Landslide hazard risk assessment method based on machine learning
By applying machine learning models in landslide disaster risk assessment and spatial prediction, the problem of inefficient landslide disaster prediction and assessment in the existing technology is solved, and more accurate and rapid risk assessment and prediction are achieved, which improves the science of disaster prevention and control work. Sex and effectiveness.
Patent Information
- Application Number
- CN202111268460.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-10-29
- Publication Date
- 2024-11-29
- Estimated Expiration
- 2041-10-29
AI Technical Summary
The existing technology is difficult to quickly and accurately conduct spatial prediction and risk assessment of landslide disasters, resulting in inefficient disaster warning and disaster prevention and mitigation.
Using machine learning-based methods, including mixed Gaussian clustering model, artificial neural network model, support vector machine model, PU-Bagging model and CA-Markov model, landslide disaster risk evaluation and spatial prediction, combined with vulnerability results risk assessment.
It has achieved rapid and accurate risk assessment and spatial prediction of landslide disasters, improved the scientificity and effectiveness of disaster warning and disaster prevention and mitigation work, and reduced the potential threats of disasters to humans and property.
Smart Images

Figure CN114330812B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of geological disasters, and in particular relates to a landslide disaster risk assessment method based on machine learning. Background Art
[0002] In recent years, the threat of landslide disasters to human living environment and residential safety has become increasingly serious, while the sharp increase in population and rapid economic development have led to excessive development and utilization of natural resources. Large-scale irrational development has destroyed the natural environment and induced a large number of geological disasters, which pose a great threat to human living environment and social economy. From the historical data of geological disaster research, it can be seen that disaster warning and disaster prevention and mitigation have always been the first task of geological disaster research. The country has also been spending a lot of material and financial resources on the research of geological disasters, hoping to reduce the occurrence of disasters and the losses caused by disasters to a certain extent. The fundamental purpose of carrying out spatial prediction and risk assessment of landslide disasters is to grasp the activities and development laws of disasters.
[0003] Through risk assessment research, we can quickly and accurately summarize the development characteristics of landslides, and then make a graded forecast of the probability of landslide disasters. The core of spatial prediction is to study the hazard of landslide disasters, and the core significance of disaster risk assessment is reflected in: predicting the areas where disasters may occur and the probability of occurrence through the laws of disaster development, using the prediction results as basic data for regional development and planning, formulating corresponding emergency plans and preventive measures for areas under disaster threat, and providing guarantee data for the scientific and effective implementation of disaster prevention and mitigation work. Summary of the invention
[0004] In order to achieve the above-mentioned purpose, the purpose of the present invention is to provide a landslide disaster risk assessment method based on machine learning, which has four main contents. First, the controlling factors and inducing factors are determined according to the collected data, and the evaluation unit is determined to analyze the distribution law and development characteristics of geological hazards; second, the accuracy test of landslide disaster hazard assessment is carried out based on the mixed Gaussian clustering model, the artificial neural network model and the support vector machine model, and finally the hazard assessment is carried out, and the first risk assessment is carried out in combination with the vulnerability results; third, the spatial prediction and dynamic prediction of landslide hazard are carried out based on the PU-Bagging model and the CA-Markov model; the second risk assessment of landslide disaster is carried out according to the landslide hazard prediction results and the vulnerability results; finally, the two risk assessment results are compared and analyzed to guide practice.
[0005] Its theoretical significance lies in that the regional landslide hazard assessment requires the study of the geological conditions that affect landslide geological hazards in the selected area, and then the quantitative classification of each factor, the relationship between different factors and the distribution of landslide hazard points, and the information value between each factor and the occurrence of landslide hazards is calculated, and the significantly correlated factors are further eliminated through factor correlation analysis, and finally a hazard assessment index system is established. Then, by analyzing the index system, multiple evaluation models are selected, and the results of different evaluation models are compared to find the optimal evaluation model and the optimal evaluation result. The comparison and analysis of the results of different models applied to landslide geological hazard evaluation have important reference value for the application of machine learning models to the geological hazard evaluation system;
[0006] Its practical significance lies in the risk assessment of landslide disasters. Its zoning results can realize focused scientific prevention and targeted management for disaster prevention and mitigation work. On the one hand, it can effectively reduce the potential threat of landslide geological disasters to the personal safety and property safety of the people. On the other hand, when conducting risk assessment of landslide geological disasters in a larger area, the present invention can also be used as a reference.
