Landslide susceptibility evaluation method for slope units based on optimized negative sample selection

Abstract

The present application provides a method for evaluating the landslide susceptibility of slope units based on optimized negative sample selection, comprising: acquiring geological data of a target area, and determining landslide hazard factors based on the geological data; determining slope units in the target area based on DEM data, and determining corresponding grid-level shallow features based on the landslide hazard factors in each slope unit; determining a number of positive samples from the target area based on landslide cataloging, and determining a number of negative samples from the target area based on the landslide hazard factors; constructing a graph convolutional neural network model, inputting all slope units in the target area as graph structures into the graph convolutional neural network model, completing model training and verification using a sample set formed by positive samples and negative samples, and evaluating the landslide susceptibility of slope units in the target area through the trained graph convolutional neural network model, obtaining and outputting the probabilistic landslide susceptibility probability of each slope unit in the target area.

Description

Technical Field

[0001] The present application relates to the technical field of landslide geological disaster assessment, and in particular to a method for assessing the susceptibility of slope units to landslides based on optimized negative sample selection. Background Art

[0002] Landslide disaster is one of the most common and serious geological disasters in the world. Regional landslide susceptibility assessment and stability analysis are effective ways to predict the spatial location and occurrence probability of potential landslides.

[0003] Selecting a reasonable basic unit for landslide evaluation is the basis for landslide evaluation. Currently, the commonly used evaluation units mainly include grid units, regional units, uniform condition units and slope units. Among them, the most commonly used are grid units and slope units. Grid units have become one of the most widely used evaluation units due to their strong applicability to infinite slope models, deep learning models, etc. However, this method cannot characterize the boundaries, terrain characteristics and physical meaning of slopes in the real world. Therefore, in recent years, landslide evaluation methods based on slope units as basic evaluation units have received more and more attention. Compared with grid units, slope units can reflect the different terrain characteristics of landslide areas.

[0004] In terms of the evaluation methods of landslide susceptibility, they are generally divided into qualitative methods and quantitative methods. Limited by data sources and algorithm technology, early landslide research mainly evaluated landslide susceptibility qualitatively based on expert knowledge and experience. With the increasing demand for social disaster prevention and control and the development of computer hardware and algorithm technology, the work of landslide susceptibility evaluation (or landslide susceptibility prediction) has gradually transitioned from qualitative analysis to quantitative analysis. In the current landslide susceptibility assessment research, due to the adaptability and good nonlinear fitting between complex geological environmental conditions and landslides, more machine learning methods such as decision trees, support vector machines, artificial neural networks and random forests are used. Due to the excellent nonlinear feature mining ability of convolutional neural networks, they are often used to mine the implicit laws of landslide inducing factors and landslide events, so they are widely used in landslide susceptibility evaluation.

[0005] In the above deep learning model training methods, positive samples are generally obtained from landslide catalogs that contain information such as the geographical location, boundaries, and types of landslides, while negative samples are often selected randomly. This selection method often ignores the impact of the geographical environment on sample selection. The credibility of negative samples collected in this way is random and unknown, so it is urgent to apply a negative sample screening method to ensure the accuracy and reliability of landslide susceptibility evaluation. Summary of the invention

[0006] The purpose of the embodiments of the present application is to provide a method for evaluating the landslide susceptibility of a slope unit based on optimizing the selection of negative samples, so as to ensure the accuracy and reliability of the landslide susceptibility evaluation by optimizing the selection of negative samples.

[0007] In order to achieve the above purpose, the embodiments of the present application are implemented in the following manner:

[0008] In a first aspect, an embodiment of the present application provides a method for evaluating the landslide susceptibility of slope units based on optimized negative sample selection, comprising: acquiring geological data of a target area, and determining landslide hazard factors based on the geological data, wherein the geological data include river networks, road networks, faults, geological maps, multispectral satellite images, and DEM data, and the landslide hazard factors include elevation, slope, slope aspect, terrain curvature, fault distance, river distance, road distance, land use, lithology, and vegetation index; based on the DEM data in the geological data, determining the slope units in the target area, and based on the landslide hazard factors in each slope unit, determining the grid-level shallow features in each slope unit; based on the landslide catalog, A number of positive samples are determined from the slope units in the target area, and, based on the landslide hazard factors, a number of negative samples are determined from the slope units in the target area, wherein the positive samples represent slope units with sample labels as having landslides, and the negative samples represent slope units with sample labels as having no landslides. A graph convolutional neural network model is constructed, and all slope units in the target area are input into the graph convolutional neural network model. The sample set formed by the positive samples and the negative samples is used to complete the training and verification of the graph convolutional neural network model. The landslide susceptibility of the slope units in the target area is evaluated through the trained graph convolutional neural network model, and the probabilistic landslide susceptibility probability of each slope unit in the target area is obtained and output.

[0009] In combination with the first aspect, in a first possible implementation method of the first aspect, multiple slope units in the target area are determined based on DEM data in the geological data, including: performing surface hydrological process analysis on the target area based on the DEM data to extract basic slope units in the target area; performing coarse extraction on the basic slope units in the target area to determine corresponding primary slope units; and correcting and optimizing the primary slope units to obtain slope units in the target area.

