencing factors includes:

[0029] Using the frequency ratio method and the statistical index method to calculate the binary statistical values of the pre-classification of the target landslide influencing factors to form several classification areas;

[0030] Based on the binary statistical values, determine new partition breakpoints by drawing curves;

[0031] Reclassify the target landslide influencing factors through the partition breakpoints to obtain the final landslide influencing factors.

[0032] A landslide susceptibility evaluation method based on a comprehensive weighting coupling model provided by the present invention, inputting the target landslide influencing factors into the comprehensive weighting coupling model and outputting the landslide susceptibility index includes:

[0033] Based on the target landslide influencing factors and the secondary state classification criteria of the target landslide influencing factors, using the evidence weight model, obtain the secondary classification weights;

[0034] Based on the target landslide influencing factors, using the BPNN-GWO model, obtain the factor weights;

[0035] Through the fusion model, multiply the secondary classification weights and the factor weights to obtain the slope susceptibility index.

[0036] According to a landslide susceptibility evaluation method based on a comprehensive weighting coupling model provided by the present invention, the evidence weight model is:

[0037]

[0038] W f =W + -W -

[0039]

[0040] Wherein, W + is the positive weight of the x class of the y factor, NLP xy is equal to the total number of landslide disaster points within the x class of the y factor, TNLP y is the sum of landslide points within each y factor, NSP xy represents the number of grid cells in a stable state in the x class of the y factor, TNSP y is the total number of grid cells of the y factor under stable conditions, W - is the negative weight to be assigned when the x class of the factor y does not exist, NLP ny is the number of landslide points in the other n categories of the factor y, NSP ny y represents the number of stable grid cells in the other n categories of the y factor, W f is the comparison weight, LSI WOfE is the landslide hazard index, W fxy is the final weight of the x class in the conditional factor y, and n is the total number of conditional factors.

[0041] According to a landslide susceptibility evaluation method based on a comprehensive weighting coupling model provided by the present invention, based on the target landslide influencing factors, using the BPNN-GWO model, obtaining the factor weights includes:

[0042] Based on collinearity analysis and importance analysis, obtain the final landslide influencing factors in the target area, and through on-site sampling and remote sensing visual interpretation, obtain landslide and non-landslide sample data;

[0043] Establishing feature data and label data based on the final landslide impact factor of the target area and the landslide and non-landslide sample data;

[0044] Constructing a BP neural network model based on the feature data and the label data, and initializing parameter weights and thresholds of the BP neural network model;

[0045] Obtaining a number of particles, determining the deviation between the output of the output layer of the GWO algorithm and the target value input by the neural network model, and adjusting the parameter weights and thresholds, thereby determining the optimal parameter combination and obtaining the position of the optimal particle;

[0046] Determining optimal parameters according to the position of the optimal particle;

[0047] A prediction model is established based on the optimal parameters, and the factor weight is obtained through the target landslide influencing factor.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] The present invention uses historical landslide cataloging data, combines expert experience and literature research, and comprehensively utilizes multi-source heterogeneous data such as field surveys, remote sensing, and basic geography to preliminarily select multiple regional landslide disaster influencing factors; analyzes the correlation of each influencing factor through a correlation matrix, and then evaluates the prediction ability of each influencing factor in combination with the Relief-F algorithm, and finally determines the landslide influencing factor to be selected, determines the weight of the category within each factor through WOE, and carries out landslide susceptibility evaluation mapping and result analysis based on the BPNN model optimized by GWO, and verifies the model performance using the receiver operating characteristic curve (ROC) and the distribution frequency statistics of landslides within the susceptibility level. The prediction of regional landslide disaster susceptibility based on the WOE-BPNN-GWO coupling model can effectively improve the prediction accuracy. Compared with traditional methods, the research process proposed by the present invention is more in line with the disaster characteristics of landslide disasters, and creatively performs weighted fusion processing within the category and between factors of landslide disaster influencing factors. The logic is clear and feasible, and the evaluation results are more reliable. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0051] Figure 1 This is a flow chart of a landslide susceptibility evaluation method based on a comprehensive weighted coupling model according to an embodiment of the present invention;

[0052] Figure 2 A schematic diagram of the location of the area to be evaluated and the information of the landslide disaster points proposed in an embodiment of the present invention;

