Shallow landslide susceptibility prediction method and system based on single landslide trigger event

Abstract

The invention belongs to the technical field of disaster prevention and reduction, discloses a shallow landslide susceptibility prediction method and system based on a single landslide trigger event, provides a new logic for predicting susceptibility, and aims to solve the problems that the research on the current shallow landslide susceptibility prediction is not rigorous, inaccurate and non-objective. According to the method, a probability risk model of a primary rainfall / earthquake triggered landslide is firstly constructed, and then influence factors of an event are eliminated, so that a real susceptibility prediction result of the shallow landslide in the region is obtained. According to the method, the reverse thinking method of reversely deducing the landslide susceptibility model from the landslide probability risk model is adopted, and the prediction precision of the regional shallow landslide susceptibility is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of disaster prevention and reduction, and in particular to a method and system for predicting the susceptibility of shallow landslides based on a single landslide triggering event. Background Art

[0002] In the existing technologies for landslide susceptibility assessment, the commonly used method is to use historical landslide data as samples and combine a variety of machine learning techniques to train prediction models. For example, multi-layer perceptron (MLP), random forest (RF), support vector machine (SVM) and deep learning models are widely used. These methods analyze the relationship between historical landslide data and environmental factors to build models to predict potential landslide areas. However, these ideas have limitations, mainly due to the non-uniform time span of historical landslide events, resulting in large differences in the level of data detail in different regions. In addition, the selection of positive and negative samples has a significant impact on the accuracy of the model, and there is an overfitting problem. These problems limit the universality and accuracy of the model in different regions and conditions.

[0003] Traditional landslide susceptibility prediction methods are difficult to make accurate judgments on the susceptibility of shallow landslides in a region. First, the landslide samples used in traditional methods are historical landslides, or landslide relics visually interpreted using high-resolution satellite images. The time type of these landslides is not clear, making it difficult to accurately predict the susceptibility of shallow landslides. Second, most of the existing methods stop at collecting landslide sample data, and then try to train various machine learning models, and finally compare the model performance, but do not consider the logical rigor of landslide susceptibility prediction. There are a large number of confusing studies on susceptibility and hazard prediction. In addition, for the prediction of shallow landslide susceptibility, the current research field has not yet found a solution to accurately predict the susceptibility of shallow landslides. Summary of the invention

[0004] In view of the problems that the current research on shallow landslide susceptibility prediction is not rigorous, accurate and objective, the present invention proposes a shallow landslide susceptibility prediction method and system based on a single landslide triggering event. The method uses the landslide data triggered by a single rainfall / earthquake event and combines multiple environmental factors to first construct a landslide probability hazard model for the area, thereby eliminating the influence of rainfall or earthquake events, so that the model can more accurately predict the shallow landslide susceptibility of the area.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions: The present invention provides a method for predicting the susceptibility of shallow landslides based on a single landslide triggering event, comprising: Step 1: Based on the high-resolution, low-cloud coverage optical remote sensing images of the area affected by a heavy rainfall / earthquake event, obtain the recent rainfall-induced landslide disaster data / earthquake-triggered landslide disaster data of the target area and complete the construction of the regional landslide database; Step 2: Select multiple landslide influencing factors based on the development characteristics of regional landslide geological hazards, and then normalize each landslide influencing factor to establish a landslide influencing factor dataset; Step 3: selecting landslide sample points and non-landslide sample points in the area where the landslide is located based on the regional landslide database; Step 4: Based on the landslide influencing factor dataset and the selected landslide sample points and non-landslide sample points, the multicollinearity degree and contribution of each landslide influencing factor are analyzed to establish the landslide training sample set and test sample set; Step 5: Based on the established landslide training sample set, train the shallow landslide probability hazard model constructed based on the machine learning model; Step 6: Evaluate the performance of the constructed landslide probability hazard model through the landslide test sample set to obtain the optimal landslide probability hazard model; Step 7: Input all the landslide influencing factor data of a single landslide triggering event into the optimal landslide probability hazard model to obtain the landslide probability hazard results in the regional heavy rainfall / earthquake event; Step 8: Set the factor variables of landslide triggering events in the landslide influencing factors to a uniform value, input them into the optimal landslide probability hazard model again for calculation, and obtain the susceptibility results of shallow landslides in the target area.

