A method and system for classifying and assessing landslide hazards and risks that takes into account surface deformation.

Abstract

This invention discloses a method and system for classifying and assessing landslide hazards and risks, taking into account surface deformation. The method includes: S1, extracting indicator data corresponding to disaster-causing factors, disaster-inducing factors, and disaster-bearing factors to generate a landslide hazard risk assessment factor database and constructing a landslide hazard risk assessment factor system; S2, generating potential landslide sliding ranges; S3, determining the hierarchical relationship of the landslide hazard risk assessment factor system, including the target layer-factor layer and the factor layer-indicator layer, and solving for the weight coefficients of the two hierarchical relationships; S4, standardizing and weighting the values ​​of each indicator to obtain the landslide hazard risk score, classifying the landslide hazard risk score to obtain the landslide hazard risk level. This invention provides technical support for the rapid screening of high-risk landslide hazards across a wide area by integrating disaster-causing factors, disaster-inducing factors, and disaster-bearing factors in the classification of potential landslide hazards.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence, remote sensing and geological disaster identification and remote sensing monitoring, specifically to a method and system for classifying and assessing landslide hazards and risks that take into account surface deformation. Background Technology

[0002] Landslides are geological hazards that severely impact my country's economic and social development. my country's complex geological conditions and frequent tectonic activity result in numerous, widely distributed, and highly unpredictable landslide hazards, causing significant casualties and economic losses annually. Once a landslide occurs, its impact is extensive; the resulting collapses, debris flows, and other sliding movements severely threaten people, houses, roads, and other infrastructure located along the landslide's path. Therefore, conducting rapid landslide risk assessments in advance, based on the distribution of the affected areas, is crucial for landslide monitoring and prevention.

[0003] Conducting landslide hazard classification and risk assessment requires addressing three key issues: first, the location of surface deformation and surrounding disaster-bearing bodies; second, the potential landslide displacement range; and third, how to classify suspected landslide hazards according to their degree of danger. Previously, disaster-bearing body identification primarily relied on manual interpretation of high-resolution optical data. However, with the development of computer and artificial intelligence technologies, deep neural network-based disaster-bearing body identification has gradually become the mainstream trend. Surface deformation is mainly extracted from InSAR data, and anomaly zones are identified based on manual interpretation, threshold segmentation, or cluster analysis. However, due to the large amount of noise in InSAR surface deformation phase data and the multiple interpretations of deformation causes, these methods are difficult to achieve efficient and accurate identification. Current landslide displacement range simulation methods are mainly divided into empirical methods and modeling methods. However, empirical methods have low prediction accuracy and rely on existing statistical data. Large-scale historical landslide statistics are lacking, making them difficult to apply effectively to engineering scenarios. Modeling methods need to consider complex local geological conditions, and the calculation process relies on complete geotechnical parameters. However, in actual engineering tasks, there are no reliable relevant parameters, making it impossible to quickly generate landslide displacement ranges on a large scale. Therefore, the above methods are difficult to achieve good results when directly applied to potential landslide hazard risk assessment work. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method for classifying and assessing landslide hazard risks that takes into account surface deformation. The method includes:

[0005] S1. Based on radar imagery and topographic data, by integrating InSAR technology, deep learning technology, and GIS spatial analysis technology, the corresponding indicator data of disaster-causing factors, disaster-inducing factors, and disaster-bearing factors, as well as the surface deformation phase and InSAR surface deformation phase anomaly areas, are extracted to generate a landslide hazard risk assessment factor database and construct a landslide hazard risk assessment factor system. Among them, the disaster-causing factors include the surface deformation rate, and the corresponding disaster-causing factor indicators include the average surface deformation rate; the disaster-inducing factors include slope, elevation, and slope units, and the corresponding disaster-inducing factor indicators include the maximum elevation difference of the hazard, the average slope, and the number of slope units in which the hazard is located; the disaster-bearing factors include buildings, roads, and rivers, and the corresponding disaster-bearing factor indicators include the number of dangerous buildings, the number of dangerous roads, and the area of ​​dangerous water bodies.

[0006] S2. Based on the InSAR surface deformation phase anomaly area, the surface deformation phase, the surface deformation rate, the slope, and the elevation, determine the potential initial slip point grid, construct a Monte Carlo landslide random walk algorithm for slopes, design slip cutoff conditions, and conduct Monte Carlo slope slip simulation until the slip cutoff conditions are met. Statistically count the frequency of each grid selected in the slip path, convert the frequency data into spatial vectors, and generate the potential landslide slip range as the potential threat range of the landslide hazard. This range is used to extract hazard-related disaster-bearing factor index data through GIS spatial analysis technology.

[0007] S3. The hierarchical relationship of the landslide hazard risk assessment factor system is determined by the analytic hierarchy process (AHP), including the target layer-factor layer and the factor layer-indicator layer. A pairwise comparison matrix is ​​constructed, and the weight coefficients of the two hierarchical relationships are solved by combining expert opinions and relevant data. The target layer is the landslide hazard risk score value, the factor layer includes disaster-causing factors, disaster-inducing factors and disaster-bearing factors, and the indicator layer includes disaster-causing factor indicators, disaster-inducing factor indicators and disaster-bearing factor indicators.

[0008] S4. Standardize the values ​​of each indicator, and sum the standardized values ​​according to the weight coefficients of the two-level relationship to obtain the landslide hazard risk score. Classify the landslide hazard risk score to obtain the landslide hazard risk level.

[0009] Preferably, in step S1, typical disaster-bearing targets, i.e., disaster-bearing factors, are extracted using the following method; and, using deformation phase data and elevation data as inputs for model training and inference, InSAR surface deformation phase anomaly areas are extracted using the following method:

[0010] (1) Collect sample data

[0011] High-resolution remote sensing images of the study area were collected as basic data. Targets with representative features were selected for manual interpretation as reference ground truth. The labeled ground truth images and the corresponding original high-resolution images were combined by band. The combined images were cut into several slices of a preset size and randomly divided into a sample training set and a sample test set according to a preset ratio.

[0012] (2) Model building

[0013] A multi-scale feature fusion network model is constructed, which adopts an encoder-decoder architecture. The encoder part extracts image features step by step through multiple Depconvgroup modules. Each Depconvgroup module consists of two sets of depthwise separable convolutional layers, BN normalization layers, and ReLU activation layers. The decoder part consists of Cbri modules at different scales. Each Cbri module consists of convolutional layers, BN normalization layers, and upsampling layers. After upsampling, features at different scales are processed into a uniform size. Finally, features at different levels are concatenated along the channel dimension, and a channel attention mechanism module is introduced to improve the model's ability to extract key features.

[0014] (3) Training the model

[0015] The multi-scale feature fusion network model is trained multiple times using the slice data of the sample training set. If the error index decreases and tends to stabilize, it indicates that the model training is complete. The multi-scale feature fusion network model is then validated using the slice data of the sample test set. If the validation model accuracy does not meet the requirements, the model training parameters are adjusted and the model is retrained until the accuracy requirements are met.

[0016] (4) Model Reasoning

[0017] The trained multi-scale feature fusion network model is used to predict all data in the study area to obtain the distribution results of typical disaster-bearing bodies in the entire area.