[0010] In combination with the first aspect, in a second possible implementation of the first aspect, based on the landslide hazard factor in each slope unit, the grid-level shallow characteristics in each slope unit are determined, including: for each slope unit: respectively calculating the mean and variance of the elevation, the mean and variance of the slope, the mean and variance of the slope aspect, the mean and variance of the terrain curvature, the mean and variance of the fault distance, the mean and variance of the river distance, the mean and variance of the road distance, and the mean of the land use conditions in the slope unit; The mean and variance of elevation, mean and variance of slope, mean and variance of aspect, mean and variance of terrain curvature, mean and variance of fault distance, mean and variance of river distance, mean and variance of road distance, mean and variance of land use, mean and variance of lithology, mean and variance of vegetation index within this slope unit are concatenated into a 20-dimensional feature vector as the grid-level shallow feature within this slope unit.

[0011] In combination with the first aspect, in a third possible implementation of the first aspect, the landslide disaster-pregnant factors include discrete influencing factors and continuous influencing factors, the discrete influencing factors include land use and lithology, and the continuous influencing factors include elevation, slope, slope aspect, terrain curvature, fault distance, river distance, road distance, and vegetation index. Based on the landslide disaster-pregnant factors, a number of negative samples are determined from the slope units in the target area, including: for each slope unit: for each landslide disaster-pregnant factor in the discrete influencing factors, the frequency ratio method is used to calculate the geographic information similarity with all positive samples to obtain the corresponding similarity component; for each landslide disaster-pregnant factor in the continuous influencing factors, the kernel density estimation method is used to calculate the geographic information similarity with all positive samples to obtain the corresponding similarity component; based on the similarity component corresponding to each landslide disaster-pregnant factor, the comprehensive similarity of the geographic environment between the slope unit and the positive sample is calculated; based on the comprehensive similarity of the geographic environment between each slope unit and the positive sample, a number of negative samples are selected, wherein the comprehensive similarity of the geographic environment between the selected negative samples and the positive samples is lower than a set value.

[0012] In combination with the third possible implementation of the first aspect, in a fourth possible implementation of the first aspect, for each landslide hazard factor in the discrete influencing factors, a frequency ratio method is used to calculate the geographic information similarity with all positive samples to obtain a corresponding similarity component, including: for each landslide hazard factor in the discrete influencing factors, the similarity component corresponding to the slope unit is calculated using the following formula:

[0013]

[0014] in, represents the similarity component corresponding to the slope unit under the i-th discrete impact factor, FRij , P ij , A ij represents the frequency ratio, number of landslides and landslide area of ​​the jth type of the ith discrete influencing factor, respectively. r , A r They represent the total number of landslides and the total landslide area in the target area, respectively.

[0015] In combination with the third possible implementation of the first aspect, in a fifth possible implementation of the first aspect, for each landslide hazard factor in the continuous influencing factors, a kernel density estimation method is used to calculate the geographic information similarity with all positive samples to obtain corresponding similarity components, including: for each landslide hazard factor in the continuous influencing factors, the similarity component corresponding to the slope unit is calculated using the following formula:

[0016]

[0017] in, represents the similarity component corresponding to the slope unit under the mth continuous impact factor, f(x m ) represents the mth continuous impact factor x m The probability density function between the landslide frequency, N is the total number of positive samples, x mn It represents the value of the nth positive sample under the mth continuous impact factor, k is the kernel function, h is the bandwidth, and they satisfy:

[0018]

[0019]

[0020] Wherein, σ represents the standard deviation of the values ​​of N positive samples under the mth continuous impact factor.

[0021] In combination with the first aspect, in a sixth possible implementation of the first aspect, a graph convolutional neural network model is constructed, all slope units in the target area are input into the graph convolutional neural network model, and a sample set formed by positive samples and negative samples is used to complete the training and verification of the graph convolutional neural network model, including: constructing a graph convolutional neural network model, wherein the graph convolutional neural network model includes a feature extraction module and a graph convolution module, the feature extraction module is used to extract object-level semantic features based on the grid-level shallow features corresponding to the slope units, and the graph convolution module is used to determine the landslide susceptibility probability corresponding to the slope units based on the slope units and their corresponding object-level semantic features; based on the grid-level shallow features corresponding to all positive samples and the negative samples, the graph convolution module extracts object-level semantic features from the slope units, and ... The grid-level shallow features corresponding to the samples are used to train and verify the feature extraction module. After the training is completed, the trained feature extraction module is used to extract the grid-level shallow features corresponding to each slope unit in the target area to obtain the object-level semantic features corresponding to each slope unit; based on each slope unit and its corresponding object-level semantic features, as well as the relationship with adjacent slope units, all slope units are processed into a graph structure and input into the graph convolution module; all positive samples and negative samples in the graph structure are composed of a sample set, and the sample set is divided into a training set and a verification set in a ratio of 7:3. The graph convolution module is trained and verified using the training set and the verification set, and finally a trained graph convolutional neural network model is obtained.

[0022] In combination with the sixth possible implementation manner of the first aspect, in the seventh possible implementation manner of the first aspect, the feature extraction module of the graph convolutional neural network model includes two convolution combination layers and two fully connected layers connected in sequence, each convolution combination layer includes a convolution layer, a batch normalization layer and a rectifier linear unit activation layer connected in sequence, and the feature extraction module is specifically used for: in the training stage of the feature extraction module, performing convolution calculation, normalization and standardization of feature intervals on the grid-level shallow features of the sample through the first convolution combination layer, and then performing convolution calculation, normalization and standardization of feature intervals through the second convolution combination layer, and then stretching the features into one-dimensional vectors and passing them into the first and second fully connected layers in sequence, using binary The crossentropy loss function calculates the error between the eigenvalue and the sample label, updates the network parameters through error back propagation, and uses the feature vector obtained after calculation by the second fully connected layer as the object-level semantic feature corresponding to the sample, where the sample label is used to reveal whether a landslide occurs in the slope unit of the sample. The sample label of the positive sample is that a landslide occurs, and the sample label of the negative sample is that a landslide does not occur. After the feature extraction module is trained, the grid-level shallow features of the input slope unit are convolved, normalized, and the feature interval is standardized through the first convolution combination layer, and then convolved, normalized, and the feature interval is standardized through the second convolution combination layer. The feature is stretched into a one-dimensional vector and passed into the first and second fully connected layers in turn. The feature vector obtained after calculation by the second fully connected layer is used as the object-level semantic feature corresponding to the slope unit.