[0007] In view of this, the present invention relies on the detailed geological disasters and field survey data provided by a geological environment monitoring station in recent years, combined with the geological environment background and engineering geological conditions such as stratum lithology distribution, topographic and geomorphic units, tectonic fault development and slope structure types, through detailed data analysis and research, to conduct spatial analysis and prediction of landslide hazard, and on this basis, combine the land use type for value assessment to obtain the vulnerability distribution, thereby providing a landslide hazard risk assessment method based on machine learning.
[0008] In order to achieve the above object, the technical solution adopted by the present invention is:
[0009] The landslide disaster risk assessment method based on machine learning is characterized by comprising the following steps:
[0010] Step 1: Collect, organize and analyze relevant geological historical data, analyze the development and distribution patterns of landslide disasters in the region in combination with the distribution of regional landslide disaster points in field surveys, and preliminarily screen out environmental factors that have a greater impact on regional landslides, including the following two aspects:
[0011] Step 1.1, selection of evaluation factors:
[0012] Firstly, the evaluation factors are preliminarily selected based on the collected data, and then the selected factors are analyzed to ensure that the selected factors have significant spatial differences in different regions of the region, and to analyze whether the factors have a strong correlation with the occurrence of regional geological disasters. Then, combined with relevant literature and previous research results, the landslide hazard evaluation factors involved in the model construction in the evaluation process are adjusted;
[0013] Step 1.2: Division of evaluation units:
[0014] When evaluating geographic space, the first task is to determine the evaluation area and divide the evaluation units. The model unit represents a part of the surface, and each attribute in each unit is homogeneous. In GIS, grid units and slope units are selected to conduct relevant experiments respectively, and the results are compared and analyzed. These two evaluation unit division methods are selected;
[0015] Step 2: Analyze geological environmental conditions, organize relevant geological environmental background data, integrate historical landslide disaster data and the latest survey disaster data, analyze the development characteristics of landslide disasters in the region, establish a landslide evaluation index system for the region, and use statistical analysis methods to analyze the occurrence frequency, classification ratio and information value of landslide disaster points under different classification conditions of each factor based on the basic development of landslide geological disasters and the landslide disaster point data of field investigations. Further calculate the correlation between the evaluation factors and appropriately delete the evaluation factors with higher correlation.
[0016] Step 3, establish a data set optimized by a mixed Gaussian clustering model (GMM) and a non-optimized randomly extracted data set, train an artificial neural network (ANN) model and a support vector machine (SVM) model using a radial basis kernel function, and evaluate the landslide hazard risk in the area after the training results reach accuracy, and test and compare the prediction accuracy and rationality of the evaluation results of different models;
[0017] Step 4: Landslide Hazard Assessment Model Analysis
[0018] When conducting spatial prediction analysis of hazards, the analysis is mainly conducted from two aspects. The first is to conduct a hazard assessment. Based on the evaluation factors selected in steps 1 and 2, and according to the disaster data that has occurred, the current probability of landslide occurrence is spatially predicted. Secondly, based on step 3, combined with the hazard prediction results of multiple periods, by analyzing the hazard levels of two adjacent periods and analyzing their development laws, a spatial prediction of landslide hazard is conducted for a certain period in the future. The process of landslide hazard assessment mainly includes:
[0019] Step 4.1: Based on Step 1.1 and Step 2, pre-process the evaluation factor data and process the data scale to process the data into sample data that can be input into the model;
[0020] Step 4.2: Based on step 1.2, combined with the PU-Bagging algorithm, the evaluation units are divided into positive samples (areas where landslides have occurred) and unlabeled samples (areas where landslides have not occurred). Through random sampling and combination, the same amount of data as the positive samples is randomly extracted from the unlabeled samples with replacement as negative samples to form a training set;
[0021] Step 4.3: Use step 3 and the obtained training set to build a model to obtain the landslide occurrence probability of all unlabeled sample areas in the evaluation unit sample, thereby realizing the spatial prediction of the landslide occurrence probability in the study area.
