A landslide susceptibility assessment method for mixed data

Abstract

The present invention relates to the technical field of landslide susceptibility evaluation, and specifically discloses a landslide susceptibility evaluation method for mixed data, which calculates the correlation matrix of numerical and categorical influencing factors, and generates a final correlation matrix through merging; the final correlation matrix is ​​used as the characteristic data of the input classification model to predict the landslide susceptibility. After the mixed data is classified, the finally obtained correlation matrix saves all the relevant information between the objects, and then the landslide sensitivity is evaluated using the commonly used classification algorithm. The accuracy of the evaluation result is calculated using the receiver operating characteristic curve ROC. After comparing with the result without considering the mixed data, it is found that the landslide susceptibility evaluation method for mixed data improves the accuracy of the susceptibility evaluation.

Description

Technical Field

[0001] The invention relates to the technical field of landslide susceptibility evaluation, in particular to a landslide susceptibility evaluation method oriented to mixed data. Background Art

[0002] Landslides caused by strong earthquakes are one of the many geological disasters that affect human life and property safety.

[0003] Landslide susceptibility assessment is an important technical means of disaster management. It predicts the spatial distribution of future earthquake-induced landslides by analyzing various influencing factors and historical landslide data in the study area. Commonly used earthquake landslide susceptibility assessment methods mainly include physical model method and data-driven method. The physical model method judges the stability of the slope through strict mechanical reasoning, and the prediction results are very accurate. However, in practical applications, the geotechnical mechanics and groundwater parameters required by this method are difficult to obtain, so it cannot be promoted on a large scale. The data-driven method uses statistical models or machine learning algorithms to analyze landslide data to evaluate landslide susceptibility. Among them, statistical models are easily affected by subjective factors in quantifying influencing factors. Data mining technology represented by machine learning can accurately establish the nonlinear relationship between landslide susceptibility and influencing factors, and can process high-dimensional and large amounts of data. This method can be used to characterize the laws of landslides, thereby obtaining high-precision prediction results. Its representative methods include logistic regression, support vector machine, random forest, etc. Domestic and foreign scholars have conducted a lot of research to improve the evaluation accuracy of data-driven methods, such as: simplification and optimization of landslide influencing factors, coupling between different connection methods and different models, and selection of non-landslide samples.

[0004] However, these methods cannot handle mixed data. In fact, the data types of influencing factors can be divided into numerical data and categorical data. Traditional landslide susceptibility evaluation reclassifies continuous numerical influencing factors, thus losing some data details and causing information loss. Moreover, they can only handle pure numerical data and cannot handle mixed data that contains both numerical and categorical data. Summary of the invention

[0005] In view of the above problems in the prior art, the present invention provides a landslide susceptibility evaluation method for mixed data.

[0006] The technical solution adopted by the present invention is as follows:

[0007] A landslide susceptibility assessment method for mixed data includes the following steps:

[0008] S1: Divide the slope units and select the landslide influencing factors;

[0009] S2: Calculate the correlation matrix corresponding to the numerical impact factor and the categorical impact factor respectively;

[0010] S3: merge the numerical correlation matrix and the categorical correlation matrix;

[0011] S4: The resulting correlation matrix is ​​used as the feature data of the input classification model to predict landslide susceptibility.

[0012] Preferably, S1 is specifically:

[0013] Divide the slope into units, collect the spatial distribution data of landslide evaluation indicators for preprocessing, and select landslide influencing factors.

[0014] Preferably, the landslide influencing factors are divided into numerical type and categorical type according to data type; wherein the numerical influencing factors include: terrain moisture index, curvature, slope aspect, elevation, rainfall, slope, distance from fault, distance from road, distance from river; categorical influencing factors include: landform type, earthquake intensity, lithology, land use type, seismic peak acceleration, soil type.