[0023] In combination with the sixth possible implementation manner of the first aspect, in an eighth possible implementation manner of the first aspect, based on each slope unit and its corresponding object-level semantic features, and the relationship between adjacent slope units, all slope units are processed into a graph structure, including: for each slope unit: taking the slope unit as a node, and taking the object-level semantic features corresponding to this slope unit as the node attribute of this node; determining the neighboring slope units adjacent to this slope unit, and establishing edges between this slope unit and each neighboring slope unit respectively; calculating the centroid distance and average height difference between this slope unit and each neighboring slope unit respectively, and taking each group of centroid distance and average height difference as the edge attribute between this slope unit and the corresponding neighboring slope unit, and finally obtaining a graph structure containing all slope units.

[0024] In combination with the sixth possible implementation manner of the first aspect, in a ninth possible implementation manner of the first aspect, the graph convolution module of the graph convolution neural network model includes two graph convolution layers and a machine learning classifier XGBoost, and the graph convolution module is specifically used for: in the training phase of the graph convolution module, performing graph convolution processing on samples in the graph structure through the two graph convolution layers to obtain graph convolution features corresponding to the samples, and then concatenating the pixel mean and standard deviation of the samples with their graph convolution features to obtain a combined feature vector corresponding to the samples, and inputting the combined feature vector into the machine learning classifier XGBoost for classification, outputting the landslide susceptibility probability of the samples, and then classifying the landslide susceptibility based on the landslide susceptibility. The error is calculated based on the probability of landslide and the sample label, and the network parameters are updated by error back propagation, where the sample label is used to reveal whether a landslide occurs in the slope unit of the sample. The sample label of the positive sample is that a landslide occurs, and the sample label of the negative sample is that a landslide does not occur. After the training of the graph convolution module is completed, the slope units in the graph structure are processed by graph convolution through two layers of graph convolution layers to obtain the graph convolution features corresponding to the slope units, and then the pixel mean and standard deviation of the slope unit are concatenated with its graph convolution features to obtain the combined feature vector corresponding to the slope unit, which is input into the machine learning classifier XGBoost for classification, and the landslide susceptibility probability of the slope unit is output.

[0025] Beneficial effects:

[0026] 1. In geographic analysis, the similarity of geographic variables in a certain area can be used to evaluate the similarity of other geographic variables or other regional activities. For landslide negative samples, the smaller the similarity of the geographic variables with the landslide area, the higher the credibility of the negative sample. Inspired by this, this scheme selects 10 landslide disaster factors. Discrete influencing factors include land use and lithology, and continuous influencing factors include elevation, slope aspect, slope, terrain curvature, fault distance, river distance, road distance, and vegetation index. The frequency ratio method and kernel density analysis method are used to calculate the similarity components of slope units in the target area, and the comprehensive similarity of the geographical environment is further calculated. The landslide negative samples for model training are screened according to the similarity. Compared with the conventional scheme of randomly selecting negative samples, the reliability of negative samples is greatly improved.

[0027] 2. This scheme transforms the pixel-level features in the slope unit into graph convolution features by constructing a graph convolution neural network model, and prepares the slope units in the entire target area into a graph structure, analyzes the landslide susceptibility in the target area, determines the probabilistic landslide susceptibility of each slope unit in the target area, and further obtains the landslide susceptibility mapping of the target area. This method is conducive to mining deep semantic features, and combining them with pixel-level features into new feature vectors by vector splicing. The purpose is to retain the semantic features while introducing the original feature vector (that is, mean and variance), which can comprehensively consider the deep features (semantic information) and shallow features (mean variance) of the slope unit, which is conducive to improving the accuracy and reliability of landslide susceptibility evaluation.

[0028] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, preferred embodiments are specifically cited below and described in detail with reference to the attached drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments of the present application will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0030] Figure 1 A flow chart of a method for evaluating the landslide susceptibility of slope units based on optimized negative sample selection provided in an embodiment of the present application. DETAILED DESCRIPTION

[0031] The technical solutions in the embodiments of the present application will be described below in conjunction with the drawings in the embodiments of the present application.

[0032] In order to facilitate the understanding of this solution, some terms and concepts involved in this embodiment are first explained here.

[0033] Landslide disaster: Landslide disaster refers to the disaster caused by the overall sliding downhill of rock or soil under the action of gravity.

[0034] Landslide susceptibility assessment (i.e. landslide susceptibility evaluation): Landslide susceptibility evaluation refers to the process of conducting a comprehensive analysis of geological, topographical, climatic and other factors to determine the degree of possibility of landslides occurring in a certain area or location.

[0035] Slope unit: Slope unit refers to a structural system built on a slope. Its main function is to bear and transmit the stress inside the slope to maintain the stability of the slope and prevent the slope from sliding or collapsing.

[0036] Digital Elevation Model (DEM): A digital elevation model is a digital simulation of the ground terrain (i.e., a digital expression of the terrain surface morphology) through limited terrain elevation data. It is a physical ground model that represents the ground elevation in the form of a set of ordered numerical arrays.