[0006] Furthermore, the landslide influencing factors include: historical earthquake parameters, elevation, slope, slope aspect, profile curvature, slope position, terrain undulation, terrain moisture index, land use type, vegetation coverage, stratum lithology, distance from active faults, distance from rivers and average annual rainfall in the past five years.

[0007] Furthermore, the step 3 comprises: Points were uniformly and randomly selected throughout the rainfall / earthquake affected area, and the points falling within the landslide were used as landslide sample points, while the points falling outside the landslide were used as non-landslide sample points.

[0008] Furthermore, the step 3 comprises: The landslides are extracted to points and all are used as landslide sample points. The selection of non-landslide sample points depends on the ratio of landslide area to non-landslide area. The ratio of the number of non-landslide sample points to the number of landslide sample points is consistent with the ratio of non-landslide area to landslide area.

[0009] Furthermore, the step 4 comprises: The variance inflation factor analysis method was used to analyze the multicollinearity of various landslide influencing factors, and the landslide influencing factors with variance inflation factors not greater than 10 were selected to establish the landslide training sample set and test sample set.

[0010] Furthermore, the step 4 comprises: The contribution of landslide influencing factors to the probability of landslide occurrence is analyzed through the multivariate linear regression model, and the importance of each landslide influencing factor is judged using the standardized regression coefficient. The larger the absolute value of the coefficient, the greater the contribution of the factor to the landslide.

[0011] Furthermore, in step 5, the machine learning model includes: gradient boosting decision tree, extreme gradient boosting tree, random forest and logistic regression algorithm.

[0012] Furthermore, the step 8 comprises: According to the natural breakpoint method, the shallow landslide susceptibility results are divided into five levels, corresponding to extremely low susceptibility area, low susceptibility area, medium susceptibility area, high susceptibility area and extremely high susceptibility area.

[0013] Another aspect of the present invention provides a shallow landslide susceptibility prediction system based on a single landslide triggering event, comprising: The first data construction module is used to obtain the recent rainfall-induced landslide disaster data / earthquake-triggered landslide disaster data of the target area based on the high-resolution low-cloud coverage optical remote sensing image of the area affected by a heavy rainfall / earthquake event, and complete the construction of the regional landslide database; The second data construction module is used to select multiple landslide influencing factors according to the development characteristics of regional landslide geological hazards, and then establish a landslide influencing factor data set after normalizing each landslide influencing factor; A sample point selection module, used for selecting landslide sample points and non-landslide sample points in the area where the landslide is located based on the regional landslide database; The data analysis module is used to analyze the multicollinearity and contribution of each landslide influencing factor based on the landslide influencing factor data set and the selected landslide sample points and non-landslide sample points, and to establish a landslide training sample set and a test sample set; A model training module is used to train a shallow landslide probability hazard model constructed based on a machine learning model based on an established landslide training sample set; The model evaluation module is used to evaluate the performance of the constructed landslide probability hazard model through the landslide test sample set to obtain the optimal landslide probability hazard model; The first model prediction module is used to input all landslide influencing factor data of a single landslide triggering event into the optimal landslide probability hazard model to obtain the landslide probability hazard results in regional heavy rainfall / earthquake events; The second model prediction module is used to set the factor variables of the landslide triggering event in the landslide influencing factors to a uniform value, and input the optimal landslide probability hazard model again for calculation to obtain the susceptibility result of shallow landslides in the target area.

[0014] Compared with the prior art, the present invention has the following beneficial effects: The present invention adopts a reverse thinking method of inferring a shallow landslide susceptibility model from a shallow landslide probability hazard model, solving the problem of non-objective and inaccurate regional shallow landslide prediction. Compared with the traditional method, the method proposed by the present invention makes the research on the landslide susceptibility model more accurate, and effectively avoids the confusion between the previous landslide susceptibility model and the landslide probability hazard model. In addition, the present invention can provide a clear technical process for landslide probability hazard prediction for different rainfall scenarios or earthquake scenarios under the condition of constructing a landslide probability hazard model. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 A schematic flow chart of a method for predicting the susceptibility of shallow landslides based on a single landslide triggering event according to an embodiment of the present invention; Figure 2 The schematic diagram is a schematic diagram of the architecture of a shallow landslide susceptibility prediction system based on a single landslide triggering event according to an embodiment of the present invention. DETAILED DESCRIPTION