[0018] Preferably, in step S2, the Monte Carlo landslide random walk algorithm is constructed, including:

[0019] Using the InSAR surface deformation phase anomaly area as the initial potential slip point of the landslide, a formula is constructed to calculate the random probability of the current slip grid sliding to the target grid. The formula is as follows:

[0020]

[0021] Among them, P px Let f be the probability that a target grid cell surrounding the current sliding grid cell becomes the next sliding grid cell, where n is the number of target grid cells, β is the slope between the current sliding grid cell and the target grid cell, and f is the slope between the target grid cells. d and fβ These are the weighting factors for adjusting the sliding direction and slope, respectively. When the direction of the next sliding grid is the same as the previous one, f... d For d 2 When the sliding direction changes by 45°, f d Let d be the value of f. When the sliding direction changes by 90°, d f is 1 β Both d and f are initialization input parameters. r The average surface deformation phase rate;

[0022] The maximum horizontal slip distance is estimated as the slip cutoff condition, and the calculation formula is as follows:

[0023] L=K1×ΔH+K2(f s +f r )+B' (3)

[0025] Where L is the maximum horizontal slip distance, ΔH is the height difference from the top of the landslide to the ground, and f s and f r K1 and K2 are the area of ​​the surface deformation phase accumulation zone and the average surface deformation phase rate, respectively. K1 and K2 are the correlation coefficients, and B' is a constant term. K1, K2 and B' can be obtained by fitting historical observation data, with values ​​of 2.54, 0.5 and -50.37, respectively.

[0026] The frequency with which each grid cell in the sliding path is selected in the sliding simulation is calculated, i.e., the probability of sliding to the target grid cell each time. The calculation formula is as follows:

[0027]

[0028] Where P is the probability of the landslide affecting the target grid, and s i n represents the number of times the target grid was selected during the simulation for the slip point, m represents the number of potential hazards associated with the slip point affecting the target grid, and n represents the number of slip points that affected the target grid. max Let i be the maximum number of iterations in the Monte Carlo random walk simulation, and m be positive integers.

[0029] Preferably, in step S3, the hierarchical relationship weights between the target layer and the factor layer, and between the factor layer and the indicator layer are solved using the analytic hierarchy process (AHP).

[0030] The weights of individual hierarchical relationships are determined using the analytic hierarchy process (AHP), including:

[0031] Construct a matrix M, where the elements are labeled a. ij Based on expert opinions and relevant data, the evaluation indicators of the two levels to be solved are compared pairwise, and the determined scales are filled into the corresponding positions in matrix M. The resulting matrix M is the pairwise comparison matrix; where a ija represents the degree of importance or scale of i compared to j. ij The corresponding position is the i-th row and j-th column in matrix M; if i = j, it means that i and j are equally important, i.e., a ij =1, the scale of all diagonal elements of matrix M is 1; a ij >0 and a ij ×a ji =1;

[0032] The weights are calculated using the arithmetic mean, geometric mean, and eigenvalue method, respectively, and the arithmetic mean of the results is calculated. Assume the pairwise comparison matrix M is:

[0033]

[0034] The weight value w of the i-th factor can be obtained using the following three methods. i ,in,

[0035] The first method is to obtain the weight vector w using the arithmetic mean. i for:

[0036]

[0037] The second method is to obtain the weight vector w using the geometric mean method. i for:

[0038]

[0039] The third method is to calculate the weight w using the eigenvalue method. i Let's first assume that the largest eigenvalue of the pairwise comparison matrix M is λ. max If the corresponding eigenvector is ξ, then the weight w can be obtained by the eigenvalue method. i for:

[0040]

[0041] The consistency of pairwise comparison matrices is checked using the following methods:

[0042] Calculate the largest eigenvalue λ of the eigenvectors of the pairwise comparison matrix. max The calculation formula is:

[0043]

[0044] Where M is the pairwise comparison matrix, w i For the indicator weights, W = [w1, ..., w n ] T , where n is the order of the comparison matrix, and i, j, k, n are positive integers;

[0045] The consistency index (CI) of the pairwise comparison matrix is ​​calculated using the following formula:

[0046]

[0047] When CI is close to 0, it is preliminarily determined that the consistency test of the constructed pairwise comparison matrix has passed.

[0048] The consistency ratio is calculated using the following formula:

[0049]

[0050] Here, RI is the average consistency index, a parameter value obtained from the lookup table. Typically, when CR < 0.1, it indicates that the consistency test of the pairwise comparison matrix has passed.

[0051] Preferably, in step S4, the Z-Score method is used to standardize the values ​​of each indicator and perform a weighted summation. The calculation formula is as follows:

[0052]

[0053] Where R is the risk score after comprehensively considering risk assessment indicators, and w j Let w be the j-th factor, m be the number of factor layers, and w be the number of factors. i Let x be the weight of the i-th indicator, n be the number of indicators in the factor layer, and x be the weight of the i-th indicator. i Let be the actual observed or measured value of the i-th indicator, μ be the average of all observed values, σ be the standard deviation of all indicator data, representing the dispersion of the data, and i, j, m, n be positive integers.

[0054] Preferably, in step S4, the k-means algorithm is used to classify the landslide hazard risk score values, including:

[0055] Predetermine the number of categories K, and randomly select K objects {a1, a2, ..., a...} K} as the landslide hazard risk score data {X1,X2,…,X n The initial cluster centers are determined by calculating the distance between each data point and each cluster center, assigning the data point to each cluster center according to the principle of proximity, and recalculating the cluster centers. The final result minimizes the sum of squared distances between each data point and its nearest cluster center. The calculation formula is as follows:

[0056]

[0057] Where T is the distance objective function for clustering, and q j Let X be the j-th initial cluster center, K be the number of sub-clusters, and X be the number of sub-clusters. i Let be the data to be clustered, n be the number of data points, and i, j, and n be positive integers.

[0058] Based on the same inventive concept, the present invention also provides a landslide hazard classification and risk assessment system that takes into account surface deformation, the system comprising:

[0059] A construction module is used to extract indicator data corresponding to disaster-causing factors, disaster-inducing factors, and disaster-bearing factors, as well as surface deformation phase and InSAR surface deformation phase anomaly areas, based on radar imagery and topographic data, by integrating InSAR technology, deep learning technology, and GIS spatial analysis technology. This generates a landslide hazard risk assessment factor database and constructs a landslide hazard risk assessment factor system. The disaster-causing factors include surface deformation rate, and the corresponding disaster-causing factor index includes the average surface deformation rate. The disaster-inducing factors include slope, elevation, and slope units, and the corresponding disaster-inducing factor indexes include the maximum elevation difference of the hazard, the average slope, and the number of slope units where the hazard is located. The disaster-bearing factors include buildings, roads, and rivers, and the corresponding disaster-bearing factor indexes include the number of dangerous buildings, the number of dangerous roads, and the area of ​​dangerous water bodies.

[0060] The generation module is used to determine potential initial slip point grids based on the InSAR surface deformation phase anomaly area, the surface deformation phase, the surface deformation rate, the slope, and the elevation; construct a Monte Carlo landslide random walk algorithm for slopes; design slip cutoff conditions; perform Monte Carlo slope slip simulation until the slip cutoff conditions are met; count the frequency of each grid selected in the slip path; convert the frequency data into spatial vectors; generate the potential landslide slip range as the potential threat range of landslide hazards; and use GIS spatial analysis technology to extract hazard-related disaster-bearing factor index data.