[0037] Law of geographical similarity: The more similar the geographical environment is, the similar the characteristics and development processes of the geographical targets are.

[0038] Continuous geographic variables: Continuous geographic variables are variables that have a continuous distribution in geographic space. These variables can take continuous values ​​at any point in geographic space, not just at limited discrete points.

[0039] Discrete geographic variables: Discrete geographic variables refer to variables with discrete distribution in geographic space. The values ​​of these variables can only be finite or countable discrete values, rather than continuous.

[0040] Frequency ratio method: The frequency ratio statistical model is a method used to estimate the association between events. It is based on the concept of frequency ratio, which determines the degree of association between two events by comparing the difference between the frequency of their co-occurrence and the frequency of their independent occurrence.

[0041] Kernel Density Estimation (KDE): Kernel density estimation is used to estimate unknown density functions in probability theory and is one of the non-parametric test methods.

[0042] Graph Convolutional Neural Network (GCN): Graph Convolutional Neural Network is a deep learning model that represents and learns node features by defining convolution operations on the graph structure. GCN considers the connection relationship between a node and its neighbor nodes, effectively propagates and aggregates node features in graph structure data, and thus achieves feature learning and prediction of the entire graph.

[0043] See also Figure 1 , Figure 1 A flow chart of a slope unit landslide susceptibility evaluation method based on optimized negative sample selection provided in an embodiment of the present application. In this embodiment, the slope unit landslide susceptibility evaluation method based on optimized negative sample selection can be applied to electronic devices (such as smart terminals, personal computers, servers, etc.), including step S10, step S20, step S30 and step S40.

[0044] First, the electronic device may execute step S10.

[0045] Step S10: Acquire geological data of the target area, and determine landslide hazard factors based on the geological data, wherein the geological data include river network, road network, fault, geological map, multispectral satellite image, DEM data, and the landslide hazard factors include elevation, slope, slope aspect, terrain curvature, fault distance, river distance, road distance, land use, lithology, and vegetation index.

[0046] In this embodiment, the electronic device can obtain the geological data of the target area, including river network, road network, fault, geological map, multi-spectral satellite image, DEM data, etc. Then, based on the geological data, through remote sensing pre-processing methods such as clipping, resampling, and Euclidean distance analysis, 10 landslide hazard factors such as elevation, slope, slope aspect, terrain curvature, fault distance, river distance, road distance, land use, lithology, and vegetation index can be obtained.

[0047] After determining the landslide hazard factor, the electronic device may execute step S20.

[0048] Step S20: Based on the DEM data in the geoscience data, the slope units in the target area are determined, and based on the landslide hazard factors in each slope unit, the grid-level shallow features in each slope unit are determined.

[0049] In this embodiment, the electronic device can determine the slope units in the target area based on DEM data in the geological data.

[0050] For example, the electronic device can analyze the surface hydrological process of the target area based on the DEM data to extract the basic slope units in the target area. For example, the basic slope units are extracted by extracting the flow direction and the accumulated amount of runoff, generating a river network, extracting the positive and negative catchment basins, merging the catchment basins, and the positive and negative DEM processes.

[0051] Then, the electronic device can roughly extract the basic slope units in the target area and determine the corresponding primary slope units. Then, the primary slope units are corrected and optimized to obtain the slope units in the target area. For example, the wrong and unreasonable slope units can be manually optimized and corrected to finally establish the fine slope units in the target area.

[0052] After determining the slope units in the target area, the electronic device can determine the grid-level shallow features in each slope unit based on the landslide hazard factors in each slope unit.

[0053] Exemplarily, for each slope unit:

[0054] The electronic device can respectively calculate the mean and variance of the elevation, the mean and variance of the slope, the mean and variance of the aspect, the mean and variance of the terrain curvature, the mean and variance of the fault distance, the mean and variance of the river distance, the mean and variance of the road distance, the mean and variance of the land use, the mean and variance of the lithology, and the mean and variance of the vegetation index in the slope unit, and then splice the mean and variance of the elevation, the mean and variance of the slope, the mean and variance of the aspect, the mean and variance of the terrain curvature, the mean and variance of the fault distance, the mean and variance of the river distance, the mean and variance of the road distance, the mean and variance of the land use, the mean and variance of the lithology, and the mean and variance of the vegetation index in the slope unit into a 20-dimensional feature vector as a grid-level shallow feature in the slope unit.

[0055] After determining the grid-level shallow features in each slope unit, the electronic device may execute step S30.

[0056] Step S30: Based on the landslide catalog, several positive samples are determined from the slope units in the target area, and based on the landslide hazard factors, several negative samples are determined from the slope units in the target area, wherein the positive samples represent slope units with sample labels as having landslides, and the negative samples represent slope units with sample labels as having no landslides.

[0057] In this embodiment, the electronic device can determine a number of positive samples from the slope units in the target area based on the landslide catalog (including the geographical location, boundary, type, etc. of the landslide). The scheme for determining positive samples is a mature technology in the prior art, that is, it is generally understood that the slope units in the target area where landslides have occurred are determined as positive samples. Therefore, the positive sample means that the sample label is the slope unit where the landslide has occurred.

[0058] As for the selection of negative samples, unlike the traditional random selection scheme (i.e., randomly selecting a part of slope units from all slope units other than positive samples as negative samples), this embodiment adopts the following method to optimize negative samples, so as to improve the credibility of negative sample selection and the accuracy of landslide susceptibility assessment. Here, negative samples refer to slope units with sample labels that do not cause landslides.