[0016] The present invention will be further explained below with reference to the accompanying drawings and specific embodiments: like Figure 1 As shown, a shallow landslide susceptibility prediction method based on a single landslide triggering event includes: Step S101: Based on the high-resolution, low-cloud coverage optical remote sensing image of the area affected by a heavy rainfall / earthquake event, obtain the recent rainfall-induced landslide disaster data / earthquake-triggered landslide disaster data of the target area, and construct a regional landslide database; Step 102: selecting a plurality of landslide influencing factors according to the development characteristics of regional landslide geological hazards, and then normalizing each landslide influencing factor to establish a landslide influencing factor data set; Step 103: Select points in the area where the landslide is located as sample points. It should be noted that the selection of landslide and non-landslide sample points has strict extraction strategy requirements. There are two specific schemes. One is to randomly select points uniformly in the entire rainfall-affected area, and the points that eventually fall within the landslide are used as landslide sample points, while the points that fall outside the landslide are used as non-landslide sample points. The second is to extract the landslide to points, and all of these points are used as landslide sample points. The selection of non-landslide sample points depends on the ratio of the landslide area to the non-landslide area. The ratio of the number of non-landslide sample points to the number of landslide sample points needs to be consistent with the ratio of the non-landslide area to the landslide area. Step 104: Based on the landslide influencing factor data set and the selected landslide and non-landslide sample points, analyze the multicollinearity and contribution degree of each landslide influencing factor. As an implementable method, select landslide influencing factors with a variance inflation factor (VIF) not greater than 10 to establish a landslide training test sample set; Step 105: Based on the established landslide training sample set and test sample set, training a landslide probability hazard model constructed based on the machine learning model; Step 106: Evaluate the performance of the constructed model using evaluation indicators such as precision, accuracy, F1 score, and ROC curve, and then construct the optimal model through methods such as model optimization; Step 107: At this time, all the influencing factor data are input into the optimal model to obtain the landslide probability hazard results under regional heavy rainfall / earthquake conditions; Step 108: Set the factor variables of the landslide triggering event in the influencing factors to a uniform value, and input the optimal model calculation result again. The result obtained at this time is the susceptibility result of shallow landslides in the area. As an implementable method, the shallow landslide susceptibility result can be divided into 5 levels as needed.

[0017] The present invention proposes a new logic for predicting susceptibility. By first constructing a probability hazard model of landslides triggered by rainfall / earthquakes, and then eliminating the influencing factors of the event, the real susceptibility prediction result of shallow landslides in the region is obtained. The present invention adopts a reverse thinking method of inferring the landslide susceptibility model from the landslide probability hazard model, which improves the prediction accuracy of regional shallow landslide susceptibility.

[0018] Furthermore, in step S101, high-resolution, low-cloud coverage optical remote sensing images are generally selected to include but are not limited to image data with a resolution greater than 6 meters, a cloud coverage rate including but not limited to less than 10%, and images without cloud or fog obstruction.

[0019] Furthermore, the landslide database described in step S101 should include and only include landslide data triggered by rainfall / earthquake, and the data should include but not be limited to spatial and attribute data, such as landslide location, range boundary, area, quantity information, etc.

[0020] Furthermore, the landslide database format described in step S101 includes but is not limited to ".shp", ".kml", and ".kmz".

[0021] Furthermore, the multiple landslide influencing factors described in step S102 include but are not limited to: historical earthquake parameters, elevation, slope, slope aspect, profile curvature, slope position, terrain undulation, terrain moisture index, land use type, vegetation coverage, stratum lithology, distance from active faults, distance from rivers and average annual rainfall in the past five years.

[0022] Furthermore, in step S102, the spatial resolution of all influencing factor data in the earthquake-landslide influencing factor data set includes but is not limited to 12.5 m, 30 m and 100 m.