[0053] Figure 3 The grading statistical chart of different factor states based on the binary statistical method proposed in the embodiment of the present invention. Among them, (a) is the grading statistical chart of the slope factor state based on the binary statistical method, (b) is the grading statistical chart of the aspect factor state based on the binary statistical method, (c) is the grading statistical chart of the plane curvature factor state based on the binary statistical method, and (d) is the grading statistical chart of the profile curvature factor state based on the binary statistical method;

[0054] Figure 4 The ROC curve graph of landslide susceptibility evaluation using two models proposed in the embodiment of the present invention. Among them, (a) is the ROC curve graph of landslide susceptibility evaluation using the BPNN model, and (b) is the ROC curve graph of landslide susceptibility evaluation using the WOE - BPNN - GWO coupling model. Detailed implementation manners

[0055] It should be noted that, without conflict, the embodiments in the present invention and the features in the embodiments can be combined with each other. The following will refer to the accompanying drawings and combine with the embodiments to detail the present application.

[0056] It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer - executable instructions. And although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in a different order than here.

[0057] In order to obtain more reliable evaluation results, starting from the background environmental conditions of landslide occurrence, screen and determine each influencing factor and its linear importance in landslide occurrence in each interval, then combine a machine learning method with better robustness to calculate the non - linear relationship between each landslide influencing factor, and further construct a regional landslide susceptibility coupling model with a clear logical process and capable of reflecting the basic conditions of regional landslide occurrence. Based on this, the present invention proposes a landslide susceptibility evaluation method based on a comprehensive weighted coupling model, which comprehensively considers the characteristics of statistical models and machine learning models, and fuses the binary statistical model weights of evidence (WOE) and a machine model based on a gray wolf optimizer (GWO) backpropagation neural network (BPNN) to obtain a coupling model, and obtains an evaluation scheme considering the linear and non - linear relationships within and between each landslide background factor.

[0058] As Figure 1 shown, this embodiment provides a landslide susceptibility evaluation method based on a comprehensive weighted coupling model, including the following steps:

[0059] Select the area to be studied and determine the sample training set and the sample validation set;

[0060] Specifically, select the study area prone to geological disasters such as landslides, collapses and debris flows as the area to be studied, obtain the spatial distribution survey samples of landslide disasters in the area to be studied, determine the non-landslide samples and landslide samples, and determine the sample training set and the sample validation set.

[0061] Obtain the landslide influencing factors through the multi-source data of the area to be studied;

[0062] Specifically, the multi-source data includes topographic and geomorphic data, hydrogeological data, rainfall data and land use data; the landslide influencing factors include slope, terrain undulation degree, slope aspect, terrain curvature (3 types), improved hydrological index, stratigraphic lithologic group, fracture zone density, land use, annual average rainfall, river incision density, distance from highway, seismic intensity and distance from earthquake epicenter.

[0063] Based on the sample training set, conduct collinearity analysis and factor importance analysis on the landslide influencing factors to obtain the target landslide influencing factors;

[0064] Specifically, conduct collinearity analysis through Pearson correlation, differential correlation coefficient and tolerance index; conduct factor importance analysis through information gain ratio; the target slope influencing factor is to use multiple collinearity analysis to judge whether there is pairwise correlation between the landslide influencing factors, eliminate the influencing factors with strong correlation, and through the prediction ability analysis of the influencing factors by the Relief-F method, finally obtain the influencing factors that are relatively independent of each other and play a significant role in the occurrence of landslides.

[0065] Based on the target landslide influencing factors, use the comprehensive weighting coupling model to predict the landslide susceptibility and obtain the output landslide susceptibility index;

[0066] Specifically, the comprehensive weighting coupling model is constructed by coupling the WOE algorithm and the BPNN-GWO model.

[0067] Classify the danger degree of the landslide susceptibility index, and use the ROC curve and the sample validation set to evaluate the accuracy of the landslide susceptibility prediction results, and complete the landslide susceptibility evaluation based on the comprehensive weighting coupling model.