[0022] Step 5: Dynamic model analysis of landslide hazard
[0023] Dynamic model analysis of landslide hazard is an important part of landslide spatial prediction, which is mainly reflected in the following aspects:
[0024] Step 5.1: Feasibility analysis of CA-Markov model in simulating the evolution of landslide hazard.
[0025] Step 5.2: Construction of dynamic prediction model of landslide hazard based on CA-Markov model and accuracy comparison experiment, select the parameter combination of the experimental scheme with the best prediction effect as the model parameter setting, and complete the model construction. Among them, the scheme with the highest consistency between the experimental results and the actual data is the best scheme.
[0026] Step 5.3: Use the most accurate experimental group parameters to predict the landslide hazard at a certain period in the future, thus achieving dynamic prediction of landslides.
[0027] Step 6: Landslide risk assessment:
[0028] Combined with the current land use status of the region, the types and distribution of disaster-prone bodies are obtained, and then 80% of the data are randomly selected and input into the weighted density model. The weighted density map of the carrier is used as the value estimate of the evaluation unit. Then, combined with the hazard index of the evaluation unit, the two are calculated to obtain the risk level of each evaluation unit. Finally, the remaining 20% of the data is used to verify the accuracy of the risk assessment.
[0029] Step 7: Compare and analyze the two landslide risk assessment methods.
[0030] The beneficial effects of the present invention are:
[0031] 1. The present invention takes the landslide geological disaster data obtained from field investigation as an example, and analyzes the spatiotemporal distribution characteristics and disaster development types of landslides in the region in combination with the regional geological environment. On this basis, the landslide disaster hazard factors suitable for the region are selected and an evaluation index system is established, and then the regional landslide disaster hazard evaluation under different evaluation models is compared. The final evaluation result can reduce the losses caused by regional landslide geological disasters and improve the efficiency of landslide geological disaster prevention and control work. It has important theoretical significance and practical value.
[0032] 2. Through risk assessment research, the development characteristics of landslides can be quickly and accurately summarized, and then the probability of landslide disasters can be predicted. The core of spatial prediction is to study the hazard of landslide disasters, and the core significance of disaster risk assessment is reflected in: predicting the areas where disasters may occur and the probability of occurrence through the laws of disaster development, using the prediction results as the basic data for regional development and planning, formulating corresponding emergency plans and preventive measures for areas under the threat of disasters, and providing guarantee data for the scientific and effective implementation of disaster prevention and mitigation work.
[0033] The present invention proposes a reasonable research process, creatively uses machine learning models to make spatial and dynamic predictions of landslide hazards, and conducts risk assessment. Compared with traditional methods, the present invention is more in line with the research characteristics of earth sciences, has clear logic, is indeed feasible, and has more reliable assessment results. BRIEF DESCRIPTION OF THE DRAWINGS
[0034] Figure 1 is a schematic diagram of a landslide hazard risk assessment method based on machine learning in this application;
[0035] Figure 2 It is a schematic diagram of the spatial prediction method of landslide hazard risk in this application.
[0036] Figure 3 It is a schematic diagram of the dynamic prediction method of landslide hazard risk in this application. DETAILED DESCRIPTION
[0037] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.