[0059] In geographic analysis, we can evaluate the similarity of other geographic variables or other regional activities based on the similarity of geographic variables in a certain area. For negative samples of landslides, the smaller the similarity of geographic variables with the area where the landslide occurred, the higher the credibility of the negative samples. Therefore, the electronic device can determine several negative samples from the slope units in the target area based on the landslide hazard factors.

[0060] Landslide disaster factors include discrete influencing factors and continuous influencing factors. Discrete influencing factors include land use and lithology, and continuous influencing factors include elevation, slope, slope aspect, terrain curvature, fault distance, river distance, road distance, and vegetation index. For example, for each slope unit:

[0061] For each landslide-predisposing factor in the discrete influencing factors, the electronic device can calculate the geographic information similarity with all positive samples using the frequency ratio method to obtain the corresponding similarity component.

[0062] Specifically, for each landslide hazard factor in the discrete influencing factors, the similarity component corresponding to the slope unit is calculated using the following formula:

[0063]

[0064] in, represents the similarity component corresponding to the slope unit under the i-th discrete impact factor, FR ij , P ij , A ij represents the frequency ratio, number of landslides and landslide area of ​​the jth type of the ith discrete influencing factor, respectively. r , A r Respectively represent the total number of landslides and the total landslide area in the target area. Thus, the similarity component corresponding to each slope unit under each discrete influencing factor can be calculated.

[0065] For each landslide-predisposing factor in the continuous influencing factors, the electronic device can use the kernel density estimation method to calculate the geographic information similarity with all positive samples to obtain the corresponding similarity component.

[0066] Specifically, for each landslide hazard factor in the continuous influencing factors, the electronic device can calculate the similarity component corresponding to the slope unit using the following formula:

[0067]

[0068] in, represents the similarity component corresponding to the slope unit under the mth continuous impact factor, f(x m ) represents the mth continuous impact factor x m The probability density function between the landslide frequency, N is the total number of positive samples, x mn It represents the value of the nth positive sample under the mth continuous impact factor, k is the kernel function, h is the bandwidth, and they satisfy:

[0069]

[0070]

[0071] Among them, σ represents the standard deviation of the values ​​of N positive samples under the mth continuous impact factor. From this, the similarity component corresponding to each slope unit under each continuous impact factor can be calculated.

[0072] Afterwards, the electronic device can calculate the comprehensive similarity of the geographical environment between the slope unit and the positive sample based on the similarity component corresponding to each landslide hazard factor.

[0073] For example, the electronic device can comprehensively measure the contribution of a single landslide-pregnant factor to the occurrence of landslides, and measure the comprehensive similarity of the geographical environment between the slope unit and the positive sample. For example, the electronic device can perform comprehensive calculations by using the average method, the minimum limiting factor method, the linear weighted method, etc. In this embodiment, the average method is used to calculate the comprehensive similarity of the geographical environment, and the following formula is used to calculate the comprehensive similarity of the geographical environment between the slope unit and the positive sample:

[0074]

[0075] Among them, S is the comprehensive similarity of the geographical environment between the slope unit and the positive sample, are the similarity components corresponding to the slope units under the first and second discrete impact factors, Represents the similarity component corresponding to the slope unit under the mth continuous impact factor.

[0076] After calculating the comprehensive similarity of the geographical environment between each slope unit and the positive sample, the electronic device can select several negative samples based on the comprehensive similarity of the geographical environment between each slope unit and the positive sample, wherein the comprehensive similarity of the geographical environment between the selected negative samples and the positive sample is lower than a set value.

[0077] For example, the electronic device may use the slope units with a geographical environment comprehensive similarity less than 0.3 as candidate areas for negative sample selection, and then select the same number of negative samples as positive samples from the candidate areas by random selection. Alternatively, the geographical environment comprehensive similarity may be used to sort the similarity of each slope unit except the positive sample, from low to high, and then select the slope unit with the lowest similarity as the same number of positive samples as the negative sample.

[0078] After determining the positive samples and the negative samples, the electronic device may execute step S40.

[0079] Step S40: construct a graph convolutional neural network model, input all slope units in the target area into the graph convolutional neural network model, use the sample set formed by positive samples and negative samples to complete the training and verification of the graph convolutional neural network model, and use the trained graph convolutional neural network model to evaluate the landslide susceptibility of the slope units in the target area, and obtain the probabilistic landslide susceptibility probability of each slope unit in the target area and output it.

[0080] In this embodiment, the electronic device can construct a graph convolutional neural network model, wherein the graph convolutional neural network model includes a feature extraction module and a graph convolution module, the feature extraction module is used to extract object-level semantic features based on the grid-level shallow features corresponding to the slope unit, and the graph convolution module is used to determine the landslide susceptibility probability corresponding to the slope unit based on the slope unit and its corresponding object-level semantic features.

[0081] For the feature extraction module, the electronic device can train and verify the feature extraction module based on the grid-level shallow features corresponding to all positive samples and the grid-level shallow features corresponding to negative samples. After the training is completed, the trained feature extraction module is used to extract the grid-level shallow features corresponding to each slope unit in the target area to obtain the object-level semantic features corresponding to each slope unit.