[0068] The present invention utilizes multi-source data, combines expert experience and literature research, and comprehensively utilizes multi-source heterogeneous data such as field surveys, remote sensing, and basic geography to initially select multiple regional landslide disaster impact factors; analyzes the correlation of each impact factor through a correlation matrix, then combines the Relief-F algorithm to evaluate the prediction ability of each impact factor, and finally determines the landslide impact factors to be selected, and conducts regional landslide disaster susceptibility prediction based on the WOE-BPNN-GWO coupling model, which can effectively improve the prediction accuracy. Compared with traditional methods, the research process proposed by the present invention is more in line with the disaster-forming characteristics of landslide disasters, creatively conducts weighted fusion processing within and between the categories of landslide disaster impact factors, with clear logic and feasibility, and the evaluation results are more reliable.

[0069] Optionally, determining the sample training set and the sample validation set includes:

[0070] Based on Google Earth images and high-resolution remote sensing images, combined with field surveys, a landslide inventory map is obtained;

[0071] The landslide inventory map is resampled to obtain landslide samples and non-landslide samples and divided into a sample training set and a sample validation set.

[0072] Specifically, through visual interpretation of Google Earth images, supplemented by high-resolution remote sensing images such as UAV images, GF-2, and Sentinel-2, and combined with field surveys of geological disasters along the line, a landslide inventory map of the study area is obtained; through resampling of the landslide inventory map, non-landslide samples and landslide samples are randomly selected in the study area and segmented. Among them, 70% of the samples are used for training, and 30% of the samples are used for validation.

[0073] Optionally, obtaining landslide impact factors through multi-source data of the area to be studied includes:

[0074] Obtain multi-source data of the area to be studied, where the multi-source data includes topographic and geological data, seismic data, meteorological data, and land use data;

[0075] The multi-source data is rasterized to obtain landslide impact factors, where the landslide impact factors include slope, terrain relief, aspect, terrain curvature, improved hydrological index, lithostratigraphic unit, fracture zone density, land use, annual average rainfall, river incision density, distance from highway, seismic intensity, and distance from epicenter.

[0076] Specifically, based on the characteristics of landslide disasters in the area to be studied and existing research data, determine the disaster-causing factors and disaster-prone factors related to landslide occurrences, collect multi-source data related to the study area, including topographic and geomorphic data, hydrogeological data, rainfall data, and land use data, and rasterize the multi-source data to unify the spatial resolution, coordinate system, and value range of each data, and determine the state classification of continuous variables.

[0077] Optionally, based on the sample training set, perform collinearity analysis and factor importance analysis on the landslide influencing factors to obtain the target landslide influencing factors, including:

[0078] Based on the sample training set, use the correlation matrix to analyze the correlation between landslide influencing factors and remove the data redundancy between landslide influencing factors;

[0079] Analyze the landslide influencing factors after removing data redundancy through the variance inflation factor and tolerance to obtain the target landslide influencing factors.

[0080] Specifically, use the correlation matrix to analyze the pairwise correlation of each influencing factor and remove the data redundancy caused by the correlation between factors; further analyze the multiple collinearity of landslide condition factors through two indicators, the variance inflation factor VIF and tolerance. VIF evaluates the change in the standard deviation between landslide influencing factors and is the reciprocal of the tolerance value, indicating that the smaller the change, the lower the multiple collinearity.

[0081] Optionally, the correlation of the landslide influencing factors is:

[0082]

[0083] Among them, r xy is a normalized parameter, which is the number of pixels of the factor, x i is the i-th pixel of factor x, y i is the i-th pixel of factor y, is the average of all pixels of factor x, is the average of all pixels of factor y.

[0084] Optionally, based on the target landslide influencing factors, use the comprehensive weighting coupling model to predict the landslide susceptibility, and obtain the landslide susceptibility index, including:

[0085] Based on the target landslide influencing factors, use the frequency ratio method and the statistical index method for reclassification to obtain the final landslide influencing factors;

[0086] Construct a comprehensive weighting coupling model, where the comprehensive weighting coupling model includes the evidence weight model, the BPNN-GWO model, and the fusion model;

[0087] Input the target landslide influencing factors into the comprehensive weighting and coupling model, and output the landslide susceptibility index.

[0088] Specifically, use the frequency ratio method and the statistical index method to determine the state classification criteria for continuous influencing factors and conduct corresponding state classifications; use the weight of evidence method to calculate the influence weights of each state classification within each influencing factor on landslides. The greater the weight of each state classification of the influencing factor, the more important the influencing factor is for landslide susceptibility; re-rasterize each influencing factor based on the state factor weights to complete the construction of a new dataset considering the internal importance of landslide influencing factors.