[0038] The landslide disaster risk assessment method based on machine learning includes the following steps:
[0039] Step 1: Collect, organize and analyze relevant geological historical data, analyze the development and distribution patterns of landslide disasters in the region in combination with the distribution of regional landslide disaster points in field surveys, and preliminarily screen out environmental factors that have a greater impact on regional landslides. This includes two aspects:
[0040] 1.1 Selection of evaluation factors
[0041] The occurrence of landslides is relatively complex, mainly including natural factors and human-induced factors. In different geographical environments, the factors that affect landslides and the importance of the factors are also different. Therefore, most of the research on landslides is an assessment of local areas. It is difficult to have a universal model to adapt to all areas or a large range. Therefore, for different areas, appropriate influencing factors should be selected to construct an evaluation model. The reasonable selection of evaluation factors can more effectively improve the accuracy of the evaluation model and reduce some unnecessary calculations and data collection work.
[0042] When selecting evaluation factors, we should first make a preliminary selection based on experience or collected data, and then analyze the selected factors to ensure that the selected factors have significant spatial differences in different regions of the area, and analyze whether the factors are strongly correlated with the occurrence of regional geological disasters. Then, combined with relevant literature and previous research results, we can adjust the landslide hazard evaluation factors involved in the model construction during the evaluation process.
[0043] 1.2 Evaluation Unit Division
[0044] When evaluating geographic space, the first task is to determine the evaluation area and divide the evaluation units. The model unit represents a part of the surface, and each attribute in each unit is homogeneous. In GIS, there are four most commonly used divisions of evaluation units: administrative division units, regular grid units, slope units, and geomorphic units. In combination with the geographical characteristics of the region, the present invention selects grid units and slope units to conduct relevant experiments respectively, and compares and analyzes the results. The two evaluation unit division methods are selected mainly because grid units are the most used. This type of grid division method is simple, easy to calculate, and convenient for spatial analysis. In addition, since the grid unit organizes data in the form of a matrix, it is convenient for computer processing, and it is also a widely used unit division method in landslide hazard; the slope unit is suitable for landslide hazard evaluation in small mountainous areas, and is not suitable for areas with large depressions and open valleys. It uses the water collection unit surrounded by each ridge line as the minimum evaluation unit, and is often used in watershed division and debris flow disaster evaluation.
[0045] Step 2: Based on the basic development of landslide geological hazards and the landslide disaster point data from field surveys, the statistical analysis method is used to analyze the occurrence frequency, classification ratio and information value of landslide disaster points under different classification conditions of each factor, further calculate the correlation between each evaluation factor, and appropriately delete the evaluation factors with higher correlation.
[0046] Step 3: Establish a data set optimized by a mixed Gaussian clustering model (GMM) and an unoptimized randomly selected data set, and train the artificial neural network (ANN) model and the support vector machine (SVM) model using the radial basis kernel function respectively. After the training results reach the accuracy, an evaluation is made on the landslide hazard risk in the area, and the prediction accuracy and rationality of the evaluation results of different models are tested and compared.
[0047] Step 4: Landslide Hazard Assessment Model Analysis
[0048] When conducting spatial prediction analysis of hazard, the analysis is mainly carried out from two aspects. The first is to conduct hazard assessment. Based on the selected evaluation factors and the disaster data that have occurred, the current probability of landslide occurrence is spatially predicted. Secondly, on the basis of hazard assessment, the hazard prediction results of multiple periods are combined, and the hazard levels of two adjacent periods are analyzed to analyze their development laws, so as to make spatial prediction of landslide hazard in a certain period in the future. The process of landslide hazard assessment mainly includes: ① Select evaluation factors, preprocess the evaluation factor data, and process the data scale to process the data into sample data that can be input into the model; ② Divide the study area into evaluation units, and then combine the PU-Bagging algorithm to divide the evaluation units into positive samples (areas where landslides have occurred) and unlabeled samples (areas where landslides have not occurred). Through random sampling and combination, the same amount of data as positive samples is randomly extracted from the unlabeled samples with replacement as negative samples to form a training set. ③ Use the obtained training set to build a model to obtain the probability of landslide occurrence in all unlabeled sample areas in the evaluation unit sample, so as to realize the spatial prediction of the probability of landslide occurrence in the study area.