[0015] Preferably, S2 is specifically:

[0016] For numerical data, calculate the Spearman correlation coefficient of all slope units of the same type of data compared with each other

[0017]

[0018] Among them, d i is the grade difference between the same type of slope unit data, and n is the number of slope units; the negative square Euclidean distance between the Spearman coefficients of different types of numerical data is calculated to obtain the correlation matrix;

[0019] For categorical data, the one-hot encoding technique is used to encode each categorical variable into a binary matrix representing category membership; then feature selection is performed on the categorical variables. After feature selection of the categorical features, the above one-hot encoding is used to calculate the average logarithmic loss of all objects compared with each other; the calculation formula for the average logarithmic loss is as follows:

[0020]

[0021] Among them, L is the loss function, N is the input sample size, y i,j When the i-th sample belongs to category j, it is 1, otherwise it is 0; p i,j is the probability that the i-th sample is predicted to be the j-th class;

[0022] The association matrix is ​​obtained by calculating the negative squared Euclidean distance between the average logarithmic losses of the classified data of different classes.

[0023] Preferably, S3 is specifically:

[0024] After calculating the numerical correlation matrix and the categorical correlation matrix, the two correlation matrices are combined with the same weights. After normalization, the values ​​at the corresponding positions of the numerical matrix and the categorical matrix are added and the average is taken to obtain the final correlation matrix.

[0025] Preferably, the normalization process is specifically as follows:

[0026] The samples of the obtained numerical correlation matrix and categorical correlation matrix are linearly normalized to between -100 and 0, where the samples with the largest difference are mapped to -100, and the most similar samples are mapped to 0; self-similarity is not considered in the normalization, and only the two most similar samples are normalized to 0.

[0027] Preferably, the classification models in S4 are:

[0028] Logistic regression model, random forest model, and support vector machine model.

[0029] The present invention has the following characteristics and advantages:

[0030] After the mixed data were classified, the resulting association matrix stored all the relevant information between the objects, and the commonly used classification algorithm was used to evaluate the landslide susceptibility. The receiver operating characteristic curve (ROC) was used to calculate the accuracy of the evaluation results. After comparing the results without considering the mixed data, it was found that the landslide susceptibility evaluation method for mixed data improved the accuracy of the susceptibility evaluation. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a flow chart of an embodiment of a mixed data landslide susceptibility evaluation method of the present invention;

[0032] Figure 2 It is a schematic diagram of the location and elevation of a study area of ​​an embodiment of a landslide susceptibility assessment method for mixed data of the present invention;

[0033] Figure 3 It is a flow chart of constructing a correlation matrix of an embodiment of a landslide susceptibility evaluation method for mixed data of the present invention;

[0034] FIG4 is a landslide susceptibility distribution diagram obtained by an embodiment of a landslide susceptibility evaluation method for mixed data of the present invention and other commonly used methods, wherein: Figure 4-a It is a landslide susceptibility map generated by logistic regression, random forest, and support vector machine without considering the influence of mixed data. Figure 4-b It is a landslide susceptibility map generated by logistic regression, random forest and support vector machine considering the influence of mixed data;

[0035] FIG5 is a receiver operating characteristic (ROC) curve obtained by an embodiment of a landslide susceptibility evaluation method for mixed data according to the present invention and other commonly used methods, wherein: Figure 5-a This is the ROC curve diagram without considering mixed data. Figure 5-b It is a ROC curve that considers the impact of mixed data. DETAILED DESCRIPTION

[0036] The present invention takes Lushan County, Ya'an City, Sichuan Province as the research area, combines commonly used machine learning models, namely, LR model, RF model, SVM model and 15 influencing factors, including 6 categorized landslide influencing factors (including earthquake intensity, lithology, geomorphic type, soil type, surface cover, earthquake peak acceleration (PGA)) and 9 numerical landslide influencing factors (including distance from fault, slope aspect, distance from road, distance from river, rainfall, elevation, slope, terrain moisture index, curvature). Then, the association matrix corresponding to the categorized data and the association matrix corresponding to the numerical data are calculated respectively, and then the categorized association matrix and the numerical association matrix are merged into a final association matrix. Since the final association matrix saves all relevant information between objects, the commonly used classification algorithm is used to evaluate the landslide sensitivity. Finally, the accuracy of the evaluation result is calculated using the ROC curve. After comparison with the result without considering the mixed data, it is found that the landslide susceptibility evaluation method for mixed data improves the accuracy of the susceptibility evaluation.