[0089] Construct a BP neural network model for landslide disaster susceptibility, and use the grey wolf optimization algorithm to optimize the parameters of the BPNN model; use the optimized model BPNN-GWO to analyze the importance of landslide disaster influencing factors in the entire study area to obtain the weight values of landslide disaster susceptibility influencing factors in the study area.

[0090] After determining the secondary classification weights and factor weights of the evidence influencing factors, multiplying the two can obtain the landslide susceptibility index determined by the weight of evidence and BP neural network coupling model (WOE-BPNN-GWO).

[0091] Optionally, based on the target landslide influencing factors, use the frequency ratio method and the statistical index method for reclassification to obtain the final landslide influencing factors including:

[0092] Use the frequency ratio method and the statistical index method to calculate the binary statistical values of the pre-classification of the target landslide influencing factors to form several classification areas.

[0093] Based on the binary statistical values, determine the new partition breakpoints by drawing curves.

[0094] Reclassify the target landslide influencing factors through the partition breakpoints to obtain the final landslide influencing factors.

[0095] Specifically, use two binary statistical methods, the frequency ratio method FR and the statistical index method SI, to calculate the binary statistical values of the pre-classification of the landslide susceptibility evaluation factors to form multiple classification areas; after calculating each factor partition, determine the new partition breakpoints by drawing curves, and then reclassify the target landslide influencing factors to obtain the final landslide influencing factors.

[0096] Optionally, input the target landslide influencing factors into the comprehensive weighting and coupling model, and the output landslide susceptibility index includes:

[0097] Based on the target landslide influencing factors and the secondary state classification criteria of the target landslide influencing factors, use the weight of evidence model to obtain the secondary classification weights.

[0098] Based on the target landslide influencing factors, use the BPNN-GWO model to obtain the factor weights;

[0099] Through the fusion model, multiply the secondary classification weights and the factor weights to obtain the slope susceptibility index.

[0100] Specifically, based on the selected factors and the determined secondary state classification criteria of the factors, use the WOE model to calculate the weights of each state category within each influencing factor to obtain the secondary classification weights; use the BPNN model optimized by GWO to calculate the influence weights of the landslide influencing factors on the occurrence of landslides to obtain the factor weights; after determining the secondary classification weights and the factor weights of the influencing factors, determine the landslide susceptibility index through the fusion model.

[0101] Optionally, the weight-of-evidence model is:

[0102]

[0103] W f =W + -W -

[0104]

[0105] Wherein, W + is the positive weight of the x class of the y factor, NLP xy is equal to the total number of landslide disaster points within the x class of the y factor, TNLP y is the sum of the landslide points within each y factor, NSP xy represents the number of grid cells in a stable state in the x class of the y factor, TNSP y is the total number of grid cells of the y factor under stable conditions, W - is the negative weight to be assigned when the x class of the factor y does not exist, NLP ny is the number of landslide points in the other n classes of the factor y, NSP ny y represents the number of stable grid cells in the other n classes of the y factor, W f is the comparison weight, LSI WOfE is the landslide hazard index, W fxy is the final weight of the x class in the conditional factor y, and n is the total number of conditional factors.

[0106] Optionally, based on the target landslide influencing factors, using the BPNN-GWO model to obtain the factor weights includes:

[0107] Based on the collinearity analysis and importance analysis, obtain the final landslide influencing factors in the target area, and through field sampling and remote sensing visual interpretation, obtain the landslide and non-landslide sample data;

[0108] Based on the final landslide impact factors in the target area and landslide and non-landslide sample data, establish characteristic data and label data;

[0109] Based on the characteristic data and label data, construct a BP neural network model, and initialize the parameter weights and thresholds of the BP neural network model;

[0110] Obtain a number of particles, calculate the deviation between the output of the output layer through the GWO algorithm and the target value input by the neural network model, and adjust the parameter weights and thresholds, thereby determining the optimal parameter combination and obtaining the position of the optimal particle;

[0111] Determine the optimal parameters based on the position of the optimal particle;

[0112] Based on the optimal parameters, establish a prediction model, and obtain the factor weights through the target landslide impact factors.