[0082] Exemplarily, the feature extraction module of the graph convolutional neural network model includes two sequentially connected convolutional combination layers and two fully connected layers, and each convolutional combination layer includes a sequentially connected convolutional layer, a batch normalization layer, and a rectified linear unit activation layer. Then, the feature extraction module is specifically used for:

[0083] In the training stage of the feature extraction module, the grid-level shallow features of the samples (positive samples and negative samples are collectively referred to as samples, and the samples have sample labels. The sample labels of the positive samples are that landslides occur, and the sample labels of the negative samples are that landslides do not occur) are convolved, normalized, and the feature intervals are standardized through the first convolution combination layer. Then, the convolution calculation, normalization, and standardization of the feature intervals are performed through the second convolution combination layer. The features are stretched into one-dimensional vectors and passed into the first and second fully connected layers in turn. The binary crossentropy loss function is used to calculate the error between the feature value and the sample label, and the network parameters are updated through error back propagation. In addition, the feature vector obtained after calculation by the second fully connected layer is used as the object-level semantic feature corresponding to the sample.

[0084] After the feature extraction module is trained, the grid-level shallow features of the input slope unit can be convolved, normalized, and the feature interval can be standardized through the first convolution combination layer, and then convolutionally calculated, normalized, and standardized through the second convolution combination layer. The features are stretched into one-dimensional vectors and passed into the first and second fully connected layers in turn. The feature vector calculated by the second fully connected layer is used as the object-level semantic feature corresponding to the slope unit.

[0085] After obtaining the object-level semantic features corresponding to each slope unit, the electronic device can process all slope units into a graph structure based on each slope unit and its corresponding object-level semantic features, as well as the relationship between adjacent slope units, and input them into the graph convolution module.

[0086] Exemplarily, for each slope unit: the electronic device can take the slope unit as a node, and use the object-level semantic features corresponding to the slope unit as the node attribute of the node; determine the adjacent slope units adjacent to the slope unit, and establish the edges between the slope unit and each adjacent slope unit; calculate the centroid distance and average height difference between the slope unit and each adjacent slope unit, and use each group of centroid distance and average height difference as the edge attribute between the slope unit and the corresponding adjacent slope unit, and finally obtain a graph structure containing all slope units. After that, the graph structure can be input into the graph convolution module.

[0087] For the graph convolution module, the electronic device can form a sample set with all positive samples and negative samples in the graph structure, divide the sample set into a training set and a verification set in a ratio of 7:3, and use the training set and the verification set to train and verify the graph convolution module, and finally obtain a trained graph convolution neural network model.

[0088] Exemplarily, the graph convolution module of the graph convolutional neural network model includes two graph convolution layers and a machine learning classifier XGBoost, and the graph convolution module is specifically used for:

[0089] During the training phase of the graph convolution module, graph convolution is performed on samples in the graph structure through two graph convolution layers to obtain the graph convolution features corresponding to the samples. The pixel mean and standard deviation of the samples are then concatenated with their graph convolution features to obtain the combined feature vector corresponding to the samples. The combined feature vector is input into the machine learning classifier XGBoost for classification, and the landslide susceptibility probability of the samples is output. The error is then calculated based on the landslide susceptibility probability and sample labels, and the network parameters are updated through error back propagation.

[0090] After the graph convolution module training is completed, the slope units in the graph structure can be processed by graph convolution through two layers of graph convolution layers to obtain the graph convolution features corresponding to the slope units, and then the pixel mean and standard deviation of the slope units are spliced ​​with their graph convolution features to obtain the combined feature vector corresponding to the slope units, which is input into the machine learning classifier XGBoost for classification and outputs the landslide susceptibility probability of the slope units. In this way, the landslide susceptibility evaluation of the slope units in the target area can be completed, and the probabilistic landslide susceptibility probability of each slope unit in the target area can be obtained.

[0091] Afterwards, the electronic device can obtain the final landslide susceptibility mapping based on the landslide susceptibility probability of the slope unit. For example, the following rules are used to evaluate and grade landslide susceptibility: if the susceptibility probability is greater than or equal to 0.9, it is determined to be an extremely high risk area; if the susceptibility probability is greater than or equal to 0.7 and less than 0.9, it is determined to be a high risk area; if the susceptibility probability is greater than or equal to 0.5 and less than 0.7, it is determined to be a medium risk area; if the susceptibility probability is greater than or equal to 0.3 and less than 0.5, it is determined to be a low risk area; if the susceptibility probability is less than 0.3, it is determined to be an extremely low risk area. Even slope units with different landslide susceptibility evaluation levels can be represented by one-to-one corresponding colors, thereby obtaining a landslide susceptibility mapping of the target area.

[0092] The slope unit landslide susceptibility assessment method based on optimized negative sample selection in this embodiment has been applied to Shimian County, Ya'an City, Sichuan Province, and the overall accuracy (OA), precision, recall, F1-Score and area under the ROC curve (AUC) were used for accuracy evaluation, reaching 0.815, 0.952, 0.815, 0.861 and 0.915 respectively, indicating that this method can be well applied to the susceptibility assessment of landslide disaster areas, and provide methodological, technical and data support for subsequent landslide prevention and post-disaster rescue work.

[0093] In summary, the embodiment of the present application provides a method for evaluating the landslide susceptibility of slope units based on optimized negative sample selection. This scheme selects 10 landslide-predisposing factors, discrete influencing factors include land use and lithology, and continuous influencing factors include elevation, slope, slope aspect, terrain curvature, fault distance, river distance, road distance, and vegetation index. The frequency ratio method and kernel density analysis method are used to calculate the similarity components of slope units in the target area, and the comprehensive similarity of the geographical environment is further calculated. The landslide negative samples for model training are screened according to the similarity. Compared with the conventional scheme of randomly selecting negative samples, the reliability of negative samples is greatly improved.