[0061] The solution module is used to determine the hierarchical relationship of the landslide hazard risk assessment factor system using the analytic hierarchy process (AHP), including the target layer-factor layer and the factor layer-indicator layer. It constructs a pairwise comparison matrix, combines expert opinions and relevant data, and solves the weight coefficients of the two hierarchical relationships. The target layer is the landslide hazard risk score value, the factor layer includes disaster-causing factors, disaster-inducing factors, and disaster-bearing factors, and the indicator layer includes disaster-causing factor indicators, disaster-inducing factor indicators, and disaster-bearing factor indicators.

[0062] The assessment module is used to standardize the values ​​of each indicator, and then perform a weighted summation of the standardized values ​​according to the weight coefficients of the two hierarchical relationships to obtain the landslide hazard risk score. The landslide hazard risk score is then classified to obtain the landslide hazard risk level.

[0063] Preferably, typical disaster-bearing targets, i.e., disaster-bearing factors, are extracted using the following methods; and, using deformation phase data and elevation data as inputs for model training and inference, InSAR surface deformation phase anomaly areas are extracted using the following methods:

[0064] (1) Collect sample data

[0065] High-resolution remote sensing images of the study area were collected as basic data. Targets with representative features were selected for manual interpretation as reference ground truth. The labeled ground truth images and the corresponding original high-resolution images were combined by band. The combined images were cut into several slices of a preset size and randomly divided into a sample training set and a sample test set according to a preset ratio.

[0066] (2) Model building

[0067] A multi-scale feature fusion network model is constructed, which adopts an encoder-decoder architecture. The encoder part extracts image features step by step through multiple Depconvgroup modules. Each Depconvgroup module consists of two sets of depthwise separable convolutional layers, BN normalization layers, and ReLU activation layers. The decoder part consists of Cbri modules at different scales. Each Cbri module consists of convolutional layers, BN normalization layers, and upsampling layers. After upsampling, features at different scales are processed into a uniform size. Finally, features at different levels are concatenated along the channel dimension, and a channel attention mechanism module is introduced to improve the model's ability to extract key features.

[0068] (3) Training the model

[0069] The multi-scale feature fusion network model is trained multiple times using the slice data of the sample training set. If the error index decreases and tends to stabilize, it indicates that the model training is complete. The multi-scale feature fusion network model is then validated using the slice data of the sample test set. If the validation model accuracy does not meet the requirements, the model training parameters are adjusted and the model is retrained until the accuracy requirements are met.

[0070] (4) Model Reasoning

[0071] The trained multi-scale feature fusion network model is used to predict all data in the study area to obtain the distribution results of typical disaster-bearing bodies in the entire area.

[0072] Preferably, the Monte Carlo landslide random walk algorithm for slopes includes:

[0073] Using the InSAR surface deformation phase anomaly area as the initial potential slip point of the landslide, a formula is constructed to calculate the random probability of the current slip grid sliding to the target grid. The formula is as follows:

[0074]

[0075] Among them, P px Let f be the probability that a target grid cell surrounding the current sliding grid cell becomes the next sliding grid cell, where n is the number of target grid cells, β is the slope between the current sliding grid cell and the target grid cell, and f is the slope between the target grid cells.d and f β These are the weighting factors for adjusting the sliding direction and slope, respectively. When the direction of the next sliding grid is the same as the previous one, f... d For d 2 When the sliding direction changes by 45°, f d Let d be the value of f. When the sliding direction changes by 90°, d f is 1 β Both d and f are initialization input parameters. r The average surface deformation phase rate;

[0076] The maximum horizontal slip distance is estimated as the slip cutoff condition, and the calculation formula is as follows:

[0077] L=K1×ΔH+K2(f s +f r )+B' (3)

[0079] Where L is the maximum horizontal slip distance, ΔH is the height difference from the top of the landslide to the ground, and f s and f r K1 and K2 are the area of ​​the surface deformation phase accumulation zone and the average surface deformation phase rate, respectively. K1 and K2 are the correlation coefficients, and B' is a constant term. K1, K2 and B' can be obtained by fitting historical observation data, with values ​​of 2.54, 0.5 and -50.37, respectively.

[0080] The frequency with which each grid cell in the sliding path is selected in the sliding simulation is calculated, i.e., the probability of sliding to the target grid cell each time. The calculation formula is as follows:

[0081]

[0082] Where P is the probability of the landslide affecting the target grid, and s i n represents the number of times the target grid was selected during the simulation for the slip point, m represents the number of potential hazards associated with the slip point affecting the target grid, and n represents the number of slip points that affected the target grid. max Let i be the maximum number of iterations in the Monte Carlo random walk simulation, and m be positive integers.

[0083] Preferably, the hierarchical relationship weights between the target layer and the factor layer, and between the factor layer and the indicator layer are solved separately using the analytic hierarchy process (AHP).

[0084] The weights of individual hierarchical relationships are determined using the analytic hierarchy process (AHP), including:

[0085] Construct a matrix M, where the elements are labeled a. ij Based on expert opinions and relevant data, the evaluation indicators of the two levels to be solved are compared pairwise, and the determined scales are filled into the corresponding positions in matrix M. The resulting matrix M is the pairwise comparison matrix; where aij a represents the degree of importance or scale of i compared to j. ij The corresponding position is the i-th row and j-th column in matrix M; if i = j, it means that i and j are equally important, i.e., a ij =1, the scale of all diagonal elements of matrix M is 1; a ij >0 and a ij ×a ji =1;

[0086] The weights are calculated using the arithmetic mean, geometric mean, and eigenvalue method, respectively, and the arithmetic mean of the results is calculated. Assume the pairwise comparison matrix M is:

[0087]

[0088] The weight value w of the i-th factor can be obtained using the following three methods. i ,in,

[0089] The first method is to obtain the weight vector w using the arithmetic mean. i for:

[0090]

[0091] The second method is to obtain the weight vector w using the geometric mean method. i for:

[0092]

[0093] The third method is to calculate the weight w using the eigenvalue method. i Let's first assume that the largest eigenvalue of the pairwise comparison matrix M is λ. max If the corresponding eigenvector is ξ, then the weight w can be obtained by the eigenvalue method. i for:

[0094]

[0095] The consistency of pairwise comparison matrices is checked using the following methods:

[0096] Calculate the largest eigenvalue λ of the eigenvectors of the pairwise comparison matrix. max The calculation formula is:

[0097]

[0098] Where M is the pairwise comparison matrix, w i For the indicator weights, W = [w1, ..., w n ] T , where n is the order of the comparison matrix, and i, j, k, n are positive integers;

[0099] The consistency index (CI) of the pairwise comparison matrix is ​​calculated using the following formula:

[0100]

[0101] When CI is close to 0, it is preliminarily determined that the consistency test of the constructed pairwise comparison matrix has passed.

[0102] The consistency ratio is calculated using the following formula:

[0103]

[0104] Here, RI is the average consistency index, a parameter value obtained from the lookup table. Typically, when CR < 0.1, it indicates that the consistency test of the pairwise comparison matrix has passed.