[0049] Step 5: Dynamic model analysis of landslide hazard
[0050] The dynamic model analysis of landslide hazard is an important part of landslide spatial prediction, which is mainly reflected in the following aspects: ① The feasibility analysis of CA-Markov model in simulating the evolution of landslide hazard. ② The construction and accuracy comparison experiment of the dynamic prediction model of landslide hazard based on CA-Markov model, the parameter combination of the experimental scheme with the best prediction effect is selected as the model parameter setting to complete the model construction. Among them, the scheme with the highest consistency between the experimental results and the actual data is the best scheme. ③ The most accurate experimental group parameters are used to predict the landslide hazard in a certain period in the future, that is, the dynamic prediction of landslide is realized.
[0051] Step 6: Landslide Risk Assessment
[0052] The risk analysis of landslides is based on the hazard research, combined with the current land use situation in the region, to obtain the types and distribution of hazard-bearing bodies, and then randomly select 80% of the data to input into the weighted density model, using the weighted density map of the bearing body as the value estimate of the evaluation unit, and then combine the hazard index of the evaluation unit to calculate the two to obtain the risk level of each evaluation unit; finally, the remaining 20% of the data is used to verify the accuracy of the risk assessment. This method takes the land value of the region into consideration, making it less likely that those places with low probability of landslides but important economic value will be ignored in governance or development and construction.
[0053] Step 7: Comparative analysis.
[0054] In summary, the present invention is mainly a landslide disaster risk assessment method based on machine learning, which mainly includes:
[0055] 1. Analyze geological environmental conditions, organize relevant geological environmental background information, integrate historical landslide disaster data and the latest survey disaster data, analyze the development characteristics of landslide disasters in the region, establish a landslide evaluation index system for the region, and evaluate the landslide hazard risk based on a machine learning evaluation model with clustering optimization.
[0056] 2. Spatial prediction of landslide hazards is carried out to obtain the landslide hazard level of the region. Combined with the hazard data of multiple periods before and after, the CA-Markov model is used to dynamically predict the hazard of any period. The model accuracy is verified by constructing comparative experiments to obtain the optimal model parameters. On this basis, the land value is comprehensively considered and the model is applied to predict the hazard results of the region in a certain period in the future, thus realizing landslide risk assessment.
[0057] like Figure 1 As shown, it should be emphasized that: as an example, the flow chart only lists the influencing factors such as slope, slope structure type, slope type, elevation, vegetation, stratum lithology, faults, river valley landforms, rainfall, road distance, settlements, human engineering, etc., but it cannot be concluded that the influencing factors of landslide disasters in the present invention are only these few items.
[0058] Basic data collection and landslide hazard data acquisition;
[0059] (1) First, complete the collection of relevant data, including statistical data on geological disasters, relevant raster and vector data, DEM, geology, geomorphology, meteorology, hydrology, and human activities. Then organize and analyze the data to obtain the processed evaluation factors.
[0060] (2) Machine learning model based on Gaussian mixture model optimization
[0061] The spatial prediction of landslide geological hazards can be regarded as a binary classification process, which generally requires the preparation of equal amounts of negative data sets (non-disasters) and positive data sets (disasters) to form a data set. Accurate use of machine learning models to predict the risk of landslide geological hazards depends largely on the quality of the sample data set. If the samples contained in the sample data set are not representative of the landslides in the region, the evaluation accuracy and evaluation results of the model trained by this data set will not be guaranteed to be effective.