[0023] Furthermore, the landslide sample selection method described in step S103 includes but is not limited to: 1) converting the landslide surface elements into raster data, and extracting each pixel of the converted raster data as a landslide sample point; 2) converting the landslide surface elements into point elements, and one landslide corresponds to one point as a landslide sample.

[0024] Furthermore, the non-landslide sample selection method described in step S103 includes but is not limited to: 1) extracting the non-landslide area in the study area and randomly selecting points within the area; 2) constructing a fishing net point matrix for the non-landslide area and randomly selecting points from it.

[0025] Further, in step S103, the sample size of the landslide samples triggered by rainfall is selected according to the ratio of landslide to non-landslide areas, including but not limited to: 1) a large number of random points are taken in the entire study area, and then the area where the points are located is statistically analyzed to see whether it belongs to the landslide area, so as to mark the landslide and non-landslide sample labels for each point respectively; 2) an appropriate amount of point data is first randomly selected in the landslide area, and then corresponding sample points are selected in the non-landslide area corresponding to the ratio of landslide to non-landslide areas. As a feasible implementation method, any landslide sampling method that can achieve the area ratio can be adopted.

[0026] Furthermore, in step S104, the multicollinearity analysis method of landslide influencing factors includes but is not limited to: 1) Variance inflation factor (VIF) analysis: By calculating the variance inflation factor (VIF) value of each influencing factor, the degree of collinearity between variables is determined. When the VIF value is higher than a specific threshold (such as 5, 10 or 70), it indicates that the variables are collinear. 2) Correlation coefficient matrix analysis: Calculate the correlation coefficients between landslide influencing factors and generate a correlation coefficient matrix. A high correlation coefficient (close to 1 or -1) indicates that there may be collinearity between variables; 3) Condition number analysis: By calculating the condition number of the characteristic matrix of landslide influencing factors, the collinearity problem is determined. When the condition number is greater than 30, it indicates that there is serious collinearity.

[0027] Furthermore, in step S104, the contribution of landslide influencing factors includes but is not limited to: 1) Multiple regression analysis: Analyze the contribution of landslide influencing factors to the probability of landslide occurrence through a multiple linear regression model, and use standardized regression coefficients to judge the importance of each factor. The larger the absolute value of the coefficient, the greater the contribution of the factor to the landslide. 2) Random forest feature importance: Calculate the feature importance of each influencing factor through a random forest algorithm. The more important the factor, the greater its contribution to the prediction of the landslide. 3) Information gain: Use the information gain method to evaluate the contribution of each influencing factor to the landslide classification information. The larger the information gain, the higher the importance of the factor to the occurrence of the landslide.

[0028] Furthermore, in step S104, the landslide training set and the test set are divided into (random and sequential methods), and the division ratios include 9:1, 8:2, 7:3, and 6:4.

[0029] Furthermore, in step S105, in the earthquake landslide training test sample set, the label marking method includes: the landslide label is marked as "1", the non-landslide label is marked as "2", or it may be: "1" and "0" or other two different characters.

[0030] Furthermore, the machine learning model described in step S105 includes but is not limited to: 1) Gradient boosted decision tree (GBDT): a classic gradient boosted decision tree algorithm that optimizes the residual of the model by gradually building multiple decision trees, and each tree learns the error of the previous prediction. GBDT performs well in regression and classification tasks. 2) Extreme gradient boosted tree (XGBoost): an improved gradient boosted tree algorithm with higher efficiency and scalability. XGBoost prevents overfitting through parallel processing and regularization, and can customize the objective function. 3) CatBoost: a gradient boosting algorithm optimized for category features, which processes category features through a special target encoding method, can reduce the bias and variance of the model, and prevent overfitting; and a combination algorithm of such algorithms. 4) Random forest: compared with other algorithms, it has the advantages of high accuracy, high generalization ability, and strong noise resistance. 5) Logistic regression algorithm: compared with other algorithms, it has the advantages of high computational efficiency, applicability to linearly separable data, and easy expansion.

[0031] Furthermore, the model evaluation method described in step S106 includes but is not limited to accuracy, precision, F1 score and ROC curve.

[0032] Furthermore, in step S107, the value range of the landslide probability hazard result includes 0-1 or other ranges greater than 0 and less than 1; the specific value is determined according to the calculation result.