[0105] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0106] This invention utilizes deep learning technology and multi-source remote sensing data to accurately identify InSAR surface deformation anomaly zones and pinpoint potential landslide hazards. Drawing on the random walk concept, it proposes an automatic prediction method for the sliding range of potential landslides that considers deformation characteristics, providing technical support for wide-area landslide hazard risk assessment. Furthermore, it proposes a classification method for potential landslide hazards that integrates disaster-causing factors, disaster-inducing factors, and disaster-bearing factors, providing technical support for the rapid screening of high-risk landslide hazards across wide areas. The technology of this invention plays a crucial role in improving the automatic identification of potential landslide hazards across wide areas and is of great significance for enhancing the intelligence level of landslide disaster monitoring and early warning. Attached Figure Description

[0107] Figure 1 A flowchart illustrating a landslide hazard classification and risk assessment method that takes into account surface deformation, provided by the present invention.

[0108] Figure 2 A technical flowchart of a landslide hazard classification and risk assessment method that takes into account surface deformation, provided by the present invention;

[0109] Figure 3 Various hidden danger risk assessment factors provided for this invention;

[0110] Figure 4 The structural framework diagram of the multi-scale feature fusion network model provided by this invention;

[0111] Figure 5-6 This is a schematic diagram of the InSAR surface deformation phase anomaly area identification results provided by the present invention;

[0112] Figure 7 This is a schematic diagram of the potential landslide risk assessment results for Zhouqu County provided by the present invention. Detailed Implementation

[0113] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0114] The present invention will now be described in further detail with reference to the accompanying drawings.

[0115] like Figure 1-2 As shown, this embodiment of the invention provides a method for classifying and assessing landslide hazard risks that takes into account surface deformation, including:

[0116] S1. Based on radar imagery and topographic data, by integrating InSAR technology, deep learning technology, and GIS spatial analysis technology, the corresponding indicator data of disaster-causing factors, disaster-inducing factors, and disaster-bearing factors, as well as the surface deformation phase and InSAR surface deformation phase anomaly areas, are extracted to generate a landslide hazard risk assessment factor database and construct a landslide hazard risk assessment factor system. Among them, disaster-causing factors include surface deformation rate, and the corresponding disaster-causing factor indicators include average surface deformation rate; disaster-inducing factors include slope, elevation, and slope units, and the corresponding disaster-inducing factor indicators include maximum elevation difference of the hazard, average slope, and number of slope units; disaster-bearing factors include buildings, roads, and rivers, and the corresponding disaster-bearing factor indicators include the number of dangerous buildings, the number of dangerous roads, and the area of ​​dangerous water bodies.

[0117] Step 1: Constructing a risk assessment factor system for landslide hazards

[0118] This step aims to establish a comprehensive landslide hazard risk assessment factor system, covering three aspects: hazard-causing factors, hazard-inducing factors, and hazard-bearing bodies. Specifically, the maximum elevation difference of the hazard, average slope, and the number of slope units where it is located are selected as hazard-inducing factor indicators; the number of dangerous buildings and roads, and the area of ​​water bodies are selected as hazard-bearing factor indicators; and the average rate of surface deformation detected based on InSAR technology is selected as a hazard-causing factor indicator. By integrating InSAR technology, deep learning technology, and GIS spatial analysis technology, the corresponding indicator data are extracted to generate a landslide hazard risk assessment factor database. Figure 3 This provides basic data for subsequent risk assessment. Figure 3 Various risk assessment factors for potential hazards are presented: (a) altitude, (b) slope, (c) deformation rate, (d) population density, (e) slope unit, (f) buildings, (g) roads, and (h) water bodies.

[0119] In this embodiment of the invention, typical disaster-bearing targets, i.e., disaster-bearing factors, are extracted using the following method; and, using deformation phase data and elevation data as inputs for model training and inference, InSAR surface deformation phase anomaly areas are extracted using the following method:

[0120] (1) Collect sample data

[0121] High-resolution remote sensing images of the study area were collected as basic data. Targets with representative features were selected for manual interpretation as reference ground truth. The labeled ground truth images and the corresponding original high-resolution images were combined by band. The combined images were cut into several slices of a preset size (256×256) and randomly divided into a sample training set and a sample test set according to a preset ratio (7:3).

[0122] (2) Model building

[0123] Construct a multi-level feature fusion network model (MFAFNet) for information extraction, such as... Figure 4 As shown, the model adopts an encoder-decoder architecture. The encoder extracts image features stepwise through multiple Depconvgroup modules. Each Depconvgroup module consists of two sets of depthwise separable convolutional layers, BN normalization layers, and ReLU activation layers, which requires less computation compared to traditional two-dimensional convolutional layers. The decoder consists of Cbri modules at different scales. Each Cbri module consists of convolutional layers, BN normalization layers, and upsampling layers. After upsampling, features at different scales are processed to a uniform size. Finally, features at different levels are concatenated (Cat) along the channel dimension, and a channel attention mechanism module (SE module) is introduced to improve the model's ability to extract key features. The model is then complete. Figure 4 In this context, Image represents the image input to the model, H represents the image height, W represents the image width, MLP (multi-layer perceptron) represents a multi-layer perceptron, which is a feedforward neural network model, and Prediction represents the prediction result.

[0124] (3) Training the model

[0125] The multi-scale feature fusion network model is trained multiple times using slice data from the training set. If the error index decreases and tends to stabilize, the model training is complete. The multi-scale feature fusion network model is then validated using slice data from the test set. If the model accuracy does not meet the requirements, the training parameters are adjusted and the model is retrained until the accuracy requirements are met.

[0126] (4) Model Reasoning

[0127] The trained multi-scale feature fusion network model was used to predict all data in the study area, and the distribution of typical disaster-bearing bodies in the whole area was obtained. Figure 5 The results of identifying anomaly zones in surface deformation phase are presented.

[0128] S2. Based on InSAR surface deformation phase anomaly areas, surface deformation phase, surface deformation rate, slope and elevation, determine potential initial slip point grids, construct a Monte Carlo landslide random walk algorithm for slopes, design slip cutoff conditions, and conduct Monte Carlo slope slip simulations until the slip cutoff conditions are met. Statistically count the frequency of each grid selected in the slip path, convert the frequency data into spatial vectors, and generate the potential landslide slip range as the potential threat range of landslide hazards. This range is used to extract hazard-related disaster-bearing factor index data through GIS spatial analysis technology.

[0129] Step 2: Estimation of the impact range of the landslide

[0130] The initial slip zone of a potential landslide was identified, and a Monte Carlo random walk algorithm for slope landslides was constructed to simulate the longitudinal and lateral movement processes of the slip. Slip cutoff conditions were designed, and numerous random walks were performed on the slip initiation point until the cutoff distance condition was met. Finally, the frequency of selection of each grid cell in the slip path was statistically analyzed, and the probability data was converted into spatial vectors to automatically generate the potential landslide range. The slip results generated in this step serve as the potential threat range of the landslide hazard, which is used to extract hazard-related disaster-bearing factor index data through geospatial analysis techniques.