[0062] Therefore, when studying the danger of landslide hazards, it is also necessary to select "non-landslide" grid units and disaster points that are equal in number to the existing landslide disaster points to form a data set for the machine learning model. Existing studies have shown that the traditional methods commonly used to select "non-landslide" grid units are: (1) randomly selecting grids other than disaster points as "non-landslide" grid units; (2) obtaining "non-landslide" units through expert experience method; (3) seed diffusion method. Among them, randomly selecting "non-landslide" grids is too arbitrary, and the data set composed is not representative, and it often cannot achieve good results when used in machine learning models. The "non-landslide" grids selected by expert experience knowledge are artificially selected "non-landslides" and are subjective. For example, in the expert experience method, some grid units with smaller slopes are selected as non-landslide grid units. However, the landslide points in the region with slopes in the range of [0-8.52) account for 8% of the total number of landslides. Therefore, the expert experience method cannot guarantee that the selected non-landslide points are representative.
[0063] In order to overcome this shortcoming, the present invention first constructs a GMM model based on cluster analysis according to existing landslide events, and then selects non-landslide grid units from the extremely low-risk grid obtained by the cluster model, so that the non-landslide grid units obtained can be guaranteed to have an extremely low probability of landslide occurrence. The reason for selecting the GMM model is that the GMM model, as an unsupervised machine learning model, can classify samples through uncertain labels, and also has the advantages of high computational efficiency and simple process.
[0064] The preliminary analysis of landslide hazard using GMM model mainly includes three steps: first, analyze the correlation of each evaluation factor, delete the factors with small impact on landslide and high correlation with other factors, and calculate the information value of each factor on landslide; second, take the information value of each factor as the input of GMM model;
[0065] The most representative non-landslide points in the region are selected through the clustering results of the GMM model. The extracted non-landslide points and known landslide points are quantified to form an optimized sample data set for the machine learning algorithm.
[0066] In order to compare with the data set optimized by GMM model clustering, a control sample data set needs to be established. The control sample data set can be constructed by removing landslide disaster points from the region through the traditional method of selecting non-landslide points, and then randomly selecting points in the region as non-landslide disaster points to form the control sample data set.
[0067] (3) The area is divided into evaluation units, mainly grid units and slope units. The grid unit is 200*200m in size. The “fishnet” tool of ArcGIS is used to construct a regular grid, and then the entire area is divided into units. The slope unit is extracted by using the hydrological analysis tool to extract the catchment basin, which is used as the minimum evaluation unit.
[0068] (4) Under each factor scale, extract the statistical data of the factor in each evaluation unit, select the mean, variance, extreme value and median as the characteristics of a unit under a certain factor dimension, and then use the K-Means clustering method to convert the extracted features into the attribute values of a unit, and classify the evaluation units under this factor scale. For example, under the DEM factor scale, extract statistical features for each evaluation unit and use them as the input feature data for cluster analysis. DEM categories are divided for all evaluation units in the entire area. Units with similar DEM regional features will be classified into the same category. Similarly, the remaining factors are classified. By analogy, each evaluation unit will belong to a category under each factor scale, and the data will be summarized after processing all other factors.
[0069] (5) The factor category obtained in the previous step is used as the factor characteristic value to construct the model. By analyzing the data composition of the entire region, it can be found that the ratio of units where disasters occurred to units where disasters did not occur is quite large, and the positive and negative samples are unbalanced. Therefore, when constructing a landslide hazard assessment model, it is necessary to solve the problem of sample balance. Therefore, in the present invention, PU-Bagging is selected as the prediction model, and proportional selection sampling is used to construct samples. Finally, the probability of landslides occurring in all evaluation units where disasters did not occur is obtained, and the predicted probability results are graded and zoned, that is, the spatial prediction of the probability of landslide disasters is achieved.