[0037] like Figure 1 As shown in the figure, an embodiment of a landslide susceptibility evaluation method for mixed data of the present invention is shown in the figure, taking Lushan County, Ya'an City, Sichuan Province as the research area. Figure 2 The superiority of the method of the present invention was verified by comparing it with classical machine learning methods (logistic regression, support vector machine, random forest).

[0038] The steps are:

[0039] 1: Divide the slope units and extract evaluation indicators.

[0040] In this paper, slope units are selected as the earthquake landslide susceptibility evaluation units. The slope units are generated based on the slope unit automatic division method of r.slopeunits software. The steps are as follows: DEM is imported into GRASS software, and then the minimum watershed unit area parameter is set to 0.8km through continuous debugging. 2 , and finally the study area was divided into 2956 slope units.

[0041] Based on the three principles of selecting influencing factors: ① Systematic principle: the influencing factors should reflect their contribution to the landslide as comprehensively as possible ② Scientific principle: select influencing factors that directly induce landslides as much as possible. ③ Operability principle: the influencing factors should be accessible and convenient for experimental operation. This paper selects 15 landslide influencing factors to construct a landslide susceptibility evaluation system. These influencing factors are divided into numerical type and categorical type according to data type. Among them, the numerical influencing factors include: terrain moisture index, curvature, slope aspect, elevation, rainfall, slope, distance from faults, distance from roads, and distance from rivers. The categorical influencing factors include: landform type, seismic fracture intensity, lithology, land use type, seismic peak acceleration, and soil type. The classic machine learning models random forest (RF), logistic regression (LR), and support vector machine (SVM) are used in parallel to participate in the landslide susceptibility evaluation and compared with the landslide susceptibility evaluation method for mixed data proposed in this invention.

[0042] 2: Correlation matrix calculation

[0043] First, the intrinsic properties of the evaluation indicators are divided into two data types: numerical type (topographic moisture index, curvature, slope aspect, elevation, rainfall, slope, distance to faults, distance to roads, distance to rivers) and categorical type (geomorphic type, seismic intensity, lithology, land use type, peak seismic acceleration, soil type).

[0044] For numerical data, calculate the Spearman correlation coefficient of all slope units of the same type of data compared with each other

[0045]

[0046] Among them, d i is the grade difference between the same type of slope unit data, and n is the number of slope units;

[0047] Calculate the negative square Euclidean distance between the Spearman coefficients of different classes of numerical data to obtain the correlation matrix;

[0048] For categorical data, the one-hot encoding technique is used to encode each categorical variable into a binary matrix representing category membership; then feature selection is performed on the categorical variables. After feature selection of the categorical features, the above one-hot encoding is used to calculate the average logarithmic loss of all objects compared with each other; the calculation formula for the average logarithmic loss is as follows:

[0049]

[0050] Among them, L is the loss function, N is the input sample size, y i,j When the i-th sample belongs to category j, it is 1, otherwise it is 0; p i,jis the probability that the i-th sample is predicted to be the j-th class;

[0051] The association matrix is ​​obtained by calculating the negative squared Euclidean distance between the average logarithmic losses of the classified data of different classes.

[0052] 3: Merge categorical and numerical incidence matrices

[0053] After calculating the numerical correlation matrix and the categorical correlation matrix, in order to combine the two correlation matrices with the same weight, this paper linearly normalizes the two correlation matrices to between -100 and 0, where the samples with the largest difference are mapped to -100, and the most similar samples are mapped to 0. Self-similarity is not considered in normalization, so only the two most similar samples are normalized to 0. After normalization, the values ​​at the corresponding positions of the categorical and numerical matrices are added and averaged to obtain the final correlation matrix.