[0113] Specifically, for the BPNN-GWO model, the specific steps are as follows:

[0114] (1) Obtain the final landslide impact factors in the target area through collinearity analysis and importance analysis, and obtain landslide and non-landslide sample data through on-site sampling and remote sensing visual interpretation. Then, based on the final landslide impact factors in the target area and the landslide and non-landslide sample data, establish training characteristic data and label data. Among them, the 12 impact factor values at the corresponding positions of the landslide samples or non-landslide samples are used as characteristic data and used as independent variables; whether it is a landslide (1 or 0) is used as the dependent variable for model training;

[0115] (2) Randomly construct a basic BP neural network model based on the sample data, and initialize the parameter weights and thresholds of the model;

[0116] (3) Randomly generate n particles, where the position of each particle represents a set of model parameters;

[0117] Use the GWO algorithm to calculate the deviation between the output of the output layer and the y value input by the neural network model, use the algorithm to correspondingly adjust parameters such as weights and thresholds, and determine the optimal parameter combination according to the distance of deviation. Update the position of the particle according to the GWO algorithm. If the requirements (maximum number of iterations) are met, stop the search;

[0118] (4) Use GWO to determine the position information of the optimal particle, and then determine the optimal number of hidden layer nodes;

[0119] (5) Use the obtained optimal parameters to establish a prediction model. Use the optimized parameters to construct a prediction model and use this model to calculate the impact factor weights.

[0120] Embodiment

[0121] This embodiment takes the mountains in the severely affected areas of Tibet caused by the Nepal earthquake in the southern part of the typical high-altitude region of the Qinghai-Tibet Plateau as the research area, and provides a landslide susceptibility evaluation method based on a comprehensive weighting coupling model. The following is a description in combination with Figures 1 - 4 as follows:

[0122] Step 1: Determination of the research area; in this example, two counties, namely Nyemo and Gyirong, under the jurisdiction of Shigatse City in the southwest border of China and the southwest of the Tibet Autonomous Region, which were affected by the "4.25" Nepal earthquake in 2015, are selected as the areas to be evaluated. The geographical location is between 84°27′ - 86°37′ east longitude and 27°53′ - 29°19′ north latitude. It is about 157 km long from south to north and about 211 km wide from east to west. The terrain has large undulations, with a maximum relative height difference of more than 7,000 meters, and it is extremely prone to geological disasters such as landslides, collapses, and debris flows.

[0123] Step 2: In this example, based on the visual interpretation of Google images after the 2015 Nepal earthquake, supplemented by remote sensing images such as UAV images, GF-2, and Sentinel-2, and combined with on-site investigations of geological disasters along the line, a total of 1,184 landslides in the research area are obtained. The same number of non-landslide samples are randomly selected in the research area, and the total number of training samples is 2,368. Among them, 70% of the samples are used for training, and 30% of the samples are used for verification.

[0124] Step 3: Obtain the topographic and geological data, seismic data, meteorological data, and land use data of the preprocessed experimental area, and calculate 15 landslide-influencing environmental factors including slope, terrain relief, aspect, terrain curvature (3 types), improved hydrological index, stratigraphic rock groups, fracture zone density, land use, annual average rainfall, river incision density, distance from the road, seismic intensity, and distance from the epicenter under the GIS platform. The continuous factors are pre-classified using the natural break method, and all factors are converted into raster data with a unified range, resolution, and projection coordinate system.

[0125] Step 4: Use the correlation matrix to analyze the pairwise correlations of each influencing factor and remove the data redundancy caused by the correlations between factors. The calculation of the correlation between two factor layers is as follows:

[0126]

[0127] where r xy is a normalized parameter, with a value range from +1 to -1. The closer r xy is to 1, the stronger the correlation between the two. Generally, if the absolute value of the correlation coefficient is greater than 0.5, it is considered that there is a strong correlation, and one is selected from the strongly correlated factors. In this embodiment, terrain relief, curvature, and distance from the epicenter are screened out.

[0128] Furthermore, the variance inflation factor VIF and tolerance are used to analyze the multicollinearity of landslide conditioning factors. VIF evaluates the change in the standard deviation between landslide influencing factors and is the reciprocal of the tolerance value, indicating that the smaller the change, the lower the multicollinearity. The calculation formula is as follows:

[0129]

[0130] where tolerance is the tolerance difference between influencing factors, and R j is the coefficient of determination of the regression of influencing factor j on all other influencing factors.