[0094] This scheme transforms the pixel-level features in the slope unit into graph convolution features by constructing a graph convolution neural network model, and prepares the slope units in the entire target area into a graph structure, analyzes the landslide susceptibility in the target area, determines the probabilistic landslide susceptibility of each slope unit in the target area, and further obtains the landslide susceptibility mapping of the target area. This method is conducive to mining deep semantic features, and concatenating them with pixel-level features into new feature vectors. The purpose is to retain the semantic features while introducing the original feature vector (that is, mean and variance), which can comprehensively consider the deep features (semantic information) and shallow features (mean variance) of the slope unit, which is conducive to improving the accuracy and reliability of landslide susceptibility evaluation.

[0095] In this document, relational terms such as first and second, etc. are used merely to distinguish one entity or operation from another entity or operation, but do not necessarily require or imply any such actual relationship or order between these entities or operations.

[0096] The above description is only an embodiment of the present application and is not intended to limit the protection scope of the present application. For those skilled in the art, the present application may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for evaluating the susceptibility of slope units to landslides based on optimized negative sample selection, characterized in that: include: Obtain the geological data of the target area and determine the landslide hazard factors based on the geological data. The geological data include river network, road network, fault, geological map, multi-spectral satellite image, DEM data. The landslide hazard factors include elevation, slope, slope aspect, terrain curvature, fault distance, river distance, road distance, land use, lithology, and vegetation index; Based on the DEM data in the geoscience data, the slope units in the target area are determined, and based on the landslide hazard factors in each slope unit, the grid-level shallow characteristics in each slope unit are determined; Based on the landslide catalog, several positive samples are determined from the slope units in the target area, and based on the landslide hazard factors, several negative samples are determined from the slope units in the target area, wherein the positive samples represent the slope units with sample labels of landslides, and the negative samples represent the slope units with sample labels of no landslides; Construct a graph convolutional neural network model, input all slope units in the target area into the graph convolutional neural network model, use the sample set formed by positive samples and negative samples to complete the training and verification of the graph convolutional neural network model, and use the trained graph convolutional neural network model to evaluate the landslide susceptibility of slope units in the target area, and obtain and output the probabilistic landslide susceptibility probability of each slope unit in the target area; Based on the landslide hazard factors in each slope unit, the grid-level shallow characteristics in each slope unit are determined, including: For each ramp unit: Calculate the mean and variance of elevation, the mean and variance of slope, the mean and variance of aspect, the mean and variance of terrain curvature, the mean and variance of fault distance, the mean and variance of river distance, the mean and variance of road distance, the mean and variance of land use, the mean and variance of lithology, and the mean and variance of vegetation index in this slope unit respectively; The mean and variance of elevation, the mean and variance of slope, the mean and variance of aspect, the mean and variance of terrain curvature, the mean and variance of fault distance, the mean and variance of river distance, the mean and variance of road distance, the mean and variance of land use, the mean and variance of lithology, and the mean and variance of vegetation index in this slope unit are concatenated into a 20-dimensional feature vector as the grid-level shallow feature in this slope unit; Construct a graph convolutional neural network model, input all slope units in the target area into the graph convolutional neural network model, and use the sample set formed by positive samples and negative samples to complete the training and verification of the graph convolutional neural network model, including: Constructing a graph convolutional neural network model, wherein the graph convolutional neural network model includes a feature extraction module and a graph convolution module, the feature extraction module is used to extract object-level semantic features based on grid-level shallow features corresponding to the slope unit, and the graph convolution module is used to determine the landslide susceptibility probability corresponding to the slope unit based on the slope unit and its corresponding object-level semantic features; Based on the grid-level shallow features corresponding to all positive samples and the grid-level shallow features corresponding to negative samples, the feature extraction module is trained and verified. After the training is completed, the trained feature extraction module is used to extract the grid-level shallow features corresponding to each slope unit in the target area to obtain the object-level semantic features corresponding to each slope unit. Based on each slope unit and its corresponding object-level semantic features, as well as the relationship between adjacent slope units, all slope units are processed into a graph structure and input into the graph convolution module; All positive samples and negative samples in the graph structure are combined into a sample set, and the sample set is divided into a training set and a validation set in a ratio of 7:

3. The training set and the validation set are used to train and verify the graph convolution module, and finally a trained graph convolutional neural network model is obtained.

2. The method for evaluating the susceptibility of slope units to landslides based on optimized negative sample selection according to claim 1, characterized in that: Based on the DEM data in the geoscience data, multiple slope units in the target area are identified, including: Analyze the surface hydrological process of the target area based on DEM data and extract the basic slope units in the target area; Roughly extract the basic slope units in the target area and determine the corresponding primary slope units; The primary slope units are corrected and optimized to obtain the slope units in the target area.

3. The method for evaluating the susceptibility of slope units to landslides based on optimized negative sample selection according to claim 1, characterized in that: Landslide hazard factors include discrete influencing factors and continuous influencing factors. Discrete influencing factors include land use and lithology, and continuous influencing factors include elevation, slope, slope aspect, terrain curvature, fault distance, river distance, road distance, and vegetation index. Based on landslide hazard factors, several negative samples are determined from the slope units in the target area, including: For each ramp unit: For each landslide-pregnant factor in the discrete influencing factors, the frequency ratio method is used to calculate the geographic information similarity with all positive samples to obtain the corresponding similarity component; For each landslide hazard factor in the continuous influencing factors, the kernel density estimation method is used to calculate the geographic information similarity with all positive samples to obtain the corresponding similarity component; Based on the similarity component corresponding to each landslide hazard factor, the comprehensive similarity of the geographical environment between the slope unit and the positive sample is calculated; Based on the comprehensive similarity of the geographical environment between each slope unit and the positive sample, several negative samples are selected, wherein the comprehensive similarity of the geographical environment between the selected negative samples and the positive sample is lower than a set value.