[0131] In this embodiment of the invention, a Monte Carlo landslide random walk algorithm is constructed, comprising:

[0132] Using InSAR surface deformation phase anomaly areas as the initial potential slip points of landslides, a formula is constructed to calculate the random probability of the current slip grid sliding to the target grid. The formula is as follows:

[0133]

[0134] Among them, P px Let f be the probability that a target grid cell surrounding the current sliding grid cell becomes the next sliding grid cell, where n is the number of target grid cells, β is the slope between the current sliding grid cell and the target grid cell, and f is the slope between the target grid cells. d and f β These are the weighting factors for adjusting the sliding direction and slope, respectively. When the direction of the next sliding grid is the same as the previous one, f... d For d 2 When the sliding direction changes by 45°, f d Let d be the value of f. When the sliding direction changes by 90°,d f is 1 β Both d and f are initialization input parameters. r The average surface deformation phase rate;

[0135] The maximum horizontal slip distance is estimated as the slip cutoff condition, and the calculation formula is as follows:

[0136] L=K1×ΔH+K2(f s +f r )+B' (3)

[0138] Where L is the maximum horizontal slip distance, ΔH is the height difference from the top of the landslide to the ground, and f s and f r K1 and K2 are the area of ​​the surface deformation phase accumulation zone and the average surface deformation phase rate, respectively. K1 and K2 are the correlation coefficients, and B' is a constant term. K1, K2 and B' can be obtained by fitting historical observation data, with values ​​of 2.54, 0.5 and -50.37, respectively.

[0139] The frequency with which each grid cell in the sliding path is selected in the sliding simulation is calculated, i.e., the probability of sliding to the target grid cell each time. The calculation formula is as follows:

[0140]

[0141] Where P is the probability of the landslide affecting the target grid, and s i n represents the number of times the target grid was selected during the simulation for the slip point, m represents the number of potential hazards associated with the slip point affecting the target grid, and n represents the number of slip points that affected the target grid. max Let i be the maximum number of iterations in the Monte Carlo random walk simulation, and m be positive integers.

[0142] S3. The hierarchical relationship of the landslide hazard risk assessment factor system is determined by the analytic hierarchy process (AHP), including the target layer-factor layer and the factor layer-indicator layer. Pairwise comparison matrices are constructed, and the weight coefficients of the two hierarchical relationships are solved by combining expert opinions and relevant data. The target layer is the landslide hazard risk score, the factor layer includes disaster-causing factors, disaster-inducing factors and disaster-bearing factors, and the indicator layer includes disaster-causing factor indicators, disaster-inducing factor indicators and disaster-bearing factor indicators.

[0143] Step 3: Determine the hierarchical relationship of the evaluation system and calculate the weights.

[0144] After constructing the evaluation factor system, the hierarchical relationship of the evaluation system was determined using the analytic hierarchy process (AHP), including the target layer, factor layer, and indicator layer. The target layer is the landslide hazard risk score; the factor layer includes disaster-inducing factors, disaster-bearing factors, and disaster-causing factors; and the indicator layer consists of the specific evaluation indicators selected in step one. By constructing a pairwise comparison matrix and combining expert opinions and relevant data, pairwise comparisons were performed on each factor to solve for the weight coefficients of the factor layer and the indicator layer.

[0145] In this embodiment of the invention, the hierarchical relationship weights between the target layer and the factor layer, and between the factor layer and the indicator layer, are solved using the analytic hierarchy process (AHP).

[0146] The weights of individual hierarchical relationships are determined using the analytic hierarchy process (AHP), including:

[0147] Construct a matrix M, where the elements are labeled a. ij Based on expert opinions and relevant data, the evaluation indicators of the two levels to be solved are compared pairwise, and the determined scales are filled into the corresponding positions in matrix M. The resulting matrix M is the pairwise comparison matrix; where a ij a represents the degree of importance or scale of i compared to j. ij The corresponding position is the i-th row and j-th column in matrix M; if i = j, it means that i and j are equally important, i.e., a ij =1, the scale of all diagonal elements of matrix M is 1; a ij >0 and a ij ×a ji =1;

[0148] The weights are calculated using the arithmetic mean, geometric mean, and eigenvalue method, respectively, and the arithmetic mean of the results is calculated. Assume the pairwise comparison matrix M is:

[0149]

[0150] The weight value w of the i-th factor can be obtained using the following three methods. i ,in,

[0151] The first method is to obtain the weight vector w using the arithmetic mean. i for:

[0152]

[0153] The second method is to obtain the weight vector w using the geometric mean method. i for:

[0154]

[0155] The third method is to calculate the weight w using the eigenvalue method. i Let's first assume that the largest eigenvalue of the pairwise comparison matrix M is λ. maxIf the corresponding eigenvector is ξ, then the weight w can be obtained by the eigenvalue method. i for:

[0156]

[0157] Since the pairwise comparison matrix is ​​determined based on expert ratings, a consistency check is required to ensure that there are no inconsistencies in the input. The following methods are used to perform a consistency check on the pairwise comparison matrix:

[0158] Calculate the largest eigenvalue λ of the eigenvectors of the pairwise comparison matrix. max The calculation formula is:

[0159]

[0160] Where M is the pairwise comparison matrix, w i For the indicator weights, W = [w1, ..., w n ] T , where n is the order of the comparison matrix, and i, j, k, n are positive integers;

[0161] The consistency index (CI) of the pairwise comparison matrix is ​​calculated using the following formula:

[0162]

[0163] When CI is close to 0, it is preliminarily determined that the consistency test of the constructed pairwise comparison matrix has passed.

[0164] The consistency ratio is calculated using the following formula:

[0165]

[0166] Here, RI is the average consistency index, a parameter value obtained from the lookup table. Typically, when CR < 0.1, it indicates that the consistency test of the pairwise comparison matrix has passed.

[0167] S4. Standardize the values ​​of each indicator, and sum the standardized values ​​according to the weight coefficients of the two-level relationship to obtain the landslide hazard risk score. Classify the landslide hazard risk score to obtain the landslide hazard risk level.

[0168] Step 4: Classification of Landslide Hazard Risk Levels

[0169] The Z-score standardization method was used to unify the numerical range of each indicator to between 0 and 1. Based on the obtained weight coefficients, the standardized values ​​were weighted and summed to calculate the landslide hazard risk score. The k-means clustering algorithm was then used to classify the calculated risk scores, thus completing the risk level classification.

[0170] In this embodiment of the invention, the Z-Score method is used to standardize the values ​​of each indicator and then perform a weighted summation. The calculation formula is as follows:

[0171]

[0172] Where R is the risk score after comprehensively considering risk assessment indicators, and w j Let w be the j-th factor, m be the number of factor layers, and w be the number of factors. i Let x be the weight of the i-th indicator, n be the number of indicators in the factor layer, and x be the weight of the i-th indicator. i Let be the actual observed or measured value of the i-th indicator, μ be the average of all observed values, σ be the standard deviation of all indicator data, representing the dispersion of the data, and i, j, m, n be positive integers.

[0173] In this embodiment of the invention, the k-means algorithm is used to classify the risk score values ​​of landslide hazards, including:

[0174] Predetermine the number of categories K, and randomly select K objects {a1, a2, ..., a...} K} as landslide hazard risk score data {X1,X2,…,X n The initial cluster centers are determined by calculating the distance between each data point and each cluster center, assigning the data point to each cluster center according to the principle of proximity, and recalculating the cluster centers. The final result minimizes the sum of squared distances between each data point and its nearest cluster center. The calculation formula is as follows:

[0175]

[0176] Where T is the distance objective function for clustering, and q j Let X be the j-th initial cluster center, K be the number of sub-clusters, and X be the number of sub-clusters. i Let be the data to be clustered, n be the number of data points, and i, j, and n be positive integers.