[0070] (6) The landslide disaster data that occurred in the region before the Nth year is used as model training data and input into the hazard assessment model to obtain the spatial prediction zoning map of landslide hazard in the region in the Nth year. The disaster data that occurred in the region before the N+2th year is then used as training data to obtain the spatial prediction zoning map of hazard in the region in the N+2th year. The accuracy of the obtained results is verified to obtain a more accurate evaluation unit division method and base learner as the optimal prediction model parameters. On this basis, the spatial prediction results of landslide hazard in the region in the Nth and N+2th years with higher accuracy are obtained. The CA-Markov model is used to take these two periods of data as input data. By continuously adjusting the model parameters, a comparative test group is constructed to predict the landslide hazard in the N+4th year. The prediction results are then compared with the landslide hazard distribution in the N+4th year obtained under the PU-Bagging model.
[0071] The similarity calculation is carried out, and the test group with the highest similarity is used as the standard. The model parameters are recorded and used to predict the spatial distribution map of landslide hazard in a specific period in the future. The spatial prediction result of landslide hazard in the N+4th year can be obtained.
[0072] (7) Taking regional economic data into consideration and combining it with land use conditions, a weighted density analysis of the hazard-prone body is conducted. The result is used as an indicator to measure regional vulnerability, and a regional vulnerability distribution map is obtained. Combined with the landslide hazard grading results in the N+4th year, a regional risk assessment is conducted to obtain the risk level of each spatial location in the current region.
[0073] Technical Effects
[0074] (1) K-Means clustering is applied to the evaluation factor processing, replacing the traditional natural breakpoint method for grade division, making up for the shortcomings of the traditional method of breakpoint division, which is highly subjective and lacks basis.
[0075] (2) The PU-Bagging algorithm is transferred to the study of geological disasters. Sample areas are randomly selected from areas where landslide disasters have not occurred for multiple iterations and marked as negative samples, which are combined with positive samples to form multiple sample sets. This replaces the contingency and randomness that exists when the traditional method randomly selects an equal number of samples as negative samples.
[0076] (3) By adopting the density ratio model, a density distribution map of the disaster-prone body is constructed as the regional vulnerability data. Compared with the traditional method of cumulative summation of values and then classification of vulnerability levels, this method reduces the workload of data collection and reduces some intermediate redundant calculations.
[0077] (4) In traditional landslide research, most of the evaluation models are constructed based on the collected data to obtain a fixed evaluation result, which has poor timeliness and limited practical value. The present invention combines the CA-Markov model to achieve dynamic research on landslide disasters.
[0078] In view of this situation, the present invention proposes a reasonable research process and creatively proposes a landslide disaster risk assessment method based on machine learning. Compared with traditional methods, the present invention is more in line with the research characteristics of earth science, has clear logic, is indeed feasible, and the assessment results are more reliable.
[0079] Although preferred embodiments of the present invention have been described, additional changes and modifications may be made to these embodiments by those skilled in the art once the basic inventive concepts are known. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention. Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.