4. The method for evaluating the susceptibility of slope units to landslides based on optimized negative sample selection according to claim 3, characterized in that: For each landslide hazard factor in the discrete influencing factors, the frequency ratio method is used to calculate the geographic information similarity with all positive samples to obtain the corresponding similarity components, including: For each landslide hazard factor in the discrete influencing factors, the similarity component corresponding to the slope unit is calculated using the following formula: in, represents the similarity component corresponding to the slope unit under the i-th discrete impact factor, FR ij , P ij , A ij represents the frequency ratio, number of landslides and landslide area of ​​the jth type of the ith discrete influencing factor, respectively. r , A r They represent the total number of landslides and the total landslide area in the target area, respectively.

5. The method for evaluating the landslide susceptibility of slope units based on optimized negative sample selection according to claim 3, characterized in that: For each landslide hazard factor in the continuous influencing factors, the kernel density estimation method is used to calculate the geographic information similarity with all positive samples to obtain the corresponding similarity components, including: For each landslide hazard factor in the continuous influencing factors, the similarity component corresponding to the slope unit is calculated using the following formula: in, represents the similarity component corresponding to the slope unit under the mth continuous impact factor, f(x m ) represents the mth continuous impact factor x m The probability density function between the landslide frequency, N is the total number of positive samples, x mn It represents the value of the nth positive sample under the mth continuous impact factor, k is the kernel function, h is the bandwidth, and they satisfy: Wherein, σ represents the standard deviation of the values ​​of N positive samples under the mth continuous impact factor.

6. The method for evaluating the susceptibility of slope units to landslides based on optimized negative sample selection according to claim 1, characterized in that: The feature extraction module of the graph convolutional neural network model includes two sequentially connected convolutional combination layers and two fully connected layers. Each convolutional combination layer includes a sequentially connected convolutional layer, a batch normalization layer, and a rectified linear unit activation layer. The feature extraction module is specifically used for: In the training phase of the feature extraction module, the grid-level shallow features of the sample are convolved, normalized, and the feature interval is standardized through the first convolution combination layer, and then convolution is performed, normalized, and the feature interval is standardized through the second convolution combination layer. The features are then stretched into one-dimensional vectors and passed to the first and second fully connected layers in sequence. The error between the feature value and the sample label is calculated using the binary crossentropy loss function, and the network parameters are updated through error back propagation. In addition, the feature vector obtained after calculation by the second fully connected layer is used as the object-level semantic feature corresponding to the sample, where the sample label is used to reveal whether a landslide occurs in the slope unit of the sample. The sample label of the positive sample is that a landslide occurs, and the sample label of the negative sample is that a landslide does not occur. After the feature extraction module is trained, the grid-level shallow features of the input slope unit are convolved, normalized, and the feature interval is standardized through the first convolution combination layer, and then convolution is performed, normalized, and the feature interval is standardized through the second convolution combination layer. The features are then stretched into one-dimensional vectors and passed into the first and second fully connected layers in turn. The feature vector calculated by the second fully connected layer is used as the object-level semantic feature corresponding to the slope unit.

7. The method for evaluating the landslide susceptibility of slope units based on optimized negative sample selection according to claim 1, characterized in that: Based on each slope unit and its corresponding object-level semantic features, as well as the relationship between adjacent slope units, all slope units are processed into a graph structure, including: For each ramp unit: The slope unit is regarded as a node, and the object-level semantic features corresponding to the slope unit are regarded as the node attributes of the node; Determine the adjacent slope units adjacent to the slope unit, and respectively establish edges between the slope unit and each adjacent slope unit; The centroid distance and average height difference between this slope unit and each adjacent slope unit are calculated respectively, and each set of centroid distance and average height difference is used as the edge attribute between this slope unit and the corresponding adjacent slope unit, and finally a graph structure containing all slope units is obtained.

8. The method for evaluating the landslide susceptibility of slope units based on optimized negative sample selection according to claim 1, characterized in that: The graph convolution module of the graph convolutional neural network model includes two graph convolution layers and a machine learning classifier XGBoost. The graph convolution module is specifically used for: In the training phase of the graph convolution module, the samples in the graph structure are processed by graph convolution through two layers of graph convolution layers to obtain the graph convolution features corresponding to the samples. Then, the pixel mean and standard deviation of the samples are concatenated with their graph convolution features to obtain the combined feature vector corresponding to the samples. The combined feature vector is input into the machine learning classifier XGBoost for classification, and the landslide susceptibility probability of the samples is output. The error is calculated based on the landslide susceptibility probability and the sample label, and the network parameters are updated through error back propagation. The sample label is used to reveal whether the slope unit of the sample has a landslide. The sample label of the positive sample is that a landslide has occurred, and the sample label of the negative sample is that a landslide has not occurred. After the training of the graph convolution module is completed, the slope units in the graph structure are convolved through two layers of graph convolution layers to obtain the graph convolution features corresponding to the slope units. Then, the pixel mean and standard deviation of the slope units are concatenated with their graph convolution features to obtain the combined feature vector corresponding to the slope units. The combined feature vector is input into the machine learning classifier XGBoost for classification, and the landslide susceptibility probability of the slope units is output.

Images