[0177] Taking Zhouqu County in Gansu Province as an example, landslide hazard assessment was conducted, with risk levels categorized into five levels: low, relatively low, medium, relatively high, and high. Thresholds for each category were calculated using the k-means clustering algorithm (Table 1). The identification results are as follows: Figure 7 As shown.

[0178] Table 1 Thresholds for Landslide Hazard Risk Assessment Categories

[0179] Risk Category Numerical range Low 0～0.187 lower 0.187～0.236 medium 0.236～0.299 higher 0.299～0.395 high 0.395～0.496

[0180] Based on the same inventive concept, embodiments of the present invention also provide a landslide hazard classification and risk assessment system that takes into account surface deformation, including:

[0181] The module is designed to extract indicator data corresponding to disaster-causing factors, disaster-inducing factors, and disaster-bearing factors based on radar imagery and topographic data, by integrating InSAR technology, deep learning technology, and GIS spatial analysis technology. It also includes surface deformation phase and InSAR surface deformation phase anomaly areas, generating a landslide hazard risk assessment factor database and constructing a landslide hazard risk assessment factor system. Disaster-causing factors include surface deformation rate, with corresponding indicators including average surface deformation rate. Disaster-inducing factors include slope, elevation, and slope units, with corresponding indicators including maximum elevation difference of the hazard, average slope, and number of slope units. Disaster-bearing factors include buildings, roads, and rivers, with corresponding indicators including the number of dangerous buildings, the number of dangerous roads, and the area of ​​dangerous water bodies.

[0182] The generation module is used to determine potential initial slip point grids based on InSAR surface deformation phase anomaly areas, surface deformation phase, surface deformation rate, slope and elevation, construct a Monte Carlo landslide random walk algorithm for slopes, design slip cutoff conditions, conduct Monte Carlo slope slip simulations until the slip cutoff conditions are met, count the frequency of each grid selected in the slip path, convert the frequency data into spatial vectors, generate the potential landslide slip range, which serves as the potential threat range of landslide hazards, and is used to extract hazard-related disaster-bearing factor index data through GIS spatial analysis technology.

[0183] The solution module is used to determine the hierarchical relationship of the landslide hazard risk assessment factor system using the analytic hierarchy process (AHP), including the target layer-factor layer and the factor layer-indicator layer. It constructs a pairwise comparison matrix and, in conjunction with expert opinions and relevant data, solves the weight coefficients of the two hierarchical relationships. The target layer is the landslide hazard risk score, the factor layer includes disaster-causing factors, disaster-inducing factors, and disaster-bearing factors, and the indicator layer includes disaster-causing factor indicators, disaster-inducing factor indicators, and disaster-bearing factor indicators.

[0184] The assessment module is used to standardize the values ​​of each indicator, and then perform a weighted summation of the standardized values ​​based on the weight coefficients of the two-level relationship to obtain a landslide hazard risk score. The landslide hazard risk score is then classified to obtain the landslide hazard risk level.

[0185] This invention first utilizes deep neural networks and multi-source remote sensing data to identify typical disaster-bearing bodies and areas with abnormal surface deformation phases, thereby determining potential landslide initiation zones. Based on InSAR data such as surface deformation range, deformation rate, and elevation, potential slip ranges are generated through random walk simulation. Finally, by combining the distribution of disaster-bearing bodies, surface deformation range and displacement, and topography, a comprehensive analysis is conducted to classify and assess the risks of suspected landslide hazards.

[0186] Compared with existing technologies, the advantages of this invention are:

[0187] 1. This invention utilizes deep learning technology and multi-source remote sensing data to accurately identify InSAR surface deformation anomaly zones, thereby enabling the accurate location of potential landslide hazards.

[0188] 2. This invention constructs a simulation method for the potential landslide impact range based on the random walk algorithm. It comprehensively considers constraints such as sliding area, sliding rate and elevation difference, and solves the problem that traditional landslide impact range simulation methods are difficult to extend to large-scale applications, thus realizing rapid and automatic prediction of the potential landslide sliding impact range.

[0189] 3. The potential landslide hazard classification method of this invention, which integrates disaster-causing factors, disaster-inducing factors, and disaster-bearing factors, provides technical support for the rapid screening of landslide hazards in a wide area with high risk of landslides.

[0190] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for classifying and assessing landslide hazard risks that takes into account surface deformation, characterized in that, The method includes: S1. Based on radar imagery and topographic data, by integrating InSAR technology, deep learning technology, and GIS spatial analysis technology, the corresponding indicator data of disaster-causing factors, disaster-inducing factors, and disaster-bearing factors, as well as the surface deformation phase and InSAR surface deformation phase anomaly areas, are extracted to generate a landslide hazard risk assessment factor database and construct a landslide hazard risk assessment factor system. Among them, the disaster-causing factors include the surface deformation rate, and the corresponding disaster-causing factor indicators include the average surface deformation rate; the disaster-inducing factors include slope, elevation, and slope units, and the corresponding disaster-inducing factor indicators include the maximum elevation difference of the hazard, the average slope, and the number of slope units in which the hazard is located; the disaster-bearing factors include buildings, roads, and rivers, and the corresponding disaster-bearing factor indicators include the number of dangerous buildings, the number of dangerous roads, and the area of ​​dangerous water bodies. S2. Based on the InSAR surface deformation phase anomaly area, the surface deformation phase, the surface deformation rate, the slope, and the elevation, determine the potential initial slip point grid, construct a Monte Carlo landslide random walk algorithm for slopes, design slip cutoff conditions, and perform Monte Carlo slope slip simulation until the slip cutoff conditions are met. Statistically count the frequency of each grid selected in the slip path, convert the frequency data into spatial vectors, and generate the potential landslide slip range as the potential threat range of landslide hazards. This range is used to extract disaster-bearing factor index data through GIS spatial analysis technology. Constructing a Monte Carlo landslide random walk algorithm for slopes, including: Using the InSAR surface deformation phase anomaly area as the initial potential slip point of the landslide, a formula is constructed to calculate the random probability of the current slip grid sliding to the target grid. The formula is as follows: Among them, P px Let f be the probability that a target grid cell surrounding the current sliding grid cell becomes the next sliding grid cell, where n is the number of target grid cells, β is the slope between the current sliding grid cell and the target grid cell, and f is the slope between the target grid cells. d and f β These are the weighting factors for adjusting the sliding direction and slope, respectively. When the direction of the next sliding grid is the same as the previous one, f... d For d 2 When the sliding direction changes by 45°, f d Let d be the value of f. When the sliding direction changes by 90°, d f is 1 β Both d and f are initialization input parameters. r The average surface deformation phase rate; The maximum horizontal slip distance is estimated as the slip cutoff condition, and the calculation formula is as follows: L=K1×ΔH+K2(f s +f r )+B' (3) Where L is the maximum horizontal slip distance, ΔH is the height difference from the top of the landslide to the ground, and f s and f r K1 and K2 are the area of ​​the surface deformation phase accumulation zone and the average surface deformation phase rate, respectively. K1 and K2 are the correlation coefficients, and B' is a constant term. K1, K2 and B' can be obtained by fitting historical observation data. The frequency with which each grid cell in the sliding path is selected in the sliding simulation is calculated, i.e., the probability of sliding to the target grid cell each time. The calculation formula is as follows: Where P is the probability of the landslide affecting the target grid, and s i n represents the number of times the target grid was selected during the simulation for the slip point, m represents the number of potential hazards associated with the slip point affecting the target grid, and n represents the number of slip points that affected the target grid. max Let i and m be the maximum number of iterations for the Monte Carlo random walk simulation, where i and m are positive integers. S3. The hierarchical relationship of the landslide hazard risk assessment factor system is determined by the analytic hierarchy process (AHP), including the target layer-factor layer and the factor layer-indicator layer. A pairwise comparison matrix is ​​constructed, and the weight coefficients of the two hierarchical relationships are solved by combining expert opinions and relevant data. The target layer is the landslide hazard risk score value, the factor layer includes disaster-causing factors, disaster-inducing factors and disaster-bearing factors, and the indicator layer includes disaster-causing factor indicators, disaster-inducing factor indicators and disaster-bearing factor indicators. S4. Standardize the values ​​of each indicator, and sum the standardized values ​​according to the weight coefficients of the two-level relationship to obtain the landslide hazard risk score. Classify the landslide hazard risk score to obtain the landslide hazard risk level.