Claims
1. The landslide hazard risk assessment method based on machine learning is characterized by the following steps: Step 1: Collect, organize and analyze relevant geological historical data, analyze the development and distribution patterns of landslide disasters in the region in combination with the distribution of regional landslide disaster points in field surveys, and preliminarily screen out environmental factors that have a greater impact on regional landslides, including the following two aspects: Step 1.1, selection of evaluation factors: Firstly, the evaluation factors are preliminarily selected based on the collected data, and then the selected factors are analyzed to ensure that the selected factors have significant spatial differences in different regions of the region, and to analyze whether the factors have a strong correlation with the occurrence of regional geological disasters. Then, combined with relevant literature and previous research results, the landslide hazard evaluation factors involved in the model construction in the evaluation process are adjusted; Step 1.2: Division of evaluation units: When evaluating geographic space, the first task is to determine the evaluation area and divide the evaluation units. The model unit represents a part of the surface, and each attribute in each unit is homogeneous. In GIS, grid units and slope units are selected to conduct relevant experiments respectively, and the results are compared and analyzed. These two evaluation unit division methods are selected; Step 2: Analyze geological environmental conditions, organize relevant geological environmental background data, integrate historical landslide disaster data and the latest survey disaster data, analyze the development characteristics of landslide disasters in the region, establish a landslide evaluation index system for the region, and use statistical analysis methods to analyze the occurrence frequency, classification ratio and information value of landslide disaster points under different classification conditions of each factor based on the basic development of landslide geological disasters and the landslide disaster point data of field investigations. Further calculate the correlation between the evaluation factors and appropriately delete the evaluation factors with higher correlation. Step 3: Establish a data set optimized by a mixed Gaussian clustering model (GMM) and a randomly selected data set that is not optimized, and train the artificial neural network (ANN) model and the support vector machine (SVM) model using the radial basis kernel function respectively. After the training results reach the accuracy, evaluate the landslide hazard risk in the area, and test and compare the prediction accuracy and rationality of the evaluation results of different models; finally, conduct the first risk assessment in combination with the landslide vulnerability results; Step 4: Landslide Hazard Assessment Model Analysis When conducting spatial prediction analysis of hazard, the analysis is mainly carried out from two aspects: first, hazard assessment is carried out, based on the evaluation factors selected in steps 1 and 2, and according to the disaster data that has occurred, the spatial prediction of the current probability of landslide occurrence is carried out; Secondly, based on step 3, combined with the hazard prediction results of multiple periods, by analyzing the hazard levels of two adjacent periods and analyzing their development laws, a spatial prediction of landslide hazard in a certain period in the future is made; The process of landslide hazard assessment mainly includes: Step 4.1: Based on Step 1.1 and Step 2, pre-process the evaluation factor data and process the data scale to process the data into sample data that can be input into the model; Step 4.2: Based on step 1.2 and combined with the PU-Bagging algorithm, the evaluation units are divided into positive samples, which are areas where landslides have occurred, and unlabeled samples, which are areas where landslides have not occurred. Through random sampling and combination, the same amount of data as the positive samples is randomly extracted from the unlabeled samples with replacement as negative samples to form a training set; Step 4.3: Use step 3 and the obtained training set to build a model to obtain the landslide occurrence probability of all unlabeled sample areas in the evaluation unit sample, thereby realizing the spatial prediction of the landslide occurrence probability in the study area; Step 5: Dynamic model analysis of landslide hazard Dynamic model analysis of landslide hazard is an important part of landslide spatial prediction, which is mainly reflected in the following aspects: Step 5.1, feasibility analysis of CA-Markov model in simulating the evolution of landslide hazard; Step 5.2: Construction of dynamic prediction model of landslide hazard based on CA-Markov model and accuracy comparison experiment, select the parameter combination of the experimental scheme with the best prediction effect as the model parameter setting, and complete the model construction; among them, the scheme with the highest consistency between the experimental results and the actual data is the best scheme; Step 5.3: Use the most accurate experimental group parameters to predict the landslide hazard in a certain period in the future, that is, to achieve dynamic prediction of landslide; Step 6: Landslide risk assessment: Combined with the current land use situation in the region, the types and distribution of disaster-prone bodies are obtained. Then, 80% of the data are randomly selected and input into the weighted density model. The weighted density map of the disaster-prone body is used as the estimated value of the evaluation unit. Then, the two are calculated in combination with the hazard index of the evaluation unit to obtain the risk level of each evaluation unit. Finally, the remaining 20% of the data is used to verify the accuracy of the risk assessment. Step 7: Compare and analyze the two landslide risk assessment methods.
Citation Information
Patent Citations
Regional landslide risk evaluation method based on slope units and machine learning
CN107463991A
Submarine landslide risk analysis method based on unsupervised machine learning
CN112036424A
Landslide risk assessment method for complex and hard mountainous area
CN113283802A
KR20200052398A
Cited By
Rock-soil composite landslide dynamic evaluation method and system based on multi-source data fusion
CN122926069A