[0054] 4: Prediction of landslide susceptibility.

[0055] Since the final association matrix stores the relevant information between all objects, the entries in the association matrix can be used as classification features of the classifier, and the final association matrix is ​​classified by the classification algorithm to obtain different landslide susceptibility levels. The generated landslide susceptibility zoning map is shown in Figure 4, and the superiority of the method of the present invention is verified by the receiver operating characteristic curve (ROC), as shown in Figure 5.

[0056] After the mixed data were classified, the resulting association matrix stored all the relevant information between the objects, and the commonly used classification algorithm was used to evaluate the landslide susceptibility. The receiver operating characteristic curve (ROC) was used to calculate the accuracy of the evaluation results. After comparing the results without considering the mixed data, it was found that the landslide susceptibility evaluation method for mixed data improved the accuracy of the susceptibility evaluation.

[0057] The embodiments described above are only descriptions of the preferred modes of the present invention, and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A landslide susceptibility assessment method for mixed data, characterized in that: The following steps are involved: S1: Divide the slope units and select the landslide influencing factors; S2: Calculate the correlation matrix corresponding to the numerical impact factor and the categorical impact factor respectively; S3: merge the numerical correlation matrix and the categorical correlation matrix; S4: The resulting correlation matrix is ​​used as the characteristic data of the input classification model to predict landslide susceptibility; S2 is specifically: For numerical data, calculate the Spearman correlation coefficient of all slope units of the same type of data compared with each other: Among them, d i is the grade difference between the same type of slope unit data, and n is the number of slope units; Calculate the negative square Euclidean distance between the Spearman coefficients of different classes of numerical data to obtain the correlation matrix; For categorical data, the one-hot encoding technique is used to encode each categorical variable into a binary matrix representing category membership; then feature selection is performed on the categorical variables. After feature selection of the categorical features, the above one-hot encoding is used to calculate the average logarithmic loss of all objects compared with each other; the calculation formula for the average logarithmic loss is as follows: Among them, L is the loss function, N is the input sample size, y i，j When the i-th sample belongs to category j, it is 1, otherwise it is 0; p i，j is the probability that the i-th sample is predicted to be the j-th class; The association matrix is ​​obtained by calculating the negative squared Euclidean distance between the average logarithmic losses of the classified data of different classes.

2. The method according to claim 1, characterized in that S1 is specifically: Divide the slope into units, collect the spatial distribution data of landslide evaluation indicators for preprocessing, and select landslide influencing factors.

3. The method according to claim 2, characterized in that The landslide influencing factors are divided into numerical type and categorical type according to data type; among them, the numerical influencing factors include: terrain moisture index, curvature, slope aspect, elevation, rainfall, slope, distance from fault, distance from road, distance from river; categorical influencing factors include: landform type, seismic intensity, lithology, land use type, seismic peak acceleration, and soil type.

4. The method according to claim 1, characterized in that: S3 is specifically: After calculating the numerical correlation matrix and the categorical correlation matrix, the two correlation matrices are combined with the same weights. After normalization, the values ​​at the corresponding positions of the numerical matrix and the categorical matrix are added and the average is taken to obtain the final correlation matrix.

5. The method according to claim 4, characterized in that The normalization process is specifically as follows: The samples of the obtained numerical correlation matrix and categorical correlation matrix are linearly normalized to between -100 and 0, where the samples with the largest difference are mapped to -100, and the most similar samples are mapped to 0; self-similarity is not considered in the normalization, and only the two most similar samples are normalized to 0.

6. The method according to claim 1, characterized in that The classification models in S4 are: Logistic regression model, random forest model, and support vector machine model.

Images