2. The method according to claim 1, characterized in that, In step S1, typical disaster-bearing targets, i.e., disaster-bearing factors, are extracted using the following methods. Furthermore, using deformation phase data and elevation data as inputs for model training and inference, InSAR surface deformation phase anomaly areas are extracted using the following methods: (1) Collect sample data High-resolution remote sensing images of the study area were collected as basic data. Targets with representative features were selected for manual interpretation as reference ground truth. The labeled ground truth images and the corresponding original high-resolution images were combined by band. The combined images were cut into several slices of a preset size and randomly divided into a sample training set and a sample test set according to a preset ratio. (2) Model building A multi-scale feature fusion network model is constructed, which adopts an encoder-decoder architecture. The encoder part extracts image features step by step through multiple Depconvgroup modules. Each Depconvgroup module consists of two sets of depthwise separable convolutional layers, BN normalization layers, and ReLU activation layers. The decoder part consists of Cbri modules at different scales. Each Cbri module consists of convolutional layers, BN normalization layers, and upsampling layers. After upsampling, features at different scales are processed into a uniform size. Finally, features at different levels are concatenated along the channel dimension, and a channel attention mechanism module is introduced to improve the model's ability to extract key features. (3) Training the model The multi-scale feature fusion network model is trained multiple times using the slice data of the sample training set. If the error index decreases and tends to stabilize, it indicates that the model training is complete. The multi-scale feature fusion network model is then validated using the slice data of the sample test set. If the validation model accuracy does not meet the requirements, the model training parameters are adjusted and the model is retrained until the accuracy requirements are met. (4) Model Reasoning The trained multi-scale feature fusion network model is used to predict all data in the study area to obtain the distribution results of typical disaster-bearing bodies in the entire area.

3. The method according to claim 1, characterized in that, In step S3, the hierarchical relationship weights between the target layer and the factor layer, and between the factor layer and the indicator layer are solved using the analytic hierarchy process. The weights of individual hierarchical relationships are determined using the analytic hierarchy process (AHP), including: Construct a matrix M, where the elements are labeled a. ij Based on expert opinions and relevant data, the evaluation indicators of the two levels to be solved are compared pairwise, and the determined scales are filled into the corresponding positions in matrix M. The resulting matrix M is the pairwise comparison matrix; where a ij a represents the degree of importance or scale of i compared to j. ij The corresponding position is the i-th row and j-th column in matrix M; if i = j, it means that i and j are equally important, i.e., a ij =1, the scale of all diagonal elements of matrix M is 1; a ij >0 and a ij ×a ji =1; The weights are calculated using the arithmetic mean, geometric mean, and eigenvalue method, respectively, and the arithmetic mean of the results is calculated. Assume the pairwise comparison matrix M is: The weight value w of the i-th factor can be obtained using the following three methods. i ,in, The first method is to obtain the weight vector w using the arithmetic mean. i for: The second method is to obtain the weight vector w using the geometric mean method. i for: The third method is to calculate the weight w using the eigenvalue method. i Let's first assume that the largest eigenvalue of the pairwise comparison matrix M is λ. max If the corresponding eigenvector is ξ, then the weight w can be obtained by the eigenvalue method. i for: The consistency of pairwise comparison matrices is checked using the following methods: Calculate the largest eigenvalue λ of the eigenvectors of the pairwise comparison matrix. max The calculation formula is: Where M is the pairwise comparison matrix, w i For the indicator weights, W = [w1, ..., w n ] T , where n is the order of the comparison matrix, and i, j, k, n are positive integers; The consistency index (CI) of the pairwise comparison matrix is ​​calculated using the following formula: When CI is close to 0, it is preliminarily determined that the consistency test of the constructed pairwise comparison matrix has passed. The consistency ratio is calculated using the following formula: Here, RI is the average consistency index, a parameter value obtained from the lookup table. Typically, when CR < 0.1, it indicates that the consistency test of the pairwise comparison matrix has passed.

4. The method according to claim 1, characterized in that, In step S4, the Z-Score method is used to standardize the values ​​of each indicator and then perform a weighted summation. The calculation formula is as follows: Where R is the risk score after comprehensively considering risk assessment indicators, and w j Let w be the j-th factor, m be the number of factor layers, and w be the number of factors. i Let x be the weight of the i-th indicator, n be the number of indicators in the factor layer, and x be the weight of the i-th indicator. i Let be the actual observed or measured value of the i-th indicator, μ be the average of all observed values, σ be the standard deviation of all indicator data, representing the dispersion of the data, and i, j, m, n be positive integers.

5. The method according to claim 1, characterized in that, In step S4, the k-means algorithm is used to classify the landslide hazard risk score values, including: Predetermine the number of categories K, and randomly select K objects {a1, a2, ..., a...} K } as the landslide hazard risk score data {X1,X2,…,X n The initial cluster centers are determined by calculating the distance between each data point and each cluster center, assigning the data point to each cluster center according to the principle of proximity, and recalculating the cluster centers. The final result minimizes the sum of squared distances between each data point and its nearest cluster center. The calculation formula is as follows: Where T is the distance objective function for clustering, and q j Let X be the j-th initial cluster center, K be the number of sub-clusters, and X be the number of sub-clusters. i Let be the data to be clustered, n be the number of data points, and i, j, and n be positive integers.