[0033] Furthermore, in step S107, all landslide impact factor data contain real parameters of a single landslide triggering event, and at this time, no modification shall be made to the input data.

[0034] Furthermore, in step S108, the factor variable of the landslide triggering event in the landslide influencing factor is set to a uniform value, which can be any value as long as it is the same, and the values ​​of other variables are kept unchanged.

[0035] Furthermore, in step S108, the shallow landslide susceptibility result is no longer the landslide susceptibility in the traditional sense, but only represents the shallow landslide susceptibility in the event impact area.

[0036] Furthermore, in step S108, the influencing factor variables input into the model again may exceed the study area, but the event-related factors still need to be kept at the same set value.

[0037] Furthermore, in step S108, the factor variables of the landslide triggering events in the landslide influencing factors are set to a uniform value, in order to eliminate the influence of the triggering events, and other methods that can achieve the same purpose are also acceptable.

[0038] Furthermore, in step S108, the zoning method for the susceptibility of shallow landslides includes but is not limited to: the natural breakpoint method, the susceptibility level is divided into 5 levels: extremely low susceptibility area, low susceptibility area, medium susceptibility area, high susceptibility area and extremely high susceptibility area; the equal interval classification method, the 0-1 range is divided into 5 susceptibility levels at intervals of 0.2.

[0039] Based on the above embodiments, Figure 2 As shown, the present invention also proposes a shallow landslide susceptibility prediction system based on a single landslide triggering event, comprising: The first data construction module is used to obtain the recent rainfall-induced landslide disaster data / earthquake-triggered landslide disaster data of the target area based on the high-resolution low-cloud coverage optical remote sensing image of the area affected by a heavy rainfall / earthquake event, and complete the construction of the regional landslide database; The second data construction module is used to select multiple landslide influencing factors according to the development characteristics of regional landslide geological hazards, and then establish a landslide influencing factor data set after normalizing each landslide influencing factor; A sample point selection module, used for selecting landslide sample points and non-landslide sample points in the area where the landslide is located based on the regional landslide database; The data analysis module is used to analyze the multicollinearity and contribution of each landslide influencing factor based on the landslide influencing factor data set and the selected landslide sample points and non-landslide sample points, and to establish a landslide training sample set and a test sample set; A model training module is used to train a shallow landslide probability hazard model constructed based on a machine learning model based on an established landslide training sample set; The model evaluation module is used to evaluate the performance of the constructed landslide probability hazard model through the landslide test sample set to obtain the optimal landslide probability hazard model; The first model prediction module is used to input all landslide influencing factor data of a single landslide triggering event into the optimal landslide probability hazard model to obtain the landslide probability hazard results in regional heavy rainfall / earthquake events; The second model prediction module is used to set the factor variables of the landslide triggering event in the landslide influencing factors to a uniform value, and input the optimal landslide probability hazard model again for calculation to obtain the susceptibility result of shallow landslides in the target area.

[0040] The above is only a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several improvements and modifications without departing from the principle of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A shallow landslide susceptibility prediction method based on a single landslide triggering event, characterized in that: include: Step 1: Based on the high-resolution, low-cloud coverage optical remote sensing images of the area affected by a heavy rainfall / earthquake event, obtain the recent rainfall-induced landslide disaster data / earthquake-triggered landslide disaster data of the target area and complete the construction of the regional landslide database; Step 2: Select multiple landslide influencing factors based on the development characteristics of regional landslide geological hazards, and then normalize each landslide influencing factor to establish a landslide influencing factor dataset; Step 3: selecting landslide sample points and non-landslide sample points in the area where the landslide is located based on the regional landslide database; Step 4: Based on the landslide influencing factor dataset and the selected landslide sample points and non-landslide sample points, the multicollinearity degree and contribution of each landslide influencing factor are analyzed to establish the landslide training sample set and test sample set; Step 5: Based on the established landslide training sample set, train the shallow landslide probability hazard model constructed based on the machine learning model; Step 6: Evaluate the performance of the constructed landslide probability hazard model through the landslide test sample set to obtain the optimal landslide probability hazard model; Step 7: Input all the landslide influencing factor data of a single landslide triggering event into the optimal landslide probability hazard model to obtain the landslide probability hazard results in the regional heavy rainfall / earthquake event; Step 8: Set the factor variables of landslide triggering events in the landslide influencing factors to a uniform value, input them into the optimal landslide probability hazard model again for calculation, and obtain the susceptibility results of shallow landslides in the target area.