[0131] The indicators of the remaining 12 landslide susceptibility factors all meet the critical values, and there is no multicollinearity.

[0132] Step 5: Use the average value of the 10-fold cross-validation evaluation criterion of the Relief-F variable selection method to evaluate the quantitative prediction ability of the above 12 landslide influencing factors. The results show that all evaluation criteria are greater than 0, indicating that all influencing factors play a positive role in the evaluation of landslide susceptibility.

[0133] Step 6: Extract the attribute information of the disaster-causing factors within the range of 70% of the training landslide sample locations in grid units to complete the collection and quantification of the sample dataset for the evaluation model.

[0134] Step 7: Use two binary statistical methods, the frequency ratio method FR and the statistical index method SI, to calculate the binary statistical values of the landslide susceptibility evaluation factors for pre-classification to form multiple classification areas. The calculation formula of the frequency ratio method FR is FR ij as follows:

[0135] FR ij = FrX ij / FrY ij

[0136] where FrX ij represents the frequency of landslides observed within the i-th class of factor j, and FrY ij represents the frequency of the percentage of landslides not observed in the i-th class of factor j.

[0137] The calculation formula of the statistical index method SI is W si as follows:

[0138]

[0139] where E xy represents the landslide density in the x-th class of factor y, E represents the landslide density in the entire study area, L xy represents the number of landslides in the x-th class of factor y, P xy the number of grid cells in the x-th class of factor y, L Trepresents the total number of landslides in the entire study area, P T represents the number of grids in the entire study area.

[0140] After calculating each factor sub-region, by drawing curves, new sub-region breakpoints are determined. This example demonstrates with the terrain factor, as shown in Figure 3 (a)-(d). The same applies to other factors, and the results are shown in Table 1.

[0141] Table 1

[0142]

[0143]

[0144] Step 8: Construct a landslide hazard susceptibility evaluation model, including three parts: determining the category weights within each factor using the Weight of Evidence (WOE) model, constructing a BP neural network artificial neural network (BPNN) model with the Grey Wolf Optimization algorithm to determine the weights of factors, and integrating and fusing these two models. The specific steps are as follows:

[0145] S81. Determine the category weights within each factor using the Weight of Evidence (WOE) model;

[0146] Based on the selected factors and the determined secondary state classification criteria of the factors, use the WOE model to calculate the category weights of each state within each influencing factor. The calculation formula of the WOE model is as follows:

[0147] During the experiment, calculate the secondary classification weights within each factor using the landslide sample frequency, as shown in the following formula:

[0148]

[0149]

[0150] Among them, W + is the positive weight of the x class of the y factor, NLP xy is equal to the total number of landslide hazard points within the x class of the y factor, TNLP y is the sum of landslide points within each y factor, NSP xy represents the number of grids in the stable state within the x class of the y factor, TNSP y is the total number of grids of the y factor under stable conditions.

[0151] Among them, W - is the negative weight to be assigned when the x class of the factor y does not exist, NLP ny is the number of landslide points in the other n classes of the factor y, NSP ny y represents the number of stable grids in the other n classes of the y factor.

[0152] From W + and W - The weight of the contrast is obtained from the difference between them, which is called the weight contrast or the final weight, and is expressed as follows: f wherein, W

[0153] W f = W + - W -

[0154]

[0155] wherein, W f is the contrast weight, LSI WOfE is the landslide hazard index, W fxy is the final weight of class x in the conditional factor y, and n is the total number of conditional factors.

[0156] S82. Calculate the influence weight of the landslide influence factors on the occurrence of landslides by using the BPNN model optimized by GWO;

[0157] The BP neural network (BPNN) can be regarded as a model method that simulates the human pattern recognition function and has strong non-linear mapping ability. Each landslide influence factor is normalized to the interval [0, 1] as the input layer of the network, and the output layer corresponds to the specific numerical expression of the danger degree. The weight calculation formula of the input layer to the output layer can be derived as follows:

[0158] To determine the weights of the elements in the input layer, the connection weights between different layers need to be determined. The convergence speed of the BP model is slow and it is easy to fall into the local minimum. Applying the GWO algorithm to the parameter optimization of BPNN, reasonable parameter setting is the key to improving the prediction ability of the model, so as to obtain the optimal model parameter combination. A parameter combination of BPNN is represented by the position information of the particle; through the optimal search of GWO, the position information of the optimal particle is obtained; the optimal number of hidden layer nodes is determined by GWO. The BPNN-GWO model has the following specific steps:

[0159] (1) Establish training feature data and label data. The 12 influence factor values corresponding to the landslide samples or non-landslide samples are used as the feature data and used as independent variables; whether it is a landslide (1 or 0) is used as the dependent variable for model training;

[0160] (2) Randomly construct a basic BP neural network model based on the sample data, and initialize the parameter weights and thresholds of the model;

[0161] (3) Randomly generate n particles, and the position of each particle represents a set of model parameters;

[0162] Calculate the deviation between the output of the output layer calculated using the GWO algorithm and the y value input by the neural network model, and use the algorithm to adjust parameters such as weights and thresholds accordingly. Determine the optimal parameter combination according to the distance of deviation. Update the position of the particle according to the GWO algorithm. If the requirements (maximum number of iterations) are met, stop the search;

[0163] (4) Use GWO to determine the position information of the optimal particle, and then determine the optimal number of hidden layer nodes;

[0164] (5) Establish a prediction model with the obtained optimal parameters. Build a prediction model using the optimized parameters and use this model to calculate the weights of the influencing factors.

[0165] S83. Perform weight coupling, where the coupling method in this implementation is the factor weight that comprehensively considers the category weights within the landslide influencing factors and the influence degree between various factors.

[0166] After determining the secondary classification weights and factor weights of the influencing factors, multiplying the two can obtain the landslide susceptibility index determined by the coupling model of weight of evidence and BP neural network (WOE - BPNN - GWO).

[0167] Step Nine: Landslide susceptibility grading map and result verification;

[0168] Evaluate the calculated landslide susceptibility index using the ROC curve and the landslide validation set respectively, and evaluate the result of the landslide susceptibility zoning using the statistical method and the landslide validation set. Specifically:

[0169] (1) The calculated landslide susceptibility index needs to be classified according to its magnitude for the degree of danger. The classification of the landslide susceptibility index results all adopts the quantile method for classification, divided into five danger zones: higher, high, medium, low, and lower, and then the corresponding landslide susceptibility map is obtained.

[0170] (2) On the landslide validation set, draw an ROC curve for the calculated landslide susceptibility index to verify the accuracy of the present invention. As shown in Figure 4 (b), the AUC (Area under ROC) value reaches 0.96, and the AUC value of the single BPNN model is 0.83. As shown in Figure 4 (a), the result has been significantly improved.

[0171] Further, count the number of landslide validations in each susceptibility level, as shown in Table 2. From the statistical results, landslide disasters are mainly distributed in the extremely high area, and the area ratios of the extremely high and extremely low susceptibility areas are both low. Therefore, the present invention has high accuracy and reliable results, and can be used as a basis for disaster control and prediction.

[0172] Table 2

[0173]

[0174] As described above, it is only a preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present application should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A landslide susceptibility assessment method based on a comprehensive weighted coupling model, characterized in that: The specific steps include: Select the area to be studied and determine the sample training set and sample verification set; Obtaining landslide influencing factors through multi-source data of the area to be studied; Based on the sample training set, collinearity analysis and factor importance analysis are performed on the landslide influencing factors to obtain target landslide influencing factors; Based on the target landslide impact factor, a comprehensive weighted coupling model is used to predict the landslide susceptibility and obtain a landslide susceptibility index, specifically including: based on the target landslide impact factor, a frequency ratio method and a statistical index method are used to perform reclassification to obtain a final landslide impact factor; Constructing the comprehensive weighted coupling model, wherein the comprehensive weighted coupling model includes an evidence weight model, a BPNN-GWO model and a fusion model; Inputting the target landslide impact factor into the comprehensive weighted coupling model, and outputting the landslide susceptibility index; Inputting the target landslide impact factor into the comprehensive weighted coupling model, and outputting the landslide susceptibility index includes: Based on the target landslide impact factor and the secondary state classification standard of the target landslide impact factor, using the evidence weight model, obtaining the secondary classification weight; Based on the target landslide impact factor, the BPNN-GWO model is used to obtain the factor weight, specifically including: obtaining the final landslide impact factor of the target area based on collinearity analysis and importance analysis, and obtaining landslide and non-landslide sample data through field sampling and remote sensing visual interpretation; Establishing feature data and label data based on the final landslide impact factor of the target area and the landslide and non-landslide sample data; Constructing a BP neural network model based on the feature data and the label data, and initializing parameter weights and thresholds of the BP neural network model; Obtaining a number of particles, determining the deviation between the output of the output layer of the GWO algorithm and the target value input by the neural network model, and adjusting the parameter weights and thresholds, thereby determining the optimal parameter combination and obtaining the position of the optimal particle; Determining optimal parameters according to the position of the optimal particle; Establishing a prediction model based on the optimal parameters, and obtaining the factor weights through the target landslide influencing factors; The landslide susceptibility index is obtained by multiplying the secondary classification weight and the factor weight through the fusion model; The landslide susceptibility index is classified according to the degree of danger, and the accuracy of the landslide susceptibility prediction result is evaluated using the ROC curve and the sample verification set, completing the landslide susceptibility evaluation based on the comprehensive weighted coupling model.