6. A landslide hazard classification and risk assessment system that takes into account surface deformation, used to implement the method described in any one of claims 1-5, characterized in that, The system includes: A construction module is used to extract indicator data corresponding to disaster-causing factors, disaster-inducing factors, and disaster-bearing factors, as well as surface deformation phase and InSAR surface deformation phase anomaly areas, based on radar imagery and topographic data, by integrating InSAR technology, deep learning technology, and GIS spatial analysis technology. This generates a landslide hazard risk assessment factor database and constructs a landslide hazard risk assessment factor system. The disaster-causing factors include surface deformation rate, and the corresponding disaster-causing factor index includes the average surface deformation rate. The disaster-inducing factors include slope, elevation, and slope units, and the corresponding disaster-inducing factor indexes include the maximum elevation difference of the hazard, the average slope, and the number of slope units where the hazard is located. The disaster-bearing factors include buildings, roads, and rivers, and the corresponding disaster-bearing factor indexes include the number of dangerous buildings, the number of dangerous roads, and the area of ​​dangerous water bodies. The generation module is used to determine potential initial slip point grids based on the InSAR surface deformation phase anomaly area, the surface deformation phase, the surface deformation rate, the slope, and the elevation; construct a Monte Carlo landslide random walk algorithm for slopes; design slip cutoff conditions; perform Monte Carlo slope slip simulation until the slip cutoff conditions are met; count the frequency of each grid selected in the slip path; convert the frequency data into spatial vectors; generate the potential landslide slip range as the potential threat range of landslide hazards; and use GIS spatial analysis technology to extract hazard-related disaster-bearing factor index data. Constructing a Monte Carlo landslide random walk algorithm for slopes, including: Using the InSAR surface deformation phase anomaly area as the initial potential slip point of the landslide, a formula is constructed to calculate the random probability of the current slip grid sliding to the target grid. The formula is as follows: Among them, P px Let f be the probability that a target grid cell surrounding the current sliding grid cell becomes the next sliding grid cell, where n is the number of target grid cells, β is the slope between the current sliding grid cell and the target grid cell, and f is the slope between the target grid cells. d and f β These are the weighting factors for adjusting the sliding direction and slope, respectively. When the direction of the next sliding grid is the same as the previous one, f... d For d 2 When the sliding direction changes by 45°, f d Let d be the value of f. When the sliding direction changes by 90°, d f is 1 β Both d and f are initialization input parameters. r The average surface deformation phase rate; The maximum horizontal slip distance is estimated as the slip cutoff condition, and the calculation formula is as follows: L=K1×ΔH+K2(f s +f r )+B' (3) Where L is the maximum horizontal slip distance, ΔH is the height difference from the top of the landslide to the ground, and f s and f r K1 and K2 are the area of ​​the surface deformation phase accumulation zone and the average surface deformation phase rate, respectively. K1 and K2 are the correlation coefficients, and B' is a constant term. K1, K2 and B' can be obtained by fitting historical observation data. The frequency with which each grid cell in the sliding path is selected in the sliding simulation is calculated, i.e., the probability of sliding to the target grid cell each time. The calculation formula is as follows: Where P is the probability of the landslide affecting the target grid, and s i n represents the number of times the target grid was selected during the simulation for the slip point, m represents the number of potential hazards associated with the slip point affecting the target grid, and n represents the number of slip points that affected the target grid. max Let i and m be the maximum number of iterations for the Monte Carlo random walk simulation, where i and m are positive integers. The solution module is used to determine the hierarchical relationship of the landslide hazard risk assessment factor system using the analytic hierarchy process (AHP), including the target layer-factor layer and the factor layer-indicator layer. It constructs a pairwise comparison matrix, combines expert opinions and relevant data, and solves the weight coefficients of the two hierarchical relationships. The target layer is the landslide hazard risk score value, the factor layer includes disaster-causing factors, disaster-inducing factors, and disaster-bearing factors, and the indicator layer includes disaster-causing factor indicators, disaster-inducing factor indicators, and disaster-bearing factor indicators. The assessment module is used to standardize the values ​​of each indicator, and then perform a weighted summation of the standardized values ​​according to the weight coefficients of the two hierarchical relationships to obtain the landslide hazard risk score. The landslide hazard risk score is then classified to obtain the landslide hazard risk level.

7. The system according to claim 6, characterized in that, Typical disaster-bearing targets, i.e., disaster-bearing factors, are extracted using the following methods. Furthermore, using deformation phase data and elevation data as inputs for model training and inference, InSAR surface deformation phase anomaly areas are extracted using the following methods: (1) Collect sample data High-resolution remote sensing images of the study area were collected as basic data. Targets with representative features were selected for manual interpretation as reference ground truth. The labeled ground truth images and the corresponding original high-resolution images were combined by band. The combined images were cut into several slices of a preset size and randomly divided into a sample training set and a sample test set according to a preset ratio. (2) Model building A multi-scale feature fusion network model is constructed, which adopts an encoder-decoder architecture. The encoder part extracts image features step by step through multiple Depconvgroup modules. Each Depconvgroup module consists of two sets of depthwise separable convolutional layers, BN normalization layers, and ReLU activation layers. The decoder part consists of Cbri modules at different scales. Each Cbri module consists of convolutional layers, BN normalization layers, and upsampling layers. After upsampling, features at different scales are processed into a uniform size. Finally, features at different levels are concatenated along the channel dimension, and a channel attention mechanism module is introduced to improve the model's ability to extract key features. (3) Training the model The multi-scale feature fusion network model is trained multiple times using the slice data of the sample training set. If the error index decreases and tends to stabilize, it indicates that the model training is complete. The multi-scale feature fusion network model is then validated using the slice data of the sample test set. If the validation model accuracy does not meet the requirements, the model training parameters are adjusted and the model is retrained until the accuracy requirements are met. (4) Model Reasoning The trained multi-scale feature fusion network model is used to predict all data in the study area to obtain the distribution results of typical disaster-bearing bodies in the entire area.

8. The system according to claim 6, characterized in that, The hierarchical relationship weights between the target layer and the factor layer, and between the factor layer and the indicator layer, were determined using the analytic hierarchy process (AHP). The weights of individual hierarchical relationships are determined using the analytic hierarchy process (AHP), including: Construct a matrix M, where the elements are labeled a. ij Based on expert opinions and relevant data, the evaluation indicators of the two levels to be solved are compared pairwise, and the determined scales are filled into the corresponding positions in matrix M. The resulting matrix M is the pairwise comparison matrix; where a ij a represents the degree of importance or scale of i compared to j. ij The corresponding position is the i-th row and j-th column in matrix M; if i = j, it means that i and j are equally important, i.e., a ij =1, the scale of all diagonal elements of matrix M is 1; a ij >0 and a ij ×a ji =1; The weights are calculated using the arithmetic mean, geometric mean, and eigenvalue method, respectively, and the arithmetic mean of the results is calculated. Assume the pairwise comparison matrix M is: The weight value w of the i-th factor can be obtained using the following three methods. i ,in, The first method is to obtain the weight vector w using the arithmetic mean. i for: The second method is to obtain the weight vector w using the geometric mean method. i for: The third method is to calculate the weight w using the eigenvalue method. i Let's first assume that the largest eigenvalue of the pairwise comparison matrix M is λ. max If the corresponding eigenvector is ξ, then the weight w can be obtained by the eigenvalue method. i for: The consistency of pairwise comparison matrices is checked using the following methods: Calculate the largest eigenvalue λ of the eigenvectors of the pairwise comparison matrix. max The calculation formula is: Where M is the pairwise comparison matrix, w i For the indicator weights, W = [w1, ..., w n ] T , where n is the order of the comparison matrix, and i, j, k, n are positive integers; The consistency index (CI) of the pairwise comparison matrix is ​​calculated using the following formula: When CI is close to 0, it is preliminarily determined that the consistency test of the constructed pairwise comparison matrix has passed. The consistency ratio is calculated using the following formula: Here, RI is the average consistency index, a parameter value obtained from the lookup table. Typically, when CR < 0.1, it indicates that the consistency test of the pairwise comparison matrix has passed.

Images