2. A shallow landslide susceptibility prediction method based on a single landslide triggering event according to claim 1, characterized in that: The landslide influencing factors include: historical earthquake parameters, elevation, slope, slope direction, profile curvature, slope position, terrain undulation, terrain moisture index, land use type, vegetation coverage, stratum lithology, distance from active faults, distance from rivers and average annual rainfall in the past five years.

3. The method for predicting shallow landslide susceptibility based on a single landslide triggering event according to claim 1, characterized in that: The step 3 comprises: Points were uniformly and randomly selected throughout the rainfall / earthquake affected area, and the points falling within the landslide were used as landslide sample points, while the points falling outside the landslide were used as non-landslide sample points.

4. The method for predicting shallow landslide susceptibility based on a single landslide triggering event according to claim 1, characterized in that: The step 3 comprises: The landslides are extracted to points and all are used as landslide sample points. The selection of non-landslide sample points depends on the ratio of landslide area to non-landslide area. The ratio of the number of non-landslide sample points to the number of landslide sample points is consistent with the ratio of non-landslide area to landslide area.

5. The method for predicting shallow landslide susceptibility based on a single landslide triggering event according to claim 1, characterized in that: The step 4 comprises: The variance inflation factor analysis method was used to analyze the multicollinearity of various landslide influencing factors, and the landslide influencing factors with variance inflation factors not greater than 10 were selected to establish the landslide training sample set and test sample set.

6. The method for predicting shallow landslide susceptibility based on a single landslide triggering event according to claim 1, characterized in that: The step 4 comprises: The contribution of landslide influencing factors to the probability of landslide occurrence is analyzed through the multivariate linear regression model, and the importance of each landslide influencing factor is judged using the standardized regression coefficient. The larger the absolute value of the coefficient, the greater the contribution of the factor to the landslide.

7. The method for predicting shallow landslide susceptibility based on a single landslide triggering event according to claim 1, characterized in that: In step 5, the machine learning models include: gradient boosting decision tree, extreme gradient boosting tree, random forest and logistic regression algorithm.

8. The method for predicting shallow landslide susceptibility based on a single landslide triggering event according to claim 1, characterized in that: The step 8 comprises: According to the natural breakpoint method, the shallow landslide susceptibility results are divided into five levels, corresponding to extremely low susceptibility area, low susceptibility area, medium susceptibility area, high susceptibility area and extremely high susceptibility area.

9. A shallow landslide susceptibility prediction system based on a single landslide triggering event, characterized in that: include: The first data construction module is used to obtain the recent rainfall-induced landslide disaster data / earthquake-triggered landslide disaster data of the target area based on the high-resolution low-cloud coverage optical remote sensing image of the area affected by a heavy rainfall / earthquake event, and complete the construction of the regional landslide database; The second data construction module is used to select multiple landslide influencing factors according to the development characteristics of regional landslide geological hazards, and then establish a landslide influencing factor data set after normalizing each landslide influencing factor; A sample point selection module, used for selecting landslide sample points and non-landslide sample points in the area where the landslide is located based on the regional landslide database; The data analysis module is used to analyze the multicollinearity and contribution of each landslide influencing factor based on the landslide influencing factor data set and the selected landslide sample points and non-landslide sample points, and to establish a landslide training sample set and a test sample set; A model training module is used to train a shallow landslide probability hazard model constructed based on a machine learning model based on an established landslide training sample set; The model evaluation module is used to evaluate the performance of the constructed landslide probability hazard model through the landslide test sample set to obtain the optimal landslide probability hazard model; The first model prediction module is used to input all landslide influencing factor data of a single landslide triggering event into the optimal landslide probability hazard model to obtain the landslide probability hazard results in regional heavy rainfall / earthquake events; The second model prediction module is used to set the factor variables of the landslide triggering event in the landslide influencing factors to a uniform value, and input the optimal landslide probability hazard model again for calculation to obtain the susceptibility result of shallow landslides in the target area.

Images