2. The landslide susceptibility assessment method based on the comprehensive weighted coupling model according to claim 1 is characterized in that: Determining the sample training set and the sample verification set includes: Based on Google Earth images and high-resolution remote sensing images, combined with field surveys, landslide catalog maps were obtained; The landslide inventory map is resampled to obtain landslide samples and non-landslide samples and divide them into the sample training set and the sample verification set.

3. The landslide susceptibility assessment method based on the comprehensive weighted coupling model according to claim 1 is characterized in that: Obtaining the landslide influencing factors through multi-source data of the area to be studied includes: Acquire multi-source data of the area to be studied, wherein the multi-source data includes topographic and geological data, seismic data, meteorological data and land use data; The multi-source data are rasterized to obtain the landslide influencing factors, wherein the landslide influencing factors include slope, terrain undulation, slope aspect, terrain curvature, improved hydrological index, stratigraphic rock group, fault zone density, land use, annual average rainfall, river cutting density, distance from highway, earthquake intensity and distance from epicenter.

4. The landslide susceptibility assessment method based on a comprehensive weighted coupling model according to claim 1, characterized in that: Based on the sample training set, collinearity analysis and factor importance analysis are performed on the landslide influencing factors to obtain the target landslide influencing factors, including: Based on the sample training set, using a correlation matrix to analyze the correlation between the landslide influencing factors, and removing data redundancy between the landslide influencing factors; The landslide impact factor after removing data redundancy is analyzed by variance expansion coefficient and tolerance to obtain the target landslide impact factor.

5. The landslide susceptibility assessment method based on the comprehensive weighted coupling model according to claim 4 is characterized in that: The correlation of the landslide influencing factors is: Among them, r xy is a normalized parameter, n is the number of pixels of the factor, x i is the i-th pixel of factor x, y i is the i-th pixel of factor y, is the average of all pixels of factor x, is the average of all pixels of factor y.

6. The landslide susceptibility assessment method based on a comprehensive weighted coupling model according to claim 1, characterized in that: Based on the target landslide impact factor, reclassification is performed using the frequency ratio method and the statistical index method to obtain the final landslide impact factor, which includes: Utilizing the frequency ratio method and the statistical index method, calculating the binary statistical values ​​of the target landslide influencing factors pre-classified into a plurality of classification areas; Based on the binary statistical value, determine a new partition breakpoint by drawing a curve; The target landslide impact factor is reclassified by the partition breakpoints to obtain the final landslide impact factor.

7. The landslide susceptibility assessment method based on a comprehensive weighted coupling model according to claim 1, characterized in that: The evidence weight model is: IN f =In + -IN - Among them, W + is the positive weight of class x for the y factor, NLP xy The total number of landslide hazard points in the x category equal to the y factor, TNLP y is the sum of landslide points within each y factor, NSP xy The number of stable grids in the x class representing the y factor, TNSP y is the total number of grids of the y factor under stable conditions, W - is the negative weight to be assigned when class x of factor y does not exist, NLP ny is the number of landslide points in the other n categories of factor y, NSP ny y represents the number of stable grids in the other n categories of the y factor, W f For comparison weight, LSI WOfE is the landslide hazard index, W fxy is the final weight of class x in the conditional factor y, and n is the total number of conditional factors.

Images